From 0ada209919c56198f9049651cb3e7723290148b9 Mon Sep 17 00:00:00 2001 From: wang-shihao21 Date: Thu, 5 Sep 2024 21:12:55 +0800 Subject: [PATCH] =?UTF-8?q?mxDriving=E4=BB=A3=E7=A0=81=E4=BB=93=E7=9B=AE?= =?UTF-8?q?=E5=BD=95=E7=BB=93=E6=9E=84=E4=BF=AE=E6=94=B9?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- CMakeLists.txt | 6 +- README.md | 27 +- bind/pybind.cpp | 7 +- docs/api/README.md | 914 +++++++++--------- include/csrc/pybind.h | 8 +- .../BEVFormer/.gitignore | 1 - .../BEVFormer/LICENSE | 0 .../BEVFormer/README.md | 0 .../BEVFormer/README_ORI.md | 0 .../BEVFormer/docs/getting_started.md | 0 .../BEVFormer/docs/install.md | 0 .../BEVFormer/docs/prepare_dataset.md | 0 .../BEVFormer/mmcv_need/base_runner.py | 0 .../BEVFormer/mmcv_need/distributed.py | 0 .../BEVFormer/mmcv_need/epoch_based_runner.py | 0 .../mmcv_need/modulated_deform_conv.py | 0 .../BEVFormer/mmcv_need/optimizer.py | 0 .../mmcv_need/points_in_polygons_npu.cpp | 0 .../BEVFormer/mmcv_need/runtime.txt | 0 .../BEVFormer/mmdet3d_need/__init__.py | 0 .../mmdet3d_need/nuscenes_dataset.py | 0 .../BEVFormer/mmdet3d_need/runtime.txt | 0 .../BEVFormer/mmdet_need/__init__.py | 0 .../BEVFormer/mmdet_need/resnet.py | 0 .../BEVFormer/projects/__init__.py | 0 .../configs/_base_/datasets/coco_instance.py | 0 .../_base_/datasets/kitti-3d-3class.py | 0 .../configs/_base_/datasets/kitti-3d-car.py | 0 .../configs/_base_/datasets/lyft-3d.py | 0 .../configs/_base_/datasets/nuim_instance.py | 0 .../configs/_base_/datasets/nus-3d.py | 0 .../configs/_base_/datasets/nus-mono3d.py | 0 .../_base_/datasets/range100_lyft-3d.py | 0 .../_base_/datasets/s3dis-3d-5class.py | 0 .../_base_/datasets/s3dis_seg-3d-13class.py | 0 .../_base_/datasets/scannet-3d-18class.py | 0 .../_base_/datasets/scannet_seg-3d-20class.py | 0 .../_base_/datasets/sunrgbd-3d-10class.py | 0 .../_base_/datasets/waymoD5-3d-3class.py | 0 .../configs/_base_/datasets/waymoD5-3d-car.py | 0 .../configs/_base_/default_runtime.py | 0 .../projects/configs/_base_/models/3dssd.py | 0 .../models/cascade_mask_rcnn_r50_fpn.py | 0 .../centerpoint_01voxel_second_secfpn_nus.py | 0 .../centerpoint_02pillar_second_secfpn_nus.py | 0 .../projects/configs/_base_/models/fcos3d.py | 0 .../configs/_base_/models/groupfree3d.py | 0 .../projects/configs/_base_/models/h3dnet.py | 0 .../_base_/models/hv_pointpillars_fpn_lyft.py | 0 .../_base_/models/hv_pointpillars_fpn_nus.py | 0 .../hv_pointpillars_fpn_range100_lyft.py | 0 .../models/hv_pointpillars_secfpn_kitti.py | 0 .../models/hv_pointpillars_secfpn_waymo.py | 0 .../_base_/models/hv_second_secfpn_kitti.py | 0 .../_base_/models/hv_second_secfpn_waymo.py | 0 .../configs/_base_/models/imvotenet_image.py | 0 .../_base_/models/mask_rcnn_r50_fpn.py | 0 .../configs/_base_/models/paconv_cuda_ssg.py | 0 .../configs/_base_/models/paconv_ssg.py | 0 .../projects/configs/_base_/models/parta2.py | 0 .../configs/_base_/models/pointnet2_msg.py | 0 .../configs/_base_/models/pointnet2_ssg.py | 0 .../projects/configs/_base_/models/votenet.py | 0 .../configs/_base_/schedules/cosine.py | 0 .../configs/_base_/schedules/cyclic_20e.py | 0 .../configs/_base_/schedules/cyclic_40e.py | 0 .../_base_/schedules/mmdet_schedule_1x.py | 0 .../configs/_base_/schedules/schedule_2x.py | 0 .../configs/_base_/schedules/schedule_3x.py | 0 .../_base_/schedules/seg_cosine_150e.py | 0 .../_base_/schedules/seg_cosine_200e.py | 0 .../_base_/schedules/seg_cosine_50e.py | 0 .../configs/bevformer/bevformer_base.py | 0 .../configs/bevformer/bevformer_small.py | 0 .../configs/bevformer/bevformer_tiny.py | 0 .../bevformer_fp16/bevformer_tiny_fp16.py | 0 .../bevformerv2/bevformerv2-r50-t1-24ep.py | 0 .../bevformerv2/bevformerv2-r50-t1-48ep.py | 0 .../bevformerv2-r50-t1-base-24ep.py | 0 .../bevformerv2-r50-t1-base-48ep.py | 0 .../bevformerv2/bevformerv2-r50-t2-24ep.py | 0 .../bevformerv2/bevformerv2-r50-t2-48ep.py | 0 .../bevformerv2/bevformerv2-r50-t8-24ep.py | 0 .../configs/datasets/custom_lyft-3d.py | 0 .../configs/datasets/custom_nus-3d.py | 0 .../configs/datasets/custom_waymo-3d.py | 0 .../projects/mmdet3d_plugin/__init__.py | 2 + .../mmdet3d_plugin/bevformer/__init__.py | 0 .../mmdet3d_plugin/bevformer/apis/__init__.py | 0 .../bevformer/apis/mmdet_train.py | 0 .../mmdet3d_plugin/bevformer/apis/test.py | 0 .../mmdet3d_plugin/bevformer/apis/train.py | 0 .../bevformer/dense_heads/__init__.py | 0 .../bevformer/dense_heads/bev_head.py | 0 .../bevformer/dense_heads/bevformer_head.py | 0 .../bevformer/detectors/__init__.py | 0 .../bevformer/detectors/bevformer.py | 0 .../bevformer/detectors/bevformerV2.py | 0 .../bevformer/detectors/bevformer_fp16.py | 0 .../bevformer/hooks/__init__.py | 0 .../bevformer/hooks/custom_hooks.py | 0 .../bevformer/modules/__init__.py | 0 .../modules/custom_base_transformer_layer.py | 0 .../bevformer/modules/decoder.py | 0 .../bevformer/modules/encoder.py | 0 .../bevformer/modules/group_attention.py | 0 .../multi_scale_deformable_attn_function.py | 0 .../modules/spatial_cross_attention.py | 35 +- .../modules/temporal_self_attention.py | 0 .../bevformer/modules/transformer.py | 0 .../bevformer/modules/transformerV2.py | 0 .../bevformer/runner/__init__.py | 0 .../bevformer/runner/epoch_based_runner.py | 0 .../core/bbox/assigners/__init__.py | 0 .../bbox/assigners/hungarian_assigner_3d.py | 0 .../core/bbox/coders/__init__.py | 0 .../core/bbox/coders/nms_free_coder.py | 0 .../core/bbox/match_costs/__init__.py | 0 .../core/bbox/match_costs/match_cost.py | 0 .../projects/mmdet3d_plugin/core/bbox/util.py | 0 .../core/evaluation/__init__.py | 0 .../core/evaluation/eval_hooks.py | 0 .../core/evaluation/kitti2waymo.py | 0 .../mmdet3d_plugin/datasets/__init__.py | 0 .../mmdet3d_plugin/datasets/builder.py | 0 .../datasets/nuscenes_dataset.py | 0 .../datasets/nuscenes_dataset_v2.py | 0 .../datasets/nuscenes_mono_dataset.py | 0 .../mmdet3d_plugin/datasets/nuscnes_eval.py | 0 .../datasets/pipelines/__init__.py | 0 .../datasets/pipelines/augmentation.py | 0 .../datasets/pipelines/dd3d_mapper.py | 0 .../datasets/pipelines/formating.py | 0 .../datasets/pipelines/loading.py | 0 .../datasets/pipelines/transform_3d.py | 0 .../datasets/samplers/__init__.py | 0 .../datasets/samplers/distributed_sampler.py | 0 .../datasets/samplers/group_sampler.py | 0 .../datasets/samplers/sampler.py | 0 .../projects/mmdet3d_plugin/dd3d/__init__.py | 0 .../mmdet3d_plugin/dd3d/datasets/__init__.py | 0 .../mmdet3d_plugin/dd3d/datasets/nuscenes.py | 0 .../dd3d/datasets/transform_utils.py | 0 .../mmdet3d_plugin/dd3d/layers/iou_loss.py | 0 .../dd3d/layers/normalization.py | 0 .../dd3d/layers/smooth_l1_loss.py | 0 .../mmdet3d_plugin/dd3d/modeling/__init__.py | 0 .../mmdet3d_plugin/dd3d/modeling/core.py | 0 .../dd3d/modeling/disentangled_box3d_loss.py | 0 .../mmdet3d_plugin/dd3d/modeling/fcos2d.py | 0 .../mmdet3d_plugin/dd3d/modeling/fcos3d.py | 0 .../dd3d/modeling/nuscenes_dd3d.py | 0 .../dd3d/modeling/prepare_targets.py | 0 .../dd3d/structures/__init__.py | 0 .../mmdet3d_plugin/dd3d/structures/boxes3d.py | 0 .../dd3d/structures/image_list.py | 0 .../mmdet3d_plugin/dd3d/structures/pose.py | 0 .../dd3d/structures/transform3d.py | 0 .../mmdet3d_plugin/dd3d/utils/comm.py | 0 .../mmdet3d_plugin/dd3d/utils/geometry.py | 0 .../mmdet3d_plugin/dd3d/utils/tasks.py | 0 .../mmdet3d_plugin/dd3d/utils/tensor2d.py | 0 .../dd3d/utils/visualization.py | 0 .../models/backbones/__init__.py | 0 .../mmdet3d_plugin/models/backbones/vovnet.py | 0 .../mmdet3d_plugin/models/hooks/__init__.py | 0 .../mmdet3d_plugin/models/hooks/hooks.py | 2 + .../mmdet3d_plugin/models/opt/__init__.py | 0 .../mmdet3d_plugin/models/opt/adamw.py | 0 .../mmdet3d_plugin/models/utils/__init__.py | 0 .../mmdet3d_plugin/models/utils/bricks.py | 0 .../mmdet3d_plugin/models/utils/grid_mask.py | 2 + .../models/utils/position_embedding.py | 0 .../mmdet3d_plugin/models/utils/visual.py | 0 .../BEVFormer/public_address_statement.md | 0 .../BEVFormer/requirements.txt | 0 model_examples/BEVFormer/test/env_npu.sh | 59 ++ .../BEVFormer/test/train_full_8p_base_fp32.sh | 97 ++ .../test/train_performance_8p_base_fp32.sh | 92 ++ .../BEVFormer/tools/__init__.py | 0 .../tools/analysis_tools/__init__.py | 0 .../tools/analysis_tools/analyze_logs.py | 0 .../tools/analysis_tools/benchmark.py | 0 .../tools/analysis_tools/get_params.py | 0 .../BEVFormer/tools/analysis_tools/visual.py | 0 .../BEVFormer/tools/create_data.py | 0 .../tools/data_converter/__init__.py | 0 .../data_converter/create_gt_database.py | 0 .../tools/data_converter/indoor_converter.py | 0 .../tools/data_converter/kitti_converter.py | 0 .../tools/data_converter/kitti_data_utils.py | 0 .../tools/data_converter/lyft_converter.py | 0 .../tools/data_converter/lyft_data_fixer.py | 0 .../tools/data_converter/nuimage_converter.py | 0 .../data_converter/nuscenes_converter.py | 0 .../tools/data_converter/s3dis_data_utils.py | 0 .../data_converter/scannet_data_utils.py | 0 .../data_converter/sunrgbd_data_utils.py | 0 .../tools/data_converter/waymo_converter.py | 0 .../BEVFormer/tools/dist_test.sh | 0 .../BEVFormer/tools/dist_train.sh | 0 .../BEVFormer/tools/fp16/dist_train.sh | 0 .../BEVFormer/tools/fp16/train.py | 0 .../BEVFormer/tools/misc/browse_dataset.py | 0 .../BEVFormer/tools/misc/fuse_conv_bn.py | 0 .../BEVFormer/tools/misc/print_config.py | 0 .../BEVFormer/tools/misc/visualize_results.py | 0 .../convert_votenet_checkpoints.py | 0 .../tools/model_converters/publish_model.py | 0 .../tools/model_converters/regnet2mmdet.py | 0 .../BEVFormer/tools/test.py | 0 .../BEVFormer/tools/train.py | 0 mx_driving/common/__init__.py | 17 - mx_driving/common/ops/csrc/functions.h | 75 +- mx_driving/common/ops/csrc/pybind.cpp | 70 +- mx_driving/data/CMakeLists.txt | 7 + mx_driving/data/__init__.py | 3 + .../{motion => data}/components/README.md | 0 mx_driving/{motion => data/ops}/__init__.py | 0 .../{common => data}/ops/csrc/PointsInBox.cpp | 0 .../ops/csrc/PointsInBoxAll.cpp | 0 mx_driving/data/ops/csrc/README.md | 6 + .../ops/csrc/RoipointPool3dForward.cpp | 0 .../vision => data}/ops/csrc/functions.h | 7 +- mx_driving/data/ops/csrc/pybind.cpp | 15 + .../ops/kernels/CMakeLists.txt | 0 mx_driving/data/ops/kernels/README.md | 13 + .../ops/kernels/framework/CMakeLists.txt | 0 .../ops/kernels/op_host/CMakeLists.txt | 0 mx_driving/data/ops/kernels/op_host/common.h | 28 + .../ops/kernels/op_host/points_in_box.cpp | 0 .../ops/kernels/op_host/points_in_box_all.cpp | 0 .../op_host/points_in_box_all_tiling.h | 0 .../kernels/op_host/points_in_box_tiling.h | 0 .../op_host/roipoint_pool3d_forward.cpp | 0 .../op_host/roipoint_pool3d_forward_tiling.h | 0 .../ops/kernels/op_kernel/CMakeLists.txt | 0 .../ops/kernels/op_kernel/points_in_box.cpp | 0 .../kernels/op_kernel/points_in_box_all.cpp | 0 .../op_kernel/roipoint_pool3d_forward.cpp | 0 .../{common => data}/ops/npu_points_in_box.py | 0 .../ops/npu_points_in_box_all.py | 0 .../ops/npu_roipoint_pool3d.py | 0 .../{motion => detection}/CMakeLists.txt | 0 mx_driving/detection/__init__.py | 5 + .../fused => detection}/components/README.md | 0 .../{perception => detection/ops}/__init__.py | 0 .../ops/boxes_overlap_bev.py | 0 .../ops/csrc/BoxesOverlapBev.cpp | 0 .../{common => detection}/ops/csrc/Nms3d.cpp | 0 .../ops/csrc/Nms3dNormal.cpp | 0 .../{motion => detection}/ops/csrc/README.md | 0 .../ops/csrc/RotatedIou.cpp | 0 .../ops/csrc/RotatedOverlaps.cpp | 0 .../fused => detection}/ops/csrc/functions.h | 26 +- .../vision => detection}/ops/csrc/pybind.cpp | 14 +- .../ops/kernels/CMakeLists.txt | 0 .../ops/kernels/README.md | 0 .../ops/kernels/op_host/CMakeLists.txt | 0 .../ops/kernels/op_host/boxes_overlap_bev.cpp | 0 .../op_host/boxes_overlap_bev_tiling.h | 0 .../op_host/gather_nms3d_mask_tiling.cpp | 0 .../op_host/gather_nms3d_mask_tiling.h | 0 .../ops/kernels/op_host/nms3d.cpp | 0 .../kernels/op_host/nms3d_normal_tiling.cpp | 0 .../ops/kernels/op_host/nms3d_normal_tiling.h | 24 + .../ops/kernels/op_host/nms3d_tiling.h | 0 .../ops/kernels/op_kernel/CMakeLists.txt | 0 .../kernels/op_kernel/boxes_overlap_bev.cpp | 0 .../kernels/op_kernel/gather_nms3d_mask.cpp | 0 .../ops/kernels/op_kernel/nms3d.cpp | 0 .../ops/kernels/op_kernel/nms3d_normal.cpp | 0 .../{common => detection}/ops/nms3d_normal.py | 0 .../{common => detection}/ops/npu_nms3d.py | 0 .../{common => detection}/ops/rotated_iou.py | 0 .../ops/rotated_overlaps.py | 0 mx_driving/fused/CMakeLists.txt | 3 + mx_driving/fused/__init__.py | 5 + .../point => fused}/components/README.md | 0 .../{perception => }/fused/ops/__init__.py | 0 .../{common => fused}/ops/csrc/AddRelu.cpp | 0 .../ops/csrc/DeformableAggregation.cpp | 0 .../ops/csrc/FusedBiasLeakyRelu.cpp | 0 .../{common => fused}/ops/csrc/MaxPool2d.cpp | 0 .../csrc/MultiScaleDeformableAttnFunction.cpp | 0 .../{perception => }/fused/ops/csrc/README.md | 0 mx_driving/fused/ops/csrc/functions.h | 57 ++ mx_driving/fused/ops/csrc/pybind.cpp | 25 + .../ops/fused_bias_leaky_relu.py | 0 .../fused/ops/kernels/CMakeLists.txt | 0 .../fused/ops/kernels/README.md | 0 .../ops/kernels/op_host/CMakeLists.txt | 0 .../ops/kernels/op_host/add_relu.cpp | 0 .../ops/kernels/op_host/add_relu_tiling.h | 0 mx_driving/fused/ops/kernels/op_host/common.h | 28 + .../op_host/deformable_aggregation.cpp | 0 .../op_host/deformable_aggregation_grad.cpp | 0 .../deformable_aggregation_grad_tiling.h | 0 .../op_host/deformable_aggregation_tiling.h | 0 .../kernels/op_host/fused_bias_leaky_relu.cpp | 0 .../op_host/fused_bias_leaky_relu_tiling.h | 0 .../ops/kernels/op_host/max_pool2d.cpp | 0 .../ops/kernels/op_host/max_pool2d.h | 0 .../op_host/multi_scale_deformable_attn.cpp | 0 .../multi_scale_deformable_attn_grad.cpp | 0 .../multi_scale_deformable_attn_grad_tiling.h | 0 ...lti_scale_deformable_attn_grad_tiling_v2.h | 0 .../multi_scale_deformable_attn_grad_v2.cpp | 0 .../multi_scale_deformable_attn_tiling.h | 0 .../ops/kernels/op_kernel/CMakeLists.txt | 0 .../ops/kernels/op_kernel/add_relu.cpp | 0 .../fused/ops/kernels/op_kernel/common.h | 0 .../op_kernel/deformable_aggregation.cpp | 0 .../op_kernel/deformable_aggregation_grad.cpp | 0 .../op_kernel/fused_bias_leaky_relu.cpp | 0 .../kernels/op_kernel/fused_bias_leaky_relu.h | 0 .../ops/kernels/op_kernel/max_pool2d.cpp | 0 .../op_kernel/ms_deform_attn_generic.h | 0 .../op_kernel/ms_deform_attn_grad_generic.h | 0 .../ms_deform_attn_grad_generic_v2.h | 0 .../op_kernel/ms_deform_attn_grad_high_perf.h | 0 .../ms_deform_attn_grad_high_perf_v2.h | 0 .../op_kernel/ms_deform_attn_high_perf.h | 0 .../op_kernel/multi_scale_deformable_attn.cpp | 0 .../multi_scale_deformable_attn_grad.cpp | 0 .../multi_scale_deformable_attn_grad_v2.cpp | 0 .../{common => fused}/ops/npu_add_relu.py | 0 .../ops/npu_deformable_aggregation.py | 0 .../{common => fused}/ops/npu_max_pool2d.py | 1 + ...pu_multi_scale_deformable_attn_function.py | 0 mx_driving/motion/ops/csrc/pybind.cpp | 5 - .../ops/kernels/op_kernel/CMakeLists.txt | 0 mx_driving/perception/CMakeLists.txt | 9 - mx_driving/perception/fused/__init__.py | 2 - .../perception/fused/ops/csrc/pybind.cpp | 13 - mx_driving/perception/point/__init__.py | 2 - mx_driving/perception/vision/__init__.py | 1 - mx_driving/perception/vision/ops/__init__.py | 0 .../perception/vision/ops/csrc/README.md | 2 - .../perception/vision/ops/kernels/README.md | 2 - mx_driving/point/CMakeLists.txt | 3 + mx_driving/point/__init__.py | 8 + .../vision => point}/components/README.md | 0 .../{perception => }/point/ops/__init__.py | 0 .../fused => point}/ops/bev_pool.py | 0 .../fused => point}/ops/bev_pool_v2.py | 0 .../fused => point}/ops/csrc/BEVPool.cpp | 0 .../ops/csrc/BEVPoolBackward.cpp | 0 .../fused => point}/ops/csrc/BEVPoolV2.cpp | 0 .../ops/csrc/BEVPoolV2Backward.cpp | 0 .../ops/csrc/DynamicScatter.cpp | 0 .../ops/csrc/DynamicVoxelization.cpp | 0 .../ops/csrc/FurthestPointSampling.cpp | 0 .../csrc/FurthestPointSamplingWithDist.cpp | 0 .../point/ops/csrc/GroupPoints.cpp | 0 .../point/ops/csrc/HardVoxelize.cpp | 0 .../point/ops/csrc/PointToVoxel.cpp | 0 .../{perception => }/point/ops/csrc/README.md | 0 .../point/ops/csrc/UniqueVoxel.cpp | 0 .../point/ops/csrc/VecPoolBackward.cpp | 0 .../ops/csrc/VoxelPoolingTrain.cpp | 0 .../point/ops/csrc/VoxelToPoint.cpp | 0 .../point/ops/csrc/functions.h | 35 + .../point/ops/csrc/pybind.cpp | 25 +- .../ops/furthest_point_sampling.py | 0 .../ops/furthest_point_sampling_with_dist.py | 0 .../point/ops/group_points.py | 0 .../ops/kernels/CMakeLists.txt | 0 .../point/ops/kernels/README.md | 0 .../ops/kernels/op_host/CMakeLists.txt | 0 .../ops/kernels/op_host/bev_pool.cpp | 0 .../ops/kernels/op_host/bev_pool_tiling.h | 0 .../ops/kernels/op_host/dynamic_scatter.cpp | 0 .../kernels/op_host/dynamic_scatter_grad.cpp | 0 .../op_host/dynamic_scatter_grad_tiling.h | 0 .../kernels/op_host/dynamic_scatter_tiling.h | 0 .../kernels/op_host/dynamic_voxelization.cpp | 0 .../op_host/dynamic_voxelization_tiling.h | 0 .../op_host/furthest_point_sampling.cpp | 0 .../op_host/furthest_point_sampling_tiling.h | 0 .../furthest_point_sampling_with_dist.cpp | 0 ...furthest_point_sampling_with_dist_tiling.h | 0 .../ops/kernels/op_host/group_points.cpp | 0 .../ops/kernels/op_host/group_points_grad.cpp | 0 .../op_host/group_points_grad_tiling.h | 0 .../ops/kernels/op_host/group_points_tiling.h | 0 .../ops/kernels/op_host/hard_voxelize.cpp | 0 .../kernels/op_host/hard_voxelize_tiling.h | 0 .../ops/kernels/op_host/point_to_voxel.cpp | 0 .../kernels/op_host/point_to_voxel_tiling.h | 0 .../ops/kernels/op_host/unique_voxel.cpp | 0 .../ops/kernels/op_host/unique_voxel_tiling.h | 0 .../ops/kernels/op_host/vec_pool_grad.cpp | 0 .../kernels/op_host/vec_pool_grad_tiling.h | 0 .../kernels/op_host/voxel_pooling_train.cpp | 0 .../op_host/voxel_pooling_train_grad.cpp | 0 .../op_host/voxel_pooling_train_grad_tiling.h | 0 .../op_host/voxel_pooling_train_tiling.h | 0 .../ops/kernels/op_kernel/CMakeLists.txt | 4 + .../ops/kernels/op_kernel/bev_pool.cpp | 0 .../ops/kernels/op_kernel/bev_pool.h | 0 .../ops/kernels/op_kernel/bev_pool_grad.cpp | 0 .../ops/kernels/op_kernel/bev_pool_v2.cpp | 0 .../ops/kernels/op_kernel/bev_pool_v2.h | 0 .../kernels/op_kernel/bev_pool_v2_grad.cpp | 0 .../point/ops/kernels/op_kernel/common.h | 46 + .../ops/kernels/op_kernel/dynamic_scatter.cpp | 0 .../kernels/op_kernel/dynamic_scatter_base.h | 0 .../op_kernel/dynamic_scatter_grad.cpp | 0 .../op_kernel/dynamic_scatter_grad_base.h | 0 .../op_kernel/dynamic_scatter_grad_max.h | 0 .../op_kernel/dynamic_scatter_grad_mean.h | 0 .../op_kernel/dynamic_scatter_grad_sum.h | 0 .../kernels/op_kernel/dynamic_scatter_max.h | 0 .../kernels/op_kernel/dynamic_scatter_mean.h | 0 .../kernels/op_kernel/dynamic_scatter_sum.h | 0 .../op_kernel/dynamic_voxelization.cpp | 0 .../op_kernel/furthest_point_sampling.cpp | 0 .../op_kernel/furthest_point_sampling.h | 0 .../furthest_point_sampling_with_dist.cpp | 0 .../ops/kernels/op_kernel/group_points.cpp | 0 .../kernels/op_kernel/group_points_grad.cpp | 0 .../ops/kernels/op_kernel/hard_voxelize.cpp | 0 .../ops/kernels/op_kernel/point_to_voxel.cpp | 0 .../ops/kernels/op_kernel/unique_voxel.cpp | 0 .../ops/kernels/op_kernel/vec_pool_grad.cpp | 0 .../kernels/op_kernel/voxel_pooling_train.cpp | 0 .../op_kernel/voxel_pooling_train_grad.cpp | 0 .../ops/kernels/op_kernel/voxel_to_point.cpp | 0 .../ops/npu_dynamic_scatter.py | 0 .../ops/voxel_pooling_train.py | 0 .../{common => point}/ops/voxelization.py | 0 setup.py | 2 +- tests/torch/test_add_relu.py | 10 +- tests/torch/test_bev_pool.py | 2 +- tests/torch/test_bev_pool_v2.py | 2 +- tests/torch/test_boxes_overlap_bev.py | 4 +- tests/torch/test_deformable_aggregation.py | 4 +- .../torch/test_deformable_aggregation_grad.py | 4 +- .../test_furthest_point_sample_with_dist.py | 4 +- tests/torch/test_furthest_point_sampling.py | 4 +- tests/torch/test_fused_bias_leaky_relu.py | 10 +- tests/torch/test_group_points.py | 2 +- ...st_multi_scale_deformable_attn_function.py | 4 +- tests/torch/test_npu_dyn_voxelization.py | 4 +- tests/torch/test_npu_dynamic_scatter.py | 6 +- tests/torch/test_npu_max_pool2d.py | 4 +- tests/torch/test_npu_nms3d.py | 4 +- tests/torch/test_npu_nms3d_normal.py | 16 +- tests/torch/test_points_in_box.py | 12 +- tests/torch/test_points_in_box_all.py | 12 +- tests/torch/test_roipoint_pool3d.py | 2 +- tests/torch/test_rotated_iou.py | 4 +- tests/torch/test_voxel_pooling_train.py | 4 +- 454 files changed, 1200 insertions(+), 758 deletions(-) rename {examples => model_examples}/BEVFormer/.gitignore (99%) rename {examples => model_examples}/BEVFormer/LICENSE (100%) rename {examples => model_examples}/BEVFormer/README.md (100%) rename {examples => model_examples}/BEVFormer/README_ORI.md (100%) rename {examples => model_examples}/BEVFormer/docs/getting_started.md (100%) rename {examples => model_examples}/BEVFormer/docs/install.md (100%) rename {examples => model_examples}/BEVFormer/docs/prepare_dataset.md (100%) rename {examples => model_examples}/BEVFormer/mmcv_need/base_runner.py (100%) rename {examples => model_examples}/BEVFormer/mmcv_need/distributed.py (100%) rename {examples => model_examples}/BEVFormer/mmcv_need/epoch_based_runner.py (100%) rename {examples => model_examples}/BEVFormer/mmcv_need/modulated_deform_conv.py (100%) rename {examples => model_examples}/BEVFormer/mmcv_need/optimizer.py (100%) rename {examples => model_examples}/BEVFormer/mmcv_need/points_in_polygons_npu.cpp (100%) rename {examples => model_examples}/BEVFormer/mmcv_need/runtime.txt (100%) rename {examples => model_examples}/BEVFormer/mmdet3d_need/__init__.py (100%) rename {examples => model_examples}/BEVFormer/mmdet3d_need/nuscenes_dataset.py (100%) rename {examples => model_examples}/BEVFormer/mmdet3d_need/runtime.txt (100%) rename {examples => model_examples}/BEVFormer/mmdet_need/__init__.py (100%) rename {examples => model_examples}/BEVFormer/mmdet_need/resnet.py (100%) rename {examples => model_examples}/BEVFormer/projects/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/coco_instance.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/kitti-3d-3class.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/kitti-3d-car.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/lyft-3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/nuim_instance.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/nus-3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/nus-mono3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/range100_lyft-3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/s3dis-3d-5class.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/s3dis_seg-3d-13class.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/scannet-3d-18class.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/scannet_seg-3d-20class.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/sunrgbd-3d-10class.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-3class.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-car.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/default_runtime.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/3dssd.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/cascade_mask_rcnn_r50_fpn.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/centerpoint_01voxel_second_secfpn_nus.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/centerpoint_02pillar_second_secfpn_nus.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/fcos3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/groupfree3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/h3dnet.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_lyft.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_nus.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_range100_lyft.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_kitti.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_waymo.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_kitti.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_waymo.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/imvotenet_image.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/mask_rcnn_r50_fpn.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/paconv_cuda_ssg.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/paconv_ssg.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/parta2.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/pointnet2_msg.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/pointnet2_ssg.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/models/votenet.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/cosine.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/cyclic_20e.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/cyclic_40e.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/mmdet_schedule_1x.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/schedule_2x.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/schedule_3x.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/seg_cosine_150e.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/seg_cosine_200e.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/_base_/schedules/seg_cosine_50e.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformer/bevformer_base.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformer/bevformer_small.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformer/bevformer_tiny.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformer_fp16/bevformer_tiny_fp16.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-24ep.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-48ep.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-24ep.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-48ep.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-24ep.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-48ep.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t8-24ep.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/datasets/custom_lyft-3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/datasets/custom_nus-3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/configs/datasets/custom_waymo-3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/__init__.py (83%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/mmdet_train.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/test.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/train.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bev_head.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bevformer_head.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformerV2.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer_fp16.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/custom_hooks.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/custom_base_transformer_layer.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/decoder.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/encoder.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/group_attention.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/multi_scale_deformable_attn_function.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/spatial_cross_attention.py (95%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/temporal_self_attention.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformer.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformerV2.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/epoch_based_runner.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/hungarian_assigner_3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/nms_free_coder.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/match_cost.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/bbox/util.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/evaluation/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/evaluation/eval_hooks.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/core/evaluation/kitti2waymo.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/builder.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset_v2.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_mono_dataset.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/nuscnes_eval.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/augmentation.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/dd3d_mapper.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/formating.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/loading.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/distributed_sampler.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/group_sampler.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/sampler.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/nuscenes.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/transform_utils.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/iou_loss.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/normalization.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/smooth_l1_loss.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/core.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/disentangled_box3d_loss.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos2d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/nuscenes_dd3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/prepare_targets.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/boxes3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/image_list.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/pose.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/transform3d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/comm.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/geometry.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tasks.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tensor2d.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/visualization.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/backbones/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/backbones/vovnet.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/hooks/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/hooks/hooks.py (81%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/opt/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/opt/adamw.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/utils/__init__.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/utils/bricks.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/utils/grid_mask.py (97%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/utils/position_embedding.py (100%) rename {examples => model_examples}/BEVFormer/projects/mmdet3d_plugin/models/utils/visual.py (100%) rename {examples => model_examples}/BEVFormer/public_address_statement.md (100%) rename {examples => model_examples}/BEVFormer/requirements.txt (100%) create mode 100644 model_examples/BEVFormer/test/env_npu.sh create mode 100644 model_examples/BEVFormer/test/train_full_8p_base_fp32.sh create mode 100644 model_examples/BEVFormer/test/train_performance_8p_base_fp32.sh rename {examples => model_examples}/BEVFormer/tools/__init__.py (100%) rename {examples => model_examples}/BEVFormer/tools/analysis_tools/__init__.py (100%) rename {examples => model_examples}/BEVFormer/tools/analysis_tools/analyze_logs.py (100%) rename {examples => model_examples}/BEVFormer/tools/analysis_tools/benchmark.py (100%) rename {examples => model_examples}/BEVFormer/tools/analysis_tools/get_params.py (100%) rename {examples => model_examples}/BEVFormer/tools/analysis_tools/visual.py (100%) rename {examples => model_examples}/BEVFormer/tools/create_data.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/__init__.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/create_gt_database.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/indoor_converter.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/kitti_converter.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/kitti_data_utils.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/lyft_converter.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/lyft_data_fixer.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/nuimage_converter.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/nuscenes_converter.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/s3dis_data_utils.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/scannet_data_utils.