Click to unfold/fold
- Contents
- [YOLOv5 Description](#YOLOv5 Description)
- [Model Architecture](#Model Architecture)
- Dataset
- [Dataset Download](#Dataset Download)
- [Dataset Structure](#Dataset Structure)
- [Dataset Conversion](#Dataset Conversion)
- Quick Start
- Script Description
- [Script and Sample Code](#Script and Sample Code)
- [Script Parameters](#Script Parameters)
- [Training Process](#Training Process)
- Training
- [Distributed Training](#Distributed Training)
- [Evaluation Process](#Evaluation Process)
- [Infer Process](#Infer Process)
- [Model Description](#Model Description)
- Performance
- Q&A
Published in April 2020 by Ultralytics, YOLOv5 achieved state-of-the-art performance on the COCO dataset for object detection. It is an important improvement of YoloV3, the implementation of a new architecture in the Backbone and the modifications in the Neck have improved the mAP(mean Average Precision) by 10% and the number of FPS(Frame per Second) by 12%.
Repository of official implementation by PyTorch：https://github.com/ultralytics/yolov5
The YOLOv5 network is mainly composed of CSP and Focus as a backbone, spatial pyramid pooling(SPP) additional module, PANet path-aggregation neck and YOLOv3 head. CSP is a novel backbone that can enhance the learning capability of CNN. The spatial pyramid pooling block is added over CSP to increase the receptive field and separate out the most significant context features. Instead of Feature pyramid networks (FPN) for object detection used in YOLOv3, the PANet is used as the method for parameter aggregation for different detector levels. To be more specific, CSPDarknet53 contains 5 CSP modules which use the convolution C with kernel size k=3x3, stride s = 2x2; Within the PANet and SPP, 1x1, 5x5, 9x9, 13x13 max poolings are applied.
YOLOv5 is trained on COCO dataset with labels of YOLO format.
Dataset:
- Raw data
- Train set: http://images.cocodataset.org/zips/train2017.zip
- Validation set: http://images.cocodataset.org/zips/val2017.zip
- Test set: http://images.cocodataset.org/zips/test2017.zip
- YOLO format labels
coco2017labels: https://github.com/ultralytics/yolov5/releases/download/v1.0/coco2017labels.zip - YOLO format segmentation labels
coco2017labels-segments: https://github.com/ultralytics/yolov5/releases/download/v1.0/coco2017labels-segments.zip
Download the raw images data, and the labels files according to target model:
| Model | Label |
|---|---|
| YOLOv5n | coco2017labels |
| YOLOv5s | coco2017labels |
| YOLOv5m | coco2017labels-segments |
| YOLOv5l | coco2017labels-segments |
| YOLOv5x | coco2017labels-segments |
After downloading the dataset and labels, you should put them in correct position
as the following text shows. The images folder saves the images and labels
folder saves the corresponding labels. The text files like train2017.txt saves
the image paths of the corresponding subset of dataset.
YOLO
├── images
| ├── train2017
| ├── val2017
| └── test2017
├── labels
| ├── train2017
| ├── val2017
├── images
| ├── train2017
| ├── val2017
| └── test2017
├── train2017.txt
├── val2017.txt
└── test2017.txtIf you want to use customized data with COCO or labelme format, you can use conversion script to convert them to YOLO format.
Conversion steps:
- Change directory to
config/data_conversion. The names of the files in this folder stand for configs of corresponding dataset. - Modify the config files of the original format and the conversion target format. Change the path in config files.
- After edit of config files, run
convert_data.pyscript. For example,python convert_data.py coco yolomeans convert dataset from coco format to yolo.
Installation
Follow the tutorial in MindSpore official website to install mindspore.
Then use the following command to install other required packages:
pip install -r requirements.txtTraining
You can use the following command to train on a single device:
# Run training example(1p) on Ascend/GPU by python command
python train.py \
--ms_strategy="StaticShape" \
--overflow_still_update=True \
--optimizer="momentum" \
--cfg="../config/network/yolov5s.yaml" \
--data="../config/data/coco.yaml" \
--hyp="../config/data/hyp.scratch-low.yaml" \
--device_target=Ascend \
--epochs=300 \
--batch_size=32 > log.txt 2>&1 &Or you can use shell scripts. The scripts support training on single or multiple devices.
The command are in the following:
# Run 1p by shell script, please change `device_target` in config file to run on Ascend/GPU, and change `T_max`, `max_epoch`, `warmup_epochs` refer to contents of notes
bash run_standalone_train_ascend.sh -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yaml
# For Ascend device, distributed training example(8p) by shell script
bash run_distribute_train_ascend.sh -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yaml -r hccl_8p_xx.jsonYou could pass --help or -H to shell script to see usage in detail.
