Fetch the repository succeeded.
Semantic segmentation is a key technique for autonomous vehicles to understand the surrounding scene. For actual autonomous vehicles, it is not desirable to spend a lot of reasoning time to obtain high-precision segmentation results. Using a lightweight architecture (encoder decoder or dual channel) or inference on low-resolution images, recent methods have enabled very fast scene resolution, even running at over 100 FPS on a single 1080Ti GPU. However, there is still a significant performance gap between these real-time approaches and models based on the bulking backbone.
To solve this problem, inspired by HRNet, the authors proposed a deep double-resolution network with deep high-resolution representation capabilities for real-time semantic segmentation of high-resolution images, especially road driving images. The authors propose a new deep dual-resolution network for real-time semantic segmentation of road scenes. DDRNet starts with one trunk and is then split into two parallel deep branches with different resolutions. One deep branch generates a relatively high-resolution feature map, and the other extracts rich contextual information through multiple downsampling operations. For efficient information fusion, multiple bilateral connections are bridged between the two branches. In addition, we also propose a new module DAPPM, which greatly increases the number of accepted domains, extracting contextual information more fully than ordinary PPM.
paper:Hong, Yuanduo, et al. "Deep dual-resolution networks for real-time and accurate semantic segmentation of road scenes."
Dataset used: ImageNet2012
The mixed precision training method accelerates the deep learning neural network training process by using both the single-precision and half-precision data types, and maintains the network precision achieved by the single-precision training at the same time. Mixed precision training can accelerate the computation process, reduce memory usage, and enable a larger model or batch size to be trained on specific hardware. For FP16 operators, if the input data type is FP32, the backend of MindSpore will automatically handle it with reduced precision. Users could check the reduced-precision operators by enabling INFO log and then searching ‘reduce precision’.
├── DDRNet
├── README_CN.md
├── README.md
├── ascend310_infer // Ascend310 infer
├── scripts
├──run_standalone_train_ascend.sh // Ascend standalone train
├──run_distribute_train_ascend.sh // Ascend distribute train
├──run_eval_ascend.sh // Ascend evaluation
├──run_standalone_train_gpu.sh // GPU standalone training
├──run_distribute_train_gpu.sh // GPU distribute training
├──run_eval_gpu.sh // GPU evaluation
└──run_infer_310.sh // 310 infer
├── src
├──configs // DDRNet configs
├──data // Dataset files
├──imagenet.py // imagenet dataset
├──augment // Augmentation
└──data_utils // modelarts utils
├──models // DDRNet model
├──trainers // TrainOneStep interface
├──tools
├──callback.py // File with callbacks
├──cell.py // Cell to amp wrapper
├──criterion.py // Functions for criterion
├──get_misc.py // Interfaces for training
├──optimizer.py // Optimizer
└──schedulers.py // Shedulers
├── train.py // Training script
├── eval.py // Evaluation script
├── export.py // Export script
├── postprocess.py // Inference Computational Precision File
└── preprocess.py // Inference preprocessing image files
Parameters for both training and evaluation can be set in yaml config in scr/configs.
arch: DDRNet23
# ===== Dataset ===== #
data_url: ./data/imagenet
set: ImageNet
num_classes: 1000
mix_up: 0.8
cutmix: 1.0
color_jitter: 0.4
auto_augment: rand-m9-mstd0.5-inc1
interpolation: bicubic
re_mode: pixel
re_count: 1
mixup_prob: 1.
switch_prob: 0.5
mixup_mode: batch
mixup_off_epoch: 0.
image_size: 224
crop_pct: 0.875
# ===== Learning Rate Policy ======== #
optimizer: momentum
use_nesterov: True
base_lr: 0.1
warmup_lr: 0.000001
min_lr: 0.00001
lr_scheduler: cosine_lr
warmup_length: 10
lr_adjust: 30 # for multistep lr
# ===== Network training config ===== #
amp_level: O2
keep_bn_fp32: True
beta: [ 0.9, 0.999 ]
clip_global_norm_value: 5.
clip_global_norm: True
is_dynamic_loss_scale: True
epochs: 300
label_smoothing: 0.1
loss_scale: 1024
weight_decay: 0.0001
decay: 0.9 # for rmsprop
momentum: 0.9
batch_size: 512
# ===== Hardware setup ===== #
num_parallel_workers: 16
device_target: Ascend
# ===== Model config ===== #
drop_path_rate: 0.1
bash ./scripts/run_standalone_train_ascend.sh [DEVICE_ID] [CONFIG_PATH]
#OR
bash ./scripts/run_distribute_train_ascend.sh [RANK_TABLE_FILE] [CONFIG_PATH]
bash ./scripts/run_eval_ascend.sh [DEVICE_ID] [CONFIG_PATH] [CHECKPOINT_PATH]
bash run_infer_310.sh [MINDIR_PATH] [DATASET_NAME(imagenet2012)] [DATASET_PATH] [DEVICE_ID(optional)]
For distributed training, you need to create an hccl configuration file in JSON format in advance. Please follow the instructions in the link below: hccl tools
bash ./scripts/run_standalone_train_gpu.sh [DEVICE_ID] [CONFIG_PATH]
#OR
bash ./scripts/run_distribute_train_gpu.sh [CONFIG_PATH]
bash ./scripts/run_eval_gpu.sh [DEVICE_ID] [CONFIG_PATH] [CHECKPOINT_PATH]
python export.py --pretrained [CKPT_FILE] --ddr_config [CONFIG_PATH] --device_target [DEVICE_TARGET]
The following shows an example of using a mindir model to perform inference.
# Ascend310 inference on ImageNet-1k
bash run_infer_310.sh [MINDIR_PATH] [DATASET_NAME] [DATASET_PATH] [DEVICE_ID]
Top1 acc: 0.76578
Top5 acc: 0.9331
| Platform | GPU | Ascend |
|---|---|---|
| Arch | DDRNet-23 | DDRNet-23 |
| Resource | GPU: 8xRTX3090 24G; CPU: Intel(R) Xeon(R) Gold 6226R; RAM: 252G | Ascend 910 |
| Upload date | 2021-03-23 | 2021-12-04 |
| MindSpore version | 1.6.0 | 1.3.0 |
| Dataset | ImageNet-1k train | ImageNet-1k train |
| Training parameters | ddrnet23_imagenet_gpu.yaml | ddrnet23_imagenet_ascend.yaml |
| Parameters (M) | batch_size=256, epoch=300 | batch_size=512, epoch=300 |
| Optimizer | Momentum | Momentum |
| Loss function | SoftmaxCrossEntropy | SoftmaxCrossEntropy |
| Output | ckpt file | ckpt file |
| Final loss | 2.44 | - |
| Velocity | eight cards: mean 5000 ms/step | eight cards: mean 940 ms/step |
| Total time | eight cards: 240 h | eight cards: 35 h |
| Platform | GPU | Ascend |
|---|---|---|
| Arch | DDRNet-23 | DDRNet-23 |
| Resource | GPU: 8xRTX3090 24G; CPU: Intel(R) Xeon(R) Gold 6226R; RAM: 252G | Ascend 310 |
| Upload date | 2021-03-23 | 2021-12-04 |
| MindSpore version | 1.6.0 | 1.3.0 |
| Dataset | ImageNet-1k val | ImageNet-1k val |
| Eval loss | 1.2 | 1.313 |
| Accuracy | eight cards: top1:76.6% top5:93.4% | eight cards: top1:76.598% top5:93.312% |
Please check the official homepage.
此处可能存在不合适展示的内容，页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容，可点击提交进行申诉，我们将尽快为您处理。
马建仓 AI 助手