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/sunrgbd_data_utils.py (100%) rename {examples => model_examples}/BEVFormer/tools/data_converter/waymo_converter.py (100%) rename {examples => model_examples}/BEVFormer/tools/dist_test.sh (100%) rename {examples => model_examples}/BEVFormer/tools/dist_train.sh (100%) rename {examples => model_examples}/BEVFormer/tools/fp16/dist_train.sh (100%) rename {examples => model_examples}/BEVFormer/tools/fp16/train.py (100%) rename {examples => model_examples}/BEVFormer/tools/misc/browse_dataset.py (100%) rename {examples => model_examples}/BEVFormer/tools/misc/fuse_conv_bn.py (100%) rename {examples => model_examples}/BEVFormer/tools/misc/print_config.py (100%) rename {examples => model_examples}/BEVFormer/tools/misc/visualize_results.py (100%) rename {examples => model_examples}/BEVFormer/tools/model_converters/convert_votenet_checkpoints.py (100%) rename {examples => model_examples}/BEVFormer/tools/model_converters/publish_model.py (100%) rename {examples => model_examples}/BEVFormer/tools/model_converters/regnet2mmdet.py (100%) rename {examples => model_examples}/BEVFormer/tools/test.py (100%) rename {examples => model_examples}/BEVFormer/tools/train.py (100%) create mode 100644 mx_driving/data/CMakeLists.txt create mode 100644 mx_driving/data/__init__.py rename mx_driving/{motion => data}/components/README.md (100%) rename mx_driving/{motion => data/ops}/__init__.py (100%) rename mx_driving/{common => data}/ops/csrc/PointsInBox.cpp (100%) rename mx_driving/{common => data}/ops/csrc/PointsInBoxAll.cpp (100%) create mode 100644 mx_driving/data/ops/csrc/README.md rename mx_driving/{common => data}/ops/csrc/RoipointPool3dForward.cpp (100%) rename mx_driving/{perception/vision => data}/ops/csrc/functions.h (69%) create mode 100644 mx_driving/data/ops/csrc/pybind.cpp rename mx_driving/{motion => data}/ops/kernels/CMakeLists.txt (100%) create mode 100644 mx_driving/data/ops/kernels/README.md rename mx_driving/{motion => data}/ops/kernels/framework/CMakeLists.txt (100%) rename mx_driving/{motion => data}/ops/kernels/op_host/CMakeLists.txt (100%) create mode 100644 mx_driving/data/ops/kernels/op_host/common.h rename mx_driving/{common => data}/ops/kernels/op_host/points_in_box.cpp (100%) rename mx_driving/{common => data}/ops/kernels/op_host/points_in_box_all.cpp (100%) rename mx_driving/{common => data}/ops/kernels/op_host/points_in_box_all_tiling.h (100%) rename mx_driving/{common => data}/ops/kernels/op_host/points_in_box_tiling.h (100%) rename mx_driving/{common => data}/ops/kernels/op_host/roipoint_pool3d_forward.cpp (100%) rename mx_driving/{common => data}/ops/kernels/op_host/roipoint_pool3d_forward_tiling.h (100%) rename mx_driving/{perception/fused => data}/ops/kernels/op_kernel/CMakeLists.txt (100%) rename mx_driving/{common => data}/ops/kernels/op_kernel/points_in_box.cpp (100%) rename mx_driving/{common => data}/ops/kernels/op_kernel/points_in_box_all.cpp (100%) rename mx_driving/{common => data}/ops/kernels/op_kernel/roipoint_pool3d_forward.cpp (100%) rename mx_driving/{common => data}/ops/npu_points_in_box.py (100%) rename mx_driving/{common => data}/ops/npu_points_in_box_all.py (100%) rename mx_driving/{common => data}/ops/npu_roipoint_pool3d.py (100%) rename mx_driving/{motion => detection}/CMakeLists.txt (100%) create mode 100644 mx_driving/detection/__init__.py rename mx_driving/{perception/fused => detection}/components/README.md (100%) rename mx_driving/{perception => detection/ops}/__init__.py (100%) rename mx_driving/{perception/vision => detection}/ops/boxes_overlap_bev.py (100%) rename mx_driving/{perception/vision => detection}/ops/csrc/BoxesOverlapBev.cpp (100%) rename mx_driving/{common => detection}/ops/csrc/Nms3d.cpp (100%) rename mx_driving/{common => detection}/ops/csrc/Nms3dNormal.cpp (100%) rename mx_driving/{motion => detection}/ops/csrc/README.md (100%) rename mx_driving/{common => detection}/ops/csrc/RotatedIou.cpp (100%) rename mx_driving/{common => detection}/ops/csrc/RotatedOverlaps.cpp (100%) rename mx_driving/{perception/fused => detection}/ops/csrc/functions.h (34%) rename mx_driving/{perception/vision => detection}/ops/csrc/pybind.cpp (37%) rename mx_driving/{perception/point => detection}/ops/kernels/CMakeLists.txt (100%) rename mx_driving/{motion => detection}/ops/kernels/README.md (100%) rename mx_driving/{perception/fused => detection}/ops/kernels/op_host/CMakeLists.txt (100%) rename mx_driving/{perception/vision => detection}/ops/kernels/op_host/boxes_overlap_bev.cpp (100%) rename mx_driving/{perception/vision => detection}/ops/kernels/op_host/boxes_overlap_bev_tiling.h (100%) rename mx_driving/{common => detection}/ops/kernels/op_host/gather_nms3d_mask_tiling.cpp (100%) rename mx_driving/{common => detection}/ops/kernels/op_host/gather_nms3d_mask_tiling.h (100%) rename mx_driving/{common => detection}/ops/kernels/op_host/nms3d.cpp (100%) rename mx_driving/{common => detection}/ops/kernels/op_host/nms3d_normal_tiling.cpp (100%) create mode 100644 mx_driving/detection/ops/kernels/op_host/nms3d_normal_tiling.h rename mx_driving/{common => detection}/ops/kernels/op_host/nms3d_tiling.h (100%) rename mx_driving/{perception/point => detection}/ops/kernels/op_kernel/CMakeLists.txt (100%) rename mx_driving/{perception/vision => detection}/ops/kernels/op_kernel/boxes_overlap_bev.cpp (100%) rename mx_driving/{common => detection}/ops/kernels/op_kernel/gather_nms3d_mask.cpp (100%) rename mx_driving/{common => detection}/ops/kernels/op_kernel/nms3d.cpp (100%) rename mx_driving/{common => detection}/ops/kernels/op_kernel/nms3d_normal.cpp (100%) rename mx_driving/{common => detection}/ops/nms3d_normal.py (100%) rename mx_driving/{common => detection}/ops/npu_nms3d.py (100%) rename mx_driving/{common => detection}/ops/rotated_iou.py (100%) rename mx_driving/{common => detection}/ops/rotated_overlaps.py (100%) create mode 100644 mx_driving/fused/CMakeLists.txt create mode 100644 mx_driving/fused/__init__.py rename mx_driving/{perception/point => fused}/components/README.md (100%) rename mx_driving/{perception => }/fused/ops/__init__.py (100%) rename mx_driving/{common => fused}/ops/csrc/AddRelu.cpp (100%) rename mx_driving/{common => fused}/ops/csrc/DeformableAggregation.cpp (100%) rename mx_driving/{common => fused}/ops/csrc/FusedBiasLeakyRelu.cpp (100%) rename mx_driving/{common => fused}/ops/csrc/MaxPool2d.cpp (100%) rename mx_driving/{common => fused}/ops/csrc/MultiScaleDeformableAttnFunction.cpp (100%) rename mx_driving/{perception => }/fused/ops/csrc/README.md (100%) create mode 100644 mx_driving/fused/ops/csrc/functions.h create mode 100644 mx_driving/fused/ops/csrc/pybind.cpp rename mx_driving/{common => fused}/ops/fused_bias_leaky_relu.py (100%) rename mx_driving/{perception => }/fused/ops/kernels/CMakeLists.txt (100%) rename mx_driving/{perception => }/fused/ops/kernels/README.md (100%) rename mx_driving/{perception/point => fused}/ops/kernels/op_host/CMakeLists.txt (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/add_relu.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/add_relu_tiling.h (100%) create mode 100644 mx_driving/fused/ops/kernels/op_host/common.h rename mx_driving/{common => fused}/ops/kernels/op_host/deformable_aggregation.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/deformable_aggregation_grad.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/deformable_aggregation_grad_tiling.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/deformable_aggregation_tiling.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/fused_bias_leaky_relu.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/fused_bias_leaky_relu_tiling.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/max_pool2d.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/max_pool2d.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/multi_scale_deformable_attn.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/multi_scale_deformable_attn_grad.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling_v2.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/multi_scale_deformable_attn_grad_v2.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_host/multi_scale_deformable_attn_tiling.h (100%) rename mx_driving/{perception/vision => fused}/ops/kernels/op_kernel/CMakeLists.txt (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/add_relu.cpp (100%) rename mx_driving/{perception => }/fused/ops/kernels/op_kernel/common.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/deformable_aggregation.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/deformable_aggregation_grad.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/fused_bias_leaky_relu.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/fused_bias_leaky_relu.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/max_pool2d.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/ms_deform_attn_generic.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/ms_deform_attn_grad_generic.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/ms_deform_attn_grad_generic_v2.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf_v2.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/ms_deform_attn_high_perf.h (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/multi_scale_deformable_attn.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/multi_scale_deformable_attn_grad.cpp (100%) rename mx_driving/{common => fused}/ops/kernels/op_kernel/multi_scale_deformable_attn_grad_v2.cpp (100%) rename mx_driving/{common => fused}/ops/npu_add_relu.py (100%) rename mx_driving/{common => fused}/ops/npu_deformable_aggregation.py (100%) rename mx_driving/{common => fused}/ops/npu_max_pool2d.py (99%) rename mx_driving/{common => fused}/ops/npu_multi_scale_deformable_attn_function.py (100%) delete mode 100644 mx_driving/motion/ops/csrc/pybind.cpp delete mode 100644 mx_driving/motion/ops/kernels/op_kernel/CMakeLists.txt delete mode 100644 mx_driving/perception/CMakeLists.txt delete mode 100644 mx_driving/perception/fused/__init__.py delete mode 100644 mx_driving/perception/fused/ops/csrc/pybind.cpp delete mode 100644 mx_driving/perception/point/__init__.py delete mode 100644 mx_driving/perception/vision/__init__.py delete mode 100644 mx_driving/perception/vision/ops/__init__.py delete mode 100644 mx_driving/perception/vision/ops/csrc/README.md delete mode 100644 mx_driving/perception/vision/ops/kernels/README.md create mode 100644 mx_driving/point/CMakeLists.txt create mode 100644 mx_driving/point/__init__.py rename mx_driving/{perception/vision => point}/components/README.md (100%) rename mx_driving/{perception => }/point/ops/__init__.py (100%) rename mx_driving/{perception/fused => point}/ops/bev_pool.py (100%) rename mx_driving/{perception/fused => point}/ops/bev_pool_v2.py (100%) rename mx_driving/{perception/fused => point}/ops/csrc/BEVPool.cpp (100%) rename mx_driving/{perception/fused => point}/ops/csrc/BEVPoolBackward.cpp (100%) rename mx_driving/{perception/fused => point}/ops/csrc/BEVPoolV2.cpp (100%) rename mx_driving/{perception/fused => point}/ops/csrc/BEVPoolV2Backward.cpp (100%) rename mx_driving/{common => point}/ops/csrc/DynamicScatter.cpp (100%) rename mx_driving/{common => point}/ops/csrc/DynamicVoxelization.cpp (100%) rename mx_driving/{common => point}/ops/csrc/FurthestPointSampling.cpp (100%) rename mx_driving/{common => point}/ops/csrc/FurthestPointSamplingWithDist.cpp (100%) rename mx_driving/{perception => }/point/ops/csrc/GroupPoints.cpp (100%) rename mx_driving/{perception => }/point/ops/csrc/HardVoxelize.cpp (100%) rename mx_driving/{perception => }/point/ops/csrc/PointToVoxel.cpp (100%) rename mx_driving/{perception => }/point/ops/csrc/README.md (100%) rename mx_driving/{perception => }/point/ops/csrc/UniqueVoxel.cpp (100%) rename mx_driving/{perception => }/point/ops/csrc/VecPoolBackward.cpp (100%) rename mx_driving/{common => point}/ops/csrc/VoxelPoolingTrain.cpp (100%) rename mx_driving/{perception => }/point/ops/csrc/VoxelToPoint.cpp (100%) rename mx_driving/{perception => }/point/ops/csrc/functions.h (43%) rename mx_driving/{perception => }/point/ops/csrc/pybind.cpp (31%) rename mx_driving/{common => point}/ops/furthest_point_sampling.py (100%) rename mx_driving/{common => point}/ops/furthest_point_sampling_with_dist.py (100%) rename mx_driving/{perception => }/point/ops/group_points.py (100%) rename mx_driving/{perception/vision => point}/ops/kernels/CMakeLists.txt (100%) rename mx_driving/{perception => }/point/ops/kernels/README.md (100%) rename mx_driving/{perception/vision => point}/ops/kernels/op_host/CMakeLists.txt (100%) rename mx_driving/{perception/fused => point}/ops/kernels/op_host/bev_pool.cpp (100%) rename mx_driving/{perception/fused => point}/ops/kernels/op_host/bev_pool_tiling.h (100%) rename mx_driving/{common => point}/ops/kernels/op_host/dynamic_scatter.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_host/dynamic_scatter_grad.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_host/dynamic_scatter_grad_tiling.h (100%) rename mx_driving/{common => point}/ops/kernels/op_host/dynamic_scatter_tiling.h (100%) rename mx_driving/{common => point}/ops/kernels/op_host/dynamic_voxelization.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_host/dynamic_voxelization_tiling.h (100%) rename mx_driving/{common => point}/ops/kernels/op_host/furthest_point_sampling.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_host/furthest_point_sampling_tiling.h (100%) rename mx_driving/{common => point}/ops/kernels/op_host/furthest_point_sampling_with_dist.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_host/furthest_point_sampling_with_dist_tiling.h (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/group_points.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/group_points_grad.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/group_points_grad_tiling.h (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/group_points_tiling.h (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/hard_voxelize.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/hard_voxelize_tiling.h (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/point_to_voxel.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/point_to_voxel_tiling.h (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/unique_voxel.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/unique_voxel_tiling.h (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/vec_pool_grad.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_host/vec_pool_grad_tiling.h (100%) rename mx_driving/{common => point}/ops/kernels/op_host/voxel_pooling_train.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_host/voxel_pooling_train_grad.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_host/voxel_pooling_train_grad_tiling.h (100%) rename mx_driving/{common => point}/ops/kernels/op_host/voxel_pooling_train_tiling.h (100%) create mode 100644 mx_driving/point/ops/kernels/op_kernel/CMakeLists.txt rename mx_driving/{perception/fused => point}/ops/kernels/op_kernel/bev_pool.cpp (100%) rename mx_driving/{perception/fused => point}/ops/kernels/op_kernel/bev_pool.h (100%) rename mx_driving/{perception/fused => point}/ops/kernels/op_kernel/bev_pool_grad.cpp (100%) rename mx_driving/{perception/fused => point}/ops/kernels/op_kernel/bev_pool_v2.cpp (100%) rename mx_driving/{perception/fused => point}/ops/kernels/op_kernel/bev_pool_v2.h (100%) rename mx_driving/{perception/fused => point}/ops/kernels/op_kernel/bev_pool_v2_grad.cpp (100%) create mode 100644 mx_driving/point/ops/kernels/op_kernel/common.h rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_base.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_grad.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_grad_base.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_grad_max.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_grad_mean.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_grad_sum.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_max.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_mean.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_scatter_sum.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/dynamic_voxelization.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/furthest_point_sampling.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/furthest_point_sampling.h (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/furthest_point_sampling_with_dist.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_kernel/group_points.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_kernel/group_points_grad.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_kernel/hard_voxelize.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_kernel/point_to_voxel.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_kernel/unique_voxel.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_kernel/vec_pool_grad.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/voxel_pooling_train.cpp (100%) rename mx_driving/{common => point}/ops/kernels/op_kernel/voxel_pooling_train_grad.cpp (100%) rename mx_driving/{perception => }/point/ops/kernels/op_kernel/voxel_to_point.cpp (100%) rename mx_driving/{common => point}/ops/npu_dynamic_scatter.py (100%) rename mx_driving/{common => point}/ops/voxel_pooling_train.py (100%) rename mx_driving/{common => point}/ops/voxelization.py (100%) diff --git a/CMakeLists.txt b/CMakeLists.txt index fb306ada..bec89142 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -8,8 +8,10 @@ include(cmake/intf.cmake) set(MX_DRIVING_DIR ${CMAKE_CURRENT_SOURCE_DIR}/mx_driving) add_subdirectory(${MX_DRIVING_DIR}/common) -add_subdirectory(${MX_DRIVING_DIR}/motion) -add_subdirectory(${MX_DRIVING_DIR}/perception) +add_subdirectory(${MX_DRIVING_DIR}/data) +add_subdirectory(${MX_DRIVING_DIR}/fused) +add_subdirectory(${MX_DRIVING_DIR}/point) +add_subdirectory(${MX_DRIVING_DIR}/detection) add_subdirectory(${MX_DRIVING_DIR}/spconv) opbuild(OPS_SRC ${ASCEND_HOST_SRC} OUT_DIR ${ASCEND_AUTOGEN_PATH}) diff --git a/README.md b/README.md index 2a25559e..2101f702 100644 --- a/README.md +++ b/README.md @@ -87,22 +87,27 @@ export LD_LIBRARY_PATH=xxx/site-packages/mx_driving/packages/vendors/customize/o │ │ ├── CMakeLists.txt │ │ ├── components # 通用组件 │ │ └── ops # 通用算子 -│ ├── motion # 运动模块 +│ ├── data # 数据预处理模块 │ │ ├── __init__.py │ │ ├── CMakeLists.txt -│ │ ├── components # 运动组件 -│ │ └── ops # 运动算子 -│ ├── perception # 感知模块 -│ | ├── __init__.py -│ | ├── CMakeLists.txt -│ | ├── fused # 融合模块 -│ | ├── point # 点云模块 -│ | └── vision # 视觉模块 -│ └── spconv # 稀疏卷积模块 +│ │ └── ops # 数据预处理算子 +│ ├── detection # 目标检测模块 +│ │ ├── __init__.py +│ │ ├── CMakeLists.txt +│ │ └── ops # 目标检测算子 +│ ├── point # 点云模块 +│ │ ├── __init__.py +│ │ ├── CMakeLists.txt +│ │ └── ops # 点云算子 +│ ├── fused # 融合模块 +│ │ ├── __init__.py +│ │ ├── CMakeLists.txt +│ │ └── ops # 融合算子 +│ └── spconv # 稀疏卷积模块 │ ├── __init__.py │ ├── CMakeLists.txt | └── ops # 稀疏卷积算子 -├── examples # 自动驾驶模型示例 +├── model_examples # 自动驾驶模型示例 │ └── BEVFormer # BEVFormer模型示例 ├── bind # torch 绑定 ├── ci # ci脚本 diff --git a/bind/pybind.cpp b/bind/pybind.cpp index 13f43d75..ade78570 100644 --- a/bind/pybind.cpp +++ b/bind/pybind.cpp @@ -3,8 +3,9 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { init_common(m); - init_perception_fused(m); - init_perception_point(m); - init_perception_vision(m); + init_fused(m); + init_point(m); + init_data(m); + init_detection(m); init_spconv(m); } diff --git a/docs/api/README.md b/docs/api/README.md index d29c7113..a581aa0b 100644 --- a/docs/api/README.md +++ b/docs/api/README.md @@ -1,5 +1,5 @@ > Note: 以prototype标注的接口，表示该接口为预发布接口，可能会有变动，不建议在生产环境中使用。 -# Common 算子 +# 通用算子 ## scatter_max ### 接口原型 ```python @@ -45,108 +45,188 @@ tensor([[0, 2, 2, 2, 2, 2, 2, 0], [3, 3, 3, 3, 3, 3, 3, 3], [1, 1, 1, 1, 1, 1, 1, 1]]) ``` -## npu_rotated_overlaps +## knn ### 接口原型 ```python -mx_driving.common.npu_rotated_overlaps(Tensor self, Tensor query_boxes, bool trans=False) -> Tensor +mx_driving.common.knn(int k, Tensor xyz, Tensor center_xyz, bool Transposed) -> Tensor ``` ### 功能描述 -计算旋转框的重叠面积。 +對center_xyz中的每個點找到xyz中對應batch中的距離最近的k個點，并且返回此k個點的索引值。 ### 参数说明 -- `self(Tensor)`：边界框张量，数据类型为`float32, float16`，形状为`[B, N, 5]`。 -- `query_boxes(Tensor)`：查询框张量，数据类型为`float32, float16`，形状为`[B, M, 5]`。 -- `trans(bool)`：是否进行坐标变换。默认值为`False`。值为`True`时，表示`xyxyt`, 值为`False`时，表示`xywht`。 +- `xyz(Tensor)`：点数据，表示(x, y, z)三維坐標，数据类型为`float32`。shape为`[B, N, 3]`(当Transposed=False)或`[B, 3, N]`(当Transposed=True)。其中`B`为batch size，`N`为点的数量。 +- `center_xyz(Tensor)`：点数据，表示(x, y, z)三維坐標，数据类型为`float32`。shape为`[B, npoint, 3]`(当Transposed=False)或`[B, 3, npoint]`(当Transposed=True)。其中`B`为batch size，`npoint`为点的数量。 +- `k(int)`：采样点的数量。 +- `Transposed(bool)`: 輸入是否需要進行轉置 ### 返回值 -- `Tensor`：重叠面积张量，数据类型为`float32, float16`，形状为`[B, N, M]`。 +- `idx(Tensor)`：采样后的索引数据，数据类型为`int32`。shape为`[B, k, npoint]`。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 ```python import torch, torch_npu -import numpy as np -from mx_driving.common import npu_rotated_overlaps -a = np.random.uniform(0, 1, (1, 3, 5)).astype(np.float16) -b = np.random.uniform(0, 1, (1, 2, 5)).astype(np.float16) -box1 = torch.from_numpy(a).npu() -box2 = torch.from_numpy(b).npu() -output = npu_rotated_overlaps(box1, box2, True) -print(output) +from mx_driving.common import knn +xyz = torch.tensor([[[1, 1, 1], [1, 1, 1], [1, 1, 1]], [[1, 1, 1], [1, 1, 1], [1, 1, 1]]], dtype=torch.float32).npu() +center_xyz = torch.tensor([[1, 2, 3]], [[1, 2, 3]], dtype=torch.float32).npu() +idx = knn(2, xyz, center_xyz, False) +print(idx) ``` ```text -tensor([[[0.0000, 0.1562, 0.0000], - [0.1562, 0.3713, 0.0611], - [0.0000, 0.0611, 0.0000]]], dtype=torch.float16) +tensor([[0, 0], [1, 1]], dtype=torch.int32) ``` -## npu_rotated_iou +### 算子約束 +1. k必須>0且<100 +2. xyz中的每個batch中的任意一個點到center_xyz對應batch中的任意一個點的距離必須在1e10f以内 +3. xyz和center_xyz的shape必须是3维，当Transposed=True时，xyz和center_xyz的shape的dim的第1位必须是3；当Transposed=False时，xyz和center_xyz的shape的dim的第2位必须是3 +4. 距离相同时索引存在不稳定排序问题,遇到距离精度通过但索引精度错误时，复用不稳定排序的CCB结论 + +## scatter_mean ### 接口原型 ```python -mx_driving.common.npu_rotated_iou(Tensor self, Tensor query_boxes, bool trans=False, int mode=0, bool is_cross=True, float v_threshold=0.0, float e_threshold=0.0) -> Tensor +mx_driving.common.scatter_mean(Tensor src, Tensor indices, int dim=0， Tensor out=None, int dim_size=None) -> Tensor ``` ### 功能描述 -计算旋转框的IoU。 +将输入张量`src`中的元素按照`indices`中的索引在指定的`dim`维进行分组，并计算每组的平均值，返回平均值。 ### 参数说明 -- `self(Tensor)`：边界框张量，数据类型为`float32, float16`，形状为`[B, N, 5]`。 -- `query_boxes(Tensor)`：查询框张量，数据类型为`float32, float16`，形状为`[B, M, 5]`。 -- `trans(bool)`：是否进行坐标变换。默认值为`False`。值为`True`时，表示`xyxyt`, 值为`False`时，表示`xywht`，其中`t`为角度制。 -- `is_cross(bool)`：值为`True`时，则对两组边界框中每个边界框之间进行计算。值为`False`时，只对对齐的边界框之间进行计算。 -- `mode(int)`：计算IoU的模式。默认值为`0`。值为`0`时，表示计算`IoU`，值为`1`时，表示计算`IoF`。 -- `v_threshold(float)`：顶点判断的容忍阈值。 -- `e_threshold(float)`：边相交判断的容忍阈值。 +- `src`：源张量，数据类型为`float32`。 +- `indices`：索引张量，数据类型为`int32`，且 + - `indices`的维度必须小于等于`src`的维度， + - `indices`每一维的长度均必须与`src`长度相同。 + - `indices`的取值必须为非负的有效索引值，参数`out`或`data_size`不为`None`时，`indices`的取值应该为输出张量在`dim`维的有效索引值。 +- `out`：被更新张量，数据类型为`float32`，可选入参，默认为`None`，输入`out`不为`None`时，`out`中的元素参与平均值的计算，且 + - `out`的维度必须与`src`的维度相同。 + - `out`除第`dim`维外其余维的长度必须与`src`相同。 +- `dim`：指定的维度，表示按照哪个维度进行分组平均计算，数据类型为`int32`，可选入参，默认取值为`0`，`dim`取值不超过`indices`的维度。 +- `dim_size`：输出张量在`dim`维的长度，数据类型为`int32`，可选入参，默认为`None`，`dim_size`的取值必须为非负的有效长度值，该参数仅在输入`out`为`None`时生效。 ### 返回值 -- `Tensor`：IoU张量，数据类型为`float32, float16`，`is_cross`为`True`时形状为`[B, N, M]，反之则为`[B, N]`。 +- `out`：求平均后的张量，数据类型为`float32`。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 + ```python import torch, torch_npu -import numpy as np -from mx_driving.common import npu_rotated_iou -a = np.random.uniform(0, 1, (2, 2, 5)).astype(np.float16) -b = np.random.uniform(0, 1, (2, 3, 5)).astype(np.float16) -box1 = torch.from_numpy(a).npu() -box2 = torch.from_numpy(b).npu() -iou = npu_rotated_iou(box1, box2, False, 0, True, 1e-5, 1e-5) -print(iou) +from mx_driving.common import scatter_mean +src = torch.randn(4, 5, 6).to(torch.float) +indices = torch.randint(5, (4, 5)).to(torch.int32) +dim = 0 +out = scatter_mean(src.npu(), indices.npu(), None, dim) +print(out) ``` -```text -tensor([[[3.3325e-01, 1.0162e-01], - [1.0162e-01, 1.0000e+00]], +### 其他说明 +- 该算子对尾块较大的场景较为亲和，对尾块很小的场景不亲和，其中，尾块表示`src`后`N`维的大小，`N = src.dim() - indices.dim()`。 - [[0.0000e+00, 0.0000e+00], - [0.0000e+00, 5.9605e-08]]], dtype=torch.float16) +## three_interpolate +### 接口原型 +```python +mx_driving.common.three_interpolate(features: torch.Tensor, indices: torch.Tensor, + weight: torch.Tensor) -> torch.Tensor: ``` -## npu_dynamic_scatter +### 功能描述 +对三维数据进行加权最近邻线性插值处理 +### 参数说明 +- `features`：需要被插值的特征，数据类型为`float32|float16`，维度为（B, C, M）。 +- `indices`：获取目标特征计算的索引，数据类型为`int32`，维度为（B, N, 3）， + - `indices`的元素值需小于`features`的第三维度，即值在[0, M)。 +- `weight`：获取目标特征计算的权重，数据类型为`float32|float16`，维度为（B, N, 3）。 + - `weight`数据类型与`features`须一致。 +- `features`，`indices`，`weights`三个参数的每个维度须小于10000。 +- `features`，`indices`，`weights`三个参数的大小请勿超过2^24。 +### 返回值 +- `output`：目标特征张量，数据类型为`float32|float16`，维度为（B, C, N）。 +### 支持的型号 +- Atlas A2 训练系列产品 +### 调用示例 +```python +import torch +from mx_driving.common import three_interpolate +features = torch.tensor( + [[[2.4350, 4.7516, 4.4995, 2.4350, 2.4350, 2.4350], + [3.1236, 2.6278, 3.0447, 3.1236, 3.1236, 3.1236], + [2.6732, 2.8677, 2.6436, 2.6732, 2.6732, 2.6732], + [0.0124, 7.0150, 7.0199, 0.0124, 0.0124, 0.0124], + [0.3207, 0.0000, 0.3411, 0.3207, 0.3207, 0.3207]], + [[0.0000, 0.9544, 2.4532, 0.0000, 0.0000, 0.0000], + [0.5346, 1.9176, 1.4715, 0.5346, 0.5346, 0.5346], + [0.0000, 0.2744, 2.0842, 0.0000, 0.0000, 0.0000], + [0.3414, 1.5063, 1.6209, 0.3414, 0.3414, 0.3414], + [0.5814, 0.0103, 0.0000, 0.5814, 0.5814, 0.5814]]], + ).npu() +idx = torch.tensor( + [[[0, 1, 2], [2, 3, 4], [2, 3, 4], [0, 1, 2], [0, 1, 2], [0, 1, 3]], + [[0, 2, 3], [1, 3, 4], [2, 1, 4], [0, 2, 4], [0, 2, 4], [0, 1, 2]]], + ).int().npu() +weight = torch.tensor( + [[[3.3333e-01, 3.3333e-01, 3.3333e-01], + [1.0000e+00, 5.8155e-08, 2.2373e-08], + [1.0000e+00, 1.7737e-08, 1.7356e-08], + [3.3333e-01, 3.3333e-01, 3.3333e-01], + [3.3333e-01, 3.3333e-01, 3.3333e-01], + [3.3333e-01, 3.3333e-01, 3.3333e-01]], + [[3.3333e-01, 3.3333e-01, 3.3333e-01], + [1.0000e+00, 1.3651e-08, 7.7312e-09], + [1.0000e+00, 1.7148e-08, 1.4070e-08], + [3.3333e-01, 3.3333e-01, 3.3333e-01], + [3.3333e-01, 3.3333e-01, 3.3333e-01], + [3.3333e-01, 3.3333e-01, 3.3333e-01]]], + ).npu() +output = three_interpolate(features, idx, weight) +print(output) +``` +```text +torch.tensor( + [[[3.8953e+00, 4.4995e+00, 4.4995e+00, 3.8953e+00, 3.8953e+00, 3.2072e+00], + [2.9320e+00, 3.0447e+00, 3.0447e+00, 2.9320e+00, 2.9320e+00, 2.9583e+00], + [2.7281e+00, 2.6436e+00, 2.6436e+00, 2.7281e+00, 2.7281e+00, 2.7380e+00], + [4.6824e+00, 7.0199e+00, 7.0199e+00, 4.6824e+00, 4.6824e+00, 2.3466e+00], + [2.2060e-01, 3.4110e-01, 3.4110e-01, 2.2060e-01, 2.2060e-01, 2.1380e-01]], + [[8.1773e-01, 9.5440e-01, 2.4532e+00,8.1773e-01, 8.1773e-01, 1.1359e+00], + [8.4689e-01, 1.9176e+00, 1.4715e+00, 8.4689e-01, 8.4689e-01, 1.3079e+00], + [6.9473e-01, 2.7440e-01, 2.0842e+00, 6.9473e-01, 6.9473e-01, 7.8619e-01], + [7.6789e-01, 1.5063e+00, 1.6209e+00, 7.6789e-01, 7.6789e-01, 1.1562e+00], + [3.8760e-01, 1.0300e-02, 8.3569e-09, 3.8760e-01, 3.8760e-01, 1.9723e-01]]], + device='npu:0' + ) +``` + +## three_nn ### 接口原型 ```python -mx_driving.common.npu_dynamic_scatter(Tensor feats, Tensor coors, str reduce_type = 'max') -> Tuple[torch.Tensor, torch.Tensor] +mx_driving.common.three_nn(Tensor target, Tensor source) -> (Tensor dist, Tensor idx) ``` ### 功能描述 -将点云特征点在对应体素中进行特征压缩。 +對target中的每個點找到source中對應batch中的距離最近的3個點，并且返回此3個點的距離和索引值。 ### 参数说明 -- `feats(Tensor)`：点云特征张量[N, C]，仅支持两维，数据类型为`float32`，特征向量`C`长度上限为2048。 -- `coors(Tensor)`：体素坐标映射张量[N, 3]，仅支持两维，数据类型为`int32`，此处以x, y, z指代体素三维坐标，其取值范围为`0 <= x, y <= 2048`, `0 <= z <= 256`。 -- `reduce_type(str)`：压缩类型。可选值为`'max'`, `'mean'`, `'sum'`。默认值为`'max'` +- `target(Tensor)`：点数据，表示(x, y, z)三維坐標，数据类型为`float32/float16`。shape为`[B, npoint, 3]`。其中`B`为batch size，`npoint`为点的数量。 +- `source(Tensor)`：点数据，表示(x, y, z)三維坐標，数据类型为`float32/float16`。shape为`[B, N, 3]`。其中`B`为batch size，`N`为点的数量。 ### 返回值 -- `voxel_feats(Tensor)`：压缩后的体素特征张量，仅支持两维，数据类型为`float32`。 -- `voxel_coors(Tensor)`：去重后的体素坐标，仅支持两维，数据类型为`int32`。 +- `dist(Tensor)`：采样后的索引数据，数据类型为`float32/float16`。shape为`[B, npoint, 3]`。 +- `idx(Tensor)`：采样后的索引数据，数据类型为`int32/int32`。shape为`[B, npoint, 3]`。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import npu_dynamic_scatter +from mx_driving.common import three_nn +source = torch.tensor([[[1, 1, 1], [1, 1, 1], [1, 1, 1]], [[1, 1, 1], [1, 1, 1], [1, 1, 1]]], dtype=torch.float32).npu() +target = torch.tensor([[1, 2, 3]], [[1, 2, 3]], dtype=torch.float32).npu() +dist, idx = three_nn(target, source) +print(dist) +print(idx) +``` +```text +tensor([[2.236, 2.236, 2.236], [2.236, 2.236, 2.236]], dtype=torch.float32) +tensor([[0, 1, 2], [0, 1, 2]], dtype=torch.int32) +``` +### 算子約束 +1. source和target的shape必须是3维，且source和target的shape的dim的第2位必须是3 +2. 距离相同时索引存在不稳定排序问题，遇到距离精度通过但索引精度错误时，复用不稳定排序的CCB结论 -feats = torch.tensor([[1, 2, 3], [3, 2, 1], [7, 8, 9], [9, 8, 7]], dtype=torch.float32).npu() -coors = torch.tensor([[1, 1, 1], [1, 1, 1], [2, 2, 2], [2, 2, 2]], dtype=torch.int32).npu() -voxel_feats, voxel_coors = npu_dynamic_scatter(feats, coors, 'max') -print(voxel_feats) -print(voxel_coors) -``` +# 数据预处理算子 ## npu_points_in_box ### 接口原型 ```python -mx_driving.common.npu_points_in_box(Tensor boxes, Tensor points) -> Tensor +mx_driving.data.npu_points_in_box(Tensor boxes, Tensor points) -> Tensor ``` ### 功能描述 判断点是否在框内。 @@ -162,7 +242,7 @@ mx_driving.common.npu_points_in_box(Tensor boxes, Tensor points) -> Tensor ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import npu_points_in_box +from mx_driving.data import npu_points_in_box boxes = torch.tensor([[[1, 2, 3, 4, 5, 6, 7], [3, 4, 5, 6, 7, 8, 9]]], dtype=torch.float32).npu() points = torch.tensor([[[1, 2, 3], [3, 4, 5]]], dtype=torch.float32).npu() out = npu_points_in_box(boxes, points) @@ -174,7 +254,7 @@ tensor([[0, 1]], dtype=torch.int32) ## npu_points_in_box_all ### 接口原型 ```python -mx_driving.common.npu_points_in_box_all(Tensor boxes, Tensor points) -> Tensor +mx_driving.data.npu_points_in_box_all(Tensor boxes, Tensor points) -> Tensor ``` ### 功能描述 判断点是否在框内。 @@ -190,7 +270,7 @@ mx_driving.common.npu_points_in_box_all(Tensor boxes, Tensor points) -> Tensor ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import npu_points_in_box +from mx_driving.data import npu_points_in_box boxes = torch.tensor([[[1, 2, 3, 4, 5, 6, 7], [3, 4, 5, 6, 7, 8, 9]]], dtype=torch.float32).npu() points = torch.tensor([[[1, 2, 5], [3, 4, 8]]], dtype=torch.float32).npu() out = npu_points_in_box(boxes, points) @@ -200,82 +280,80 @@ print(out) tensor([[[1, 0], [0, 1]]], dtype=torch.int32) ``` -## npu_multi_scale_deformable_attn_function +## RoipointPool3d ### 接口原型 ```python -mx_driving.common.npu_multi_scale_deformable_attn_function(Tensor value, Tensor shape, Tensor offset, Tensor locations, Tensor weight) -> Tensor +mx_driving.data.RoipointPool3d(int num_sampled_points, Tensor points, Tensor point_features, Tensor boxes3d) -> (Tensor pooled_features, Tensor pooled_empty_flag) ``` ### 功能描述 -多尺度可变形注意力机制, 将多个视角的特征图进行融合。 +对每个3D方案的几何特定特征进行编码。 ### 参数说明 -- `value(Tensor)`：特征张量，数据类型为`float32, float16`。shape为`[bs, num_keys, num_heads, embed_dims]`。其中`bs`为batch size，`num_keys`为特征图的大小，`num_heads`为头的数量，`embed_dims`为特征图的维度，需要为8的倍数。 -- `shape(Tensor)`：特征图的形状，数据类型为`int32, int64`。shape为`[num_levels, 2]`。其中`num_levels`为特征图的数量，`2`分别代表`H, W`。 -- `offset(Tensor)`：偏移量张量，数据类型为`int32, int64`。shape为`[num_levels]`。 -- `locations(Tensor)`：位置张量，数据类型为`float32, float16`。shape为`[bs, num_queries, num_heads, num_levels, num_points, 2]`。其中`bs`为batch size，`num_queries`为查询的数量，`num_heads`为头的数量，`num_levels`为特征图的数量，`num_points`为采样点的数量，`2`分别代表`y, x`。 -- `weight(Tensor)`：权重张量，数据类型为`float32, float16`。shape为`[bs, num_queries, num_heads, num_levels, num_points]`。其中`bs`为batch size，`num_queries`为查询的数量，`num_heads`为头的数量，`num_levels`为特征图的数量，`num_points`为采样点的数量。 +- `num_sampled_points(int)`：特征点的数量，正整数。 +- `points(Tensor)`：点张量，数据类型为`float32, float16`。shape 为`[B, N, 3]`。`3`分别代表`x, y, z`。 +- `point_features(Tensor)`：点特征张量，数据类型为`float32, float16`。shape 为`[B, N, C]`。`C`分别代表`x, y, z`。 +- `boxes3d(Tensor)`：框张量，数据类型为`float32, float16`。shape 为`[B, M, 7]`。`7`分别代表`x, y, z, x_size, y_size, z_size, rz`。 ### 返回值 -- `output(Tensor)`：融合后的特征张量，数据类型为`float32, float16`。shape为`[bs, num_queries, num_heads*embed_dims]`。 +- `pooled_features(Tensor)`：点在框内的特征张量，数据类型为`float32, float16`。shape 为`[B, M, num, 3+C]`。 +- `pooled_empty_flag(Tensor)`：所有点不在框内的空标记张量，数据类型为`int32`。shape 为`[B, M]`。 +### 约束说明 +- `points`、`point_features`和`boxes3d`的数据类型必须相同，以及`B`也必须相同。 +- `num_sampled_points`必须小于等于`N`。 +- 数据类型为`float32`时，建议`B`小于100、`N`小于等于2640、`M`小于等于48、`num_sampled_points`小于等于48，个别shape值略微超过建议值无影响，但所有shape值均大于建议值时，算子执行会发生错误。 +- 数据类型为`float16`时，建议`B`小于100、`N`小于等于3360、`M`小于等于60、`num_sampled_points`小于等于60，个别shape值略微超过建议值无影响，但所有shape值均大于建议值时，算子执行会发生错误。 +- `N`/`M`的值越大，性能劣化越严重，建议`N`小于`M`的六百倍，否则性能可能会低于0.1x A100。 ### 支持的型号 - Atlas A2 训练系列产品 -### 约束说明 -- `locations`的值在`[0, 1]`之间。 ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import npu_multi_scale_deformable_attn_function -bs, num_levels, num_heads, num_points, num_queries, embed_dims = 1, 1, 4, 8, 16, 32 - -shapes = torch.as_tensor([(100, 100)], dtype=torch.long) -num_keys = sum((H * W).item() for H, W in shapes) - -value = torch.rand(bs, num_keys, num_heads, embed_dims) * 0.01 -sampling_locations = torch.ones(bs, num_queries, num_heads, num_levels, num_points, 2) * 0.005 -attention_weights = torch.rand(bs, num_queries, num_heads, num_levels, num_points) + 1e-5 -level_start_index = torch.cat((shapes.new_zeros((1, )), shapes.prod(1).cumsum(0)[:-1])) - -out = npu_multi_scale_deformable_attn_function(value.npu(), shapes.npu(), level_start_index.npu(), sampling_locations.npu(), attention_weights.npu()) -print(out) +from mx_driving.data import RoIPointPool3d +num_sampled_points = 1 +points = torch.tensor([[[1, 2, 3]]], dtype=torch.float).npu() +point_features = points.clone() +boxes3d = torch.tensor([[[1, 2, 3, 4, 5, 6, 1]]], dtype=torch.float).npu() +roipoint_pool3d = RoIPointPool3d(num_sampled_points) +pooled_features, pooled_empty_flag = roipoint_pool3d(points, point_features, boxes3d) +print(pooled_features) +print(pooled_empty_flag) ``` ```text -tensor([[[9.3002, 11.1603, 0.0000, 0.0000]]], dtype=torch.float32) +tensor([[[[1., 2., 3., 1., 2., 3.]]]]) +tensor([[0]], dtype=torch.int32) ``` -## voxelization + +# 目标检测算子 +## boxes_overlap_bev ### 接口原型 ```python -mx_driving.common.voxelization(Tensor points, List[float] voxel_size, List[float] coors_range, int max_points=-1, int max_voxels=-1, bool deterministic=True) -> Tensor +mx_driving.detection.boxes_overlap_bev(Tensor boxes_a, Tensor boxes_b) -> Tensor ``` ### 功能描述 -将点云数据进行体素化。 +Calculates the intersection of bounding boxes in Bird's Eye View. ### 参数说明 -- `points(Tensor)`：点云数据，数据类型为`float32`。shape为`[N, F]`。其中`N`为点的数量，`F`分别代表每个点的特征维度，其中`N > 0, F >= 3`。 -- `voxel_size(List[float])`：体素大小，数据类型为`float32`。shape为`[3]`。其中`3`分别代表`x, y, z`。 -- `coors_range(List[float])`：体素范围，数据类型为`float32`。shape为`[6]`。其中`6`分别代表`x_min, y_min, z_min, x_max, y_max, z_max`。 -- `max_points(int)`：每个体素的最大点数。默认值为`-1`。 -- `max_voxels(int)`：最大体素数。默认值为`-1`。 -- `deterministic(bool)`：是否确定性。默认值为`True`。 +- `boxes_a (Tensor)`：第一组bounding boxes，数据类型为`float32`。shape为`[M, 5]`。其中`5`分别代表`x1, y1, x2, y2, angle`, `x1, y1, x2, y2`代表box四个顶点的横纵坐标，`angle`代表box的弧度制旋转角。 +- `boxes_b (Tensor)`：第二组bounding boxes，数据类型为`float32`。shape为`[N, 5]`。其中`5`分别代表`x1, y1, x2, y2, angle`, `x1, y1, x2, y2`代表box四个顶点的横纵坐标，`angle`代表box的弧度制旋转角。 ### 返回值 -- `coors(Tensor)`：每个点所属的体素坐标，数据类型为`int32`。shape为`[N, 3]`。 +- `area_overlap(Tensor)`：包含两组bounding boxes交叠面积的张量，数据类型为`float32`。shape为`[M, N]`。 +### 约束说明 + ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import Voxelization -points = torch.randint(-20, 100, [16, 3], dtype=torch.float32).npu() -coors_range = [0, -40, -3, 70.4, 40, 1] -max_points = -1 -voxel_size = [0.5, 0.5, 0.5] -dynamic_voxelization = Voxelization(voxel_size, coors_range, max_points) -out = dynamic_voxelization.forward(points) -print(out) +from mx_driving.detection import boxes_overlap_bev +boxes_a = torch.tensor([[0, 0, 2, 2, 0]], dtype=torch.float32).npu() +boxes_b = torch.tensor([[1, 1, 3, 3, 0]], dtype=torch.float32).npu() +area_overlap = boxes_overlap_bev(boxes_a, boxes_b) +print(area_overlap) ``` -## npu_nms3d_normal +## npu_nms3d ### 接口原型 ```python -mx_driving.common.npu_nms3d_normal(Tensor boxes, Tensor scores, float: iou_threshold) -> Tensor +mx_driving.detection.npu_nms3d(Tensor boxes, Tensor scores, float: iou_threshold) -> Tensor ``` ### 功能描述 -3D非极大值抑制。 +3D非极大值抑制，在bev视角下剔除多个3d box交并比大于阈值的box。 ### 参数说明 - `boxes(Tensor)`：框张量，数据类型为`float32, float16`。shape 为`[N, 7]`。`7`分别代表`x, y, z, x_size, y_size, z_size, rz`。 - `scores(Tensor)`：评分张量，数据类型为`float32, float16`。shape 为`[N]`。 @@ -287,22 +365,23 @@ mx_driving.common.npu_nms3d_normal(Tensor boxes, Tensor scores, float: iou_thres ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import npu_nms3d_normal +from mx_driving.detection import npu_nms3d boxes = torch.tensor([[1, 2, 3, 4, 5, 6, 7], [3, 4, 5, 6, 7, 8, 9]], dtype=torch.float32).npu() scores = torch.tensor([1, 2], dtype=torch.float32).npu() -out = npu_nms3d_normal(boxes, scores, 0.5) +out = npu_nms3d(boxes, scores, 0.5) print(out) ``` ```text -tensor([[1, 0]], dtype=torch.int32) +tensor([[1]], dtype=torch.int32) ``` -## npu_nms3d + +## npu_nms3d_normal ### 接口原型 ```python -mx_driving.common.npu_nms3d(Tensor boxes, Tensor scores, float: iou_threshold) -> Tensor +mx_driving.detection.npu_nms3d_normal(Tensor boxes, Tensor scores, float: iou_threshold) -> Tensor ``` ### 功能描述 -3D非极大值抑制，在bev视角下剔除多个3d box交并比大于阈值的box。 +3D非极大值抑制。 ### 参数说明 - `boxes(Tensor)`：框张量，数据类型为`float32, float16`。shape 为`[N, 7]`。`7`分别代表`x, y, z, x_size, y_size, z_size, rz`。 - `scores(Tensor)`：评分张量，数据类型为`float32, float16`。shape 为`[N]`。 @@ -314,241 +393,157 @@ mx_driving.common.npu_nms3d(Tensor boxes, Tensor scores, float: iou_threshold) - ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import npu_nms3d +from mx_driving.detection import npu_nms3d_normal boxes = torch.tensor([[1, 2, 3, 4, 5, 6, 7], [3, 4, 5, 6, 7, 8, 9]], dtype=torch.float32).npu() scores = torch.tensor([1, 2], dtype=torch.float32).npu() -out = npu_nms3d(boxes, scores, 0.5) -print(out) -``` -```text -tensor([[1]], dtype=torch.int32) -``` -## npu_furthest_point_sampling -### 接口原型 -```python -mx_driving.common.npu_furthest_point_sampling(Tensor points, int num_points) -> Tensor -``` -### 功能描述 -点云数据的最远点采样。 -### 参数说明 -- `points(Tensor)`：点云数据，数据类型为`float32`。shape为`[B, N, 3]`。其中`B`为batch size，`N`为点的数量，`3`分别代表`x, y, z`。 -- `num_points(int)`：采样点的数量。 -### 返回值 -- `Tensor`：采样后的点云数据，数据类型为`float32`。shape为`[B, num_points]`。 -### 支持的型号 -- Atlas A2 训练系列产品 -### 调用示例 -```python -import torch, torch_npu -from mx_driving.common import npu_furthest_point_sampling -points = torch.tensor([[[1, 2, 3], [4, 5, 6], [7, 8, 9]]], dtype=torch.float32).npu() -out = npu_furthest_point_sampling(points, 2) +out = npu_nms3d_normal(boxes, scores, 0.5) print(out) ``` ```text -tensor([[0, 2]], dtype=torch.int32) +tensor([[1, 0]], dtype=torch.int32) ``` -### 算子約束 -1. points输入shape[B, N, 3]的总大小(B x N x 3)不应该超过383166 -## furthest_point_sample_with_dist +## npu_rotated_iou ### 接口原型 ```python -mx_driving.common.furthest_point_sample_with_dist(Tensor points, int num_points) -> (Tensor, Tensor) +mx_driving.detection.npu_rotated_iou(Tensor self, Tensor query_boxes, bool trans=False, int mode=0, bool is_cross=True, float v_threshold=0.0, float e_threshold=0.0) -> Tensor ``` ### 功能描述 -与`npu_furthest_point_sampling`功能相同，但输入略有不同。 -### 参数说明 -- `points(Tensor)`：点云数据，表示各点间的距离，数据类型为`float32`。shape为`[B, N, N]`。其中`B`为batch size，`N`为点的数量。 -- `num_points(int)`：采样点的数量。 +计算旋转框的IoU。 +### 参数说明 +- `self(Tensor)`：边界框张量，数据类型为`float32, float16`，形状为`[B, N, 5]`。 +- `query_boxes(Tensor)`：查询框张量，数据类型为`float32, float16`，形状为`[B, M, 5]`。 +- `trans(bool)`：是否进行坐标变换。默认值为`False`。值为`True`时，表示`xyxyt`, 值为`False`时，表示`xywht`，其中`t`为角度制。 +- `is_cross(bool)`：值为`True`时，则对两组边界框中每个边界框之间进行计算。值为`False`时，只对对齐的边界框之间进行计算。 +- `mode(int)`：计算IoU的模式。默认值为`0`。值为`0`时，表示计算`IoU`，值为`1`时，表示计算`IoF`。 +- `v_threshold(float)`：顶点判断的容忍阈值。 +- `e_threshold(float)`：边相交判断的容忍阈值。 ### 返回值 -- `Tensor`：采样后的点云数据，数据类型为`float32`。shape为`[B, num_points]`。 +- `Tensor`：IoU张量，数据类型为`float32, float16`，`is_cross`为`True`时形状为`[B, N, M]，反之则为`[B, N]`。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import furthest_point_sample_with_dist -points = torch.tensor([[[1, 2, 3], [4, 5, 6], [7, 8, 9]]], dtype=torch.float32).npu() -out = furthest_point_sample_with_dist(points, 2) -print(out) +import numpy as np +from mx_driving.detection import npu_rotated_iou +a = np.random.uniform(0, 1, (2, 2, 5)).astype(np.float16) +b = np.random.uniform(0, 1, (2, 3, 5)).astype(np.float16) +box1 = torch.from_numpy(a).npu() +box2 = torch.from_numpy(b).npu() +iou = npu_rotated_iou(box1, box2, False, 0, True, 1e-5, 1e-5) +print(iou) ``` ```text -tensor([[0, 2]], dtype=torch.int32) -``` +tensor([[[3.3325e-01, 1.0162e-01], + [1.0162e-01, 1.0000e+00]], -## three_interpolate + [[0.0000e+00, 0.0000e+00], + [0.0000e+00, 5.9605e-08]]], dtype=torch.float16) +``` +## npu_rotated_overlaps ### 接口原型 ```python -mx_driving.common.three_interpolate(features: torch.Tensor, indices: torch.Tensor, - weight: torch.Tensor) -> torch.Tensor: +mx_driving.detection.npu_rotated_overlaps(Tensor self, Tensor query_boxes, bool trans=False) -> Tensor ``` ### 功能描述 -对三维数据进行加权最近邻线性插值处理 +计算旋转框的重叠面积。 ### 参数说明 -- `features`：需要被插值的特征，数据类型为`float32|float16`，维度为（B, C, M）。 -- `indices`：获取目标特征计算的索引，数据类型为`int32`，维度为（B, N, 3）， - - `indices`的元素值需小于`features`的第三维度，即值在[0, M)。 -- `weight`：获取目标特征计算的权重，数据类型为`float32|float16`，维度为（B, N, 3）。 - - `weight`数据类型与`features`须一致。 -- `features`，`indices`，`weights`三个参数的每个维度须小于10000。 -- `features`，`indices`，`weights`三个参数的大小请勿超过2^24。 +- `self(Tensor)`：边界框张量，数据类型为`float32, float16`，形状为`[B, N, 5]`。 +- `query_boxes(Tensor)`：查询框张量，数据类型为`float32, float16`，形状为`[B, M, 5]`。 +- `trans(bool)`：是否进行坐标变换。默认值为`False`。值为`True`时，表示`xyxyt`, 值为`False`时，表示`xywht`。 ### 返回值 -- `output`：目标特征张量，数据类型为`float32|float16`，维度为（B, C, N）。 +- `Tensor`：重叠面积张量，数据类型为`float32, float16`，形状为`[B, N, M]`。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 ```python -import torch -from mx_driving.common import three_interpolate -features = torch.tensor( - [[[2.4350, 4.7516, 4.4995, 2.4350, 2.4350, 2.4350], - [3.1236, 2.6278, 3.0447, 3.1236, 3.1236, 3.1236], - [2.6732, 2.8677, 2.6436, 2.6732, 2.6732, 2.6732], - [0.0124, 7.0150, 7.0199, 0.0124, 0.0124, 0.0124], - [0.3207, 0.0000, 0.3411, 0.3207, 0.3207, 0.3207]], - [[0.0000, 0.9544, 2.4532, 0.0000, 0.0000, 0.0000], - [0.5346, 1.9176, 1.4715, 0.5346, 0.5346, 0.5346], - [0.0000, 0.2744, 2.0842, 0.0000, 0.0000, 0.0000], - [0.3414, 1.5063, 1.6209, 0.3414, 0.3414, 0.3414], - [0.5814, 0.0103, 0.0000, 0.5814, 0.5814, 0.5814]]], - ).npu() -idx = torch.tensor( - [[[0, 1, 2], [2, 3, 4], [2, 3, 4], [0, 1, 2], [0, 1, 2], [0, 1, 3]], - [[0, 2, 3], [1, 3, 4], [2, 1, 4], [0, 2, 4], [0, 2, 4], [0, 1, 2]]], - ).int().npu() -weight = torch.tensor( - [[[3.3333e-01, 3.3333e-01, 3.3333e-01], - [1.0000e+00, 5.8155e-08, 2.2373e-08], - [1.0000e+00, 1.7737e-08, 1.7356e-08], - [3.3333e-01, 3.3333e-01, 3.3333e-01], - [3.3333e-01, 3.3333e-01, 3.3333e-01], - [3.3333e-01, 3.3333e-01, 3.3333e-01]], - [[3.3333e-01, 3.3333e-01, 3.3333e-01], - [1.0000e+00, 1.3651e-08, 7.7312e-09], - [1.0000e+00, 1.7148e-08, 1.4070e-08], - [3.3333e-01, 3.3333e-01, 3.3333e-01], - [3.3333e-01, 3.3333e-01, 3.3333e-01], - [3.3333e-01, 3.3333e-01, 3.3333e-01]]], - ).npu() -output = three_interpolate(features, idx, weight) +import torch, torch_npu +import numpy as np +from mx_driving.detection import npu_rotated_overlaps +a = np.random.uniform(0, 1, (1, 3, 5)).astype(np.float16) +b = np.random.uniform(0, 1, (1, 2, 5)).astype(np.float16) +box1 = torch.from_numpy(a).npu() +box2 = torch.from_numpy(b).npu() +output = npu_rotated_overlaps(box1, box2, True) print(output) ``` ```text -torch.tensor( - [[[3.8953e+00, 4.4995e+00, 4.4995e+00, 3.8953e+00, 3.8953e+00, 3.2072e+00], - [2.9320e+00, 3.0447e+00, 3.0447e+00, 2.9320e+00, 2.9320e+00, 2.9583e+00], - [2.7281e+00, 2.6436e+00, 2.6436e+00, 2.7281e+00, 2.7281e+00, 2.7380e+00], - [4.6824e+00, 7.0199e+00, 7.0199e+00, 4.6824e+00, 4.6824e+00, 2.3466e+00], - [2.2060e-01, 3.4110e-01, 3.4110e-01, 2.2060e-01, 2.2060e-01, 2.1380e-01]], - [[8.1773e-01, 9.5440e-01, 2.4532e+00,8.1773e-01, 8.1773e-01, 1.1359e+00], - [8.4689e-01, 1.9176e+00, 1.4715e+00, 8.4689e-01, 8.4689e-01, 1.3079e+00], - [6.9473e-01, 2.7440e-01, 2.0842e+00, 6.9473e-01, 6.9473e-01, 7.8619e-01], - [7.6789e-01, 1.5063e+00, 1.6209e+00, 7.6789e-01, 7.6789e-01, 1.1562e+00], - [3.8760e-01, 1.0300e-02, 8.3569e-09, 3.8760e-01, 3.8760e-01, 1.9723e-01]]], - device='npu:0' - ) +tensor([[[0.0000, 0.1562, 0.0000], + [0.1562, 0.3713, 0.0611], + [0.0000, 0.0611, 0.0000]]], dtype=torch.float16) ``` - -## voxel_pooling_train +# 融合算子 +## npu_multi_scale_deformable_attn_function ### 接口原型 ```python -mx_driving.common.npu_voxel_pooling_train(Tensor geom_xyz, Tensor input_features, List[int] voxel_num) -> Tensor +mx_driving.fused.npu_multi_scale_deformable_attn_function(Tensor value, Tensor shape, Tensor offset, Tensor locations, Tensor weight) -> Tensor ``` ### 功能描述 -点云数据体素化。 +多尺度可变形注意力机制, 将多个视角的特征图进行融合。 ### 参数说明 -- `geom_xyz`：体素坐标，数据类型为`int32`，维度为（B, N, 3）, 3表示x, y, z。 -- `input_features`：点云数据，数据类型为`float32|float16`，维度为（B, N, C）。 -- `voxel_num`：体素格子长宽高，数据类型为`int32`，维度为（3），3表示体素格子的长宽高。 +- `value(Tensor)`：特征张量，数据类型为`float32, float16`。shape为`[bs, num_keys, num_heads, embed_dims]`。其中`bs`为batch size，`num_keys`为特征图的大小，`num_heads`为头的数量，`embed_dims`为特征图的维度，需要为8的倍数。 +- `shape(Tensor)`：特征图的形状，数据类型为`int32, int64`。shape为`[num_levels, 2]`。其中`num_levels`为特征图的数量，`2`分别代表`H, W`。 +- `offset(Tensor)`：偏移量张量，数据类型为`int32, int64`。shape为`[num_levels]`。 +- `locations(Tensor)`：位置张量，数据类型为`float32, float16`。shape为`[bs, num_queries, num_heads, num_levels, num_points, 2]`。其中`bs`为batch size，`num_queries`为查询的数量，`num_heads`为头的数量，`num_levels`为特征图的数量，`num_points`为采样点的数量，`2`分别代表`y, x`。 +- `weight(Tensor)`：权重张量，数据类型为`float32, float16`。shape为`[bs, num_queries, num_heads, num_levels, num_points]`。其中`bs`为batch size，`num_queries`为查询的数量，`num_heads`为头的数量，`num_levels`为特征图的数量，`num_points`为采样点的数量。 ### 返回值 -- `output(Tensor)`：输出结果，数据类型为`float32|float16`。shape为`[B, num_voxel_y, num_voxel_x, C]`。 -### 约束说明 -- B <= 128 -- N <= 100000 -- C <= 256 -- num_voxel_x <= 1000 -- num_voxel_y <= 1000 -- num_voxel_z <= 10 -- B * num_voxel_y * num_voxel_x * C <= 100000000 -- B * N * C <= 100000000 -- 反向具有相同约束。 +- `output(Tensor)`：融合后的特征张量，数据类型为`float32, float16`。shape为`[bs, num_queries, num_heads*embed_dims]`。 ### 支持的型号 - Atlas A2 训练系列产品 +### 约束说明 +- `locations`的值在`[0, 1]`之间。 ### 调用示例 ```python -import torch -import torch_npu -import mx_driving.common - -def gen_data(geom_shape, feature_shape, coeff, batch_size, num_channels, dtype): - geom_xyz = torch.rand(geom_shape) * coeff - geom_xyz = geom_xyz.reshape(batch_size, -1, 3) - geom_xyz[:, :, 2] /= 100 - geom_xyz_cpu = geom_xyz.int() - features = torch.rand(feature_shape, dtype=dtype) - 0.5 - features_cpu = features.reshape(batch_size, -1, num_channels) - - return geom_xyz_cpu, features_cpu - -dtype = torch.float32 -coeff = 90 -voxel_num = [128, 128, 1] -batch_size = 2 -num_points = 40 -num_channel = 80 -xyz = 3 - -geom_shape = [batch_size, num_points, xyz] -feature_shape = [batch_size, num_points, num_channel] +import torch, torch_npu +from mx_driving.fused import npu_multi_scale_deformable_attn_function +bs, num_levels, num_heads, num_points, num_queries, embed_dims = 1, 1, 4, 8, 16, 32 -geom_cpu, feature_cpu = gen_data(geom_shape, feature_shape, coeff, batch_size, num_channel, dtype) +shapes = torch.as_tensor([(100, 100)], dtype=torch.long) +num_keys = sum((H * W).item() for H, W in shapes) -geom_npu = geom_cpu.npu() -feature_npu = feature_cpu.npu() +value = torch.rand(bs, num_keys, num_heads, embed_dims) * 0.01 +sampling_locations = torch.ones(bs, num_queries, num_heads, num_levels, num_points, 2) * 0.005 +attention_weights = torch.rand(bs, num_queries, num_heads, num_levels, num_points) + 1e-5 +level_start_index = torch.cat((shapes.new_zeros((1, )), shapes.prod(1).cumsum(0)[:-1])) -result_npu = mx_driving.common.npu_voxel_pooling_train(geom_npu, feature_npu, voxel_num) +out = npu_multi_scale_deformable_attn_function(value.npu(), shapes.npu(), level_start_index.npu(), sampling_locations.npu(), attention_weights.npu()) +print(out) +``` +```text +tensor([[[9.3002, 11.1603, 0.0000, 0.0000]]], dtype=torch.float32) ``` -## scatter_mean +## npu_max_pool2d ### 接口原型 ```python -mx_driving.common.scatter_mean(Tensor src, Tensor indices, int dim=0， Tensor out=None, int dim_size=None) -> Tensor +mx_driving.fused.npu_max_pool2d(Tensor x, int kernel_size, int stride, int padding) -> Tensor ``` ### 功能描述 -将输入张量`src`中的元素按照`indices`中的索引在指定的`dim`维进行分组，并计算每组的平均值，返回平均值。 +Performs max pooling on the input and outputs max values. ### 参数说明 -- `src`：源张量，数据类型为`float32`。 -- `indices`：索引张量，数据类型为`int32`，且 - - `indices`的维度必须小于等于`src`的维度， - - `indices`每一维的长度均必须与`src`长度相同。 - - `indices`的取值必须为非负的有效索引值，参数`out`或`data_size`不为`None`时，`indices`的取值应该为输出张量在`dim`维的有效索引值。 -- `out`：被更新张量，数据类型为`float32`，可选入参，默认为`None`，输入`out`不为`None`时，`out`中的元素参与平均值的计算，且 - - `out`的维度必须与`src`的维度相同。 - - `out`除第`dim`维外其余维的长度必须与`src`相同。 -- `dim`：指定的维度，表示按照哪个维度进行分组平均计算，数据类型为`int32`，可选入参，默认取值为`0`，`dim`取值不超过`indices`的维度。 -- `dim_size`：输出张量在`dim`维的长度，数据类型为`int32`，可选入参，默认为`None`，`dim_size`的取值必须为非负的有效长度值，该参数仅在输入`out`为`None`时生效。 +- `x (Tensor)`：一组待池化对象，数据类型为`float32`，format为NCHW。 ### 返回值 -- `out`：求平均后的张量，数据类型为`float32`。 +- `y (Tensor)`：池化后的最大值，数据类型为`float32`，format为NCHW。 +### 约束说明 +kernel_size仅支持3，stride仅支持2，padding仅支持1，且输入C轴数据量要求为8的倍数，H和W需要大于100。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 - ```python import torch, torch_npu -from mx_driving.common import scatter_mean -src = torch.randn(4, 5, 6).to(torch.float) -indices = torch.randint(5, (4, 5)).to(torch.int32) -dim = 0 -out = scatter_mean(src.npu(), indices.npu(), None, dim) -print(out) +from mx_driving.fused import npu_max_pool2d +kernel_size = 3 +stride = 2 +padding = 1 +x = torch.randn(18, 64, 464, 800).npu() +res = npu_max_pool2d(x, kernel_size, stride, padding) +print(res) ``` -### 其他说明 -- 该算子对尾块较大的场景较为亲和，对尾块很小的场景不亲和，其中，尾块表示`src`后`N`维的大小，`N = src.dim() - indices.dim()`。 ## npu_deformable_aggregation ### 接口原型 ```python -mx_driving.common.npu_deformable_aggregation(Tensor feature_maps, Tensor spatial_shape, Tensor scale_start_index, Tensor sample_locations, Tensor weight) -> Tensor +mx_driving.fused.npu_deformable_aggregation(Tensor feature_maps, Tensor spatial_shape, Tensor scale_start_index, Tensor sample_locations, Tensor weight) -> Tensor ``` ### 功能描述 可变形聚合，对于每个锚点实例，对多个关键点的多时间戳、视图、缩放特征进行稀疏采样后分层融合为实例特征，实现精确的锚点细化。 @@ -577,7 +572,7 @@ mx_driving.common.npu_deformable_aggregation(Tensor feature_maps, Tensor spatial ### 调用示例 ```python import torch, torch_npu -import mx_driving.common +import mx_driving.fused bs, num_feat, c, cam, anchor, pts, scale, group = 1, 2816, 256, 1, 10, 2000, 1, 8 @@ -587,15 +582,15 @@ scale_start_index = torch.tensor([[[0]]]) sampling_location = torch.rand(bs, anchor, pts, cam, 2) weights = torch.randn(bs, anchor, pts, cam, scale, group) -out = mx_driving.common.npu_deformable_aggregation(feature_maps.npu(), spatial_shape.npu(), scale_start_index.npu(), sampling_location.npu(), weights.npu()) +out = mx_driving.fused.npu_deformable_aggregation(feature_maps.npu(), spatial_shape.npu(), scale_start_index.npu(), sampling_location.npu(), weights.npu()) print(out) ``` -# Perception 算子 +# 点云算子 ## bev_pool ### 接口原型 ```python -mx_driving.perception.fused.bev_pool(Tensor feat, Tensor geom_feat, int B, int D, int H, int W) -> Tensor +mx_driving.point.bev_pool(Tensor feat, Tensor geom_feat, int B, int D, int H, int W) -> Tensor ``` ### 功能描述 BEV池化。可参考论文`BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation` @@ -619,7 +614,7 @@ BEV池化。可参考论文`BEVFusion: Multi-Task Multi-Sensor Fusion with Unifi ### 调用示例 ```python import torch, torch_npu -from mx_driving.perception.fused import bev_pool +from mx_driving.point import bev_pool feat = torch.rand(4, 256).npu() feat.requires_grad_() geom_feat = torch.tensor([[0, 0, 0, 0], [0, 0, 0, 1], [0, 0, 0, 2], [0, 0, 0, 3]], dtype=torch.int32).npu() @@ -628,11 +623,10 @@ loss = bev_pooled_feat.sum() loss.backward() print(feat.grad) ``` - ## bev_pool_v2 ### 接口原型 ```python -mx_driving.perception.fused.bev_pool_v2(Tensor depth, feat, Tensor ranks_depth, Tensor ranks_feat, Tensor ranks_bev, +mx_driving.point.bev_pool_v2(Tensor depth, feat, Tensor ranks_depth, Tensor ranks_feat, Tensor ranks_bev, List[int] bev_feat_shape, Tensor interval_starts, Tensor interval_lengths) -> Tensor ``` ### 功能描述 @@ -647,41 +641,92 @@ BEV池化优化版。可参考论文`BEVDet: High-performance Multi-camera 3D Ob - `interval_starts(Tensor)`：间隔开始张量，数据类型为`int32`。shape为`[N_INTERVALS]`。 - `interval_lengths(Tensor)`：间隔长度张量，数据类型为`int32`。shape为`[N_INTERVALS]`。 ### 返回值 -- `bev_pooled_feat(Tensor)`：BEV池化后的特征张量，数据类型为`float32`。shape为`[B, D, H, W, C]`。 -### 约束说明 -- `ranks_depth`的值必须在`[0, B*B*D*H*W]`之间。 -- `ranks_feat`的值必须在`[0, B*N*H*W]`之间。 -- `ranks_bev`的值必须在`[0, B*D*H*W]`之间。 -- C <= 1024 -- B * D * H * W * C <= 2^31, B, D <= 8, H, W <= 256 -- N_RANKS <= 2^21 -- 对于反向也是同样的约束。 +- `bev_pooled_feat(Tensor)`：BEV池化后的特征张量，数据类型为`float32`。shape为`[B, D, H, W, C]`。 +### 约束说明 +- `ranks_depth`的值必须在`[0, B*B*D*H*W]`之间。 +- `ranks_feat`的值必须在`[0, B*N*H*W]`之间。 +- `ranks_bev`的值必须在`[0, B*D*H*W]`之间。 +- C <= 1024 +- B * D * H * W * C <= 2^31, B, D <= 8, H, W <= 256 +- N_RANKS <= 2^21 +- 对于反向也是同样的约束。 +### 支持的型号 +- Atlas A2 训练系列产品 +### 调用示例 +```python +import torch, torch_npu +from mx_driving.point import bev_pool_v2 +depth = torch.rand(2, 1, 8, 256, 256).npu() +feat = torch.rand(2, 1, 256, 256, 64).npu() +feat.requires_grad_() +ranks_depth = torch.tensor([0, 1], dtype=torch.int32).npu() +ranks_feat = torch.tensor([0, 1], dtype=torch.int32).npu() +ranks_bev = torch.tensor([0, 1], dtype=torch.int32).npu() +bev_feat_shape = [2, 8, 256, 256, 64] +interval_starts = torch.tensor([0], dtype=torch.int32).npu() +interval_lengths = torch.tensor([2], dtype=torch.int32).npu() +bev_pooled_feat = bev_pool_v2(depth, feat, ranks_depth, ranks_feat, ranks_bev, bev_feat_shape, interval_starts, interval_lengths) +loss = bev_pooled_feat.sum() +loss.backward() +print(loss) +print(feat.grad) +``` +## furthest_point_sample_with_dist +### 接口原型 +```python +mx_driving.point.furthest_point_sample_with_dist(Tensor points, int num_points) -> (Tensor, Tensor) +``` +### 功能描述 +与`npu_furthest_point_sampling`功能相同，但输入略有不同。 +### 参数说明 +- `points(Tensor)`：点云数据，表示各点间的距离，数据类型为`float32`。shape为`[B, N, N]`。其中`B`为batch size，`N`为点的数量。 +- `num_points(int)`：采样点的数量。 +### 返回值 +- `Tensor`：采样后的点云数据，数据类型为`float32`。shape为`[B, num_points]`。 +### 支持的型号 +- Atlas A2 训练系列产品 +### 调用示例 +```python +import torch, torch_npu +from mx_driving.point import furthest_point_sample_with_dist +points = torch.tensor([[[1, 2, 3], [4, 5, 6], [7, 8, 9]]], dtype=torch.float32).npu() +out = furthest_point_sample_with_dist(points, 2) +print(out) +``` +```text +tensor([[0, 2]], dtype=torch.int32) +``` +## npu_furthest_point_sampling +### 接口原型 +```python +mx_driving.point.npu_furthest_point_sampling(Tensor points, int num_points) -> Tensor +``` +### 功能描述 +点云数据的最远点采样。 +### 参数说明 +- `points(Tensor)`：点云数据，数据类型为`float32`。shape为`[B, N, 3]`。其中`B`为batch size，`N`为点的数量，`3`分别代表`x, y, z`。 +- `num_points(int)`：采样点的数量。 +### 返回值 +- `Tensor`：采样后的点云数据，数据类型为`float32`。shape为`[B, num_points]`。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 ```python import torch, torch_npu -from mx_driving.perception.fused import bev_pool_v2 -depth = torch.rand(2, 1, 8, 256, 256).npu() -feat = torch.rand(2, 1, 256, 256, 64).npu() -feat.requires_grad_() -ranks_depth = torch.tensor([0, 1], dtype=torch.int32).npu() -ranks_feat = torch.tensor([0, 1], dtype=torch.int32).npu() -ranks_bev = torch.tensor([0, 1], dtype=torch.int32).npu() -bev_feat_shape = [2, 8, 256, 256, 64] -interval_starts = torch.tensor([0], dtype=torch.int32).npu() -interval_lengths = torch.tensor([2], dtype=torch.int32).npu() -bev_pooled_feat = bev_pool_v2(depth, feat, ranks_depth, ranks_feat, ranks_bev, bev_feat_shape, interval_starts, interval_lengths) -loss = bev_pooled_feat.sum() -loss.backward() -print(loss) -print(feat.grad) +from mx_driving.point import npu_furthest_point_sampling +points = torch.tensor([[[1, 2, 3], [4, 5, 6], [7, 8, 9]]], dtype=torch.float32).npu() +out = npu_furthest_point_sampling(points, 2) +print(out) ``` - +```text +tensor([[0, 2]], dtype=torch.int32) +``` +### 算子约束 +1. points输入shape[B, N, 3]的总大小(B x N x 3)不应该超过383166 ## group_points ### 接口原型 ```python -mx_driving.perception.point.npu_group_points(Tensor features, Tensor indices) -> Tensor +mx_driving.point.npu_group_points(Tensor features, Tensor indices) -> Tensor ``` ### 功能描述 点云数据按照索引重新分组。 @@ -700,7 +745,7 @@ mx_driving.perception.point.npu_group_points(Tensor features, Tensor indices) -> ```python import torch import torch_npu -import mx_driving.perception.point +import mx_driving.point indices = torch.tensor([[[0, 2, 5, 5], [1, 0, 5, 0], [2, 1, 4, 4]]]).int().npu() features = torch.tensor([[[0.9178, -0.7250, -1.6587, 0.0715, -0.2252, 0.4994], [0.6190, 0.1755, -1.7902, -0.5852, -0.3311, 1.9764], @@ -709,7 +754,7 @@ features = torch.tensor([[[0.9178, -0.7250, -1.6587, 0.0715, -0.2252, 0.4994], [0.7239, 0.2321, -0.6578, -1.1395, -2.3874, 1.1281]]], dtype=torch.float32).npu() features.requires_grad = True -output = mx_driving.perception.point.npu_group_points(features, indices) +output = mx_driving.point.npu_group_points(features, indices) output.backward(output) grad_features = features.grad ``` @@ -741,166 +786,64 @@ expected_grad_features = tensor( [-0.7029, 0.2388, 0.8612, 0.0000, -2.8564, -1.9131], [ 2.1717, 0.4642, -1.3156, 0.0000, -4.7748, 3.3843]]], dtype=torch.float32) ``` - -## knn -### 接口原型 -```python -mx_driving.common.knn(int k, Tensor xyz, Tensor center_xyz, bool Transposed) -> Tensor -``` -### 功能描述 -對center_xyz中的每個點找到xyz中對應batch中的距離最近的k個點，并且返回此k個點的索引值。 -### 参数说明 -- `xyz(Tensor)`：点数据，表示(x, y, z)三維坐標，数据类型为`float32`。shape为`[B, N, 3]`(当Transposed=False)或`[B, 3, N]`(当Transposed=True)。其中`B`为batch size，`N`为点的数量。 -- `center_xyz(Tensor)`：点数据，表示(x, y, z)三維坐標，数据类型为`float32`。shape为`[B, npoint, 3]`(当Transposed=False)或`[B, 3, npoint]`(当Transposed=True)。其中`B`为batch size，`npoint`为点的数量。 -- `k(int)`：采样点的数量。 -- `Transposed(bool)`: 輸入是否需要進行轉置 -### 返回值 -- `idx(Tensor)`：采样后的索引数据，数据类型为`int32`。shape为`[B, k, npoint]`。 -### 支持的型号 -- Atlas A2 训练系列产品 -### 调用示例 -```python -import torch, torch_npu -from mx_driving.common import knn -xyz = torch.tensor([[[1, 1, 1], [1, 1, 1], [1, 1, 1]], [[1, 1, 1], [1, 1, 1], [1, 1, 1]]], dtype=torch.float32).npu() -center_xyz = torch.tensor([[1, 2, 3]], [[1, 2, 3]], dtype=torch.float32).npu() -idx = knn(2, xyz, center_xyz, False) -print(idx) -``` -```text -tensor([[0, 0], [1, 1]], dtype=torch.int32) -``` -### 算子約束 -1. k必須>0且<100 -2. xyz中的每個batch中的任意一個點到center_xyz對應batch中的任意一個點的距離必須在1e10f以内 -3. xyz和center_xyz的shape必须是3维，当Transposed=True时，xyz和center_xyz的shape的dim的第1位必须是3；当Transposed=False时，xyz和center_xyz的shape的dim的第2位必须是3 -4. 距离相同时索引存在不稳定排序问题,遇到距离精度通过但索引精度错误时，复用不稳定排序的CCB结论 - -## three_nn +## voxelization ### 接口原型 ```python -mx_driving.common.three_nn(Tensor target, Tensor source) -> (Tensor dist, Tensor idx) +mx_driving.point.voxelization(Tensor points, List[float] voxel_size, List[float] coors_range, int max_points=-1, int max_voxels=-1, bool deterministic=True) -> Tensor ``` ### 功能描述 -對target中的每個點找到source中對應batch中的距離最近的3個點，并且返回此3個點的距離和索引值。 +将点云数据进行体素化。 ### 参数说明 -- `target(Tensor)`：点数据，表示(x, y, z)三維坐標，数据类型为`float32/float16`。shape为`[B, npoint, 3]`。其中`B`为batch size，`npoint`为点的数量。 -- `source(Tensor)`：点数据，表示(x, y, z)三維坐標，数据类型为`float32/float16`。shape为`[B, N, 3]`。其中`B`为batch size，`N`为点的数量。 +- `points(Tensor)`：点云数据，数据类型为`float32`。shape为`[N, F]`。其中`N`为点的数量，`F`分别代表每个点的特征维度，其中`N > 0, F >= 3`。 +- `voxel_size(List[float])`：体素大小，数据类型为`float32`。shape为`[3]`。其中`3`分别代表`x, y, z`。 +- `coors_range(List[float])`：体素范围，数据类型为`float32`。shape为`[6]`。其中`6`分别代表`x_min, y_min, z_min, x_max, y_max, z_max`。 +- `max_points(int)`：每个体素的最大点数。默认值为`-1`。 +- `max_voxels(int)`：最大体素数。默认值为`-1`。 +- `deterministic(bool)`：是否确定性。默认值为`True`。 ### 返回值 -- `dist(Tensor)`：采样后的索引数据，数据类型为`float32/float16`。shape为`[B, npoint, 3]`。 -- `idx(Tensor)`：采样后的索引数据，数据类型为`int32/int32`。shape为`[B, npoint, 3]`。 +- `coors(Tensor)`：每个点所属的体素坐标，数据类型为`int32`。shape为`[N, 3]`。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import three_nn -source = torch.tensor([[[1, 1, 1], [1, 1, 1], [1, 1, 1]], [[1, 1, 1], [1, 1, 1], [1, 1, 1]]], dtype=torch.float32).npu() -target = torch.tensor([[1, 2, 3]], [[1, 2, 3]], dtype=torch.float32).npu() -dist, idx = three_nn(target, source) -print(dist) -print(idx) -``` -```text -tensor([[2.236, 2.236, 2.236], [2.236, 2.236, 2.236]], dtype=torch.float32) -tensor([[0, 1, 2], [0, 1, 2]], dtype=torch.int32) +from mx_driving.point import Voxelization +points = torch.randint(-20, 100, [16, 3], dtype=torch.float32).npu() +coors_range = [0, -40, -3, 70.4, 40, 1] +max_points = -1 +voxel_size = [0.5, 0.5, 0.5] +dynamic_voxelization = Voxelization(voxel_size, coors_range, max_points) +out = dynamic_voxelization.forward(points) +print(out) ``` -### 算子約束 -1. source和target的shape必须是3维，且source和target的shape的dim的第2位必须是3 -2. 距离相同时索引存在不稳定排序问题，遇到距离精度通过但索引精度错误时，复用不稳定排序的CCB结论 -## RoipointPool3d +## npu_dynamic_scatter ### 接口原型 ```python -mx_driving.common.RoipointPool3d(int num_sampled_points, Tensor points, Tensor point_features, Tensor boxes3d) -> (Tensor pooled_features, Tensor pooled_empty_flag) +mx_driving.point.npu_dynamic_scatter(Tensor feats, Tensor coors, str reduce_type = 'max') -> Tuple[torch.Tensor, torch.Tensor] ``` ### 功能描述 -对每个3D方案的几何特定特征进行编码。 +将点云特征点在对应体素中进行特征压缩。 ### 参数说明 -- `num_sampled_points(int)`：特征点的数量，正整数。 -- `points(Tensor)`：点张量，数据类型为`float32, float16`。shape 为`[B, N, 3]`。`3`分别代表`x, y, z`。 -- `point_features(Tensor)`：点特征张量，数据类型为`float32, float16`。shape 为`[B, N, C]`。`C`分别代表`x, y, z`。 -- `boxes3d(Tensor)`：框张量，数据类型为`float32, float16`。shape 为`[B, M, 7]`。`7`分别代表`x, y, z, x_size, y_size, z_size, rz`。 +- `feats(Tensor)`：点云特征张量[N, C]，仅支持两维，数据类型为`float32`，特征向量`C`长度上限为2048。 +- `coors(Tensor)`：体素坐标映射张量[N, 3]，仅支持两维，数据类型为`int32`，此处以x, y, z指代体素三维坐标，其取值范围为`0 <= x, y <= 2048`, `0 <= z <= 256`。 +- `reduce_type(str)`：压缩类型。可选值为`'max'`, `'mean'`, `'sum'`。默认值为`'max'` ### 返回值 -- `pooled_features(Tensor)`：点在框内的特征张量，数据类型为`float32, float16`。shape 为`[B, M, num, 3+C]`。 -- `pooled_empty_flag(Tensor)`：所有点不在框内的空标记张量，数据类型为`int32`。shape 为`[B, M]`。 -### 约束说明 -- `points`、`point_features`和`boxes3d`的数据类型必须相同，以及`B`也必须相同。 -- `num_sampled_points`必须小于等于`N`。 -- 数据类型为`float32`时，建议`B`小于100、`N`小于等于2640、`M`小于等于48、`num_sampled_points`小于等于48，个别shape值略微超过建议值无影响，但所有shape值均大于建议值时，算子执行会发生错误。 -- 数据类型为`float16`时，建议`B`小于100、`N`小于等于3360、`M`小于等于60、`num_sampled_points`小于等于60，个别shape值略微超过建议值无影响，但所有shape值均大于建议值时，算子执行会发生错误。 -- `N`/`M`的值越大，性能劣化越严重，建议`N`小于`M`的六百倍，否则性能可能会低于0.1x A100。 +- `voxel_feats(Tensor)`：压缩后的体素特征张量，仅支持两维，数据类型为`float32`。 +- `voxel_coors(Tensor)`：去重后的体素坐标，仅支持两维，数据类型为`int32`。 ### 支持的型号 - Atlas A2 训练系列产品 ### 调用示例 ```python import torch, torch_npu -from mx_driving.common import RoIPointPool3d -num_sampled_points = 1 -points = torch.tensor([[[1, 2, 3]]], dtype=torch.float).npu() -point_features = points.clone() -boxes3d = torch.tensor([[[1, 2, 3, 4, 5, 6, 1]]], dtype=torch.float).npu() -roipoint_pool3d = RoIPointPool3d(num_sampled_points) -pooled_features, pooled_empty_flag = roipoint_pool3d(points, point_features, boxes3d) -print(pooled_features) -print(pooled_empty_flag) -``` -```text -tensor([[[[1., 2., 3., 1., 2., 3.]]]]) -tensor([[0]], dtype=torch.int32) -``` -## boxes_overlap_bev -### 接口原型 -```python -mx_driving.perception.vision.boxes_overlap_bev(Tensor boxes_a, Tensor boxes_b) -> Tensor -``` -### 功能描述 -Calculates the intersection of bounding boxes in Bird's Eye View. -### 参数说明 -- `boxes_a (Tensor)`：第一组bounding boxes，数据类型为`float32`。shape为`[M, 5]`。其中`5`分别代表`x1, y1, x2, y2, angle`, `x1, y1, x2, y2`代表box四个顶点的横纵坐标，`angle`代表box的弧度制旋转角。 -- `boxes_b (Tensor)`：第二组bounding boxes，数据类型为`float32`。shape为`[N, 5]`。其中`5`分别代表`x1, y1, x2, y2, angle`, `x1, y1, x2, y2`代表box四个顶点的横纵坐标，`angle`代表box的弧度制旋转角。 -### 返回值 -- `area_overlap(Tensor)`：包含两组bounding boxes交叠面积的张量，数据类型为`float32`。shape为`[M, N]`。 -### 约束说明 +from mx_driving.point import npu_dynamic_scatter -### 支持的型号 -- Atlas A2 训练系列产品 -### 调用示例 -```python -import torch, torch_npu -from mx_driving.perception.vision import boxes_overlap_bev -boxes_a = torch.tensor([[0, 0, 2, 2, 0]], dtype=torch.float32).npu() -boxes_b = torch.tensor([[1, 1, 3, 3, 0]], dtype=torch.float32).npu() -area_overlap = boxes_overlap_bev(boxes_a, boxes_b) -print(area_overlap) -``` +feats = torch.tensor([[1, 2, 3], [3, 2, 1], [7, 8, 9], [9, 8, 7]], dtype=torch.float32).npu() +coors = torch.tensor([[1, 1, 1], [1, 1, 1], [2, 2, 2], [2, 2, 2]], dtype=torch.int32).npu() +voxel_feats, voxel_coors = npu_dynamic_scatter(feats, coors, 'max') -## npu_max_pool2d -### 接口原型 -```python -mx_driving.common.npu_max_pool2d(Tensor x, int kernel_size, int stride, int padding) -> Tensor -``` -### 功能描述 -Performs max pooling on the input and outputs max values. -### 参数说明 -- `x (Tensor)`：一组待池化对象，数据类型为`float32`，format为NCHW。 -### 返回值 -- `y (Tensor)`：池化后的最大值，数据类型为`float32`，format为NCHW。 -### 约束说明 -kernel_size仅支持3，stride仅支持2，padding仅支持1，且输入C轴数据量要求为8的倍数，H和W需要大于100。 -### 支持的型号 -- Atlas A2 训练系列产品 -### 调用示例 -```python -import torch, torch_npu -from mx_driving.common import npu_max_pool2d -kernel_size = 3 -stride = 2 -padding = 1 -x = torch.randn(18, 64, 464, 800).npu() -res = npu_max_pool2d(x, kernel_size, stride, padding) -print(res) +print(voxel_feats) +print(voxel_coors) ``` - ## unique_voxel ### 接口原型 ```python @@ -936,3 +879,66 @@ print(argsort_indices) print(uin_argsort_indices) ``` + + +## voxel_pooling_train +### 接口原型 +```python +mx_driving.point.npu_voxel_pooling_train(Tensor geom_xyz, Tensor input_features, List[int] voxel_num) -> Tensor +``` +### 功能描述 +点云数据体素化。 +### 参数说明 +- `geom_xyz`：体素坐标，数据类型为`int32`，维度为（B, N, 3）, 3表示x, y, z。 +- `input_features`：点云数据，数据类型为`float32|float16`，维度为（B, N, C）。 +- `voxel_num`：体素格子长宽高，数据类型为`int32`，维度为（3），3表示体素格子的长宽高。 +### 返回值 +- `output(Tensor)`：输出结果，数据类型为`float32|float16`。shape为`[B, num_voxel_y, num_voxel_x, C]`。 +### 约束说明 +- B <= 128 +- N <= 100000 +- C <= 256 +- num_voxel_x <= 1000 +- num_voxel_y <= 1000 +- num_voxel_z <= 10 +- B * num_voxel_y * num_voxel_x * C <= 100000000 +- B * N * C <= 100000000 +- 反向具有相同约束。 +### 支持的型号 +- Atlas A2 训练系列产品 +### 调用示例 +```python +import torch +import torch_npu +import mx_driving.point + +def gen_data(geom_shape, feature_shape, coeff, batch_size, num_channels, dtype): + geom_xyz = torch.rand(geom_shape) * coeff + geom_xyz = geom_xyz.reshape(batch_size, -1, 3) + geom_xyz[:, :, 2] /= 100 + geom_xyz_cpu = geom_xyz.int() + features = torch.rand(feature_shape, dtype=dtype) - 0.5 + features_cpu = features.reshape(batch_size, -1, num_channels) + + return geom_xyz_cpu, features_cpu + +dtype = torch.float32 +coeff = 90 +voxel_num = [128, 128, 1] +batch_size = 2 +num_points = 40 +num_channel = 80 +xyz = 3 + +geom_shape = [batch_size, num_points, xyz] +feature_shape = [batch_size, num_points, num_channel] + +geom_cpu, feature_cpu = gen_data(geom_shape, feature_shape, coeff, batch_size, num_channel, dtype) + +geom_npu = geom_cpu.npu() +feature_npu = feature_cpu.npu() + +result_npu = mx_driving.point.npu_voxel_pooling_train(geom_npu, feature_npu, voxel_num) +``` +# 稀疏卷积算子 + diff --git a/include/csrc/pybind.h b/include/csrc/pybind.h index 27ccf462..c2851f26 100644 --- a/include/csrc/pybind.h +++ b/include/csrc/pybind.h @@ -18,9 +18,9 @@ #include void init_common(pybind11::module& m); -void init_motion(pybind11::module& m); -void init_perception_fused(pybind11::module& m); -void init_perception_point(pybind11::module& m); -void init_perception_vision(pybind11::module& m); +void init_fused(pybind11::module& m); +void init_point(pybind11::module& m); +void init_data(pybind11::module& m); +void init_detection(pybind11::module& m); void init_spconv(pybind11::module& m); #endif // CSRC_PYBIND_H_ diff --git a/examples/BEVFormer/.gitignore b/model_examples/BEVFormer/.gitignore similarity index 99% rename from examples/BEVFormer/.gitignore rename to model_examples/BEVFormer/.gitignore index 4b6213ed..95ec5be9 100644 --- a/examples/BEVFormer/.gitignore +++ b/model_examples/BEVFormer/.gitignore @@ -13,7 +13,6 @@ ckpts/ data/ ckpts data -test/ val/ work_dirs/ develop-eggs/ diff --git a/examples/BEVFormer/LICENSE b/model_examples/BEVFormer/LICENSE similarity index 100% rename from examples/BEVFormer/LICENSE rename to model_examples/BEVFormer/LICENSE diff --git a/examples/BEVFormer/README.md b/model_examples/BEVFormer/README.md similarity index 100% rename from examples/BEVFormer/README.md rename to model_examples/BEVFormer/README.md diff --git a/examples/BEVFormer/README_ORI.md b/model_examples/BEVFormer/README_ORI.md similarity index 100% rename from examples/BEVFormer/README_ORI.md rename to model_examples/BEVFormer/README_ORI.md diff --git a/examples/BEVFormer/docs/getting_started.md b/model_examples/BEVFormer/docs/getting_started.md similarity index 100% rename from examples/BEVFormer/docs/getting_started.md rename to model_examples/BEVFormer/docs/getting_started.md diff --git a/examples/BEVFormer/docs/install.md b/model_examples/BEVFormer/docs/install.md similarity index 100% rename from examples/BEVFormer/docs/install.md rename to model_examples/BEVFormer/docs/install.md diff --git a/examples/BEVFormer/docs/prepare_dataset.md b/model_examples/BEVFormer/docs/prepare_dataset.md similarity index 100% rename from examples/BEVFormer/docs/prepare_dataset.md rename to model_examples/BEVFormer/docs/prepare_dataset.md diff --git a/examples/BEVFormer/mmcv_need/base_runner.py b/model_examples/BEVFormer/mmcv_need/base_runner.py similarity index 100% rename from examples/BEVFormer/mmcv_need/base_runner.py rename to model_examples/BEVFormer/mmcv_need/base_runner.py diff --git a/examples/BEVFormer/mmcv_need/distributed.py b/model_examples/BEVFormer/mmcv_need/distributed.py similarity index 100% rename from examples/BEVFormer/mmcv_need/distributed.py rename to model_examples/BEVFormer/mmcv_need/distributed.py diff --git a/examples/BEVFormer/mmcv_need/epoch_based_runner.py b/model_examples/BEVFormer/mmcv_need/epoch_based_runner.py similarity index 100% rename from examples/BEVFormer/mmcv_need/epoch_based_runner.py rename to model_examples/BEVFormer/mmcv_need/epoch_based_runner.py diff --git a/examples/BEVFormer/mmcv_need/modulated_deform_conv.py b/model_examples/BEVFormer/mmcv_need/modulated_deform_conv.py similarity index 100% rename from examples/BEVFormer/mmcv_need/modulated_deform_conv.py rename to model_examples/BEVFormer/mmcv_need/modulated_deform_conv.py diff --git a/examples/BEVFormer/mmcv_need/optimizer.py b/model_examples/BEVFormer/mmcv_need/optimizer.py similarity index 100% rename from examples/BEVFormer/mmcv_need/optimizer.py rename to model_examples/BEVFormer/mmcv_need/optimizer.py diff --git a/examples/BEVFormer/mmcv_need/points_in_polygons_npu.cpp b/model_examples/BEVFormer/mmcv_need/points_in_polygons_npu.cpp similarity index 100% rename from examples/BEVFormer/mmcv_need/points_in_polygons_npu.cpp rename to model_examples/BEVFormer/mmcv_need/points_in_polygons_npu.cpp diff --git a/examples/BEVFormer/mmcv_need/runtime.txt b/model_examples/BEVFormer/mmcv_need/runtime.txt similarity index 100% rename from examples/BEVFormer/mmcv_need/runtime.txt rename to model_examples/BEVFormer/mmcv_need/runtime.txt diff --git a/examples/BEVFormer/mmdet3d_need/__init__.py b/model_examples/BEVFormer/mmdet3d_need/__init__.py similarity index 100% rename from examples/BEVFormer/mmdet3d_need/__init__.py rename to model_examples/BEVFormer/mmdet3d_need/__init__.py diff --git a/examples/BEVFormer/mmdet3d_need/nuscenes_dataset.py b/model_examples/BEVFormer/mmdet3d_need/nuscenes_dataset.py similarity index 100% rename from examples/BEVFormer/mmdet3d_need/nuscenes_dataset.py rename to model_examples/BEVFormer/mmdet3d_need/nuscenes_dataset.py diff --git a/examples/BEVFormer/mmdet3d_need/runtime.txt b/model_examples/BEVFormer/mmdet3d_need/runtime.txt similarity index 100% rename from examples/BEVFormer/mmdet3d_need/runtime.txt rename to model_examples/BEVFormer/mmdet3d_need/runtime.txt diff --git a/examples/BEVFormer/mmdet_need/__init__.py b/model_examples/BEVFormer/mmdet_need/__init__.py similarity index 100% rename from examples/BEVFormer/mmdet_need/__init__.py rename to model_examples/BEVFormer/mmdet_need/__init__.py diff --git a/examples/BEVFormer/mmdet_need/resnet.py b/model_examples/BEVFormer/mmdet_need/resnet.py similarity index 100% rename from examples/BEVFormer/mmdet_need/resnet.py rename to model_examples/BEVFormer/mmdet_need/resnet.py diff --git a/examples/BEVFormer/projects/__init__.py b/model_examples/BEVFormer/projects/__init__.py similarity index 100% rename from examples/BEVFormer/projects/__init__.py rename to model_examples/BEVFormer/projects/__init__.