If you want to use custom dataset，you can use compute_anchors.py to compute new anchors,
then use the output anchors to update the anchors item in corresponding model config files.
Evaluation
You can use the following command to evaluate a model:
# Run evaluation on Ascend/GPU by python command
python val.py \
--weights="path/to/weights.ckpt" \
--cfg="../config/network/yolov5s.yaml" \
--data="../config/data/coco.yaml" \
--hyp="../config/data/hyp.scratch-low.yaml" \
--device_target=Ascend \
--img_size=640 \
--conf=0.001 \
--rect=False \
--iou_thres=0.60 \
--batch_size=32 > log.txt 2>&1 &The rect switch can increase mAP of evaluation result.
The results in official repository is evaluated with this switch on.
Please note this difference when you compare evaluation results of two repositories.
Or you can also use shell scripts to do evaluation:
# Run distributed evaluation by shell script
bash run_distribute_test_ascend.sh -w path/to/weights.ckpt -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yaml -r hccl_8p_xx.json
# Run standalone evaluation by shell script
bash run_standalone_test_ascend.sh -w path/to/weights.ckpt -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yamlThe corresponding config files are in config folder. The coco.yaml in config/data folder is about dataset configs.
The hyp.scratch-low.yaml are hyperparameters settings. The yolov5s.yaml saves model architecture configs.
Click to unfold/fold
yolov5
├── README.md // descriptions about yolov5
├── README_CN.md // Chinese descriptions about yolov5
├── __init__.py
├── config
│ ├── args.py // get config parameters from command line
│ ├── data
│ │ ├── coco.yaml // configs about dataset
│ │ ├── hyp.scratch-high.yaml // configs about hyper-parameters
│ │ ├── hyp.scratch-low.yaml
│ │ └── hyp.scratch-med.yaml
│ ├── data_conversion
│ │ ├── coco.yaml // config of coco format dataset
│ │ ├── labelme.yaml // config of labelme format dataset
│ │ └── yolo.yaml // config of yolo format dataset
│ └── network // configs of model architecture
│ ├── yolov5l.yaml
│ ├── yolov5m.yaml
│ ├── yolov5n.yaml
│ ├── yolov5s.yaml
│ └── yolov5x.yaml
├── compute_anchors.py // compute anchors for specified data
├── convert_data.py // convert dataset format
├── deploy // code for inference
│ ├── __init__.py
│ └── infer_engine
│ ├── __init__.py
│ ├── lite.py // code for inference with MindSporeLite
│ ├── mindx.py // code for inference with mindx
│ └── model_base.py
├── export.py
├── preprocess.py
├── scripts
│ ├── common.sh // common functions used in shell scripts
│ ├── get_coco.sh
│ ├── hccl_tools.py // generate rank table files for distributed training or evaluation
│ ├── mpirun_test.sh // launch evaluation with OpenMPI
│ ├── mpirun_train.sh // launch training with OpenMPI
│ ├── run_distribute_test_ascend.sh // launch distributed evaluation(8p) on Ascend
│ ├── run_distribute_train_ascend.sh // launch distributed training(8p) on Ascend
│ ├── run_standalone_test_ascend.sh // launch 1p evaluation on Ascend
│ └── run_standalone_train_ascend.sh // launch 1p training on Ascend
├── src
│ ├── __init__.py
│ ├── all_finite.py
│ ├── augmentations.py // data augmentations
│ ├── autoanchor.py
│ ├── boost.py
│ ├── callback.py
│ ├── checkpoint_fuse.py
│ ├── coco_visual.py
│ ├── data // code for dataset format conversion
│ │ ├── __init__.py
│ │ ├── base.py // base class for data conversion
│ │ ├── coco.py // transfer dataset with coco format to others
│ │ ├── labelme.py // transfer dataset with labelme format to others
│ │ └── yolo.py // transfer dataset with yolo format to others
│ ├── dataset.py // create dataset
│ ├── general.py // general functions used in other scripts
│ ├── loss_scale.py
│ ├── metrics.py
│ ├── modelarts.py
│ ├── ms2pt.py // transfer weights from MindSpore to PyTorch
│ ├── network
│ │ ├── __init__.py
│ │ ├── common.py // common code for building network
│ │ ├── loss.py // loss
│ │ └── yolo.py // YOLOv5 network
│ ├── optimizer.py // optimizer
│ ├── plots.py
│ └── pt2ms.py // transfer weights from PyTorch to MindSpore
├── test.py // script for evaluation
├── third_party // third-party code
│ ├── __init__.py
│ ├── fast_coco // faster coco mAP computation
│ │ ├── __init__.py
│ │ ├── build.sh
│ │ ├── cocoeval
│ │ │ ├── cocoeval.cpp
│ │ │ └── cocoeval.h
│ │ ├── fast_coco_eval_api.py
│ │ └── setup.py
│ ├── fast_nms // faster nms computation
