代码拉取完成，页面将自动刷新
ICNet(Image Cascade Network) propose a full convolution network which incorporates multi-resolution branches under proper label guidance to address the challenge of real-time semantic segmentation.
paper from ECCV2018
ICNet takes cascade image inputs (i.e., low-, medium- and high resolution images), adopts cascade feature fusion unit and is trained with cascade label guidance.The input image with full resolution (e.g., 1024×2048 in Cityscapes) is downsampled by factors of 2 and 4, forming cascade input to medium- and high-resolution branches.
used Dataset :Cityscape Dataset Website (please download 1st and 3rd zip)
It contains 5,000 finely annotated images split into training, validation and testing sets with 2,975, 500, and 1,525 images respectively.
.
└─ICNet
├── ascend310_infer
│ ├── build.sh
│ ├── CMakeLists.txt
│ ├── inc
│ │ └── utils.h
│ └── src
│ ├── main.cc
│ └── utils.cc
├── eval.py # validation
├── export.py # export mindir
├── postprocess.py # 310 infer calculate accuracy
├── README.md # descriptions about ICNet
├── Res50V1_PRE # scripts for pretrain
│ ├── scripts
│ │ └── run_distribute_train.sh
│ ├── src
│ │ ├── config.py
│ │ ├── CrossEntropySmooth.py
│ │ ├── dataset.py
│ │ ├── lr_generator.py
│ │ └── resnet50_v1.py
│ └── train.py
├── scripts
│ ├── run_distribute_train8p.sh # multi cards distributed training in ascend
│ ├── run_eval.sh # validation script
│ └── run_infer_310.sh # 310 infer script
├── src
│ ├── cityscapes_mindrecord.py # create mindrecord dataset
│ ├── __init__.py
│ ├── logger.py # logger
│ ├── losses.py # used losses
│ ├── loss.py # loss
│ ├── lr_scheduler.py # lr
│ ├── metric.py # metric
│ ├── models
│ │ ├── icnet_1p.py # net single card
│ │ ├── icnet_dc.py # net multi cards
│ │ ├── icnet.py # validation card
│ │ └── resnet50_v1.py # backbone
│ ├── model_utils
│ │ └── icnet.yaml # config
│ └── visualize.py # inference visualization
└── train.py # train
Set script parameters in src/model_utils/icnet.yaml .
name: "icnet"
backbone: "resnet50v1"
base_size: 1024 # during augmentation, shorter size will be resized between [base_size*0.5, base_size*2.0]
crop_size: 960 # end of augmentation, crop to training
init_lr: 0.02
momentum: 0.9
weight_decay: 0.0001
train_batch_size_percard: 4
valid_batch_size: 1
cityscapes_root: "/data/cityscapes/" # set dataset path
epochs: 160
val_epoch: 1
mindrecord_dir: '' # set mindrecord path
pretrained_model_path: '/root/ResNet50V1B-150_625.ckpt' # use the latest checkpoint file after pre-training
save_checkpoint_epochs: 5
keep_checkpoint_max: 10
python cityscapes_mindrecord.py [DATASET_PATH] [MINDRECORD_PATH]
[MINDRCORD_PATH] in script should be consistent with 'mindrecord_dir' in config file.
The folder Res50V1_PRE contains the scripts for pre-training and its dataset is image net. More details in GENet_Res50
bash run_distribute_train.sh [RANK_TABLE_FILE] [DATASET_PATH]
The hccl.json file specified by [RANK_TABLE_FILE] is used when running distributed tasks. You can use hccl_tools to generate this file.
bash scripts/run_train1p.sh [PROJECT_PATH] [DEVICE_ID]
bash scripts/run_distribute_train8p.sh [RANK_TABLE_FILE] [PROJECT_PATH]
The training results will be saved in the example path, The folder name starts with "ICNet-".You can find the checkpoint file and similar results below in LOG(0-7)/log.txt.
# distributed training result(8p)
epoch: 1 step: 93, loss is 0.5659234
epoch time: 672111.671 ms, per step time: 7227.007 ms
epoch: 2 step: 93, loss is 1.0220546
epoch time: 66850.354 ms, per step time: 718.821 ms
epoch: 3 step: 93, loss is 0.49694514
epoch time: 70490.510 ms, per step time: 757.962 ms
epoch: 4 step: 93, loss is 0.74727297
epoch time: 73657.396 ms, per step time: 792.015 ms
epoch: 5 step: 93, loss is 0.45953503
epoch time: 97117.785 ms, per step time: 1044.277 ms
Check the checkpoint path used for evaluation before running the following command.
bash run_eval.sh [DATASET_PATH] [CHECKPOINT_PATH] [PROJECT_PATH] [DEVICE_ID]
The results at eval/log were as follows:
Found 500 images in the folder /data/cityscapes/leftImg8bit/val
pretrained....
2021-06-01 19:03:54,570 semantic_segmentation INFO: Start validation, Total sample: 500
avgmiou 0.69962835
avg_pixacc 0.94285786
avgtime 0.19648232793807982
Before inference, please refer to MindSpore Inference with C++ Deployment Guide to set environment variables.
bash run_infer_310.sh [The path of the MINDIR for 310 infer] [The path of the dataset for 310 infer] 0
| Parameter | ICNet |
|---|---|
| resources | Ascend 910；CPU 2.60GHz, 192core；memory：755G |
| Upload date | 2021.6.1 |
| mindspore version | mindspore1.2.0 |
| training parameter | epoch=160,batch_size=4 |
| optimizer | SGD optimizer，momentum=0.9,weight_decay=0.0001 |
| loss function | SoftmaxCrossEntropyLoss |
| training speed | epoch time：21469.152 ms(8pcs) per step time :330.851 ms(8pcs) |
| total time | 1h20m34s(8pcs) |
| Script URL | |
| Random number seed | set_seed = 1234 |
The seed in the create_icnet_dataset function is set in cityscapes_mindrecord.py, and the random seed in train.py is also used for weight initialization.
Please visit the official website 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