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/coco_instance.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/coco_instance.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/coco_instance.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/coco_instance.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/kitti-3d-3class.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/kitti-3d-3class.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/kitti-3d-3class.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/kitti-3d-3class.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/kitti-3d-car.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/kitti-3d-car.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/kitti-3d-car.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/kitti-3d-car.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/lyft-3d.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/lyft-3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/lyft-3d.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/lyft-3d.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/nuim_instance.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/nuim_instance.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/nuim_instance.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/nuim_instance.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/nus-3d.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/nus-3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/nus-3d.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/nus-3d.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/nus-mono3d.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/nus-mono3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/nus-mono3d.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/nus-mono3d.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/range100_lyft-3d.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/range100_lyft-3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/range100_lyft-3d.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/range100_lyft-3d.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/s3dis-3d-5class.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/s3dis-3d-5class.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/s3dis-3d-5class.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/s3dis-3d-5class.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/s3dis_seg-3d-13class.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/s3dis_seg-3d-13class.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/s3dis_seg-3d-13class.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/s3dis_seg-3d-13class.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/scannet-3d-18class.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/scannet-3d-18class.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/scannet-3d-18class.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/scannet-3d-18class.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/scannet_seg-3d-20class.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/scannet_seg-3d-20class.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/scannet_seg-3d-20class.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/scannet_seg-3d-20class.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/sunrgbd-3d-10class.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/sunrgbd-3d-10class.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/sunrgbd-3d-10class.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/sunrgbd-3d-10class.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-3class.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-3class.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-3class.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-3class.py diff --git a/examples/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-car.py b/model_examples/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-car.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-car.py rename to model_examples/BEVFormer/projects/configs/_base_/datasets/waymoD5-3d-car.py diff --git a/examples/BEVFormer/projects/configs/_base_/default_runtime.py b/model_examples/BEVFormer/projects/configs/_base_/default_runtime.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/default_runtime.py rename to model_examples/BEVFormer/projects/configs/_base_/default_runtime.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/3dssd.py b/model_examples/BEVFormer/projects/configs/_base_/models/3dssd.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/3dssd.py rename to model_examples/BEVFormer/projects/configs/_base_/models/3dssd.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/cascade_mask_rcnn_r50_fpn.py b/model_examples/BEVFormer/projects/configs/_base_/models/cascade_mask_rcnn_r50_fpn.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/cascade_mask_rcnn_r50_fpn.py rename to model_examples/BEVFormer/projects/configs/_base_/models/cascade_mask_rcnn_r50_fpn.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/centerpoint_01voxel_second_secfpn_nus.py b/model_examples/BEVFormer/projects/configs/_base_/models/centerpoint_01voxel_second_secfpn_nus.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/centerpoint_01voxel_second_secfpn_nus.py rename to model_examples/BEVFormer/projects/configs/_base_/models/centerpoint_01voxel_second_secfpn_nus.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/centerpoint_02pillar_second_secfpn_nus.py b/model_examples/BEVFormer/projects/configs/_base_/models/centerpoint_02pillar_second_secfpn_nus.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/centerpoint_02pillar_second_secfpn_nus.py rename to model_examples/BEVFormer/projects/configs/_base_/models/centerpoint_02pillar_second_secfpn_nus.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/fcos3d.py b/model_examples/BEVFormer/projects/configs/_base_/models/fcos3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/fcos3d.py rename to model_examples/BEVFormer/projects/configs/_base_/models/fcos3d.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/groupfree3d.py b/model_examples/BEVFormer/projects/configs/_base_/models/groupfree3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/groupfree3d.py rename to model_examples/BEVFormer/projects/configs/_base_/models/groupfree3d.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/h3dnet.py b/model_examples/BEVFormer/projects/configs/_base_/models/h3dnet.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/h3dnet.py rename to model_examples/BEVFormer/projects/configs/_base_/models/h3dnet.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_lyft.py b/model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_lyft.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_lyft.py rename to model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_lyft.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_nus.py b/model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_nus.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_nus.py rename to model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_nus.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_range100_lyft.py b/model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_range100_lyft.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_range100_lyft.py rename to model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_fpn_range100_lyft.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_kitti.py b/model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_kitti.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_kitti.py rename to model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_kitti.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_waymo.py b/model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_waymo.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_waymo.py rename to model_examples/BEVFormer/projects/configs/_base_/models/hv_pointpillars_secfpn_waymo.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_kitti.py b/model_examples/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_kitti.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_kitti.py rename to model_examples/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_kitti.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_waymo.py b/model_examples/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_waymo.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_waymo.py rename to model_examples/BEVFormer/projects/configs/_base_/models/hv_second_secfpn_waymo.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/imvotenet_image.py b/model_examples/BEVFormer/projects/configs/_base_/models/imvotenet_image.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/imvotenet_image.py rename to model_examples/BEVFormer/projects/configs/_base_/models/imvotenet_image.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/mask_rcnn_r50_fpn.py b/model_examples/BEVFormer/projects/configs/_base_/models/mask_rcnn_r50_fpn.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/mask_rcnn_r50_fpn.py rename to model_examples/BEVFormer/projects/configs/_base_/models/mask_rcnn_r50_fpn.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/paconv_cuda_ssg.py b/model_examples/BEVFormer/projects/configs/_base_/models/paconv_cuda_ssg.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/paconv_cuda_ssg.py rename to model_examples/BEVFormer/projects/configs/_base_/models/paconv_cuda_ssg.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/paconv_ssg.py b/model_examples/BEVFormer/projects/configs/_base_/models/paconv_ssg.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/paconv_ssg.py rename to model_examples/BEVFormer/projects/configs/_base_/models/paconv_ssg.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/parta2.py b/model_examples/BEVFormer/projects/configs/_base_/models/parta2.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/parta2.py rename to model_examples/BEVFormer/projects/configs/_base_/models/parta2.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/pointnet2_msg.py b/model_examples/BEVFormer/projects/configs/_base_/models/pointnet2_msg.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/pointnet2_msg.py rename to model_examples/BEVFormer/projects/configs/_base_/models/pointnet2_msg.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/pointnet2_ssg.py b/model_examples/BEVFormer/projects/configs/_base_/models/pointnet2_ssg.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/pointnet2_ssg.py rename to model_examples/BEVFormer/projects/configs/_base_/models/pointnet2_ssg.py diff --git a/examples/BEVFormer/projects/configs/_base_/models/votenet.py b/model_examples/BEVFormer/projects/configs/_base_/models/votenet.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/models/votenet.py rename to model_examples/BEVFormer/projects/configs/_base_/models/votenet.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/cosine.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/cosine.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/cosine.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/cosine.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/cyclic_20e.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/cyclic_20e.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/cyclic_20e.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/cyclic_20e.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/cyclic_40e.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/cyclic_40e.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/cyclic_40e.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/cyclic_40e.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/mmdet_schedule_1x.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/mmdet_schedule_1x.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/mmdet_schedule_1x.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/mmdet_schedule_1x.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/schedule_2x.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/schedule_2x.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/schedule_2x.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/schedule_2x.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/schedule_3x.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/schedule_3x.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/schedule_3x.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/schedule_3x.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_150e.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_150e.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_150e.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_150e.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_200e.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_200e.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_200e.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_200e.py diff --git a/examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_50e.py b/model_examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_50e.py similarity index 100% rename from examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_50e.py rename to model_examples/BEVFormer/projects/configs/_base_/schedules/seg_cosine_50e.py diff --git a/examples/BEVFormer/projects/configs/bevformer/bevformer_base.py b/model_examples/BEVFormer/projects/configs/bevformer/bevformer_base.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformer/bevformer_base.py rename to model_examples/BEVFormer/projects/configs/bevformer/bevformer_base.py diff --git a/examples/BEVFormer/projects/configs/bevformer/bevformer_small.py b/model_examples/BEVFormer/projects/configs/bevformer/bevformer_small.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformer/bevformer_small.py rename to model_examples/BEVFormer/projects/configs/bevformer/bevformer_small.py diff --git a/examples/BEVFormer/projects/configs/bevformer/bevformer_tiny.py b/model_examples/BEVFormer/projects/configs/bevformer/bevformer_tiny.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformer/bevformer_tiny.py rename to model_examples/BEVFormer/projects/configs/bevformer/bevformer_tiny.py diff --git a/examples/BEVFormer/projects/configs/bevformer_fp16/bevformer_tiny_fp16.py b/model_examples/BEVFormer/projects/configs/bevformer_fp16/bevformer_tiny_fp16.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformer_fp16/bevformer_tiny_fp16.py rename to model_examples/BEVFormer/projects/configs/bevformer_fp16/bevformer_tiny_fp16.py diff --git a/examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-24ep.py b/model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-24ep.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-24ep.py rename to model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-24ep.py diff --git a/examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-48ep.py b/model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-48ep.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-48ep.py rename to model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-48ep.py diff --git a/examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-24ep.py b/model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-24ep.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-24ep.py rename to model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-24ep.py diff --git a/examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-48ep.py b/model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-48ep.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-48ep.py rename to model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t1-base-48ep.py diff --git a/examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-24ep.py b/model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-24ep.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-24ep.py rename to model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-24ep.py diff --git a/examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-48ep.py b/model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-48ep.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-48ep.py rename to model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t2-48ep.py diff --git a/examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t8-24ep.py b/model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t8-24ep.py similarity index 100% rename from examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t8-24ep.py rename to model_examples/BEVFormer/projects/configs/bevformerv2/bevformerv2-r50-t8-24ep.py diff --git a/examples/BEVFormer/projects/configs/datasets/custom_lyft-3d.py b/model_examples/BEVFormer/projects/configs/datasets/custom_lyft-3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/datasets/custom_lyft-3d.py rename to model_examples/BEVFormer/projects/configs/datasets/custom_lyft-3d.py diff --git a/examples/BEVFormer/projects/configs/datasets/custom_nus-3d.py b/model_examples/BEVFormer/projects/configs/datasets/custom_nus-3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/datasets/custom_nus-3d.py rename to model_examples/BEVFormer/projects/configs/datasets/custom_nus-3d.py diff --git a/examples/BEVFormer/projects/configs/datasets/custom_waymo-3d.py b/model_examples/BEVFormer/projects/configs/datasets/custom_waymo-3d.py similarity index 100% rename from examples/BEVFormer/projects/configs/datasets/custom_waymo-3d.py rename to model_examples/BEVFormer/projects/configs/datasets/custom_waymo-3d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/__init__.py similarity index 83% rename from examples/BEVFormer/projects/mmdet3d_plugin/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/__init__.py index 1c81f6a2..267b2d4f 100644 --- a/examples/BEVFormer/projects/mmdet3d_plugin/__init__.py +++ b/model_examples/BEVFormer/projects/mmdet3d_plugin/__init__.py @@ -1,3 +1,5 @@ +# Copyright (c) OpenMMLab. All rights reserved. +# Copyright 2024 Huawei Technologies Co., Ltd import models.utils import bevformer import dd3d diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/mmdet_train.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/mmdet_train.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/mmdet_train.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/mmdet_train.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/test.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/test.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/test.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/test.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/train.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/train.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/train.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/apis/train.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bev_head.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bev_head.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bev_head.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bev_head.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bevformer_head.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bevformer_head.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bevformer_head.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/dense_heads/bevformer_head.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformerV2.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformerV2.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformerV2.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformerV2.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer_fp16.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer_fp16.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer_fp16.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/detectors/bevformer_fp16.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/custom_hooks.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/custom_hooks.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/custom_hooks.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/hooks/custom_hooks.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/custom_base_transformer_layer.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/custom_base_transformer_layer.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/custom_base_transformer_layer.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/custom_base_transformer_layer.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/decoder.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/decoder.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/decoder.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/decoder.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/encoder.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/encoder.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/encoder.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/encoder.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/group_attention.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/group_attention.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/group_attention.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/group_attention.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/multi_scale_deformable_attn_function.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/multi_scale_deformable_attn_function.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/multi_scale_deformable_attn_function.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/multi_scale_deformable_attn_function.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/spatial_cross_attention.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/spatial_cross_attention.py similarity index 95% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/spatial_cross_attention.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/spatial_cross_attention.py index d7dccdb6..c72099ed 100644 --- a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/spatial_cross_attention.py +++ b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/spatial_cross_attention.py @@ -3,20 +3,20 @@ # Copyright (c) OpenMMLab. All rights reserved. # Copyright 2024 Huawei Technologies Co., Ltd # --------------------------------------------- -# Modified by ZheXu Liu +# Modified by Zhiqi Li # --------------------------------------------- -from mmcv.ops.multi_scale_deform_attn import multi_scale_deformable_attn_pytorch import warnings +import math import torch import torch.nn as nn import torch.nn.functional as F +from mmcv.ops.multi_scale_deform_attn import multi_scale_deformable_attn_pytorch from mmcv.cnn import xavier_init, constant_init from mmcv.cnn.bricks.registry import (ATTENTION, TRANSFORMER_LAYER, TRANSFORMER_LAYER_SEQUENCE) from mmcv.cnn.bricks.transformer import build_attention -import math from mmcv.runner import force_fp32 from mmcv.runner.base_module import BaseModule, ModuleList, Sequential from projects.mmdet3d_plugin.models.utils.bricks import run_time @@ -26,6 +26,8 @@ bev_mask_global = torch.tensor([]).npu() indexes_global = None max_len_global = None bev_mask_id_global = -1 + + @ATTENTION.register_module() class SpatialCrossAttention(BaseModule): """An attention module used in BEVFormer. @@ -133,23 +135,16 @@ class SpatialCrossAttention(BaseModule): D = reference_points_cam.size(3) indexes = [] - global bev_mask_global, indexes_global, max_len_global, bev_mask_id_global, count_global + global bev_mask_global, indexes_global, max_len_global, bev_mask_id_global bev_mask_id = id(bev_mask) if bev_mask_id == bev_mask_id_global: indexes = indexes_global max_len = max_len_global - count = count_global else: - count = torch.any(bev_mask, 3) - bev_mask_ = count.squeeze() - for i, mask_per_img in enumerate(bev_mask_): - index_query_per_img = mask_per_img.nonzero().squeeze(-1) + for i, mask_per_img in enumerate(bev_mask): + index_query_per_img = mask_per_img[0].sum(-1).to(torch.float).nonzero().squeeze(-1) indexes.append(index_query_per_img) - max_len = max([len(each) for each in indexes]) - count = count.permute(1, 2, 0).sum(-1) - count = torch.clamp(count, min=1.0) - count_global = count bev_mask_global = bev_mask.clone() indexes_global = indexes max_len_global = max_len @@ -161,9 +156,9 @@ class SpatialCrossAttention(BaseModule): reference_points_rebatch = reference_points_cam.new_zeros( [bs, self.num_cams, max_len, D, 2]) - for i, reference_points_per_img in enumerate(reference_points_cam): - index_query_per_img = indexes[i] - for j in range(bs): + for j in range(bs): + for i, reference_points_per_img in enumerate(reference_points_cam): + index_query_per_img = indexes[i] queries_rebatch[j, i, :len(index_query_per_img)] = query[j, index_query_per_img] reference_points_rebatch[j, i, :len(index_query_per_img)] = reference_points_per_img[j, index_query_per_img] @@ -174,14 +169,16 @@ class SpatialCrossAttention(BaseModule): value = value.permute(2, 0, 1, 3).reshape( bs * self.num_cams, l, self.embed_dims) - queries = self.deformable_attention(query=queries_rebatch.view(bs*self.num_cams, max_len, self.embed_dims), key=key, value=value, - reference_points=reference_points_rebatch.view(bs*self.num_cams, max_len, D, 2), spatial_shapes=spatial_shapes, + queries = self.deformable_attention(query=queries_rebatch.view(bs * self.num_cams, max_len, self.embed_dims), key=key, value=value, + reference_points=reference_points_rebatch.view(bs * self.num_cams, max_len, D, 2), spatial_shapes=spatial_shapes, level_start_index=level_start_index).view(bs, self.num_cams, max_len, self.embed_dims) for j in range(bs): for i, index_query_per_img in enumerate(indexes): slots[j, index_query_per_img] += queries[j, i, :len(index_query_per_img)] - + count = bev_mask.sum(-1) > 0 + count = count.permute(1, 2, 0).sum(-1) + count = torch.clamp(count, min=1.0) slots = slots / count[..., None] slots = self.output_proj(slots) diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/temporal_self_attention.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/temporal_self_attention.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/temporal_self_attention.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/temporal_self_attention.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformer.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformer.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformer.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformer.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformerV2.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformerV2.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformerV2.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/modules/transformerV2.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/epoch_based_runner.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/epoch_based_runner.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/epoch_based_runner.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/bevformer/runner/epoch_based_runner.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/hungarian_assigner_3d.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/hungarian_assigner_3d.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/hungarian_assigner_3d.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/assigners/hungarian_assigner_3d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/nms_free_coder.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/nms_free_coder.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/nms_free_coder.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/coders/nms_free_coder.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/match_cost.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/match_cost.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/match_cost.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/match_costs/match_cost.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/util.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/util.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/util.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/bbox/util.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/eval_hooks.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/eval_hooks.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/eval_hooks.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/eval_hooks.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/kitti2waymo.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/kitti2waymo.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/kitti2waymo.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/core/evaluation/kitti2waymo.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/builder.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/builder.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/builder.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/builder.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset_v2.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset_v2.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset_v2.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_dataset_v2.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_mono_dataset.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_mono_dataset.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_mono_dataset.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscenes_mono_dataset.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscnes_eval.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscnes_eval.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscnes_eval.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/nuscnes_eval.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/augmentation.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/augmentation.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/augmentation.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/augmentation.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/dd3d_mapper.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/dd3d_mapper.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/dd3d_mapper.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/dd3d_mapper.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/formating.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/formating.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/formating.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/formating.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/loading.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/loading.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/loading.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/loading.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/distributed_sampler.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/distributed_sampler.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/distributed_sampler.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/distributed_sampler.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/group_sampler.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/group_sampler.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/group_sampler.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/group_sampler.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/sampler.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/sampler.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/sampler.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/datasets/samplers/sampler.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/nuscenes.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/nuscenes.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/nuscenes.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/nuscenes.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/transform_utils.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/transform_utils.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/transform_utils.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/datasets/transform_utils.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/iou_loss.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/iou_loss.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/iou_loss.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/iou_loss.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/normalization.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/normalization.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/normalization.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/normalization.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/smooth_l1_loss.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/smooth_l1_loss.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/smooth_l1_loss.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/layers/smooth_l1_loss.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/core.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/core.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/core.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/core.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/disentangled_box3d_loss.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/disentangled_box3d_loss.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/disentangled_box3d_loss.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/disentangled_box3d_loss.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos2d.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos2d.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos2d.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos2d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos3d.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos3d.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos3d.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/fcos3d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/nuscenes_dd3d.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/nuscenes_dd3d.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/nuscenes_dd3d.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/nuscenes_dd3d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/prepare_targets.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/prepare_targets.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/prepare_targets.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/modeling/prepare_targets.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/boxes3d.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/boxes3d.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/boxes3d.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/boxes3d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/image_list.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/image_list.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/image_list.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/image_list.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/pose.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/pose.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/pose.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/pose.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/transform3d.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/transform3d.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/transform3d.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/structures/transform3d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/comm.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/comm.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/comm.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/comm.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/geometry.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/geometry.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/geometry.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/geometry.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tasks.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tasks.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tasks.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tasks.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tensor2d.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tensor2d.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tensor2d.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/tensor2d.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/visualization.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/visualization.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/visualization.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/dd3d/utils/visualization.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/backbones/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/backbones/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/backbones/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/backbones/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/backbones/vovnet.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/backbones/vovnet.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/backbones/vovnet.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/backbones/vovnet.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/hooks.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/hooks.py similarity index 81% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/hooks.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/hooks.py index f41f0a7a..b6274b94 100644 --- a/examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/hooks.py +++ b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/hooks/hooks.py @@ -1,3 +1,5 @@ +# Copyright (c) OpenMMLab. All rights reserved. +# Copyright 2024 Huawei Technologies Co., Ltd from mmcv.runner.hooks.hook import HOOKS, Hook from projects.mmdet3d_plugin.models.utils import run_time diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/opt/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/opt/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/opt/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/opt/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/opt/adamw.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/opt/adamw.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/opt/adamw.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/opt/adamw.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/utils/__init__.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/__init__.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/utils/__init__.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/__init__.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/utils/bricks.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/bricks.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/utils/bricks.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/bricks.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/utils/grid_mask.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/grid_mask.py similarity index 97% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/utils/grid_mask.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/grid_mask.py index e97cd766..f5be94e2 100644 --- a/examples/BEVFormer/projects/mmdet3d_plugin/models/utils/grid_mask.py +++ b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/grid_mask.py @@ -1,3 +1,5 @@ +# Copyright (c) OpenMMLab. All rights reserved. +# Copyright 2024 Huawei Technologies Co., Ltd import torch import torch.nn as nn import numpy as np diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/utils/position_embedding.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/position_embedding.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/utils/position_embedding.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/position_embedding.py diff --git a/examples/BEVFormer/projects/mmdet3d_plugin/models/utils/visual.py b/model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/visual.py similarity index 100% rename from examples/BEVFormer/projects/mmdet3d_plugin/models/utils/visual.py rename to model_examples/BEVFormer/projects/mmdet3d_plugin/models/utils/visual.py diff --git a/examples/BEVFormer/public_address_statement.md b/model_examples/BEVFormer/public_address_statement.md similarity index 100% rename from examples/BEVFormer/public_address_statement.md rename to model_examples/BEVFormer/public_address_statement.md diff --git a/examples/BEVFormer/requirements.txt b/model_examples/BEVFormer/requirements.txt similarity index 100% rename from examples/BEVFormer/requirements.txt rename to model_examples/BEVFormer/requirements.txt diff --git a/model_examples/BEVFormer/test/env_npu.sh b/model_examples/BEVFormer/test/env_npu.sh new file mode 100644 index 00000000..aa4642ab --- /dev/null +++ b/model_examples/BEVFormer/test/env_npu.sh @@ -0,0 +1,59 @@ +#!/bin/bash +CANN_INSTALL_PATH_CONF='/etc/Ascend/ascend_cann_install.info' + +if [ -f $CANN_INSTALL_PATH_CONF ]; then + CANN_INSTALL_PATH=$(cat $CANN_INSTALL_PATH_CONF | grep Install_Path | cut -d "=" -f 2) +else + CANN_INSTALL_PATH="/usr/local/Ascend" +fi + +if [ -d ${CANN_INSTALL_PATH}/ascend-toolkit/latest ]; then + source ${CANN_INSTALL_PATH}/ascend-toolkit/set_env.sh +else + source ${CANN_INSTALL_PATH}/nnae/set_env.sh +fi +msnpureport -g error -d 0 +msnpureport -g error -d 1 +msnpureport -g error -d 2 +msnpureport -g error -d 3 +msnpureport -g error -d 4 +msnpureport -g error -d 5 +msnpureport -g error -d 6 +msnpureport -g error -d 7 + +#设置Shape数据缓存 +export HOST_CACHE_CAPACITY=20 +#将Host日志输出到串口,0-关闭/1-开启 +export ASCEND_SLOG_PRINT_TO_STDOUT=0 +#设置默认日志级别,0-debug/1-info/2-warning/3-error +export ASCEND_GLOBAL_LOG_LEVEL=3 +#设置Event日志开启标志,0-关闭/1-开启 +export ASCEND_GLOBAL_EVENT_ENABLE=0 +#设置是否开启taskque,0-关闭/1-开启 +export TASK_QUEUE_ENABLE=1 +#设置是否开启combined标志,0-关闭/1-开启 +export COMBINED_ENABLE=1 +#HCCL白名单开关,1-关闭/0-开启 +export HCCL_WHITELIST_DISABLE=1 +export HCCL_IF_IP=$(hostname -I |awk '{print $1}') +export HCCL_CONNECT_TIMEOUT=1200 + + +path_lib=$(python3 -c """ +import sys +import re +result='' +for index in range(len(sys.path)): + match_sit = re.search('-packages', sys.path[index]) + if match_sit is not None: + match_lib = re.search('lib', sys.path[index]) + + if match_lib is not None: + end=match_lib.span()[1] + result += sys.path[index][0:end] + ':' + + result+=sys.path[index] + '/torch/lib:' +print(result)""" +) + +echo ${path_lib} diff --git a/model_examples/BEVFormer/test/train_full_8p_base_fp32.sh b/model_examples/BEVFormer/test/train_full_8p_base_fp32.sh new file mode 100644 index 00000000..d88070b2 --- /dev/null +++ b/model_examples/BEVFormer/test/train_full_8p_base_fp32.sh @@ -0,0 +1,97 @@ +# 网络名称,同目录名称,需要模型审视修改 +Network="BEVFormer_Base" +batch_size=1 +world_size=8 +epochs=24 + +# cd到与test文件夹同层级目录下执行脚本，提高兼容性；test_path_dir为包含test文件夹的路径 +cur_path=$(pwd) +cur_path_last_dirname=${cur_path##*/} +if [ x"${cur_path_last_dirname}" == x"test" ]; then + test_path_dir=${cur_path} + cd .. + cur_path=$(pwd) +else + test_path_dir=${cur_path}/test +fi + +source ${test_path_dir}/env_npu.sh + +#创建DeviceID输出目录，不需要修改 +output_path=${cur_path}/test/output/ + +if [ -d ${output_path} ]; then + rm -rf ${output_path} +fi + +mkdir -p ${output_path} + +for para in $* +do + if [[ $para == --epochs* ]];then + epochs=`echo ${para#*=}` + fi +done + +#训练开始时间，不需要修改 +start_time=$(date +%s) + +sed -i "s|log_config = dict(interval=1,|log_config = dict(interval=50,|g" projects/configs/bevformer/bevformer_base.py +sed -i "s|total_epochs = .*|total_epochs = ${epochs}|g" projects/configs/bevformer/bevformer_base.py +sed -i "s|runner = dict(type='EpochBasedRunner', max_epochs=total_epochs, stop_iters=100)|runner = dict(type='EpochBasedRunner', max_epochs=total_epochs)|g" projects/configs/bevformer/bevformer_base.py + +bash ./tools/dist_train.sh ./projects/configs/bevformer/bevformer_base.py ${world_size} > ${test_path_dir}/output/train_full_8p_base_fp32.log 2>&1 & + +wait + +#训练结束时间，不需要修改 +end_time=$(date +%s) +e2e_time=$(($end_time - $start_time)) + +#结果打印，不需要修改 +echo "------------------ Final result ------------------" + +#获取性能数据，不需要修改 +#单迭代训练时长，不需要修改 +TrainingTime=$(grep -o ", time: [0-9.]*" ${test_path_dir}/output/train_full_8p_base_fp32.log | tail -n 10 | grep -o "[0-9.]*" | awk '{sum += $1} END {print sum/NR}') + +#吞吐量 +ActualFPS=$(awk BEGIN'{print ('$batch_size' * '$world_size') / '$TrainingTime'}') + +#打印，不需要修改 +echo "Final Performance images/sec : $ActualFPS" + +#loss值，不需要修改 +ActualLoss=$(grep -o "loss: [0-9.]*" ${test_path_dir}/output/train_full_8p_base_fp32.log | awk 'END {print $NF}') + +#NDS值 +NDS=$(grep -o "pts_bbox_NuScenes/NDS: [0-9.]*" ${test_path_dir}/output/train_full_8p_base_fp32.log | awk 'END {print $NF}') + +#mAP值 +mAP=$(grep -o "pts_bbox_NuScenes/mAP: [0-9.]*" ${test_path_dir}/output/train_full_8p_base_fp32.log | awk 'END {print $NF}') + +#打印，不需要修改 +echo "Final Train Loss : ${ActualLoss}" +echo "NDS : ${NDS}" +echo "mAP : ${mAP}" +echo "E2E Training Duration sec : $e2e_time" + +#性能看护结果汇总 +#训练用例信息，不需要修改 +BatchSize=${batch_size} +WORLD_SIZE=${world_size} +DeviceType=$(uname -m) +CaseName=${Network}_bs${BatchSize}_${WORLD_SIZE}'p'_'acc' + +#关键信息打印到${CaseName}.log中，不需要修改 +echo "Network = ${Network}" >${test_path_dir}/output/${CaseName}.log +echo "RankSize = ${WORLD_SIZE}" >>${test_path_dir}/output/${CaseName}.log +echo "BatchSize = ${BatchSize}" >>${test_path_dir}/output/${CaseName}.log +echo "DeviceType = ${DeviceType}" >>${test_path_dir}/output/${CaseName}.log +echo "CaseName = ${CaseName}" >>${test_path_dir}/output/${CaseName}.log +echo "ActualFPS = ${ActualFPS}" >>${test_path_dir}/output/${CaseName}.log +echo "TrainingTime = ${TrainingTime}" >>${test_path_dir}/output/${CaseName}.log +echo "ActualLoss = ${ActualLoss}" >>${test_path_dir}/output/${CaseName}.log +echo "NDS = ${NDS}" >>${test_path_dir}/output/${CaseName}.log +echo "mAP = ${mAP}" >>${test_path_dir}/output/${CaseName}.log +echo "E2ETrainingTime = ${e2e_time}" >>${test_path_dir}/output/${CaseName}.log \ No newline at end of file diff --git a/model_examples/BEVFormer/test/train_performance_8p_base_fp32.sh b/model_examples/BEVFormer/test/train_performance_8p_base_fp32.sh new file mode 100644 index 00000000..7088b69c --- /dev/null +++ b/model_examples/BEVFormer/test/train_performance_8p_base_fp32.sh @@ -0,0 +1,92 @@ +# 网络名称,同目录名称,需要模型审视修改 +Network="BEVFormer_Base" +batch_size=1 +world_size=8 + +# cd到与test文件夹同层级目录下执行脚本，提高兼容性；test_path_dir为包含test文件夹的路径 +cur_path=$(pwd) +cur_path_last_dirname=${cur_path##*/} +if [ x"${cur_path_last_dirname}" == x"test" ]; then + test_path_dir=${cur_path} + cd .. + cur_path=$(pwd) +else + test_path_dir=${cur_path}/test +fi + +source ${test_path_dir}/env_npu.sh + +#创建DeviceID输出目录，不需要修改 +output_path=${cur_path}/test/output/ + +if [ -d ${output_path} ]; then + rm -rf ${output_path} +fi + +mkdir -p ${output_path} + +sed -i "s|log_config = dict(interval=50,|log_config = dict(interval=1,|g" projects/configs/bevformer/bevformer_base.py +sed -i "s|runner = dict(type='EpochBasedRunner', max_epochs=total_epochs)|runner = dict(type='EpochBasedRunner', max_epochs=total_epochs, stop_iters=1500)|g" projects/configs/bevformer/bevformer_base.py + + +#训练开始时间，不需要修改 +start_time=$(date +%s) + +bash ./tools/dist_train.sh ./projects/configs/bevformer/bevformer_base.py ${world_size} > ${test_path_dir}/output/train_performance_8p_base_fp32.log 2>&1 & + +wait + +sed -i "s|log_config = dict(interval=1,|log_config = dict(interval=50,|g" projects/configs/bevformer/bevformer_base.py +sed -i "s|runner = dict(type='EpochBasedRunner', max_epochs=total_epochs, stop_iters=1500)|runner = dict(type='EpochBasedRunner', max_epochs=total_epochs)|g" projects/configs/bevformer/bevformer_base.py + +#训练结束时间，不需要修改 +end_time=$(date +%s) +e2e_time=$(($end_time - $start_time)) + +#结果打印，不需要修改 +echo "------------------ Final result ------------------" + +#获取性能数据，不需要修改 +#单迭代训练时长，不需要修改 +TrainingTime=$(grep -o ", time: [0-9.]*" ${test_path_dir}/output/train_performance_8p_base_fp32.log | tail -n 30 | grep -o "[0-9.]*" | awk '{sum += $1} END {print sum/NR}') + +#吞吐量 +ActualFPS=$(awk BEGIN'{print ('$batch_size' * '$world_size') / '$TrainingTime'}') + +#打印，不需要修改 +echo "Final Performance images/sec : $ActualFPS" + +#loss值，不需要修改 +ActualLoss=$(grep -o "loss: [0-9.]*" ${test_path_dir}/output/train_performance_8p_base_fp32.log | awk 'END {print $NF}') + +#NDS值 +NDS=$(grep -o "pts_bbox_NuScenes/NDS: [0-9.]*" ${test_path_dir}/output/train_performance_8p_base_fp32.log | awk 'END {print $NF}') + +#mAP值 +mAP=$(grep -o "pts_bbox_NuScenes/mAP: [0-9.]*" ${test_path_dir}/output/train_performance_8p_base_fp32.log | awk 'END {print $NF}') + +#打印，不需要修改 +echo "Final Train Loss : ${ActualLoss}" +echo "NDS : ${NDS}" +echo "mAP : ${mAP}" +echo "E2E Training Duration sec : $e2e_time" + +#性能看护结果汇总 +#训练用例信息，不需要修改 +BatchSize=${batch_size} +WORLD_SIZE=${world_size} +DeviceType=$(uname -m) +CaseName=${Network}_bs${BatchSize}_${WORLD_SIZE}'p'_'performance' + +#关键信息打印到${CaseName}.log中，不需要修改 +echo "Network = ${Network}" >${test_path_dir}/output/${CaseName}.log +echo "RankSize = ${WORLD_SIZE}" >>${test_path_dir}/output/${CaseName}.log +echo "BatchSize = ${BatchSize}" >>${test_path_dir}/output/${CaseName}.log +echo "DeviceType = ${DeviceType}" >>${test_path_dir}/output/${CaseName}.log +echo "CaseName = ${CaseName}" >>${test_path_dir}/output/${CaseName}.log +echo "ActualFPS = ${ActualFPS}" >>${test_path_dir}/output/${CaseName}.log +echo "TrainingTime = ${TrainingTime}" >>${test_path_dir}/output/${CaseName}.log +echo "ActualLoss = ${ActualLoss}" >>${test_path_dir}/output/${CaseName}.log +echo "NDS = ${NDS}" >>${test_path_dir}/output/${CaseName}.log +echo "mAP = ${mAP}" >>${test_path_dir}/output/${CaseName}.log +echo "E2ETrainingTime = ${e2e_time}" >>${test_path_dir}/output/${CaseName}.log \ No newline at end of file diff --git a/examples/BEVFormer/tools/__init__.py b/model_examples/BEVFormer/tools/__init__.py similarity index 100% rename from examples/BEVFormer/tools/__init__.py rename to model_examples/BEVFormer/tools/__init__.py diff --git a/examples/BEVFormer/tools/analysis_tools/__init__.py b/model_examples/BEVFormer/tools/analysis_tools/__init__.py similarity index 100% rename from examples/BEVFormer/tools/analysis_tools/__init__.py rename to model_examples/BEVFormer/tools/analysis_tools/__init__.py diff --git a/examples/BEVFormer/tools/analysis_tools/analyze_logs.py b/model_examples/BEVFormer/tools/analysis_tools/analyze_logs.py similarity index 100% rename from examples/BEVFormer/tools/analysis_tools/analyze_logs.py rename to model_examples/BEVFormer/tools/analysis_tools/analyze_logs.py diff --git a/examples/BEVFormer/tools/analysis_tools/benchmark.py b/model_examples/BEVFormer/tools/analysis_tools/benchmark.py similarity index 100% rename from examples/BEVFormer/tools/analysis_tools/benchmark.py rename to model_examples/BEVFormer/tools/analysis_tools/benchmark.py diff --git a/examples/BEVFormer/tools/analysis_tools/get_params.py b/model_examples/BEVFormer/tools/analysis_tools/get_params.py similarity index 100% rename from examples/BEVFormer/tools/analysis_tools/get_params.py rename to model_examples/BEVFormer/tools/analysis_tools/get_params.py diff --git a/examples/BEVFormer/tools/analysis_tools/visual.py b/model_examples/BEVFormer/tools/analysis_tools/visual.py similarity index 100% rename from examples/BEVFormer/tools/analysis_tools/visual.py rename to model_examples/BEVFormer/tools/analysis_tools/visual.py diff --git a/examples/BEVFormer/tools/create_data.py b/model_examples/BEVFormer/tools/create_data.py similarity index 100% rename from examples/BEVFormer/tools/create_data.py rename to model_examples/BEVFormer/tools/create_data.py diff --git a/examples/BEVFormer/tools/data_converter/__init__.py b/model_examples/BEVFormer/tools/data_converter/__init__.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/__init__.py rename to model_examples/BEVFormer/tools/data_converter/__init__.py diff --git a/examples/BEVFormer/tools/data_converter/create_gt_database.py b/model_examples/BEVFormer/tools/data_converter/create_gt_database.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/create_gt_database.py rename to model_examples/BEVFormer/tools/data_converter/create_gt_database.py diff --git a/examples/BEVFormer/tools/data_converter/indoor_converter.py b/model_examples/BEVFormer/tools/data_converter/indoor_converter.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/indoor_converter.py rename to model_examples/BEVFormer/tools/data_converter/indoor_converter.py diff --git a/examples/BEVFormer/tools/data_converter/kitti_converter.py b/model_examples/BEVFormer/tools/data_converter/kitti_converter.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/kitti_converter.py rename to model_examples/BEVFormer/tools/data_converter/kitti_converter.py diff --git a/examples/BEVFormer/tools/data_converter/kitti_data_utils.py b/model_examples/BEVFormer/tools/data_converter/kitti_data_utils.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/kitti_data_utils.py rename to model_examples/BEVFormer/tools/data_converter/kitti_data_utils.py diff --git a/examples/BEVFormer/tools/data_converter/lyft_converter.py b/model_examples/BEVFormer/tools/data_converter/lyft_converter.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/lyft_converter.py rename to model_examples/BEVFormer/tools/data_converter/lyft_converter.py diff --git a/examples/BEVFormer/tools/data_converter/lyft_data_fixer.py b/model_examples/BEVFormer/tools/data_converter/lyft_data_fixer.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/lyft_data_fixer.py rename to model_examples/BEVFormer/tools/data_converter/lyft_data_fixer.py diff --git a/examples/BEVFormer/tools/data_converter/nuimage_converter.py b/model_examples/BEVFormer/tools/data_converter/nuimage_converter.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/nuimage_converter.py rename to model_examples/BEVFormer/tools/data_converter/nuimage_converter.py diff --git a/examples/BEVFormer/tools/data_converter/nuscenes_converter.py b/model_examples/BEVFormer/tools/data_converter/nuscenes_converter.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/nuscenes_converter.py rename to model_examples/BEVFormer/tools/data_converter/nuscenes_converter.py diff --git a/examples/BEVFormer/tools/data_converter/s3dis_data_utils.py b/model_examples/BEVFormer/tools/data_converter/s3dis_data_utils.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/s3dis_data_utils.py rename to model_examples/BEVFormer/tools/data_converter/s3dis_data_utils.py diff --git a/examples/BEVFormer/tools/data_converter/scannet_data_utils.py b/model_examples/BEVFormer/tools/data_converter/scannet_data_utils.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/scannet_data_utils.py rename to model_examples/BEVFormer/tools/data_converter/scannet_data_utils.py diff --git a/examples/BEVFormer/tools/data_converter/sunrgbd_data_utils.py b/model_examples/BEVFormer/tools/data_converter/sunrgbd_data_utils.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/sunrgbd_data_utils.py rename to model_examples/BEVFormer/tools/data_converter/sunrgbd_data_utils.py diff --git a/examples/BEVFormer/tools/data_converter/waymo_converter.py b/model_examples/BEVFormer/tools/data_converter/waymo_converter.py similarity index 100% rename from examples/BEVFormer/tools/data_converter/waymo_converter.py rename to model_examples/BEVFormer/tools/data_converter/waymo_converter.py diff --git a/examples/BEVFormer/tools/dist_test.sh b/model_examples/BEVFormer/tools/dist_test.sh similarity index 100% rename from examples/BEVFormer/tools/dist_test.sh rename to model_examples/BEVFormer/tools/dist_test.sh diff --git a/examples/BEVFormer/tools/dist_train.sh b/model_examples/BEVFormer/tools/dist_train.sh similarity index 100% rename from examples/BEVFormer/tools/dist_train.sh rename to model_examples/BEVFormer/tools/dist_train.sh diff --git a/examples/BEVFormer/tools/fp16/dist_train.sh b/model_examples/BEVFormer/tools/fp16/dist_train.sh similarity index 100% rename from examples/BEVFormer/tools/fp16/dist_train.sh rename to model_examples/BEVFormer/tools/fp16/dist_train.sh diff --git a/examples/BEVFormer/tools/fp16/train.py b/model_examples/BEVFormer/tools/fp16/train.py similarity index 100% rename from examples/BEVFormer/tools/fp16/train.py rename to model_examples/BEVFormer/tools/fp16/train.py diff --git a/examples/BEVFormer/tools/misc/browse_dataset.py b/model_examples/BEVFormer/tools/misc/browse_dataset.py similarity index 100% rename from examples/BEVFormer/tools/misc/browse_dataset.py rename to model_examples/BEVFormer/tools/misc/browse_dataset.py diff --git a/examples/BEVFormer/tools/misc/fuse_conv_bn.py b/model_examples/BEVFormer/tools/misc/fuse_conv_bn.py similarity index 100% rename from examples/BEVFormer/tools/misc/fuse_conv_bn.py rename to model_examples/BEVFormer/tools/misc/fuse_conv_bn.py diff --git a/examples/BEVFormer/tools/misc/print_config.py b/model_examples/BEVFormer/tools/misc/print_config.py similarity index 100% rename from examples/BEVFormer/tools/misc/print_config.py rename to model_examples/BEVFormer/tools/misc/print_config.py diff --git a/examples/BEVFormer/tools/misc/visualize_results.py b/model_examples/BEVFormer/tools/misc/visualize_results.py similarity index 100% rename from examples/BEVFormer/tools/misc/visualize_results.py rename to model_examples/BEVFormer/tools/misc/visualize_results.py diff --git a/examples/BEVFormer/tools/model_converters/convert_votenet_checkpoints.py b/model_examples/BEVFormer/tools/model_converters/convert_votenet_checkpoints.py similarity index 100% rename from examples/BEVFormer/tools/model_converters/convert_votenet_checkpoints.py rename to model_examples/BEVFormer/tools/model_converters/convert_votenet_checkpoints.py diff --git a/examples/BEVFormer/tools/model_converters/publish_model.py b/model_examples/BEVFormer/tools/model_converters/publish_model.py similarity index 100% rename from examples/BEVFormer/tools/model_converters/publish_model.py rename to model_examples/BEVFormer/tools/model_converters/publish_model.py diff --git a/examples/BEVFormer/tools/model_converters/regnet2mmdet.py b/model_examples/BEVFormer/tools/model_converters/regnet2mmdet.py similarity index 100% rename from examples/BEVFormer/tools/model_converters/regnet2mmdet.py rename to model_examples/BEVFormer/tools/model_converters/regnet2mmdet.py diff --git a/examples/BEVFormer/tools/test.py b/model_examples/BEVFormer/tools/test.py similarity index 100% rename from examples/BEVFormer/tools/test.py rename to model_examples/BEVFormer/tools/test.py diff --git a/examples/BEVFormer/tools/train.py b/model_examples/BEVFormer/tools/train.py similarity index 100% rename from examples/BEVFormer/tools/train.py rename to model_examples/BEVFormer/tools/train.py diff --git a/mx_driving/common/__init__.py b/mx_driving/common/__init__.py index 8f7b0296..747cb6b7 100644 --- a/mx_driving/common/__init__.py +++ b/mx_driving/common/__init__.py @@ -1,25 +1,8 @@ from .ops.three_interpolate import three_interpolate from .ops.scatter_max import scatter_max -from .ops.rotated_iou import npu_rotated_iou -from .ops.furthest_point_sampling_with_dist import furthest_point_sample_with_dist -from .ops.npu_dynamic_scatter import npu_dynamic_scatter -from .ops.npu_points_in_box import npu_points_in_box -from .ops.npu_points_in_box_all import npu_points_in_box_all -from .ops.npu_multi_scale_deformable_attn_function import npu_multi_scale_deformable_attn_function -from .ops.voxelization import voxelization, Voxelization -from .ops.nms3d_normal import npu_nms3d_normal -from .ops.furthest_point_sampling import npu_furthest_point_sampling -from .ops.npu_nms3d import npu_nms3d -from .ops.rotated_overlaps import npu_rotated_overlaps from .ops.npu_scatter_mean_grad import npu_scatter_mean_grad -from .ops.voxel_pooling_train import npu_voxel_pooling_train from .ops.knn import knn from .ops.threeNN import three_nn -from .ops.npu_roipoint_pool3d import RoipointPool3d as RoIPointPool3d -from .ops.npu_max_pool2d import npu_max_pool2d -from .ops.npu_add_relu import npu_add_relu from .ops.scatter_mean import scatter_mean from .ops.sort_pairs import sort_pairs -from .ops.fused_bias_leaky_relu import npu_fused_bias_leaky_relu -from .ops.npu_deformable_aggregation import npu_deformable_aggregation diff --git a/mx_driving/common/ops/csrc/functions.h b/mx_driving/common/ops/csrc/functions.h index be3ee366..e24f832a 100644 --- a/mx_driving/common/ops/csrc/functions.h +++ b/mx_driving/common/ops/csrc/functions.h @@ -16,6 +16,8 @@ #include +at::Tensor knn(const at::Tensor& xyz, const at::Tensor& center_xyz, int32_t nsample, bool is_from_knn); + at::Tensor npu_three_interpolate( int b, int c, int m, int n, const at::Tensor& points, const at::Tensor& idx, const at::Tensor& weight); @@ -27,86 +29,13 @@ std::tuple scatter_max_with_argmax_v2( at::Tensor npu_scatter_max_backward(const at::Tensor& x, const at::Tensor& segment_ids, const at::Tensor& num_segments); -at::Tensor npu_rotated_overlaps(const at::Tensor& self, const at::Tensor& query_boxes, bool trans); - -at::Tensor npu_rotated_iou(const at::Tensor& boxes, const at::Tensor& query_boxes, bool trans, int64_t mode, - bool is_cross, double v_threshold, double e_threshold); - at::Tensor npu_scatter(const at::Tensor& self, const at::Tensor& indices, const at::Tensor& updates, int64_t dim); -at::Tensor furthest_point_sampling_with_dist( - const at::Tensor& points_dist, const at::Tensor& nearest_temp, int32_t num_points); - -std::tuple npu_dynamic_scatter(const at::Tensor& feats, const at::Tensor& coors, - const at::Tensor& prefix_sum_point_per_voxel, const at::Tensor& argsort_coor, int32_t num_voxels, - const char* reduce_type); - -void npu_dynamic_scatter_grad(at::Tensor& grad_point_feats, const at::Tensor& grad_voxel_feats, - const at::Tensor& prefix_sum_point_per_voxel, const at::Tensor& argsort_coor, const at::Tensor& compare_mask, - const char* reduce_type); - -at::Tensor npu_max_pool2d(const at::Tensor& x, int kernel_size, int stride, int padding); - -at::Tensor npu_points_in_box(const at::Tensor& boxes, const at::Tensor& pts); - -at::Tensor npu_points_in_box_all(const at::Tensor& boxes, const at::Tensor& pts); - -at::Tensor npu_multi_scale_deformable_attn_function(const at::Tensor& value, const at::Tensor& value_spatial_shapes, - const at::Tensor& value_level_start_index, const at::Tensor& sampling_locations, - const at::Tensor& attention_weights); - -std::tuple multi_scale_deformable_attn_grad(const at::Tensor& value, - const at::Tensor& shape, const at::Tensor& level_start_index, const at::Tensor& location_trans, - const at::Tensor& attn_weight_trans, const at::Tensor& grad_output); - -std::tuple multi_scale_deformable_attn_grad_v2(const at::Tensor& value, - const at::Tensor& shape, const at::Tensor& level_start_index, const at::Tensor& location_trans, - const at::Tensor& attn_weight_trans, const at::Tensor& grad_output); - -at::Tensor npu_furthest_point_sampling(const at::Tensor& point_xyz, const at::Tensor& nearset_temp, int32_t num_points); - -at::Tensor dynamic_voxelization(const at::Tensor& points, at::Tensor& coors, int grid_x, int grid_y, int grid_z, - double voxel_x, double voxel_y, double voxel_z, double coors_min_x, double coors_min_y, double coorsMinZ); - -std::tuple nms3d_normal(const at::Tensor& boxes, double nms_overlap_thresh); - -std::tuple nms3d(const at::Tensor& boxes, double threshold); - at::Tensor npu_scatter_mean_grad(const at::Tensor& grad_out, const at::Tensor& index, int32_t dim); -std::tuple voxel_pooling_train(const at::Tensor& inputFeatures, const at::Tensor& geom, - at::Tensor& outputFeatures, at::Tensor& posMemo, int batchSize, int numPoints, int numChannels, int numVoxelX, - int numVoxelY, int numVoxelZ); - -at::Tensor voxel_pool_train_backward(const at::Tensor& grad_out, const at::Tensor& posMemo, const int64_t batchSize, - const int64_t numPoints, const int64_t numChannels, const int64_t h, const int64_t w); - -at::Tensor knn(const at::Tensor& xyz, const at::Tensor& center_xyz, int32_t nsample, bool is_from_knn); - -std::tuple npu_roipoint_pool3d_forward(const int32_t num_sampled_points, - const at::Tensor& points, const at::Tensor& point_features, const at::Tensor& boxes3d); - -at::Tensor npu_add_relu(at::Tensor& x, const at::Tensor& y); - -at::Tensor npu_add_relu_grad(at::Tensor& self, at::Tensor& grad_output); std::tuple npu_scatter_mean(at::Tensor& src, at::Tensor& index, c10::optional out, c10::optional dim, c10::optional dim_size); std::tuple npu_sort_pairs(const at::Tensor &keys_in, const at::Tensor &values_in, int64_t dim, bool descending); -at::Tensor fused_bias_leaky_relu(const at::Tensor& x, const at::Tensor& bias, const double negative_slop, const double scale); - -at::Tensor deformable_aggregation(const at::Tensor& mc_ms_feat, const at::Tensor& spatial_shape, - const at::Tensor& scale_start_index, const at::Tensor& sampling_location, const at::Tensor& weights); -std::tuple deformable_aggregation_grad( - const at::Tensor& mc_ms_feat, - const at::Tensor& spatial_shape, - const at::Tensor& scale_start_index, - const at::Tensor& sampling_location, - const at::Tensor& weights, - const at::Tensor& grad_output, - const at::Tensor& grad_mc_ms_feat, - const at::Tensor& grad_sampling_location, - const at::Tensor& grad_weights -); #endif // COMMON_OPS_CSRC_FUNCTIONS_H_ diff --git a/mx_driving/common/ops/csrc/pybind.cpp b/mx_driving/common/ops/csrc/pybind.cpp index a246f9cd..e09c4920 100644 --- a/mx_driving/common/ops/csrc/pybind.cpp +++ b/mx_driving/common/ops/csrc/pybind.cpp @@ -22,75 +22,23 @@ void init_common(pybind11::module& m) { - // three_interpolate - m.def("npu_three_interpolate", &npu_three_interpolate); - m.def("npu_three_interpolate_backward", &npu_three_interpolate_backward); - - // scatter_max - m.def("scatter_max_with_argmax_v2", &scatter_max_with_argmax_v2); - m.def("npu_scatter_max_backward", &npu_scatter_max_backward); - - // roated overlap - m.def("npu_rotated_overlaps", &npu_rotated_overlaps, "npu_rotated_overlap NPU version"); - - // rotated iou - m.def("npu_rotated_iou", &npu_rotated_iou); - - // furthest_points_sampling_with_dist - m.def("furthest_point_sampling_with_dist", &furthest_point_sampling_with_dist); - - // npu_points_in_box - m.def("npu_points_in_box", &npu_points_in_box); - // npu_points_in_box_all - m.def("npu_points_in_box_all", &npu_points_in_box_all); - // npu_points_in_box - m.def("npu_max_pool2d", &npu_max_pool2d); - // npu_multi_scale_deformable_attn_function - m.def("npu_multi_scale_deformable_attn_function", &npu_multi_scale_deformable_attn_function); - m.def("multi_scale_deformable_attn_grad", &multi_scale_deformable_attn_grad); - m.def("multi_scale_deformable_attn_grad_v2", &multi_scale_deformable_attn_grad_v2); - - // npu_dynamic_scatter - m.def("npu_dynamic_scatter", &npu_dynamic_scatter); - m.def("npu_dynamic_scatter_grad", &npu_dynamic_scatter_grad); - - // dyn_voxelization - m.def("dynamic_voxelization", &dynamic_voxelization); - - // nms3d_normal - m.def("nms3d_normal", &nms3d_normal); - - // nms3d - m.def("nms3d", &nms3d); - - // npu_furthest_point_sampling - m.def("npu_furthest_point_sampling", &npu_furthest_point_sampling); - - // npu_scatter_mean_grad - m.def("npu_scatter_mean_grad", &npu_scatter_mean_grad); - - // voxel_pooling - m.def("voxel_pooling_train", &voxel_pooling_train); - m.def("voxel_pool_train_backward", &voxel_pool_train_backward); - // knn m.def("knn", &knn); - // npu_roipoint_pool3d_forward - m.def("npu_roipoint_pool3d_forward", &npu_roipoint_pool3d_forward); + // npu_scatter_mean_grad + m.def("npu_scatter_mean_grad", &npu_scatter_mean_grad); - // npu_add_relu - m.def("npu_add_relu", &npu_add_relu); - m.def("npu_add_relu_grad", &npu_add_relu_grad); + // three_interpolate + m.def("npu_three_interpolate", &npu_three_interpolate); + m.def("npu_three_interpolate_backward", &npu_three_interpolate_backward); // scatter_mean m.def("npu_scatter_mean", &npu_scatter_mean, "npu_scatter_mean NPU version"); + + // scatter_max + m.def("scatter_max_with_argmax_v2", &scatter_max_with_argmax_v2); + m.def("npu_scatter_max_backward", &npu_scatter_max_backward); // npu_sort_pairs m.def("npu_sort_pairs", &npu_sort_pairs, "sort_pairs NPU version"); - // fused_bias_leaky_relu - m.def("fused_bias_leaky_relu", &fused_bias_leaky_relu); - // npu_deformable_aggregation - m.def("npu_deformable_aggregation", &deformable_aggregation); - m.def("npu_deformable_aggregation_grad", &deformable_aggregation_grad); } diff --git a/mx_driving/data/CMakeLists.txt b/mx_driving/data/CMakeLists.txt new file mode 100644 index 00000000..3f1ac043 --- /dev/null +++ b/mx_driving/data/CMakeLists.txt @@ -0,0 +1,7 @@ +if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/ops/kernels) + add_subdirectory(ops/kernels) +endif() + +if (${ENABLE_ONNX} AND EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/ops/onnx) + add_subdirectory(ops/onnx/plugin) +endif() diff --git a/mx_driving/data/__init__.py b/mx_driving/data/__init__.py new file mode 100644 index 00000000..e8973220 --- /dev/null +++ b/mx_driving/data/__init__.py @@ -0,0 +1,3 @@ +from .ops.npu_points_in_box import npu_points_in_box +from .ops.npu_points_in_box_all import npu_points_in_box_all +from .ops.npu_roipoint_pool3d import RoipointPool3d as RoIPointPool3d \ No newline at end of file diff --git a/mx_driving/motion/components/README.md b/mx_driving/data/components/README.md similarity index 100% rename from mx_driving/motion/components/README.md rename to mx_driving/data/components/README.md diff --git a/mx_driving/motion/__init__.py b/mx_driving/data/ops/__init__.py similarity index 100% rename from mx_driving/motion/__init__.py rename to mx_driving/data/ops/__init__.py diff --git a/mx_driving/common/ops/csrc/PointsInBox.cpp b/mx_driving/data/ops/csrc/PointsInBox.cpp similarity index 100% rename from mx_driving/common/ops/csrc/PointsInBox.cpp rename to mx_driving/data/ops/csrc/PointsInBox.cpp diff --git a/mx_driving/common/ops/csrc/PointsInBoxAll.cpp b/mx_driving/data/ops/csrc/PointsInBoxAll.cpp similarity index 100% rename from mx_driving/common/ops/csrc/PointsInBoxAll.cpp rename to mx_driving/data/ops/csrc/PointsInBoxAll.cpp diff --git a/mx_driving/data/ops/csrc/README.md b/mx_driving/data/ops/csrc/README.md new file mode 100644 index 00000000..8073915f --- /dev/null +++ b/mx_driving/data/ops/csrc/README.md @@ -0,0 +1,6 @@ +## Description +The `csrc` lib implements python interface, which use `pybind11` to wrap the C++ code. +There are 3 files you need to focus: +1. `pybind.cpp`: Define the python interface. +2. `functions.cpp`: Define the C++ interface. +3. The file naming in `Pascal` style: The implementation of the C++ interface. \ No newline at end of file diff --git a/mx_driving/common/ops/csrc/RoipointPool3dForward.cpp b/mx_driving/data/ops/csrc/RoipointPool3dForward.cpp similarity index 100% rename from mx_driving/common/ops/csrc/RoipointPool3dForward.cpp rename to mx_driving/data/ops/csrc/RoipointPool3dForward.cpp diff --git a/mx_driving/perception/vision/ops/csrc/functions.h b/mx_driving/data/ops/csrc/functions.h similarity index 69% rename from mx_driving/perception/vision/ops/csrc/functions.h rename to mx_driving/data/ops/csrc/functions.h index b268f4a7..e509d755 100644 --- a/mx_driving/perception/vision/ops/csrc/functions.h +++ b/mx_driving/data/ops/csrc/functions.h @@ -17,5 +17,10 @@ #include #include -at::Tensor npu_boxes_overlap_bev(const at::Tensor &boxes_a, const at::Tensor &boxes_b); +at::Tensor npu_points_in_box(const at::Tensor& boxes, const at::Tensor& pts); + +at::Tensor npu_points_in_box_all(const at::Tensor& boxes, const at::Tensor& pts); + +std::tuple npu_roipoint_pool3d_forward(const int32_t num_sampled_points, + const at::Tensor& points, const at::Tensor& point_features, const at::Tensor& boxes3d); #endif // PERCEPTION_VISION_OPS_CSRC_FUNCTIONS_H_ diff --git a/mx_driving/data/ops/csrc/pybind.cpp b/mx_driving/data/ops/csrc/pybind.cpp new file mode 100644 index 00000000..21a84a7d --- /dev/null +++ b/mx_driving/data/ops/csrc/pybind.cpp @@ -0,0 +1,15 @@ +#include +#include "csrc/pybind.h" +#include "functions.h" + +void init_data(pybind11::module& m) +{ + // npu_points_in_box + m.def("npu_points_in_box", &npu_points_in_box); + + // npu_points_in_box_all + m.def("npu_points_in_box_all", &npu_points_in_box_all); + + // npu_roipoint_pool3d_forward + m.def("npu_roipoint_pool3d_forward", &npu_roipoint_pool3d_forward); +} diff --git a/mx_driving/motion/ops/kernels/CMakeLists.txt b/mx_driving/data/ops/kernels/CMakeLists.txt similarity index 100% rename from mx_driving/motion/ops/kernels/CMakeLists.txt rename to mx_driving/data/ops/kernels/CMakeLists.txt diff --git a/mx_driving/data/ops/kernels/README.md b/mx_driving/data/ops/kernels/README.md new file mode 100644 index 00000000..1e664555 --- /dev/null +++ b/mx_driving/data/ops/kernels/README.md @@ -0,0 +1,13 @@ +## 算子原型 + + + + + + + + + + + +
算子类型(OpType)Add
算子输入nameshapedata typeformat
x-floatND
y-floatND
算子输出z-floatND
核函数名add_custom
\ No newline at end of file diff --git a/mx_driving/motion/ops/kernels/framework/CMakeLists.txt b/mx_driving/data/ops/kernels/framework/CMakeLists.txt similarity index 100% rename from mx_driving/motion/ops/kernels/framework/CMakeLists.txt rename to mx_driving/data/ops/kernels/framework/CMakeLists.txt diff --git a/mx_driving/motion/ops/kernels/op_host/CMakeLists.txt b/mx_driving/data/ops/kernels/op_host/CMakeLists.txt similarity index 100% rename from mx_driving/motion/ops/kernels/op_host/CMakeLists.txt rename to mx_driving/data/ops/kernels/op_host/CMakeLists.txt diff --git a/mx_driving/data/ops/kernels/op_host/common.h b/mx_driving/data/ops/kernels/op_host/common.h new file mode 100644 index 00000000..4580dff5 --- /dev/null +++ b/mx_driving/data/ops/kernels/op_host/common.h @@ -0,0 +1,28 @@ +/* + * Copyright (c) Huawei Technologies Co., Ltd. 2023-2024. All rights reserved. + */ +#ifndef COMMON_H +#define COMMON_H + +#include "register/op_def_registry.h" +#include "tiling/platform/platform_ascendc.h" +#include "tiling/tiling_api.h" +#include "register/tilingdata_base.h" + +inline uint32_t ceil_multiple(uint32_t num, uint32_t block) +{ + if (block == 0) { + return 0; + } + return (num + block - 1) / block; +} + +inline uint32_t ceil_value(uint32_t num, uint32_t block) +{ + if (block == 0) { + return 0; + } + return ((num + block - 1) / block) * block; +} + +#endif // COMMON_H diff --git a/mx_driving/common/ops/kernels/op_host/points_in_box.cpp b/mx_driving/data/ops/kernels/op_host/points_in_box.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/points_in_box.cpp rename to mx_driving/data/ops/kernels/op_host/points_in_box.cpp diff --git a/mx_driving/common/ops/kernels/op_host/points_in_box_all.cpp b/mx_driving/data/ops/kernels/op_host/points_in_box_all.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/points_in_box_all.cpp rename to mx_driving/data/ops/kernels/op_host/points_in_box_all.cpp diff --git a/mx_driving/common/ops/kernels/op_host/points_in_box_all_tiling.h b/mx_driving/data/ops/kernels/op_host/points_in_box_all_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/points_in_box_all_tiling.h rename to mx_driving/data/ops/kernels/op_host/points_in_box_all_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/points_in_box_tiling.h b/mx_driving/data/ops/kernels/op_host/points_in_box_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/points_in_box_tiling.h rename to mx_driving/data/ops/kernels/op_host/points_in_box_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/roipoint_pool3d_forward.cpp b/mx_driving/data/ops/kernels/op_host/roipoint_pool3d_forward.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/roipoint_pool3d_forward.cpp rename to mx_driving/data/ops/kernels/op_host/roipoint_pool3d_forward.cpp diff --git a/mx_driving/common/ops/kernels/op_host/roipoint_pool3d_forward_tiling.h b/mx_driving/data/ops/kernels/op_host/roipoint_pool3d_forward_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/roipoint_pool3d_forward_tiling.h rename to mx_driving/data/ops/kernels/op_host/roipoint_pool3d_forward_tiling.h diff --git a/mx_driving/perception/fused/ops/kernels/op_kernel/CMakeLists.txt b/mx_driving/data/ops/kernels/op_kernel/CMakeLists.txt similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_kernel/CMakeLists.txt rename to mx_driving/data/ops/kernels/op_kernel/CMakeLists.txt diff --git a/mx_driving/common/ops/kernels/op_kernel/points_in_box.cpp b/mx_driving/data/ops/kernels/op_kernel/points_in_box.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/points_in_box.cpp rename to mx_driving/data/ops/kernels/op_kernel/points_in_box.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/points_in_box_all.cpp b/mx_driving/data/ops/kernels/op_kernel/points_in_box_all.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/points_in_box_all.cpp rename to mx_driving/data/ops/kernels/op_kernel/points_in_box_all.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/roipoint_pool3d_forward.cpp b/mx_driving/data/ops/kernels/op_kernel/roipoint_pool3d_forward.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/roipoint_pool3d_forward.cpp rename to mx_driving/data/ops/kernels/op_kernel/roipoint_pool3d_forward.cpp diff --git a/mx_driving/common/ops/npu_points_in_box.py b/mx_driving/data/ops/npu_points_in_box.py similarity index 100% rename from mx_driving/common/ops/npu_points_in_box.py rename to mx_driving/data/ops/npu_points_in_box.py diff --git a/mx_driving/common/ops/npu_points_in_box_all.py b/mx_driving/data/ops/npu_points_in_box_all.py similarity index 100% rename from mx_driving/common/ops/npu_points_in_box_all.py rename to mx_driving/data/ops/npu_points_in_box_all.py diff --git a/mx_driving/common/ops/npu_roipoint_pool3d.py b/mx_driving/data/ops/npu_roipoint_pool3d.py similarity index 100% rename from mx_driving/common/ops/npu_roipoint_pool3d.py rename to mx_driving/data/ops/npu_roipoint_pool3d.py diff --git a/mx_driving/motion/CMakeLists.txt b/mx_driving/detection/CMakeLists.txt similarity index 100% rename from mx_driving/motion/CMakeLists.txt rename to mx_driving/detection/CMakeLists.txt diff --git a/mx_driving/detection/__init__.py b/mx_driving/detection/__init__.py new file mode 100644 index 00000000..118eebed --- /dev/null +++ b/mx_driving/detection/__init__.py @@ -0,0 +1,5 @@ +from .ops.boxes_overlap_bev import boxes_overlap_bev +from .ops.nms3d_normal import npu_nms3d_normal +from .ops.npu_nms3d import npu_nms3d +from .ops.rotated_iou import npu_rotated_iou +from .ops.rotated_overlaps import npu_rotated_overlaps \ No newline at end of file diff --git a/mx_driving/perception/fused/components/README.md b/mx_driving/detection/components/README.md similarity index 100% rename from mx_driving/perception/fused/components/README.md rename to mx_driving/detection/components/README.md diff --git a/mx_driving/perception/__init__.py b/mx_driving/detection/ops/__init__.py similarity index 100% rename from mx_driving/perception/__init__.py rename to mx_driving/detection/ops/__init__.py diff --git a/mx_driving/perception/vision/ops/boxes_overlap_bev.py b/mx_driving/detection/ops/boxes_overlap_bev.py similarity index 100% rename from mx_driving/perception/vision/ops/boxes_overlap_bev.py rename to mx_driving/detection/ops/boxes_overlap_bev.py diff --git a/mx_driving/perception/vision/ops/csrc/BoxesOverlapBev.cpp b/mx_driving/detection/ops/csrc/BoxesOverlapBev.cpp similarity index 100% rename from mx_driving/perception/vision/ops/csrc/BoxesOverlapBev.cpp rename to mx_driving/detection/ops/csrc/BoxesOverlapBev.cpp diff --git a/mx_driving/common/ops/csrc/Nms3d.cpp b/mx_driving/detection/ops/csrc/Nms3d.cpp similarity index 100% rename from mx_driving/common/ops/csrc/Nms3d.cpp rename to mx_driving/detection/ops/csrc/Nms3d.cpp diff --git a/mx_driving/common/ops/csrc/Nms3dNormal.cpp b/mx_driving/detection/ops/csrc/Nms3dNormal.cpp similarity index 100% rename from mx_driving/common/ops/csrc/Nms3dNormal.cpp rename to mx_driving/detection/ops/csrc/Nms3dNormal.cpp diff --git a/mx_driving/motion/ops/csrc/README.md b/mx_driving/detection/ops/csrc/README.md similarity index 100% rename from mx_driving/motion/ops/csrc/README.md rename to mx_driving/detection/ops/csrc/README.md diff --git a/mx_driving/common/ops/csrc/RotatedIou.cpp b/mx_driving/detection/ops/csrc/RotatedIou.cpp similarity index 100% rename from mx_driving/common/ops/csrc/RotatedIou.cpp rename to mx_driving/detection/ops/csrc/RotatedIou.cpp diff --git a/mx_driving/common/ops/csrc/RotatedOverlaps.cpp b/mx_driving/detection/ops/csrc/RotatedOverlaps.cpp similarity index 100% rename from mx_driving/common/ops/csrc/RotatedOverlaps.cpp rename to mx_driving/detection/ops/csrc/RotatedOverlaps.cpp diff --git a/mx_driving/perception/fused/ops/csrc/functions.h b/mx_driving/detection/ops/csrc/functions.h similarity index 34% rename from mx_driving/perception/fused/ops/csrc/functions.h rename to mx_driving/detection/ops/csrc/functions.h index e8335acf..56ca5a6c 100644 --- a/mx_driving/perception/fused/ops/csrc/functions.h +++ b/mx_driving/detection/ops/csrc/functions.h @@ -11,22 +11,20 @@ // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. +#ifndef PERCEPTION_VISION_OPS_CSRC_FUNCTIONS_H_ +#define PERCEPTION_VISION_OPS_CSRC_FUNCTIONS_H_ -#ifndef PERCEPTION_FUSED_OPS_CSRC_FUNCTIONS_H_ -#define PERCEPTION_FUSED_OPS_CSRC_FUNCTIONS_H_ -#include +#include #include +std::tuple nms3d_normal(const at::Tensor& boxes, double nms_overlap_thresh); -at::Tensor npu_bev_pool(const at::Tensor& feat, const at::Tensor& geom_feat, const at::Tensor& interval_lengths, - const at::Tensor& interval_starts, int64_t b, int64_t d, int64_t h, int64_t w); -at::Tensor npu_bev_pool_backward(const at::Tensor& grad_out, const at::Tensor& geom_feat, - const at::Tensor& interval_lengths, const at::Tensor& interval_starts, int64_t b, int64_t d, int64_t h, int64_t w); +std::tuple nms3d(const at::Tensor& boxes, double threshold); -at::Tensor npu_bev_pool_v2(const at::Tensor& depth, const at::Tensor& feat, const at::Tensor& ranks_depth, - const at::Tensor& ranks_feat, const at::Tensor& ranks_bev, const at::Tensor& interval_lengths, - const at::Tensor& interval_starts, int64_t b, int64_t d, int64_t h, int64_t w); -std::tuple npu_bev_pool_v2_backward(const at::Tensor& grad_out, const at::Tensor& depth, - const at::Tensor& feat, const at::Tensor& ranks_depth, const at::Tensor& ranks_feat, const at::Tensor& ranks_bev, - const at::Tensor& interval_lengths, const at::Tensor& interval_starts, int64_t b, int64_t d, int64_t h, int64_t w); -#endif // PERCEPTION_FUSED_OPS_CSRC_FUNCTIONS_H_ +at::Tensor npu_rotated_overlaps(const at::Tensor& self, const at::Tensor& query_boxes, bool trans); + +at::Tensor npu_rotated_iou(const at::Tensor& boxes, const at::Tensor& query_boxes, bool trans, int64_t mode, + bool is_cross, double v_threshold, double e_threshold); + +at::Tensor npu_boxes_overlap_bev(const at::Tensor &boxes_a, const at::Tensor &boxes_b); +#endif // PERCEPTION_VISION_OPS_CSRC_FUNCTIONS_H_ diff --git a/mx_driving/perception/vision/ops/csrc/pybind.cpp b/mx_driving/detection/ops/csrc/pybind.cpp similarity index 37% rename from mx_driving/perception/vision/ops/csrc/pybind.cpp rename to mx_driving/detection/ops/csrc/pybind.cpp index ff49f49e..c2941001 100644 --- a/mx_driving/perception/vision/ops/csrc/pybind.cpp +++ b/mx_driving/detection/ops/csrc/pybind.cpp @@ -2,8 +2,20 @@ #include "csrc/pybind.h" #include "functions.h" -void init_perception_vision(pybind11::module &m) +void init_detection(pybind11::module& m) { + // nms3d_normal + m.def("nms3d_normal", &nms3d_normal); + + // nms3d + m.def("nms3d", &nms3d); + + // roated overlap + m.def("npu_rotated_overlaps", &npu_rotated_overlaps, "npu_rotated_overlap NPU version"); + + // rotated iou + m.def("npu_rotated_iou", &npu_rotated_iou); + // npu_boxes_overlap_bev m.def("npu_boxes_overlap_bev", &npu_boxes_overlap_bev, "boxes_overlap_bev NPU version"); } diff --git a/mx_driving/perception/point/ops/kernels/CMakeLists.txt b/mx_driving/detection/ops/kernels/CMakeLists.txt similarity index 100% rename from mx_driving/perception/point/ops/kernels/CMakeLists.txt rename to mx_driving/detection/ops/kernels/CMakeLists.txt diff --git a/mx_driving/motion/ops/kernels/README.md b/mx_driving/detection/ops/kernels/README.md similarity index 100% rename from mx_driving/motion/ops/kernels/README.md rename to mx_driving/detection/ops/kernels/README.md diff --git a/mx_driving/perception/fused/ops/kernels/op_host/CMakeLists.txt b/mx_driving/detection/ops/kernels/op_host/CMakeLists.txt similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_host/CMakeLists.txt rename to mx_driving/detection/ops/kernels/op_host/CMakeLists.txt diff --git a/mx_driving/perception/vision/ops/kernels/op_host/boxes_overlap_bev.cpp b/mx_driving/detection/ops/kernels/op_host/boxes_overlap_bev.cpp similarity index 100% rename from mx_driving/perception/vision/ops/kernels/op_host/boxes_overlap_bev.cpp rename to mx_driving/detection/ops/kernels/op_host/boxes_overlap_bev.cpp diff --git a/mx_driving/perception/vision/ops/kernels/op_host/boxes_overlap_bev_tiling.h b/mx_driving/detection/ops/kernels/op_host/boxes_overlap_bev_tiling.h similarity index 100% rename from mx_driving/perception/vision/ops/kernels/op_host/boxes_overlap_bev_tiling.h rename to mx_driving/detection/ops/kernels/op_host/boxes_overlap_bev_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/gather_nms3d_mask_tiling.cpp b/mx_driving/detection/ops/kernels/op_host/gather_nms3d_mask_tiling.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/gather_nms3d_mask_tiling.cpp rename to mx_driving/detection/ops/kernels/op_host/gather_nms3d_mask_tiling.cpp diff --git a/mx_driving/common/ops/kernels/op_host/gather_nms3d_mask_tiling.h b/mx_driving/detection/ops/kernels/op_host/gather_nms3d_mask_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/gather_nms3d_mask_tiling.h rename to mx_driving/detection/ops/kernels/op_host/gather_nms3d_mask_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/nms3d.cpp b/mx_driving/detection/ops/kernels/op_host/nms3d.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/nms3d.cpp rename to mx_driving/detection/ops/kernels/op_host/nms3d.cpp diff --git a/mx_driving/common/ops/kernels/op_host/nms3d_normal_tiling.cpp b/mx_driving/detection/ops/kernels/op_host/nms3d_normal_tiling.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/nms3d_normal_tiling.cpp rename to mx_driving/detection/ops/kernels/op_host/nms3d_normal_tiling.cpp diff --git a/mx_driving/detection/ops/kernels/op_host/nms3d_normal_tiling.h b/mx_driving/detection/ops/kernels/op_host/nms3d_normal_tiling.h new file mode 100644 index 00000000..9976c248 --- /dev/null +++ b/mx_driving/detection/ops/kernels/op_host/nms3d_normal_tiling.h @@ -0,0 +1,24 @@ +/* + * Copyright (c) Huawei Technologies Co., Ltd. 2023-2024. All rights reserved. + */ +#ifndef NMS3D_NORMAL_TILING_H +#define NMS3D_NORMAL_TILING_H + +#include "register/tilingdata_base.h" + +namespace optiling { +BEGIN_TILING_DATA_DEF(Nms3dNormalTilingData) + TILING_DATA_FIELD_DEF(uint32_t, usedCoreNum) // used cores + TILING_DATA_FIELD_DEF(uint32_t, boxNum) // count of boxes + TILING_DATA_FIELD_DEF(uint32_t, loopTime) // loop times + TILING_DATA_FIELD_DEF(uint32_t, eachSum) // count of each core, = loop_time * 8 + TILING_DATA_FIELD_DEF(uint32_t, tailSum) // count of tail core + TILING_DATA_FIELD_DEF(uint32_t, tailNum) // last time count of tail core + TILING_DATA_FIELD_DEF(uint32_t, maskNum) // mask align 32bit + TILING_DATA_FIELD_DEF(float, overlapThresh) +END_TILING_DATA_DEF; + +REGISTER_TILING_DATA_CLASS(Nms3dNormal, Nms3dNormalTilingData) +} // namespace optiling + +#endif // NMS3D_NORMAL_TILING_H diff --git a/mx_driving/common/ops/kernels/op_host/nms3d_tiling.h b/mx_driving/detection/ops/kernels/op_host/nms3d_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/nms3d_tiling.h rename to mx_driving/detection/ops/kernels/op_host/nms3d_tiling.h diff --git a/mx_driving/perception/point/ops/kernels/op_kernel/CMakeLists.txt b/mx_driving/detection/ops/kernels/op_kernel/CMakeLists.txt similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_kernel/CMakeLists.txt rename to mx_driving/detection/ops/kernels/op_kernel/CMakeLists.txt diff --git a/mx_driving/perception/vision/ops/kernels/op_kernel/boxes_overlap_bev.cpp b/mx_driving/detection/ops/kernels/op_kernel/boxes_overlap_bev.cpp similarity index 100% rename from mx_driving/perception/vision/ops/kernels/op_kernel/boxes_overlap_bev.cpp rename to mx_driving/detection/ops/kernels/op_kernel/boxes_overlap_bev.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/gather_nms3d_mask.cpp b/mx_driving/detection/ops/kernels/op_kernel/gather_nms3d_mask.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/gather_nms3d_mask.cpp rename to mx_driving/detection/ops/kernels/op_kernel/gather_nms3d_mask.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/nms3d.cpp b/mx_driving/detection/ops/kernels/op_kernel/nms3d.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/nms3d.cpp rename to mx_driving/detection/ops/kernels/op_kernel/nms3d.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/nms3d_normal.cpp b/mx_driving/detection/ops/kernels/op_kernel/nms3d_normal.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/nms3d_normal.cpp rename to mx_driving/detection/ops/kernels/op_kernel/nms3d_normal.cpp diff --git a/mx_driving/common/ops/nms3d_normal.py b/mx_driving/detection/ops/nms3d_normal.py similarity index 100% rename from mx_driving/common/ops/nms3d_normal.py rename to mx_driving/detection/ops/nms3d_normal.py diff --git a/mx_driving/common/ops/npu_nms3d.py b/mx_driving/detection/ops/npu_nms3d.py similarity index 100% rename from mx_driving/common/ops/npu_nms3d.py rename to mx_driving/detection/ops/npu_nms3d.py diff --git a/mx_driving/common/ops/rotated_iou.py b/mx_driving/detection/ops/rotated_iou.py similarity index 100% rename from mx_driving/common/ops/rotated_iou.py rename to mx_driving/detection/ops/rotated_iou.py diff --git a/mx_driving/common/ops/rotated_overlaps.py b/mx_driving/detection/ops/rotated_overlaps.py similarity index 100% rename from mx_driving/common/ops/rotated_overlaps.py rename to mx_driving/detection/ops/rotated_overlaps.py diff --git a/mx_driving/fused/CMakeLists.txt b/mx_driving/fused/CMakeLists.txt new file mode 100644 index 00000000..621d1fa9 --- /dev/null +++ b/mx_driving/fused/CMakeLists.txt @@ -0,0 +1,3 @@ +if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/ops/kernels) + add_subdirectory(ops/kernels) +endif() diff --git a/mx_driving/fused/__init__.py b/mx_driving/fused/__init__.py new file mode 100644 index 00000000..3c5a5f99 --- /dev/null +++ b/mx_driving/fused/__init__.py @@ -0,0 +1,5 @@ +from .ops.npu_max_pool2d import npu_max_pool2d +from .ops.npu_add_relu import npu_add_relu +from .ops.npu_multi_scale_deformable_attn_function import npu_multi_scale_deformable_attn_function +from .ops.fused_bias_leaky_relu import npu_fused_bias_leaky_relu +from .ops.npu_deformable_aggregation import npu_deformable_aggregation \ No newline at end of file diff --git a/mx_driving/perception/point/components/README.md b/mx_driving/fused/components/README.md similarity index 100% rename from mx_driving/perception/point/components/README.md rename to mx_driving/fused/components/README.md diff --git a/mx_driving/perception/fused/ops/__init__.py b/mx_driving/fused/ops/__init__.py similarity index 100% rename from mx_driving/perception/fused/ops/__init__.py rename to mx_driving/fused/ops/__init__.py diff --git a/mx_driving/common/ops/csrc/AddRelu.cpp b/mx_driving/fused/ops/csrc/AddRelu.cpp similarity index 100% rename from mx_driving/common/ops/csrc/AddRelu.cpp rename to mx_driving/fused/ops/csrc/AddRelu.cpp diff --git a/mx_driving/common/ops/csrc/DeformableAggregation.cpp b/mx_driving/fused/ops/csrc/DeformableAggregation.cpp similarity index 100% rename from mx_driving/common/ops/csrc/DeformableAggregation.cpp rename to mx_driving/fused/ops/csrc/DeformableAggregation.cpp diff --git a/mx_driving/common/ops/csrc/FusedBiasLeakyRelu.cpp b/mx_driving/fused/ops/csrc/FusedBiasLeakyRelu.cpp similarity index 100% rename from mx_driving/common/ops/csrc/FusedBiasLeakyRelu.cpp rename to mx_driving/fused/ops/csrc/FusedBiasLeakyRelu.cpp diff --git a/mx_driving/common/ops/csrc/MaxPool2d.cpp b/mx_driving/fused/ops/csrc/MaxPool2d.cpp similarity index 100% rename from mx_driving/common/ops/csrc/MaxPool2d.cpp rename to mx_driving/fused/ops/csrc/MaxPool2d.cpp diff --git a/mx_driving/common/ops/csrc/MultiScaleDeformableAttnFunction.cpp b/mx_driving/fused/ops/csrc/MultiScaleDeformableAttnFunction.cpp similarity index 100% rename from mx_driving/common/ops/csrc/MultiScaleDeformableAttnFunction.cpp rename to mx_driving/fused/ops/csrc/MultiScaleDeformableAttnFunction.cpp diff --git a/mx_driving/perception/fused/ops/csrc/README.md b/mx_driving/fused/ops/csrc/README.md similarity index 100% rename from mx_driving/perception/fused/ops/csrc/README.md rename to mx_driving/fused/ops/csrc/README.md diff --git a/mx_driving/fused/ops/csrc/functions.h b/mx_driving/fused/ops/csrc/functions.h new file mode 100644 index 00000000..97ae736d --- /dev/null +++ b/mx_driving/fused/ops/csrc/functions.h @@ -0,0 +1,57 @@ +// Copyright (c) 2024, Huawei Technologies.All rights reserved. +// +// Licensed under the BSD 3-Clause License (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// https://opensource.org/licenses/BSD-3-Clause +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#ifndef PERCEPTION_FUSED_OPS_CSRC_FUNCTIONS_H_ +#define PERCEPTION_FUSED_OPS_CSRC_FUNCTIONS_H_ +#include +#include + +at::Tensor npu_max_pool2d(const at::Tensor& x, int kernel_size, int stride, int padding); + +at::Tensor npu_multi_scale_deformable_attn_function(const at::Tensor& value, const at::Tensor& value_spatial_shapes, + const at::Tensor& value_level_start_index, const at::Tensor& sampling_locations, + const at::Tensor& attention_weights); + +std::tuple multi_scale_deformable_attn_grad(const at::Tensor& value, + const at::Tensor& shape, const at::Tensor& level_start_index, const at::Tensor& location_trans, + const at::Tensor& attn_weight_trans, const at::Tensor& grad_output); + +std::tuple multi_scale_deformable_attn_grad_v2(const at::Tensor& value, + const at::Tensor& shape, const at::Tensor& level_start_index, const at::Tensor& location_trans, + const at::Tensor& attn_weight_trans, const at::Tensor& grad_output); + +at::Tensor npu_add_relu(at::Tensor& x, const at::Tensor& y); + +at::Tensor npu_add_relu_grad(at::Tensor& self, at::Tensor& grad_output); +std::tuple npu_scatter_mean(at::Tensor& src, at::Tensor& index, + c10::optional out, c10::optional dim, + c10::optional dim_size); + +at::Tensor fused_bias_leaky_relu(const at::Tensor& x, const at::Tensor& bias, const double negative_slop, const double scale); + +at::Tensor deformable_aggregation(const at::Tensor& mc_ms_feat, const at::Tensor& spatial_shape, + const at::Tensor& scale_start_index, const at::Tensor& sampling_location, const at::Tensor& weights); +std::tuple deformable_aggregation_grad( + const at::Tensor& mc_ms_feat, + const at::Tensor& spatial_shape, + const at::Tensor& scale_start_index, + const at::Tensor& sampling_location, + const at::Tensor& weights, + const at::Tensor& grad_output, + const at::Tensor& grad_mc_ms_feat, + const at::Tensor& grad_sampling_location, + const at::Tensor& grad_weights +); + +#endif // PERCEPTION_FUSED_OPS_CSRC_FUNCTIONS_H_ diff --git a/mx_driving/fused/ops/csrc/pybind.cpp b/mx_driving/fused/ops/csrc/pybind.cpp new file mode 100644 index 00000000..80819b12 --- /dev/null +++ b/mx_driving/fused/ops/csrc/pybind.cpp @@ -0,0 +1,25 @@ +#include "csrc/pybind.h" + +#include + +#include "functions.h" +void init_fused(pybind11::module& m) +{ + // nnpu_max_pool2d + m.def("npu_max_pool2d", &npu_max_pool2d); + // npu_multi_scale_deformable_attn_function + m.def("npu_multi_scale_deformable_attn_function", &npu_multi_scale_deformable_attn_function); + m.def("multi_scale_deformable_attn_grad", &multi_scale_deformable_attn_grad); + m.def("multi_scale_deformable_attn_grad_v2", &multi_scale_deformable_attn_grad_v2); + + // npu_add_relu + m.def("npu_add_relu", &npu_add_relu); + m.def("npu_add_relu_grad", &npu_add_relu_grad); + + // fused_bias_leaky_relu + m.def("fused_bias_leaky_relu", &fused_bias_leaky_relu); + + // npu_deformable_aggregation + m.def("npu_deformable_aggregation", &deformable_aggregation); + m.def("npu_deformable_aggregation_grad", &deformable_aggregation_grad); +} diff --git a/mx_driving/common/ops/fused_bias_leaky_relu.py b/mx_driving/fused/ops/fused_bias_leaky_relu.py similarity index 100% rename from mx_driving/common/ops/fused_bias_leaky_relu.py rename to mx_driving/fused/ops/fused_bias_leaky_relu.py diff --git a/mx_driving/perception/fused/ops/kernels/CMakeLists.txt b/mx_driving/fused/ops/kernels/CMakeLists.txt similarity index 100% rename from mx_driving/perception/fused/ops/kernels/CMakeLists.txt rename to mx_driving/fused/ops/kernels/CMakeLists.txt diff --git a/mx_driving/perception/fused/ops/kernels/README.md b/mx_driving/fused/ops/kernels/README.md similarity index 100% rename from mx_driving/perception/fused/ops/kernels/README.md rename to mx_driving/fused/ops/kernels/README.md diff --git a/mx_driving/perception/point/ops/kernels/op_host/CMakeLists.txt b/mx_driving/fused/ops/kernels/op_host/CMakeLists.txt similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/CMakeLists.txt rename to mx_driving/fused/ops/kernels/op_host/CMakeLists.txt diff --git a/mx_driving/common/ops/kernels/op_host/add_relu.cpp b/mx_driving/fused/ops/kernels/op_host/add_relu.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/add_relu.cpp rename to mx_driving/fused/ops/kernels/op_host/add_relu.cpp diff --git a/mx_driving/common/ops/kernels/op_host/add_relu_tiling.h b/mx_driving/fused/ops/kernels/op_host/add_relu_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/add_relu_tiling.h rename to mx_driving/fused/ops/kernels/op_host/add_relu_tiling.h diff --git a/mx_driving/fused/ops/kernels/op_host/common.h b/mx_driving/fused/ops/kernels/op_host/common.h new file mode 100644 index 00000000..4580dff5 --- /dev/null +++ b/mx_driving/fused/ops/kernels/op_host/common.h @@ -0,0 +1,28 @@ +/* + * Copyright (c) Huawei Technologies Co., Ltd. 2023-2024. All rights reserved. + */ +#ifndef COMMON_H +#define COMMON_H + +#include "register/op_def_registry.h" +#include "tiling/platform/platform_ascendc.h" +#include "tiling/tiling_api.h" +#include "register/tilingdata_base.h" + +inline uint32_t ceil_multiple(uint32_t num, uint32_t block) +{ + if (block == 0) { + return 0; + } + return (num + block - 1) / block; +} + +inline uint32_t ceil_value(uint32_t num, uint32_t block) +{ + if (block == 0) { + return 0; + } + return ((num + block - 1) / block) * block; +} + +#endif // COMMON_H diff --git a/mx_driving/common/ops/kernels/op_host/deformable_aggregation.cpp b/mx_driving/fused/ops/kernels/op_host/deformable_aggregation.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/deformable_aggregation.cpp rename to mx_driving/fused/ops/kernels/op_host/deformable_aggregation.cpp diff --git a/mx_driving/common/ops/kernels/op_host/deformable_aggregation_grad.cpp b/mx_driving/fused/ops/kernels/op_host/deformable_aggregation_grad.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/deformable_aggregation_grad.cpp rename to mx_driving/fused/ops/kernels/op_host/deformable_aggregation_grad.cpp diff --git a/mx_driving/common/ops/kernels/op_host/deformable_aggregation_grad_tiling.h b/mx_driving/fused/ops/kernels/op_host/deformable_aggregation_grad_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/deformable_aggregation_grad_tiling.h rename to mx_driving/fused/ops/kernels/op_host/deformable_aggregation_grad_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/deformable_aggregation_tiling.h b/mx_driving/fused/ops/kernels/op_host/deformable_aggregation_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/deformable_aggregation_tiling.h rename to mx_driving/fused/ops/kernels/op_host/deformable_aggregation_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/fused_bias_leaky_relu.cpp b/mx_driving/fused/ops/kernels/op_host/fused_bias_leaky_relu.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/fused_bias_leaky_relu.cpp rename to mx_driving/fused/ops/kernels/op_host/fused_bias_leaky_relu.cpp diff --git a/mx_driving/common/ops/kernels/op_host/fused_bias_leaky_relu_tiling.h b/mx_driving/fused/ops/kernels/op_host/fused_bias_leaky_relu_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/fused_bias_leaky_relu_tiling.h rename to mx_driving/fused/ops/kernels/op_host/fused_bias_leaky_relu_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/max_pool2d.cpp b/mx_driving/fused/ops/kernels/op_host/max_pool2d.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/max_pool2d.cpp rename to mx_driving/fused/ops/kernels/op_host/max_pool2d.cpp diff --git a/mx_driving/common/ops/kernels/op_host/max_pool2d.h b/mx_driving/fused/ops/kernels/op_host/max_pool2d.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/max_pool2d.h rename to mx_driving/fused/ops/kernels/op_host/max_pool2d.h diff --git a/mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn.cpp b/mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn.cpp rename to mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn.cpp diff --git a/mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_grad.cpp b/mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_grad.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_grad.cpp rename to mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_grad.cpp diff --git a/mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling.h b/mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling.h rename to mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling_v2.h b/mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling_v2.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling_v2.h rename to mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_grad_tiling_v2.h diff --git a/mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_grad_v2.cpp b/mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_grad_v2.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_grad_v2.cpp rename to mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_grad_v2.cpp diff --git a/mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_tiling.h b/mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/multi_scale_deformable_attn_tiling.h rename to mx_driving/fused/ops/kernels/op_host/multi_scale_deformable_attn_tiling.h diff --git a/mx_driving/perception/vision/ops/kernels/op_kernel/CMakeLists.txt b/mx_driving/fused/ops/kernels/op_kernel/CMakeLists.txt similarity index 100% rename from mx_driving/perception/vision/ops/kernels/op_kernel/CMakeLists.txt rename to mx_driving/fused/ops/kernels/op_kernel/CMakeLists.txt diff --git a/mx_driving/common/ops/kernels/op_kernel/add_relu.cpp b/mx_driving/fused/ops/kernels/op_kernel/add_relu.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/add_relu.cpp rename to mx_driving/fused/ops/kernels/op_kernel/add_relu.cpp diff --git a/mx_driving/perception/fused/ops/kernels/op_kernel/common.h b/mx_driving/fused/ops/kernels/op_kernel/common.h similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_kernel/common.h rename to mx_driving/fused/ops/kernels/op_kernel/common.h diff --git a/mx_driving/common/ops/kernels/op_kernel/deformable_aggregation.cpp b/mx_driving/fused/ops/kernels/op_kernel/deformable_aggregation.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/deformable_aggregation.cpp rename to mx_driving/fused/ops/kernels/op_kernel/deformable_aggregation.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/deformable_aggregation_grad.cpp b/mx_driving/fused/ops/kernels/op_kernel/deformable_aggregation_grad.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/deformable_aggregation_grad.cpp rename to mx_driving/fused/ops/kernels/op_kernel/deformable_aggregation_grad.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/fused_bias_leaky_relu.cpp b/mx_driving/fused/ops/kernels/op_kernel/fused_bias_leaky_relu.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/fused_bias_leaky_relu.cpp rename to mx_driving/fused/ops/kernels/op_kernel/fused_bias_leaky_relu.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/fused_bias_leaky_relu.h b/mx_driving/fused/ops/kernels/op_kernel/fused_bias_leaky_relu.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/fused_bias_leaky_relu.h rename to mx_driving/fused/ops/kernels/op_kernel/fused_bias_leaky_relu.h diff --git a/mx_driving/common/ops/kernels/op_kernel/max_pool2d.cpp b/mx_driving/fused/ops/kernels/op_kernel/max_pool2d.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/max_pool2d.cpp rename to mx_driving/fused/ops/kernels/op_kernel/max_pool2d.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_generic.h b/mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_generic.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_generic.h rename to mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_generic.h diff --git a/mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_grad_generic.h b/mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_grad_generic.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_grad_generic.h rename to mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_grad_generic.h diff --git a/mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_grad_generic_v2.h b/mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_grad_generic_v2.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_grad_generic_v2.h rename to mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_grad_generic_v2.h diff --git a/mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf.h b/mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf.h rename to mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf.h diff --git a/mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf_v2.h b/mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf_v2.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf_v2.h rename to mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_grad_high_perf_v2.h diff --git a/mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_high_perf.h b/mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_high_perf.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/ms_deform_attn_high_perf.h rename to mx_driving/fused/ops/kernels/op_kernel/ms_deform_attn_high_perf.h diff --git a/mx_driving/common/ops/kernels/op_kernel/multi_scale_deformable_attn.cpp b/mx_driving/fused/ops/kernels/op_kernel/multi_scale_deformable_attn.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/multi_scale_deformable_attn.cpp rename to mx_driving/fused/ops/kernels/op_kernel/multi_scale_deformable_attn.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/multi_scale_deformable_attn_grad.cpp b/mx_driving/fused/ops/kernels/op_kernel/multi_scale_deformable_attn_grad.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/multi_scale_deformable_attn_grad.cpp rename to mx_driving/fused/ops/kernels/op_kernel/multi_scale_deformable_attn_grad.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/multi_scale_deformable_attn_grad_v2.cpp b/mx_driving/fused/ops/kernels/op_kernel/multi_scale_deformable_attn_grad_v2.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/multi_scale_deformable_attn_grad_v2.cpp rename to mx_driving/fused/ops/kernels/op_kernel/multi_scale_deformable_attn_grad_v2.cpp diff --git a/mx_driving/common/ops/npu_add_relu.py b/mx_driving/fused/ops/npu_add_relu.py similarity index 100% rename from mx_driving/common/ops/npu_add_relu.py rename to mx_driving/fused/ops/npu_add_relu.py diff --git a/mx_driving/common/ops/npu_deformable_aggregation.py b/mx_driving/fused/ops/npu_deformable_aggregation.py similarity index 100% rename from mx_driving/common/ops/npu_deformable_aggregation.py rename to mx_driving/fused/ops/npu_deformable_aggregation.py diff --git a/mx_driving/common/ops/npu_max_pool2d.py b/mx_driving/fused/ops/npu_max_pool2d.py similarity index 99% rename from mx_driving/common/ops/npu_max_pool2d.py rename to mx_driving/fused/ops/npu_max_pool2d.py index 981b2888..38b68afb 100644 --- a/mx_driving/common/ops/npu_max_pool2d.py +++ b/mx_driving/fused/ops/npu_max_pool2d.py @@ -9,6 +9,7 @@ Modification 1. Add support for Ascend NPU from torch.autograd import Function import ads_c + class MaxPool2d(Function): @staticmethod # 'pylint: disable=too-many-arguments,huawei-too-many-arguments diff --git a/mx_driving/common/ops/npu_multi_scale_deformable_attn_function.py b/mx_driving/fused/ops/npu_multi_scale_deformable_attn_function.py similarity index 100% rename from mx_driving/common/ops/npu_multi_scale_deformable_attn_function.py rename to mx_driving/fused/ops/npu_multi_scale_deformable_attn_function.py diff --git a/mx_driving/motion/ops/csrc/pybind.cpp b/mx_driving/motion/ops/csrc/pybind.cpp deleted file mode 100644 index 7b362419..00000000 --- a/mx_driving/motion/ops/csrc/pybind.cpp +++ /dev/null @@ -1,5 +0,0 @@ -#include -#include "csrc/pybind.h" - -void init_motion(pybind11::module& m) { -} diff --git a/mx_driving/motion/ops/kernels/op_kernel/CMakeLists.txt b/mx_driving/motion/ops/kernels/op_kernel/CMakeLists.txt deleted file mode 100644 index e69de29b..00000000 diff --git a/mx_driving/perception/CMakeLists.txt b/mx_driving/perception/CMakeLists.txt deleted file mode 100644 index c8777acb..00000000 --- a/mx_driving/perception/CMakeLists.txt +++ /dev/null @@ -1,9 +0,0 @@ -if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/fused/ops/kernels) - add_subdirectory(fused/ops/kernels) -endif() -if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/point/ops/kernels) - add_subdirectory(point/ops/kernels) -endif() -if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/vision/ops/kernels) - add_subdirectory(vision/ops/kernels) -endif() diff --git a/mx_driving/perception/fused/__init__.py b/mx_driving/perception/fused/__init__.py deleted file mode 100644 index 30f8aefb..00000000 --- a/mx_driving/perception/fused/__init__.py +++ /dev/null @@ -1,2 +0,0 @@ -from .ops.bev_pool import bev_pool -from .ops.bev_pool_v2 import bev_pool_v2 \ No newline at end of file diff --git a/mx_driving/perception/fused/ops/csrc/pybind.cpp b/mx_driving/perception/fused/ops/csrc/pybind.cpp deleted file mode 100644 index d4001c57..00000000 --- a/mx_driving/perception/fused/ops/csrc/pybind.cpp +++ /dev/null @@ -1,13 +0,0 @@ -#include "csrc/pybind.h" - -#include - -#include "functions.h" -void init_perception_fused(pybind11::module& m) -{ - // bev_pool - m.def("npu_bev_pool", &npu_bev_pool, "npu_bev_pool NPU version"); - m.def("npu_bev_pool_backward", &npu_bev_pool_backward, "npu_bev_pool_backward NPU version"); - m.def("npu_bev_pool_v2", &npu_bev_pool_v2, "npu_bev_pool_v2 NPU version"); - m.def("npu_bev_pool_v2_backward", &npu_bev_pool_v2_backward, "npu_bev_pool_v2_backward NPU version"); -} diff --git a/mx_driving/perception/point/__init__.py b/mx_driving/perception/point/__init__.py deleted file mode 100644 index 218b2f73..00000000 --- a/mx_driving/perception/point/__init__.py +++ /dev/null @@ -1,2 +0,0 @@ -from .ops.group_points import npu_group_points - diff --git a/mx_driving/perception/vision/__init__.py b/mx_driving/perception/vision/__init__.py deleted file mode 100644 index 586c264f..00000000 --- a/mx_driving/perception/vision/__init__.py +++ /dev/null @@ -1 +0,0 @@ -from .ops.boxes_overlap_bev import boxes_overlap_bev diff --git a/mx_driving/perception/vision/ops/__init__.py b/mx_driving/perception/vision/ops/__init__.py deleted file mode 100644 index e69de29b..00000000 diff --git a/mx_driving/perception/vision/ops/csrc/README.md b/mx_driving/perception/vision/ops/csrc/README.md deleted file mode 100644 index 0bbe4f39..00000000 --- a/mx_driving/perception/vision/ops/csrc/README.md +++ /dev/null @@ -1,2 +0,0 @@ -## Description -+ The folder contains some cpp source files, which provide code for adaptation of ascend kernels. It provide links for kernels and cpp interfaces. \ No newline at end of file diff --git a/mx_driving/perception/vision/ops/kernels/README.md b/mx_driving/perception/vision/ops/kernels/README.md deleted file mode 100644 index 214fb0a6..00000000 --- a/mx_driving/perception/vision/ops/kernels/README.md +++ /dev/null @@ -1,2 +0,0 @@ -## Description -+ The folder contains some ascend-kernel source files, which are like cuda-kernels and supply some ops that can be run on ascend device. \ No newline at end of file diff --git a/mx_driving/point/CMakeLists.txt b/mx_driving/point/CMakeLists.txt new file mode 100644 index 00000000..621d1fa9 --- /dev/null +++ b/mx_driving/point/CMakeLists.txt @@ -0,0 +1,3 @@ +if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/ops/kernels) + add_subdirectory(ops/kernels) +endif() diff --git a/mx_driving/point/__init__.py b/mx_driving/point/__init__.py new file mode 100644 index 00000000..16f62bef --- /dev/null +++ b/mx_driving/point/__init__.py @@ -0,0 +1,8 @@ +from .ops.group_points import npu_group_points +from .ops.bev_pool import bev_pool +from .ops.bev_pool_v2 import bev_pool_v2 +from .ops.furthest_point_sampling_with_dist import furthest_point_sample_with_dist +from .ops.furthest_point_sampling import npu_furthest_point_sampling +from .ops.npu_dynamic_scatter import npu_dynamic_scatter +from .ops.voxelization import voxelization, Voxelization +from .ops.voxel_pooling_train import npu_voxel_pooling_train \ No newline at end of file diff --git a/mx_driving/perception/vision/components/README.md b/mx_driving/point/components/README.md similarity index 100% rename from mx_driving/perception/vision/components/README.md rename to mx_driving/point/components/README.md diff --git a/mx_driving/perception/point/ops/__init__.py b/mx_driving/point/ops/__init__.py similarity index 100% rename from mx_driving/perception/point/ops/__init__.py rename to mx_driving/point/ops/__init__.py diff --git a/mx_driving/perception/fused/ops/bev_pool.py b/mx_driving/point/ops/bev_pool.py similarity index 100% rename from mx_driving/perception/fused/ops/bev_pool.py rename to mx_driving/point/ops/bev_pool.py diff --git a/mx_driving/perception/fused/ops/bev_pool_v2.py b/mx_driving/point/ops/bev_pool_v2.py similarity index 100% rename from mx_driving/perception/fused/ops/bev_pool_v2.py rename to mx_driving/point/ops/bev_pool_v2.py diff --git a/mx_driving/perception/fused/ops/csrc/BEVPool.cpp b/mx_driving/point/ops/csrc/BEVPool.cpp similarity index 100% rename from mx_driving/perception/fused/ops/csrc/BEVPool.cpp rename to mx_driving/point/ops/csrc/BEVPool.cpp diff --git a/mx_driving/perception/fused/ops/csrc/BEVPoolBackward.cpp b/mx_driving/point/ops/csrc/BEVPoolBackward.cpp similarity index 100% rename from mx_driving/perception/fused/ops/csrc/BEVPoolBackward.cpp rename to mx_driving/point/ops/csrc/BEVPoolBackward.cpp diff --git a/mx_driving/perception/fused/ops/csrc/BEVPoolV2.cpp b/mx_driving/point/ops/csrc/BEVPoolV2.cpp similarity index 100% rename from mx_driving/perception/fused/ops/csrc/BEVPoolV2.cpp rename to mx_driving/point/ops/csrc/BEVPoolV2.cpp diff --git a/mx_driving/perception/fused/ops/csrc/BEVPoolV2Backward.cpp b/mx_driving/point/ops/csrc/BEVPoolV2Backward.cpp similarity index 100% rename from mx_driving/perception/fused/ops/csrc/BEVPoolV2Backward.cpp rename to mx_driving/point/ops/csrc/BEVPoolV2Backward.cpp diff --git a/mx_driving/common/ops/csrc/DynamicScatter.cpp b/mx_driving/point/ops/csrc/DynamicScatter.cpp similarity index 100% rename from mx_driving/common/ops/csrc/DynamicScatter.cpp rename to mx_driving/point/ops/csrc/DynamicScatter.cpp diff --git a/mx_driving/common/ops/csrc/DynamicVoxelization.cpp b/mx_driving/point/ops/csrc/DynamicVoxelization.cpp similarity index 100% rename from mx_driving/common/ops/csrc/DynamicVoxelization.cpp rename to mx_driving/point/ops/csrc/DynamicVoxelization.cpp diff --git a/mx_driving/common/ops/csrc/FurthestPointSampling.cpp b/mx_driving/point/ops/csrc/FurthestPointSampling.cpp similarity index 100% rename from mx_driving/common/ops/csrc/FurthestPointSampling.cpp rename to mx_driving/point/ops/csrc/FurthestPointSampling.cpp diff --git a/mx_driving/common/ops/csrc/FurthestPointSamplingWithDist.cpp b/mx_driving/point/ops/csrc/FurthestPointSamplingWithDist.cpp similarity index 100% rename from mx_driving/common/ops/csrc/FurthestPointSamplingWithDist.cpp rename to mx_driving/point/ops/csrc/FurthestPointSamplingWithDist.cpp diff --git a/mx_driving/perception/point/ops/csrc/GroupPoints.cpp b/mx_driving/point/ops/csrc/GroupPoints.cpp similarity index 100% rename from mx_driving/perception/point/ops/csrc/GroupPoints.cpp rename to mx_driving/point/ops/csrc/GroupPoints.cpp diff --git a/mx_driving/perception/point/ops/csrc/HardVoxelize.cpp b/mx_driving/point/ops/csrc/HardVoxelize.cpp similarity index 100% rename from mx_driving/perception/point/ops/csrc/HardVoxelize.cpp rename to mx_driving/point/ops/csrc/HardVoxelize.cpp diff --git a/mx_driving/perception/point/ops/csrc/PointToVoxel.cpp b/mx_driving/point/ops/csrc/PointToVoxel.cpp similarity index 100% rename from mx_driving/perception/point/ops/csrc/PointToVoxel.cpp rename to mx_driving/point/ops/csrc/PointToVoxel.cpp diff --git a/mx_driving/perception/point/ops/csrc/README.md b/mx_driving/point/ops/csrc/README.md similarity index 100% rename from mx_driving/perception/point/ops/csrc/README.md rename to mx_driving/point/ops/csrc/README.md diff --git a/mx_driving/perception/point/ops/csrc/UniqueVoxel.cpp b/mx_driving/point/ops/csrc/UniqueVoxel.cpp similarity index 100% rename from mx_driving/perception/point/ops/csrc/UniqueVoxel.cpp rename to mx_driving/point/ops/csrc/UniqueVoxel.cpp diff --git a/mx_driving/perception/point/ops/csrc/VecPoolBackward.cpp b/mx_driving/point/ops/csrc/VecPoolBackward.cpp similarity index 100% rename from mx_driving/perception/point/ops/csrc/VecPoolBackward.cpp rename to mx_driving/point/ops/csrc/VecPoolBackward.cpp diff --git a/mx_driving/common/ops/csrc/VoxelPoolingTrain.cpp b/mx_driving/point/ops/csrc/VoxelPoolingTrain.cpp similarity index 100% rename from mx_driving/common/ops/csrc/VoxelPoolingTrain.cpp rename to mx_driving/point/ops/csrc/VoxelPoolingTrain.cpp diff --git a/mx_driving/perception/point/ops/csrc/VoxelToPoint.cpp b/mx_driving/point/ops/csrc/VoxelToPoint.cpp similarity index 100% rename from mx_driving/perception/point/ops/csrc/VoxelToPoint.cpp rename to mx_driving/point/ops/csrc/VoxelToPoint.cpp diff --git a/mx_driving/perception/point/ops/csrc/functions.h b/mx_driving/point/ops/csrc/functions.h similarity index 43% rename from mx_driving/perception/point/ops/csrc/functions.h rename to mx_driving/point/ops/csrc/functions.h index 00127f9c..8c8a43bf 100644 --- a/mx_driving/perception/point/ops/csrc/functions.h +++ b/mx_driving/point/ops/csrc/functions.h @@ -39,4 +39,39 @@ std::tuple unique_voxel std::tuple hard_voxelize(const at::Tensor& points, const std::vector voxel_sizes, const std::vector coor_ranges, int64_t max_points, int64_t max_voxels); +at::Tensor npu_bev_pool(const at::Tensor& feat, const at::Tensor& geom_feat, const at::Tensor& interval_lengths, + const at::Tensor& interval_starts, int64_t b, int64_t d, int64_t h, int64_t w); +at::Tensor npu_bev_pool_backward(const at::Tensor& grad_out, const at::Tensor& geom_feat, + const at::Tensor& interval_lengths, const at::Tensor& interval_starts, int64_t b, int64_t d, int64_t h, int64_t w); + +at::Tensor npu_bev_pool_v2(const at::Tensor& depth, const at::Tensor& feat, const at::Tensor& ranks_depth, + const at::Tensor& ranks_feat, const at::Tensor& ranks_bev, const at::Tensor& interval_lengths, + const at::Tensor& interval_starts, int64_t b, int64_t d, int64_t h, int64_t w); +std::tuple npu_bev_pool_v2_backward(const at::Tensor& grad_out, const at::Tensor& depth, + const at::Tensor& feat, const at::Tensor& ranks_depth, const at::Tensor& ranks_feat, const at::Tensor& ranks_bev, + const at::Tensor& interval_lengths, const at::Tensor& interval_starts, int64_t b, int64_t d, int64_t h, int64_t w); + +at::Tensor furthest_point_sampling_with_dist( + const at::Tensor& points_dist, const at::Tensor& nearest_temp, int32_t num_points); + +std::tuple npu_dynamic_scatter(const at::Tensor& feats, const at::Tensor& coors, + const at::Tensor& prefix_sum_point_per_voxel, const at::Tensor& argsort_coor, int32_t num_voxels, + const char* reduce_type); + +void npu_dynamic_scatter_grad(at::Tensor& grad_point_feats, const at::Tensor& grad_voxel_feats, + const at::Tensor& prefix_sum_point_per_voxel, const at::Tensor& argsort_coor, const at::Tensor& compare_mask, + const char* reduce_type); + +at::Tensor npu_furthest_point_sampling(const at::Tensor& point_xyz, const at::Tensor& nearset_temp, int32_t num_points); + +std::tuple voxel_pooling_train(const at::Tensor& inputFeatures, const at::Tensor& geom, + at::Tensor& outputFeatures, at::Tensor& posMemo, int batchSize, int numPoints, int numChannels, int numVoxelX, + int numVoxelY, int numVoxelZ); + +at::Tensor voxel_pool_train_backward(const at::Tensor& grad_out, const at::Tensor& posMemo, const int64_t batchSize, + const int64_t numPoints, const int64_t numChannels, const int64_t h, const int64_t w); + +at::Tensor dynamic_voxelization(const at::Tensor& points, at::Tensor& coors, int grid_x, int grid_y, int grid_z, + double voxel_x, double voxel_y, double voxel_z, double coors_min_x, double coors_min_y, double coorsMinZ); + #endif // PERCEPTION_POINT_OPS_CSRC_FUNCTIONS_H_ diff --git a/mx_driving/perception/point/ops/csrc/pybind.cpp b/mx_driving/point/ops/csrc/pybind.cpp similarity index 31% rename from mx_driving/perception/point/ops/csrc/pybind.cpp rename to mx_driving/point/ops/csrc/pybind.cpp index e4a5fe44..98fe18c7 100644 --- a/mx_driving/perception/point/ops/csrc/pybind.cpp +++ b/mx_driving/point/ops/csrc/pybind.cpp @@ -4,7 +4,7 @@ #include "functions.h" -void init_perception_point(pybind11::module& m) +void init_point(pybind11::module& m) { // group_points m.def("group_points", &group_points); @@ -20,4 +20,27 @@ void init_perception_point(pybind11::module& m) m.def("unique_voxel", &unique_voxel); m.def("hard_voxelize", &hard_voxelize); + + // bev_pool + m.def("npu_bev_pool", &npu_bev_pool, "npu_bev_pool NPU version"); + m.def("npu_bev_pool_backward", &npu_bev_pool_backward, "npu_bev_pool_backward NPU version"); + m.def("npu_bev_pool_v2", &npu_bev_pool_v2, "npu_bev_pool_v2 NPU version"); + m.def("npu_bev_pool_v2_backward", &npu_bev_pool_v2_backward, "npu_bev_pool_v2_backward NPU version"); + + // furthest_points_sampling_with_dist + m.def("furthest_point_sampling_with_dist", &furthest_point_sampling_with_dist); + + // npu_dynamic_scatter + m.def("npu_dynamic_scatter", &npu_dynamic_scatter); + m.def("npu_dynamic_scatter_grad", &npu_dynamic_scatter_grad); + + // dyn_voxelization + m.def("dynamic_voxelization", &dynamic_voxelization); + + // npu_furthest_point_sampling + m.def("npu_furthest_point_sampling", &npu_furthest_point_sampling); + + // voxel_pooling + m.def("voxel_pooling_train", &voxel_pooling_train); + m.def("voxel_pool_train_backward", &voxel_pool_train_backward); } diff --git a/mx_driving/common/ops/furthest_point_sampling.py b/mx_driving/point/ops/furthest_point_sampling.py similarity index 100% rename from mx_driving/common/ops/furthest_point_sampling.py rename to mx_driving/point/ops/furthest_point_sampling.py diff --git a/mx_driving/common/ops/furthest_point_sampling_with_dist.py b/mx_driving/point/ops/furthest_point_sampling_with_dist.py similarity index 100% rename from mx_driving/common/ops/furthest_point_sampling_with_dist.py rename to mx_driving/point/ops/furthest_point_sampling_with_dist.py diff --git a/mx_driving/perception/point/ops/group_points.py b/mx_driving/point/ops/group_points.py similarity index 100% rename from mx_driving/perception/point/ops/group_points.py rename to mx_driving/point/ops/group_points.py diff --git a/mx_driving/perception/vision/ops/kernels/CMakeLists.txt b/mx_driving/point/ops/kernels/CMakeLists.txt similarity index 100% rename from mx_driving/perception/vision/ops/kernels/CMakeLists.txt rename to mx_driving/point/ops/kernels/CMakeLists.txt diff --git a/mx_driving/perception/point/ops/kernels/README.md b/mx_driving/point/ops/kernels/README.md similarity index 100% rename from mx_driving/perception/point/ops/kernels/README.md rename to mx_driving/point/ops/kernels/README.md diff --git a/mx_driving/perception/vision/ops/kernels/op_host/CMakeLists.txt b/mx_driving/point/ops/kernels/op_host/CMakeLists.txt similarity index 100% rename from mx_driving/perception/vision/ops/kernels/op_host/CMakeLists.txt rename to mx_driving/point/ops/kernels/op_host/CMakeLists.txt diff --git a/mx_driving/perception/fused/ops/kernels/op_host/bev_pool.cpp b/mx_driving/point/ops/kernels/op_host/bev_pool.cpp similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_host/bev_pool.cpp rename to mx_driving/point/ops/kernels/op_host/bev_pool.cpp diff --git a/mx_driving/perception/fused/ops/kernels/op_host/bev_pool_tiling.h b/mx_driving/point/ops/kernels/op_host/bev_pool_tiling.h similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_host/bev_pool_tiling.h rename to mx_driving/point/ops/kernels/op_host/bev_pool_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/dynamic_scatter.cpp b/mx_driving/point/ops/kernels/op_host/dynamic_scatter.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/dynamic_scatter.cpp rename to mx_driving/point/ops/kernels/op_host/dynamic_scatter.cpp diff --git a/mx_driving/common/ops/kernels/op_host/dynamic_scatter_grad.cpp b/mx_driving/point/ops/kernels/op_host/dynamic_scatter_grad.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/dynamic_scatter_grad.cpp rename to mx_driving/point/ops/kernels/op_host/dynamic_scatter_grad.cpp diff --git a/mx_driving/common/ops/kernels/op_host/dynamic_scatter_grad_tiling.h b/mx_driving/point/ops/kernels/op_host/dynamic_scatter_grad_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/dynamic_scatter_grad_tiling.h rename to mx_driving/point/ops/kernels/op_host/dynamic_scatter_grad_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/dynamic_scatter_tiling.h b/mx_driving/point/ops/kernels/op_host/dynamic_scatter_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/dynamic_scatter_tiling.h rename to mx_driving/point/ops/kernels/op_host/dynamic_scatter_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/dynamic_voxelization.cpp b/mx_driving/point/ops/kernels/op_host/dynamic_voxelization.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/dynamic_voxelization.cpp rename to mx_driving/point/ops/kernels/op_host/dynamic_voxelization.cpp diff --git a/mx_driving/common/ops/kernels/op_host/dynamic_voxelization_tiling.h b/mx_driving/point/ops/kernels/op_host/dynamic_voxelization_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/dynamic_voxelization_tiling.h rename to mx_driving/point/ops/kernels/op_host/dynamic_voxelization_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/furthest_point_sampling.cpp b/mx_driving/point/ops/kernels/op_host/furthest_point_sampling.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/furthest_point_sampling.cpp rename to mx_driving/point/ops/kernels/op_host/furthest_point_sampling.cpp diff --git a/mx_driving/common/ops/kernels/op_host/furthest_point_sampling_tiling.h b/mx_driving/point/ops/kernels/op_host/furthest_point_sampling_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/furthest_point_sampling_tiling.h rename to mx_driving/point/ops/kernels/op_host/furthest_point_sampling_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/furthest_point_sampling_with_dist.cpp b/mx_driving/point/ops/kernels/op_host/furthest_point_sampling_with_dist.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/furthest_point_sampling_with_dist.cpp rename to mx_driving/point/ops/kernels/op_host/furthest_point_sampling_with_dist.cpp diff --git a/mx_driving/common/ops/kernels/op_host/furthest_point_sampling_with_dist_tiling.h b/mx_driving/point/ops/kernels/op_host/furthest_point_sampling_with_dist_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/furthest_point_sampling_with_dist_tiling.h rename to mx_driving/point/ops/kernels/op_host/furthest_point_sampling_with_dist_tiling.h diff --git a/mx_driving/perception/point/ops/kernels/op_host/group_points.cpp b/mx_driving/point/ops/kernels/op_host/group_points.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/group_points.cpp rename to mx_driving/point/ops/kernels/op_host/group_points.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_host/group_points_grad.cpp b/mx_driving/point/ops/kernels/op_host/group_points_grad.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/group_points_grad.cpp rename to mx_driving/point/ops/kernels/op_host/group_points_grad.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_host/group_points_grad_tiling.h b/mx_driving/point/ops/kernels/op_host/group_points_grad_tiling.h similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/group_points_grad_tiling.h rename to mx_driving/point/ops/kernels/op_host/group_points_grad_tiling.h diff --git a/mx_driving/perception/point/ops/kernels/op_host/group_points_tiling.h b/mx_driving/point/ops/kernels/op_host/group_points_tiling.h similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/group_points_tiling.h rename to mx_driving/point/ops/kernels/op_host/group_points_tiling.h diff --git a/mx_driving/perception/point/ops/kernels/op_host/hard_voxelize.cpp b/mx_driving/point/ops/kernels/op_host/hard_voxelize.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/hard_voxelize.cpp rename to mx_driving/point/ops/kernels/op_host/hard_voxelize.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_host/hard_voxelize_tiling.h b/mx_driving/point/ops/kernels/op_host/hard_voxelize_tiling.h similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/hard_voxelize_tiling.h rename to mx_driving/point/ops/kernels/op_host/hard_voxelize_tiling.h diff --git a/mx_driving/perception/point/ops/kernels/op_host/point_to_voxel.cpp b/mx_driving/point/ops/kernels/op_host/point_to_voxel.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/point_to_voxel.cpp rename to mx_driving/point/ops/kernels/op_host/point_to_voxel.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_host/point_to_voxel_tiling.h b/mx_driving/point/ops/kernels/op_host/point_to_voxel_tiling.h similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/point_to_voxel_tiling.h rename to mx_driving/point/ops/kernels/op_host/point_to_voxel_tiling.h diff --git a/mx_driving/perception/point/ops/kernels/op_host/unique_voxel.cpp b/mx_driving/point/ops/kernels/op_host/unique_voxel.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/unique_voxel.cpp rename to mx_driving/point/ops/kernels/op_host/unique_voxel.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_host/unique_voxel_tiling.h b/mx_driving/point/ops/kernels/op_host/unique_voxel_tiling.h similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/unique_voxel_tiling.h rename to mx_driving/point/ops/kernels/op_host/unique_voxel_tiling.h diff --git a/mx_driving/perception/point/ops/kernels/op_host/vec_pool_grad.cpp b/mx_driving/point/ops/kernels/op_host/vec_pool_grad.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/vec_pool_grad.cpp rename to mx_driving/point/ops/kernels/op_host/vec_pool_grad.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_host/vec_pool_grad_tiling.h b/mx_driving/point/ops/kernels/op_host/vec_pool_grad_tiling.h similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_host/vec_pool_grad_tiling.h rename to mx_driving/point/ops/kernels/op_host/vec_pool_grad_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/voxel_pooling_train.cpp b/mx_driving/point/ops/kernels/op_host/voxel_pooling_train.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/voxel_pooling_train.cpp rename to mx_driving/point/ops/kernels/op_host/voxel_pooling_train.cpp diff --git a/mx_driving/common/ops/kernels/op_host/voxel_pooling_train_grad.cpp b/mx_driving/point/ops/kernels/op_host/voxel_pooling_train_grad.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_host/voxel_pooling_train_grad.cpp rename to mx_driving/point/ops/kernels/op_host/voxel_pooling_train_grad.cpp diff --git a/mx_driving/common/ops/kernels/op_host/voxel_pooling_train_grad_tiling.h b/mx_driving/point/ops/kernels/op_host/voxel_pooling_train_grad_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/voxel_pooling_train_grad_tiling.h rename to mx_driving/point/ops/kernels/op_host/voxel_pooling_train_grad_tiling.h diff --git a/mx_driving/common/ops/kernels/op_host/voxel_pooling_train_tiling.h b/mx_driving/point/ops/kernels/op_host/voxel_pooling_train_tiling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_host/voxel_pooling_train_tiling.h rename to mx_driving/point/ops/kernels/op_host/voxel_pooling_train_tiling.h diff --git a/mx_driving/point/ops/kernels/op_kernel/CMakeLists.txt b/mx_driving/point/ops/kernels/op_kernel/CMakeLists.txt new file mode 100644 index 00000000..c51870f1 --- /dev/null +++ b/mx_driving/point/ops/kernels/op_kernel/CMakeLists.txt @@ -0,0 +1,4 @@ +file(GLOB KERNEL_SRC ${CMAKE_CURRENT_SOURCE_DIR}/*.cpp ${CMAKE_CURRENT_SOURCE_DIR}/*.h) +set(ASCEND_KERNEL_SRC + ${ASCEND_KERNEL_SRC} ${KERNEL_SRC} + CACHE INTERNAL "") diff --git a/mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool.cpp b/mx_driving/point/ops/kernels/op_kernel/bev_pool.cpp similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool.cpp rename to mx_driving/point/ops/kernels/op_kernel/bev_pool.cpp diff --git a/mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool.h b/mx_driving/point/ops/kernels/op_kernel/bev_pool.h similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool.h rename to mx_driving/point/ops/kernels/op_kernel/bev_pool.h diff --git a/mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool_grad.cpp b/mx_driving/point/ops/kernels/op_kernel/bev_pool_grad.cpp similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool_grad.cpp rename to mx_driving/point/ops/kernels/op_kernel/bev_pool_grad.cpp diff --git a/mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool_v2.cpp b/mx_driving/point/ops/kernels/op_kernel/bev_pool_v2.cpp similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool_v2.cpp rename to mx_driving/point/ops/kernels/op_kernel/bev_pool_v2.cpp diff --git a/mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool_v2.h b/mx_driving/point/ops/kernels/op_kernel/bev_pool_v2.h similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool_v2.h rename to mx_driving/point/ops/kernels/op_kernel/bev_pool_v2.h diff --git a/mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool_v2_grad.cpp b/mx_driving/point/ops/kernels/op_kernel/bev_pool_v2_grad.cpp similarity index 100% rename from mx_driving/perception/fused/ops/kernels/op_kernel/bev_pool_v2_grad.cpp rename to mx_driving/point/ops/kernels/op_kernel/bev_pool_v2_grad.cpp diff --git a/mx_driving/point/ops/kernels/op_kernel/common.h b/mx_driving/point/ops/kernels/op_kernel/common.h new file mode 100644 index 00000000..2041af49 --- /dev/null +++ b/mx_driving/point/ops/kernels/op_kernel/common.h @@ -0,0 +1,46 @@ +#ifndef COMMON_H_ +#define COMMON_H_ + +#include "kernel_operator.h" + +constexpr int32_t TILING_ALIGN32B_FLAG = 1; +constexpr int32_t TILING_FP32_BIT = 1; +constexpr int32_t TILING_FP16_BIT = 2; +constexpr int32_t TILING_BF16_BIT = 3; + +class TaskIterator { +public: + __aicore__ inline TaskIterator( + int32_t blkIdx, int32_t blkDim, int32_t avgTaskNum, int32_t tailTaskNum, int32_t totalTaskNum) + : blkIdx_(blkIdx), blkDim_(blkDim), totalTaskNum_(totalTaskNum) + { + nextIdx_ = blkIdx * avgTaskNum + (blkIdx < tailTaskNum ? blkIdx : tailTaskNum); + endIdx_ = nextIdx_ + avgTaskNum + (blkIdx < tailTaskNum ? 