1. Version description

yolov5 version Tags=v2.0, python version 3.7.5

2. Prepare the data set

2.1 Download the coco2017 data set and decompress it. The decompressed directory is as follows:

├── coco_data: #root directory
     ├── train2017 #Training set pictures, about 118287
     ├── val2017 #Validation set pictures, about 5000
     └── annotations # annotation directory
     ├── instances_train2017.json #The training set annotation file corresponding to target detection and segmentation tasks
     ├── instances_val2017.json #Validation set annotation file corresponding to target detection and segmentation tasks
     ├── captions_train2017.json
     ├── captions_val2017.json
     ├── person_keypoints_train2017.json
     └── person_keypoints_val2017.json

2.2 Generate yolov5 special annotation file

(1) Copy coco/coco2yolo.py and coco/coco_class.txt in the code warehouse to coco_dataroot directory

(2) Run coco2yolo.py

python3 coco2yolo.py

(3) After running the above script, train2017.txt and val2017.txt will be generated in coco_data root directory

3. Configure the data set path

Modify the train field and val field in the data/coco.yaml file to point to the train2017.txt and val2017.txt generated in the previous section respectively, such as:

train: /data/coco_data/train2017.txt
val: /data/coco_data/val2017.txt

4.GPU, CPU dependency

Install python dependency package according to requirements-GPU.txt

5. NPU dependency

Install the python dependency package according to requirements.txt, and also need to install (NPU-driver.run, NPU-firmware.run, NPU-toolkit.run, torch-ascend.whl, apex.whl)

6. Compile and install Opencv-python

In order to get the best image processing performance, Please compile and install opencv-python instead of directly installing. The compilation and installation steps are as follows:

export GIT_SSL_NO_VERIFY=true
git clone https://github.com/opencv/opencv.git
cd opencv
mkdir -p build
cd build
cmake -D BUILD_opencv_python3=yes -D BUILD_opencv_python2=no -D PYTHON3_EXECUTABLE=/usr/local/python3.7.5/bin/python3.7m -D PYTHON3_INCLUDE_DIR=/usr/local/python3.7.5/include/python3.7m -D PYTHONARY =/usr/local/python3.7.5/lib/libpython3.7m.so -D PYTHON3_NUMPY_INCLUDE_DIRS=/usr/local/python3.7.5/lib/python3.7/site-packages/numpy/core/include -D PYTHON3_PACKAGES_PATH=/ usr/local/python3.7.5/lib/python3.7/site-packages -D PYTHON3_DEFAULT_EXECUTABLE=/usr/local/python3.7.5/bin/python3.7m ..
make -j$nproc
make install

7. NPU stand-alone single-card training instruction

bash train_npu_1p.sh

8. NPU stand-alone eight-card training instruction

bash train_npu_8p_mp.sh

9. NPU evalution instruction

(1) Modify the parameter --coco_instance_path in evaluation_npu_1p.sh to the actual path in the data set. For example, modify the script to

python3.7 test.py --data /data/coco.yaml --coco_instance_path /data/coco/annotations/instances_val2017.json --img-size 672 --weight'yolov5_0.pt' --batch-size 32 - device npu --npu 0

(2) Start the evaluation

bash evaluation_npu_1p.sh

10. GPU stand-alone single-card training instructions

python train.py --data coco.yaml --cfg yolov5x.yaml --weights'' --batch-size 32 --device 0

11.GPU stand-alone eight-card training instruction

python -m torch.distributed.launch --nproc_per_node 8 train.py --data coco.yaml --cfg yolov5x.yaml --weights'' --batch-size 256

12.CPU instructions

python train.py --data coco.yaml --cfg yolov5x.yaml --weights'' --batch-size 32 --device cpu

13. Export onnx instruction

python export_onnx.py --weights ./xxx.pt --img-size 640 --batch-size 1

13. Reults

13.1 NPU Traininig Performance

Model	Size (pixels)	NPU Nums	Dataset	Training Data Nums	Validation Data Nums	Batch Size	Epochs	FPS	Total Training Time (H)
yolo5x	640	1	COCO-2017	118287	5000	32	300	51.8	188.528
yolo5x	640	8	COCO-2017	118287	5000	256	300	400.5	27.287

13.2 NPU Inference Performance

Model	Size (pixels)	mAPval 0.0 : 1.0	Speed Atlas 300T (ms)
yolo5x-1p	640	yolo5x-1p	48.5
yolo5x-8p	640	yolo5x-8p	50.3