│ │ ├── __init__.py
│ │ ├── build.sh
│ │ ├── nms.pyx
│ │ └── setup.py
│ └── yolo2coco // yolo data format to coco format converter
│ ├── __init__.py
│ └── yolo2coco.py
└── train.py // script for training
Major parameters in train.py are:
optional arguments:
--ms_strategy Training strategy. Default: "StaticShape"
--distributed_train Distributed training or not. Default: False
--device_target Device where the code will be executed. Default: "Ascend"
--cfg Model architecture yaml config file path. Default: "./config/network/yolov5s.yaml"
--data Dataset yaml config file path. Default: "./config/data/data.yaml"
--hyp Hyper-parameters yaml config file path. Default: "./config/data/hyp.scratch-low.yaml"
--epochs Training epochs. Default: 300
--batch_size Batch size per device. Default: 32
--save_checkpoint Whether save checkpoint. Default: True
--start_save_epoch Epoch index after which checkpoint will be saved. Default: 1
--save_interval Epoch interval to save checkpoints. Default: 1
--max_ckpt_num Maximum number of saved checkpoints. Default: 10
--cache_images Whether cache images for faster training. Default: False
--optimizer Optimizer used for training. Default: "sgd"
--sync_bn Whether use SyncBatchNorm, only available in DDP mode. Default: False
--project Folder path to save output data. Default: "runs/train"
--linear_lr Whether use linear learning rate. Default: True
--run_eval Whether do evaluation after a training epoch. Default: True
--eval_start_epoch Epoch index after which model will do evaluation. Default: 200
--eval_epoch_interval Epoch interval to do evaluation. Default: 10
--distributed_eval Distributed evaluation or not. Default: False
For Ascend device, you can use shell scripts. The scripts support training on single or multiple devices.
The command are in the following:
# Run 1p by shell script, please change `device_target` in config file to run on Ascend/GPU, and change `T_max`, `max_epoch`, `warmup_epochs` refer to contents of notes
bash run_standalone_train_ascend.sh -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yaml
# For Ascend device, distributed training example(8p) by shell script
bash run_distribute_train_ascend.sh -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yaml -r hccl_8p_xx.jsonOr you can use the following command to start standalone training:
# Run training example(1p) on Ascend/GPU by python command
python train.py \
--ms_strategy="StaticShape" \
--optimizer="momentum" \
--cfg="../config/network/yolov5s.yaml" \
--data="../config/data/coco.yaml" \
--hyp="../config/data/hyp.scratch-low.yaml" \
--device_target=Ascend \
--epochs=300 \
--batch_size=32 > log.txt 2>&1 &We recommend do training by running shell script.
You should fine tune the parameters when run training for custom dataset.
The python command above will run in the background.
Distributed training example(8p) by shell script:
# For Ascend device, distributed training example(8p) by shell script
bash run_distribute_train_ascend.sh -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yaml -r hccl_8p_xx.json
# For GPU device, distributed training example(8p) by shell script
bash run_distribute_train_gpu.sh ../config/network/yolov5s.yaml ../config/data/coco.yaml \
../config/data/hyp.scratch-low.yamlYou can also use OpenMPI to run distributed training. You should follow the official tutorial to configure OpenMPI environment，then execute the following command：
bash mpirun_train.sh -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yamlBefore running the command below, please check the checkpoint path used for evaluation.
# Run evaluation by python command
python val.py \
--weights="path/to/weights.ckpt" \
--cfg="../config/network/yolov5s.yaml" \
--data="../config/data/coco.yaml" \
--hyp="../config/data/hyp.scratch-low.yaml" \
--device_target=Ascend \
--img_size=640 \
--conf=0.001 \
--rect=False \
--iou_thres=0.65 \
--batch_size=32 > log.txt 2>&1 &
# OR
# Run evaluation(8p) by shell script
bash run_distribute_test_ascend.sh -w path/to/weights.ckpt -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yaml -r hccl_8p_xx.json
# OR
# Run standalone evaluation by shell script
bash run_standalone_test_ascend.sh --w path/to/weights.ckpt -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yamlThe above python command will run in the background. You can view the results through the file "log.txt".