1 : 0); + } + + __aicore__ inline bool HasNext() const + { + return nextIdx_ < endIdx_; + } + + __aicore__ inline int32_t Next() + { + return nextIdx_++; + } + + __aicore__ inline int32_t GetNext() const + { + return nextIdx_; + } + + __aicore__ inline int32_t GetTaskNum() const + { + return totalTaskNum_; + } + +private: + int32_t blkIdx_, blkDim_; + int32_t nextIdx_, endIdx_; + int32_t totalTaskNum_; +}; +#endif // COMMON_H_ \ No newline at end of file diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter.cpp b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter.cpp rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_base.h b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_base.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_base.h rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_base.h diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad.cpp b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad.cpp rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad_base.h b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad_base.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad_base.h rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad_base.h diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad_max.h b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad_max.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad_max.h rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad_max.h diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad_mean.h b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad_mean.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad_mean.h rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad_mean.h diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad_sum.h b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad_sum.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_grad_sum.h rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_grad_sum.h diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_max.h b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_max.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_max.h rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_max.h diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_mean.h b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_mean.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_mean.h rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_mean.h diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_sum.h b/mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_sum.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_scatter_sum.h rename to mx_driving/point/ops/kernels/op_kernel/dynamic_scatter_sum.h diff --git a/mx_driving/common/ops/kernels/op_kernel/dynamic_voxelization.cpp b/mx_driving/point/ops/kernels/op_kernel/dynamic_voxelization.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/dynamic_voxelization.cpp rename to mx_driving/point/ops/kernels/op_kernel/dynamic_voxelization.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/furthest_point_sampling.cpp b/mx_driving/point/ops/kernels/op_kernel/furthest_point_sampling.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/furthest_point_sampling.cpp rename to mx_driving/point/ops/kernels/op_kernel/furthest_point_sampling.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/furthest_point_sampling.h b/mx_driving/point/ops/kernels/op_kernel/furthest_point_sampling.h similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/furthest_point_sampling.h rename to mx_driving/point/ops/kernels/op_kernel/furthest_point_sampling.h diff --git a/mx_driving/common/ops/kernels/op_kernel/furthest_point_sampling_with_dist.cpp b/mx_driving/point/ops/kernels/op_kernel/furthest_point_sampling_with_dist.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/furthest_point_sampling_with_dist.cpp rename to mx_driving/point/ops/kernels/op_kernel/furthest_point_sampling_with_dist.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_kernel/group_points.cpp b/mx_driving/point/ops/kernels/op_kernel/group_points.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_kernel/group_points.cpp rename to mx_driving/point/ops/kernels/op_kernel/group_points.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_kernel/group_points_grad.cpp b/mx_driving/point/ops/kernels/op_kernel/group_points_grad.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_kernel/group_points_grad.cpp rename to mx_driving/point/ops/kernels/op_kernel/group_points_grad.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_kernel/hard_voxelize.cpp b/mx_driving/point/ops/kernels/op_kernel/hard_voxelize.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_kernel/hard_voxelize.cpp rename to mx_driving/point/ops/kernels/op_kernel/hard_voxelize.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_kernel/point_to_voxel.cpp b/mx_driving/point/ops/kernels/op_kernel/point_to_voxel.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_kernel/point_to_voxel.cpp rename to mx_driving/point/ops/kernels/op_kernel/point_to_voxel.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_kernel/unique_voxel.cpp b/mx_driving/point/ops/kernels/op_kernel/unique_voxel.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_kernel/unique_voxel.cpp rename to mx_driving/point/ops/kernels/op_kernel/unique_voxel.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_kernel/vec_pool_grad.cpp b/mx_driving/point/ops/kernels/op_kernel/vec_pool_grad.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_kernel/vec_pool_grad.cpp rename to mx_driving/point/ops/kernels/op_kernel/vec_pool_grad.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/voxel_pooling_train.cpp b/mx_driving/point/ops/kernels/op_kernel/voxel_pooling_train.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/voxel_pooling_train.cpp rename to mx_driving/point/ops/kernels/op_kernel/voxel_pooling_train.cpp diff --git a/mx_driving/common/ops/kernels/op_kernel/voxel_pooling_train_grad.cpp b/mx_driving/point/ops/kernels/op_kernel/voxel_pooling_train_grad.cpp similarity index 100% rename from mx_driving/common/ops/kernels/op_kernel/voxel_pooling_train_grad.cpp rename to mx_driving/point/ops/kernels/op_kernel/voxel_pooling_train_grad.cpp diff --git a/mx_driving/perception/point/ops/kernels/op_kernel/voxel_to_point.cpp b/mx_driving/point/ops/kernels/op_kernel/voxel_to_point.cpp similarity index 100% rename from mx_driving/perception/point/ops/kernels/op_kernel/voxel_to_point.cpp rename to mx_driving/point/ops/kernels/op_kernel/voxel_to_point.cpp diff --git a/mx_driving/common/ops/npu_dynamic_scatter.py b/mx_driving/point/ops/npu_dynamic_scatter.py similarity index 100% rename from mx_driving/common/ops/npu_dynamic_scatter.py rename to mx_driving/point/ops/npu_dynamic_scatter.py diff --git a/mx_driving/common/ops/voxel_pooling_train.py b/mx_driving/point/ops/voxel_pooling_train.py similarity index 100% rename from mx_driving/common/ops/voxel_pooling_train.py rename to mx_driving/point/ops/voxel_pooling_train.py diff --git a/mx_driving/common/ops/voxelization.py b/mx_driving/point/ops/voxelization.py similarity index 100% rename from mx_driving/common/ops/voxelization.py rename to mx_driving/point/ops/voxelization.py diff --git a/setup.py b/setup.py index a317752d..1effcb4a 100644 --- a/setup.py +++ b/setup.py @@ -12,7 +12,7 @@ from utils import extension BASE_DIR = os.path.dirname(os.path.realpath(__file__)) VERSION = torch.__version__ -full_components = ["common", "motion", "perception/fused", "perception/point", "perception/vision", "spconv"] +full_components = ["common", "data", "fused", "point", "detection", "spconv"] source_file = glob.glob(os.path.join("./bind/", "*.cpp")) include_dirs = [os.path.join(BASE_DIR, "include")] for part in full_components: diff --git a/tests/torch/test_add_relu.py b/tests/torch/test_add_relu.py index 210d72da..2a9901bb 100644 --- a/tests/torch/test_add_relu.py +++ b/tests/torch/test_add_relu.py @@ -5,7 +5,7 @@ import torch_npu import torch.nn.functional as F from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.fused DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -18,7 +18,7 @@ class TestPointsInBox(TestCase): y = np.random.uniform(2.0, 2.0, [1, 100, 3]).astype(np.float32) y = torch.from_numpy(y) cpu_result = F.relu(x + y) - x = mx_driving.common.npu_add_relu(x.npu(), y.npu()).cpu().numpy() + x = mx_driving.fused.npu_add_relu(x.npu(), y.npu()).cpu().numpy() self.assertRtolEqual(x, cpu_result.numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `AddRelu` is only supported on 910B, skip this ut!") @@ -28,7 +28,7 @@ class TestPointsInBox(TestCase): y = np.random.uniform(2.0, 2.0, [18, 256, 232, 400]).astype(np.float32) y = torch.from_numpy(y) cpu_result = F.relu(x + y) - x = mx_driving.common.npu_add_relu(x.npu(), y.npu()).cpu().numpy() + x = mx_driving.fused.npu_add_relu(x.npu(), y.npu()).cpu().numpy() self.assertRtolEqual(x, cpu_result.numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `AddRelu` is only supported on 910B, skip this ut!") @@ -38,7 +38,7 @@ class TestPointsInBox(TestCase): y = np.random.uniform(2.0, 2.0, [18, 256, 232, 400]).astype(np.float16) y = torch.from_numpy(y) cpu_result = F.relu(x.float() + y.float()) - x = mx_driving.common.npu_add_relu(x.npu(), y.npu()).cpu().numpy() + x = mx_driving.fused.npu_add_relu(x.npu(), y.npu()).cpu().numpy() self.assertRtolEqual(x, cpu_result.half().numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `AddRelu` is only supported on 910B, skip this ut!") @@ -48,7 +48,7 @@ class TestPointsInBox(TestCase): y = np.random.uniform(2.0, 2.0, [18]).astype(np.float16) y = torch.from_numpy(y) cpu_result = F.relu(x.float() + y.float()) - x = mx_driving.common.npu_add_relu(x.npu(), y.npu()).cpu().numpy() + x = mx_driving.fused.npu_add_relu(x.npu(), y.npu()).cpu().numpy() self.assertRtolEqual(x, cpu_result.half().numpy()) diff --git a/tests/torch/test_bev_pool.py b/tests/torch/test_bev_pool.py index f7cbf4f3..f08e92ad 100644 --- a/tests/torch/test_bev_pool.py +++ b/tests/torch/test_bev_pool.py @@ -3,7 +3,7 @@ import torch import numpy as np import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -from mx_driving.perception.fused import bev_pool +from mx_driving.point import bev_pool DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] diff --git a/tests/torch/test_bev_pool_v2.py b/tests/torch/test_bev_pool_v2.py index ac8c4939..9870e4ff 100644 --- a/tests/torch/test_bev_pool_v2.py +++ b/tests/torch/test_bev_pool_v2.py @@ -6,7 +6,7 @@ import torch_npu from ads_c import npu_bev_pool_v2_backward from torch_npu.testing.testcase import TestCase, run_tests -from mx_driving.perception.fused import bev_pool_v2 +from mx_driving.point import bev_pool_v2 DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] diff --git a/tests/torch/test_boxes_overlap_bev.py b/tests/torch/test_boxes_overlap_bev.py index 41ce509d..209cbde5 100644 --- a/tests/torch/test_boxes_overlap_bev.py +++ b/tests/torch/test_boxes_overlap_bev.py @@ -6,7 +6,7 @@ import numpy as np import torch import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.perception.vision +import mx_driving.detection DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -284,7 +284,7 @@ class TestBoxesOverlapBev(TestCase): def npu_to_exec(self, npu_inputs): npu_boxes_a = npu_inputs.boxes_a npu_boxes_b = npu_inputs.boxes_b - npu_ans_overlap = mx_driving.perception.vision.boxes_overlap_bev(npu_boxes_a, npu_boxes_b) + npu_ans_overlap = mx_driving.detection.boxes_overlap_bev(npu_boxes_a, npu_boxes_b) return npu_ans_overlap.cpu().float().numpy() def check_precision(self, actual, expected, rtol=1e-4, atol=1e-4, msg=None): diff --git a/tests/torch/test_deformable_aggregation.py b/tests/torch/test_deformable_aggregation.py index 2c22b0a1..a274b789 100644 --- a/tests/torch/test_deformable_aggregation.py +++ b/tests/torch/test_deformable_aggregation.py @@ -4,7 +4,7 @@ import numpy as np import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.fused DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -151,7 +151,7 @@ class TestDeformableAggregation(TestCase): num_scale, num_embeds, num_groups, num_feat, feature_maps, spatial_shape, scale_start_index, sample_location, weights) - out_npu = mx_driving.common.npu_deformable_aggregation(torch_feature_maps, + out_npu = mx_driving.fused.npu_deformable_aggregation(torch_feature_maps, torch_spatial_shape, torch_scale_start_index, torch_sample_location, diff --git a/tests/torch/test_deformable_aggregation_grad.py b/tests/torch/test_deformable_aggregation_grad.py index aec4c8be..8c2bbae4 100644 --- a/tests/torch/test_deformable_aggregation_grad.py +++ b/tests/torch/test_deformable_aggregation_grad.py @@ -5,7 +5,7 @@ import torch import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.fused DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -227,7 +227,7 @@ class TestDeformableAggregation(TestCase): grad_weights, ) - out_npu = mx_driving.common.npu_deformable_aggregation( + out_npu = mx_driving.fused.npu_deformable_aggregation( torch_feature_maps, torch_spatial_shape, torch_scale_start_index, diff --git a/tests/torch/test_furthest_point_sample_with_dist.py b/tests/torch/test_furthest_point_sample_with_dist.py index e6302881..9f13d8c3 100644 --- a/tests/torch/test_furthest_point_sample_with_dist.py +++ b/tests/torch/test_furthest_point_sample_with_dist.py @@ -18,7 +18,7 @@ import numpy as np import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.point DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -65,7 +65,7 @@ class TestFurthestPointSampleWithDist(TestCase): def custom_op_exec(self, point_dist, point_num, input_dtype): point_dist_npu = torch.tensor(point_dist, dtype=input_dtype).npu() - output = mx_driving.common.furthest_point_sample_with_dist(point_dist_npu, point_num) + output = mx_driving.point.furthest_point_sample_with_dist(point_dist_npu, point_num) return output.cpu().numpy() @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `FurthestPointSampleWithDist` is only supported on 910B, skip this ut!") diff --git a/tests/torch/test_furthest_point_sampling.py b/tests/torch/test_furthest_point_sampling.py index a822d2fe..a637d8eb 100644 --- a/tests/torch/test_furthest_point_sampling.py +++ b/tests/torch/test_furthest_point_sampling.py @@ -20,7 +20,7 @@ import torch import torch_npu from torch_npu.testing.testcase import TestCase, run_tests from torch_npu.testing.common_utils import create_common_tensor -import mx_driving.common +import mx_driving.point DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -142,7 +142,7 @@ class TestFurthestPointSample(TestCase): return myTest.getCpuRes() def npu_op_exec(self, myTest): - return mx_driving.common.npu_furthest_point_sampling(myTest.point.clone().permute(0, 2, 1).npu(), myTest.numPoints) + return mx_driving.point.npu_furthest_point_sampling(myTest.point.clone().permute(0, 2, 1).npu(), myTest.numPoints) def compare_res(self, myTest): myTest.createData() diff --git a/tests/torch/test_fused_bias_leaky_relu.py b/tests/torch/test_fused_bias_leaky_relu.py index 731a6635..cfd7f15e 100644 --- a/tests/torch/test_fused_bias_leaky_relu.py +++ b/tests/torch/test_fused_bias_leaky_relu.py @@ -5,7 +5,7 @@ import torch_npu import torch.nn.functional as F from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.fused DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -27,7 +27,7 @@ class TestFusedBiasLeakyRelu(TestCase): cpu_result = F.leaky_relu(x + bias, negative_slop) cpu_result = cpu_result * scale - npu_result = mx_driving.common.npu_fused_bias_leaky_relu(x.npu(), bias.npu(), negative_slop, scale).cpu().numpy() + npu_result = mx_driving.fused.npu_fused_bias_leaky_relu(x.npu(), bias.npu(), negative_slop, scale).cpu().numpy() self.assertRtolEqual(npu_result, cpu_result.numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `FusedBiasLeakyRelu` is only supported on 910B, skip this ut!") @@ -40,7 +40,7 @@ class TestFusedBiasLeakyRelu(TestCase): cpu_result = F.leaky_relu(x + bias, negative_slop) cpu_result = cpu_result * scale - npu_result = mx_driving.common.npu_fused_bias_leaky_relu(x.npu(), bias.npu(), negative_slop, scale).cpu().numpy() + npu_result = mx_driving.fused.npu_fused_bias_leaky_relu(x.npu(), bias.npu(), negative_slop, scale).cpu().numpy() self.assertRtolEqual(npu_result, cpu_result.numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `FusedBiasLeakyRelu` is only supported on 910B, skip this ut!") @@ -53,7 +53,7 @@ class TestFusedBiasLeakyRelu(TestCase): cpu_result = F.leaky_relu(x.float() + bias.float(), negative_slop) cpu_result = cpu_result * scale - npu_result = mx_driving.common.npu_fused_bias_leaky_relu(x.npu(), bias.npu(), negative_slop, scale).cpu().numpy() + npu_result = mx_driving.fused.npu_fused_bias_leaky_relu(x.npu(), bias.npu(), negative_slop, scale).cpu().numpy() self.assertRtolEqual(npu_result, cpu_result.half().numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `FusedBiasLeakyRelu` is only supported on 910B, skip this ut!") @@ -66,7 +66,7 @@ class TestFusedBiasLeakyRelu(TestCase): cpu_result = F.leaky_relu(x.float() + bias.float(), negative_slop) cpu_result = cpu_result * scale - npu_result = mx_driving.common.npu_fused_bias_leaky_relu(x.npu(), bias.npu(), negative_slop, scale).cpu().numpy() + npu_result = mx_driving.fused.npu_fused_bias_leaky_relu(x.npu(), bias.npu(), negative_slop, scale).cpu().numpy() self.assertRtolEqual(npu_result, cpu_result.half().numpy()) diff --git a/tests/torch/test_group_points.py b/tests/torch/test_group_points.py index 013062e3..a9bf4c3e 100644 --- a/tests/torch/test_group_points.py +++ b/tests/torch/test_group_points.py @@ -4,7 +4,7 @@ import numpy as np import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -from mx_driving.perception.point import npu_group_points +from mx_driving.point import npu_group_points DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] diff --git a/tests/torch/test_multi_scale_deformable_attn_function.py b/tests/torch/test_multi_scale_deformable_attn_function.py index 776b084c..96aee415 100644 --- a/tests/torch/test_multi_scale_deformable_attn_function.py +++ b/tests/torch/test_multi_scale_deformable_attn_function.py @@ -3,7 +3,7 @@ from collections import namedtuple import torch import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.fused DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -130,7 +130,7 @@ class TestMultiScaleDeformableAttnFunction(TestCase): npu_sampling_locations = npu_inputs.sampling_locations npu_attention_weights = npu_inputs.attention_weights npu_grad_output = npu_inputs.grad_output - npu_output = mx_driving.common.npu_multi_scale_deformable_attn_function(npu_value, npu_shapes, npu_offset, npu_sampling_locations, npu_attention_weights) + npu_output = mx_driving.fused.npu_multi_scale_deformable_attn_function(npu_value, npu_shapes, npu_offset, npu_sampling_locations, npu_attention_weights) npu_output.backward(npu_grad_output) return ExecResults( output=npu_output.detach().cpu().numpy(), diff --git a/tests/torch/test_npu_dyn_voxelization.py b/tests/torch/test_npu_dyn_voxelization.py index 0f8c4e19..09ac9f12 100644 --- a/tests/torch/test_npu_dyn_voxelization.py +++ b/tests/torch/test_npu_dyn_voxelization.py @@ -5,7 +5,7 @@ import torch import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.point DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -54,7 +54,7 @@ class TestDynVoxelization(TestCase): def npu_to_exec(self, points, coors_range, voxel_size): max_num_points = -1 - dynamic_voxelization_npu = mx_driving.common.Voxelization(voxel_size, coors_range, max_num_points) + dynamic_voxelization_npu = mx_driving.point.Voxelization(voxel_size, coors_range, max_num_points) coors = dynamic_voxelization_npu.forward(points) return coors diff --git a/tests/torch/test_npu_dynamic_scatter.py b/tests/torch/test_npu_dynamic_scatter.py index 0d071665..bca00da3 100644 --- a/tests/torch/test_npu_dynamic_scatter.py +++ b/tests/torch/test_npu_dynamic_scatter.py @@ -7,7 +7,7 @@ from torch_npu.testing.testcase import TestCase, run_tests from torch_npu.testing.common_utils import create_common_tensor import ads_c -import mx_driving.common +import mx_driving.point DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -70,11 +70,11 @@ class TestDynamicScatter(TestCase): grad_point_feats[point_idx, :] = torch.where(mask_bit, grad_voxel_feats[voxel_idx, :], zero_tensor) def npu_op_exec(self, feats, coors, reduce_type): - output_feats, output_coors = mx_driving.common.npu_dynamic_scatter(feats, coors, reduce_type) + output_feats, output_coors = mx_driving.point.npu_dynamic_scatter(feats, coors, reduce_type) return output_feats.cpu().numpy(), output_coors.cpu().numpy() def grad_npu_op_exec(self, feats, coors, reduce_type): - output_feats, output_coors = mx_driving.common.npu_dynamic_scatter(feats, coors, reduce_type) + output_feats, output_coors = mx_driving.point.npu_dynamic_scatter(feats, coors, reduce_type) return output_feats.cpu().numpy(), output_coors.cpu().numpy() @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `DynamicScatter` is only supported on 910B, skip this ut!") diff --git a/tests/torch/test_npu_max_pool2d.py b/tests/torch/test_npu_max_pool2d.py index d2e2adec..e9294f13 100644 --- a/tests/torch/test_npu_max_pool2d.py +++ b/tests/torch/test_npu_max_pool2d.py @@ -1,7 +1,7 @@ import torch import torch.nn as nn from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.fused def gen_inputs(shape, dtype): @@ -17,7 +17,7 @@ def cpu_to_exec(x_data_cpu): def npu_to_exec(x_data_cpu): - npu_output = mx_driving.common.npu_max_pool2d(x_data_cpu.npu(), 3, 2, 1) + npu_output = mx_driving.fused.npu_max_pool2d(x_data_cpu.npu(), 3, 2, 1) return npu_output diff --git a/tests/torch/test_npu_nms3d.py b/tests/torch/test_npu_nms3d.py index 60c039c0..bcc9fc0a 100644 --- a/tests/torch/test_npu_nms3d.py +++ b/tests/torch/test_npu_nms3d.py @@ -8,7 +8,7 @@ import torch_npu from torch_npu.testing.common_utils import create_common_tensor from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.detection torch.npu.config.allow_internal_format = False torch_npu.npu.set_compile_mode(jit_compile=False) @@ -246,7 +246,7 @@ class TestNms3d(TestCase): return keep, num_out def npu_to_exec(self, boxes, scores, threshold=0.0): - keep = mx_driving.common.npu_nms3d(boxes, scores, threshold) + keep = mx_driving.detection.npu_nms3d(boxes, scores, threshold) return keep.cpu() @unittest.skipIf(DEVICE_NAME != True, "OP `Nms3d` is only supported on 910B, skip this ut!") diff --git a/tests/torch/test_npu_nms3d_normal.py b/tests/torch/test_npu_nms3d_normal.py index 88036555..450b1e0c 100644 --- a/tests/torch/test_npu_nms3d_normal.py +++ b/tests/torch/test_npu_nms3d_normal.py @@ -5,7 +5,7 @@ import numpy as np import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.detection DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -24,7 +24,7 @@ class TestNms3dNormal(TestCase): np_inds = np.array([1, 0, 3]) boxes = torch.from_numpy(np_boxes) scores = torch.from_numpy(np_scores) - inds = mx_driving.common.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) + inds = mx_driving.detection.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) self.assertRtolEqual(inds.cpu().numpy(), np_inds) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `Nms3d_Normal` is only supported on 910B, skip this ut!") @@ -36,7 +36,7 @@ class TestNms3dNormal(TestCase): np_scores = np.random.rand(10).astype(np.float32) boxes = torch.from_numpy(np_boxes) scores = torch.from_numpy(np_scores) - inds = mx_driving.common.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) + inds = mx_driving.detection.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) self.assertRtolEqual(len(inds.cpu().numpy()), 9) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `Nms3d_Normal` is only supported on 910B, skip this ut!") @@ -47,7 +47,7 @@ class TestNms3dNormal(TestCase): np_scores = np.random.rand(200).astype(np.float32) boxes = torch.from_numpy(np_boxes) scores = torch.from_numpy(np_scores) - inds = mx_driving.common.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) + inds = mx_driving.detection.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) self.assertRtolEqual(len(inds.cpu().numpy()), 79) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `Nms3d_Normal` is only supported on 910B, skip this ut!") @@ -58,7 +58,7 @@ class TestNms3dNormal(TestCase): np_scores = np.random.rand(369).astype(np.float32) boxes = torch.from_numpy(np_boxes) scores = torch.from_numpy(np_scores) - inds = mx_driving.common.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) + inds = mx_driving.detection.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) self.assertRtolEqual(len(inds.cpu().numpy()), 109) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `Nms3d_Normal` is only supported on 910B, skip this ut!") @@ -69,7 +69,7 @@ class TestNms3dNormal(TestCase): np_scores = np.random.rand(555).astype(np.float32) boxes = torch.from_numpy(np_boxes) scores = torch.from_numpy(np_scores) - inds = mx_driving.common.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) + inds = mx_driving.detection.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) self.assertRtolEqual(len(inds.cpu().numpy()), 148) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `Nms3d_Normal` is only supported on 910B, skip this ut!") @@ -80,7 +80,7 @@ class TestNms3dNormal(TestCase): np_scores = np.random.rand(300).astype(np.float32) boxes = torch.from_numpy(np_boxes) scores = torch.from_numpy(np_scores) - inds = mx_driving.common.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) + inds = mx_driving.detection.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) self.assertRtolEqual(len(inds.cpu().numpy()), 102) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `Nms3d_Normal` is only supported on 910B, skip this ut!") @@ -91,7 +91,7 @@ class TestNms3dNormal(TestCase): np_scores = np.random.rand(600).astype(np.float32) boxes = torch.from_numpy(np_boxes) scores = torch.from_numpy(np_scores) - inds = mx_driving.common.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) + inds = mx_driving.detection.npu_nms3d_normal(boxes.npu(), scores.npu(), 0.3) self.assertRtolEqual(len(inds.cpu().numpy()), 161) if __name__ == "__main__": diff --git a/tests/torch/test_points_in_box.py b/tests/torch/test_points_in_box.py index 195116fe..15dbc3bd 100644 --- a/tests/torch/test_points_in_box.py +++ b/tests/torch/test_points_in_box.py @@ -17,7 +17,7 @@ import numpy as np import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.data DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -75,7 +75,7 @@ class TestPointsInBox(TestCase): [[[4, 6.928, 0], [6.928, 4, 0], [4, -6.928, 0], [6.928, -4, 0], [-4, 6.928, 0], [-6.928, 4, 0], [-4, -6.928, 0], [-6.928, -4, 0]]], dtype=torch.float32).npu() - point_indices = mx_driving.common.npu_points_in_box(boxes, pts).cpu().numpy() + point_indices = mx_driving.data.npu_points_in_box(boxes, pts).cpu().numpy() expected_point_indices = torch.tensor([[-1, -1, 0, -1, 0, -1, -1, -1]], dtype=torch.int32).cpu().numpy() self.assertRtolEqual(point_indices, expected_point_indices) @@ -96,7 +96,7 @@ class TestPointsInBox(TestCase): points[b].float(), point_indices[b]) - point_indices_npu = mx_driving.common.npu_points_in_box(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box(boxes.npu(), points.npu()) self.assertRtolEqual(point_indices.numpy(), point_indices_npu.cpu().numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `PointsInBox` is only supported on 910B, skip this ut!") @@ -116,7 +116,7 @@ class TestPointsInBox(TestCase): point_indices[b]) with self.assertRaisesRegex(RuntimeError, "boxes is larger than 200"): - point_indices_npu = mx_driving.common.npu_points_in_box(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box(boxes.npu(), points.npu()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `PointsInBox` is only supported on 910B, skip this ut!") def test_points_in_box_shape_large_points(self, device="npu"): @@ -134,7 +134,7 @@ class TestPointsInBox(TestCase): points[b].float(), point_indices[b]) - point_indices_npu = mx_driving.common.npu_points_in_box(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box(boxes.npu(), points.npu()) self.assertRtolEqual(point_indices.numpy(), point_indices_npu.cpu().numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `PointsInBox` is only supported on 910B, skip this ut!") @@ -154,7 +154,7 @@ class TestPointsInBox(TestCase): point_indices[b]) with self.assertRaisesRegex(RuntimeError, "points_in_box npu only support batch size = 1"): - point_indices_npu = mx_driving.common.npu_points_in_box(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box(boxes.npu(), points.npu()) if __name__ == "__main__": diff --git a/tests/torch/test_points_in_box_all.py b/tests/torch/test_points_in_box_all.py index 8952f9c8..23ca047f 100644 --- a/tests/torch/test_points_in_box_all.py +++ b/tests/torch/test_points_in_box_all.py @@ -16,7 +16,7 @@ import torch import numpy as np import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.data DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -72,7 +72,7 @@ class TestPointsInBoxAll(TestCase): for b in range(batch_size): point_indices[b] = points_in_boxes_all_cpu_forward(boxes[b].float(), points[b].float()) - point_indices_npu = mx_driving.common.npu_points_in_box_all(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box_all(boxes.npu(), points.npu()) self.assertRtolEqual(point_indices.numpy(), point_indices_npu.cpu().numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `PointsInBoxAll` is only supported on 910B, skip this ut!") @@ -89,7 +89,7 @@ class TestPointsInBoxAll(TestCase): for b in range(batch_size): point_indices[b] = points_in_boxes_all_cpu_forward(boxes[b].float(), points[b].float()) - point_indices_npu = mx_driving.common.npu_points_in_box_all(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box_all(boxes.npu(), points.npu()) self.assertRtolEqual(point_indices.numpy(), point_indices_npu.cpu().numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `PointsInBoxAll` is only supported on 910B, skip this ut!") @@ -106,7 +106,7 @@ class TestPointsInBoxAll(TestCase): for b in range(batch_size): point_indices[b] = points_in_boxes_all_cpu_forward(boxes[b].float(), points[b].float()) - point_indices_npu = mx_driving.common.npu_points_in_box_all(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box_all(boxes.npu(), points.npu()) self.assertRtolEqual(point_indices.numpy(), point_indices_npu.cpu().numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `PointsInBoxAll` is only supported on 910B, skip this ut!") @@ -123,7 +123,7 @@ class TestPointsInBoxAll(TestCase): for b in range(batch_size): point_indices[b] = points_in_boxes_all_cpu_forward(boxes[b].float(), points[b].float()) - point_indices_npu = mx_driving.common.npu_points_in_box_all(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box_all(boxes.npu(), points.npu()) self.assertRtolEqual(point_indices.numpy(), point_indices_npu.cpu().numpy()) @unittest.skipIf(DEVICE_NAME != 'Ascend910B', "OP `PointsInBoxAll` is only supported on 910B, skip this ut!") @@ -140,7 +140,7 @@ class TestPointsInBoxAll(TestCase): for b in range(batch_size): point_indices[b] = points_in_boxes_all_cpu_forward(boxes[b].float(), points[b].float()) - point_indices_npu = mx_driving.common.npu_points_in_box_all(boxes.npu(), points.npu()) + point_indices_npu = mx_driving.data.npu_points_in_box_all(boxes.npu(), points.npu()) self.assertRtolEqual(point_indices.numpy(), point_indices_npu.cpu().numpy()) if __name__ == "__main__": diff --git a/tests/torch/test_roipoint_pool3d.py b/tests/torch/test_roipoint_pool3d.py index 53eef4de..f902ea28 100644 --- a/tests/torch/test_roipoint_pool3d.py +++ b/tests/torch/test_roipoint_pool3d.py @@ -18,7 +18,7 @@ import torch import torch_npu import numpy as np from torch_npu.testing.testcase import TestCase, run_tests -from mx_driving.common import RoIPointPool3d +from mx_driving.data import RoIPointPool3d sys.path.append("../utils") from random_matrix import random_value diff --git a/tests/torch/test_rotated_iou.py b/tests/torch/test_rotated_iou.py index f66f8f08..2692d579 100644 --- a/tests/torch/test_rotated_iou.py +++ b/tests/torch/test_rotated_iou.py @@ -6,7 +6,7 @@ import numpy as np import torch import torch_npu from torch_npu.testing.testcase import TestCase, run_tests -import mx_driving.common +import mx_driving.detection DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -288,7 +288,7 @@ class TestNpuRotatedIou(TestCase): def npu_to_exec(self, npu_inputs): npu_boxes_a = npu_inputs.boxes_a npu_boxes_b = npu_inputs.boxes_b - npu_ans_overlap = mx_driving.common.npu_rotated_iou(npu_boxes_a, npu_boxes_b, False, 0, True, 1e-5, 1e-5) + npu_ans_overlap = mx_driving.detection.npu_rotated_iou(npu_boxes_a, npu_boxes_b, False, 0, True, 1e-5, 1e-5) return npu_ans_overlap.cpu().float().numpy() def check_precision(self, actual, expected, rtol=1e-4, atol=1e-4, msg=None): diff --git a/tests/torch/test_voxel_pooling_train.py b/tests/torch/test_voxel_pooling_train.py index 42cd23d2..ac372802 100644 --- a/tests/torch/test_voxel_pooling_train.py +++ b/tests/torch/test_voxel_pooling_train.py @@ -6,7 +6,7 @@ import torch_npu from torch_npu.testing.testcase import TestCase, run_tests import ads_c import numpy as np -import mx_driving.common +import mx_driving.point DEVICE_NAME = torch_npu.npu.get_device_name(0)[:10] @@ -71,7 +71,7 @@ class TestVoxelPoolingTrain(TestCase): return pos, result, grad_features_cpu def npu_to_exec(self, geom_xyz, input_features, voxel_num): - result = mx_driving.common.npu_voxel_pooling_train( + result = mx_driving.point.npu_voxel_pooling_train( geom_xyz, input_features, voxel_num) result.backward(result) -- Gitee