YOLOv4-pytorch for ModelArts

This is a pytorch implementation of YOLOv4 based on argusswift/YOLOv4-pytorch. You can train model on your own dataset and deploy it to the ModelArts easily.

1. Environment

Windows
python 3.6

Run the installation script to install all the dependencies.

pip install -r requirements.txt

2. Preparations

2.1 Dataset

This project supports datasets in Pascal VOC format. You need to place your data as follows:

ModelArts_Yolov4
├───data
│   └───Your_dataset
│       ├───Annotations
│       │   ├──1.xml
│       │   └──...
│       ├───JPEGImages
│       │   ├──1.jpg
│       │   └──...

Then:

Update the "DATASET_NAME" and "Customer_DATA" in the config/yolov4_config.py.
Split the data into trainset and testset with data/gen_img_index_file.py. After this you will get two files: train.txt and test.txt in your dataset folder.
Convert the pascal voc *.xml format annotation to *.txt format (Image_path xmin0,ymin0,xmax0,ymax0,class0 ) using data/convert_voc_to_txt.py. You will get data/train_annotation.txt and data/test_annotation.txt.
Generate annotation files for each class with data/gen_cls_anno.py. These files are generated in the data/your_dataset/ClassAnnos/ directory and are used to calculate APs.
Run utils/anchor_kmeans.py, which performs kmeans algrithom on the ground truth bboxes to get the most general anchor boxes. Update the "MODEL['ANCHORS']" in the config/yolov4_config.py.

2.2 Download Weight File

Mobilenetv3 pre-trained weight: mobilenetv3(code: yolo)
Make dir weights/ in the ModelArts_Yolov4 and put the weight file in it.

3. Training

Run the following command to start training and see the details in the config/yolov4_config.py.

python -u train.py

During training, backups of model will be saved in weights/*.pth. You can interrupt training and resume training from these backups at any time using the following command.

python -u train.py --weight_file  your_backup.pth  --resume

4. Detection

Run predict.py and you can predict images from testset one by one.

5. Deployment

Copy weights/best.pth to ModelArts/model/best.pth. Then upload the entire ModelArts folder to the ModelArts platform.

Liuyvjin / ModelArts_Yolov4