You can also use OpenMPI to run distributed test. You should follow the official tutorial to configure OpenMPI environment，then execute the following command：
bash mpirun_test.sh --w path/to/weights.ckpt -c ../config/network/yolov5s.yaml -d ../config/data/coco.yaml \
-h ../config/data/hyp.scratch-low.yamlDownload Ascend-mindxsdk-mxmanufacture package of community version from MindX SDK community according to architecture of your device.
We recommend package with .run suffix. We are now support MindX SDK 3.0 version.
When downloading complete, please firstly make sure you have configured related Ascend environment variables, then use the following command to install package:
bash Ascend-mindxsdk-mxmanufacture_xxx.run --installAfter installation, you can use python -c "import mindx" to test whether installation is successful。
If you see error related to libgobject.so.2, you need to configure environment variable for library libffi.so.7:
- Firstly, use
find / -nane "libffi.so.7"to find the location of this library file； - Then use
export LD_PRELOAD=/path/to/libffi.so.7to configure environment variable.
The model of ckpt format can be transformed to om format by atc tool
for doing inference on inference server. The following are steps:
- Export model with
AIRformat：python export.py --weights /path/to/model.ckpt --file_format AIR; - Transform model with
AIRformat toomformat byatctool：/usr/local/Ascend/latest/atc/bin/atc --model=yolov5s.air --framework=1 --output=./yolov5s --input_format=NCHW --input_shape="Inputs:1,3,640,640" --soc_version=Ascend310, the--soc_versionoption can be got bynpu-smi infocommand. Supported option choices areAscend310，Ascend310P3; - Infer by executing
infer.pyscript：python infer.py --batch_size 1 --om yolov5s.om
Note that, because dynamic shape is not supported for om format, the rect switch can not be set. So the mAP is lower than
the result of checkpoint with rect enabled.
| Model | size (pixels) |
mAPval 50-95 rect=True |
mAPval 50 rect=True |
mAPval 50-95 rect=False |
mAPval 50 rect=False |
Epoch Time(s) | Throughput (images/s) |
|---|---|---|---|---|---|---|---|
| YOLOv5n | 640 | 0.279 | 0.459 | 0.277 | 0.455 | 66 | 224.00 |
| YOLOv5s | 640 | 0.375 | 0.572 | 0.373 | 0.57 | 79 | 187.14 |
| YOLOv5m | 640 | 0.453 | 0.637 | 0.451 | 0.637 | 133 | 111.16 |
| YOLOv5l | 640 | 0.489 | 0.675 | 0.486 | 0.671 | 163 | 90.70 |
| YOLOv5x | 640 | 0.505 | 0.686 | 0.506 | 0.687 | 221 | 66.90 |
Note
- All models are trained to 300 epochs with default settings. Nano and Small models use hyp.scratch-low.yaml hyper-parameters, all others use hyp.scratch-high.yaml.
- The following are settings used for different models:
--data coco.yaml --epochs 300 --weights '' --cfg yolov5n.yaml --batch-size 16
yolov5s.yaml 32
yolov5m.yaml 24
yolov5l.yaml 24
yolov5x.yaml 24- The result of
Epoch Timeis evaluated on 8 Ascend 910A with batch_size 32 per device. - The result of
Throughputis of single Ascend 910A device. - mAPval values are for single-model single-scale on COCO val2017 dataset.
The key configs are--img_size 640 --conf_thres 0.001 --iou_thres 0.65 - When data preprocessing is the bottleneck, you can set
--cache_imagestoramormemoryto accelerate preprocessing. Note thatrammay cause out of memory. yolov5nneed enable--sync_bn.
- cannot allocate memory in static TLS block
ImportError: /xxx/scikit_image.libs/libgomp-xxx.so: cannot allocate memory in static TLS block
It seems that scikit-image has not been built correctly.This error is not caused by our code, but some packages depend on scikit-image package. Generally, you can change the
import order to solve this error by adding import sklearn or import skimage at the beginning of
the train.py.
If this still cannot solve the problem, you can search for this error to find other solution.
- During the training,
losssuddenly increases to a large value(generallylobj losscauses this), or some loss isnan。
This problem is caused by overflow during the training process. Overflow makes loss becomes nan, and after updating
the the model, the value weights will become very large. This usually appears when training small dataset with just one class.
If you come into this problem, you can change the enable_clip_grad to True in hyp-scratch.xx.yaml to enable gradient clip.
Besides, in our updated code, we add overflow detection. When we detect that the overflow happens, we will skip the update of this step,
which can avoid this.
mAPis not good
Well, there are many possible reasons making mAP not good enough, like overflow mentioned in the 2nd question.
If you use method in the above, you should see mAP will become good.
You can also try to adjust the lr0 in config/data/hyp-scratch-xx.yaml, or change --batch_size to finetune the model.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent