[PYTHON] Try to implement yolact

environment

• CUDA V10.0.130
• Python 3.7.3
• gcc 7.3.0
• Pytorch 1.2.0

setup

git clone https://github.com/dbolya/yolact.git
cd yolact
pip install opencv-python pillow==6.2.1 pycocotools matplotlib
pip install cython
pip install torch==1.2.0 torchvision==0.4.0    # CUDA10.Stable at 0

Compile DCNv2 when using yolact ++

cd external/DCNv2
python setup.py develop

Inference with trained model

1. Download the trained model https://drive.google.com/file/d/15id0Qq5eqRbkD-N3ZjDZXdCvRyIaHpFB/view?usp=sharing
2. Perform inference

• Image

  python eval.py 
      --trained_model=weights/yolact_base_54_800000.pth  
      --score_threshold=0.15 
      --top_k=15 
      --images=path/to/input/folder:path/to/output/folder
  

• Video

  python eval.py 
      --trained_model=weights/yolact_base_54_800000.pth 
      --score_threshold=0.15 
      --top_k=15 
      --video_multiframe=4 
      --video=input_video.mp4:output_video.mp4
  

Learn with original data

1. Prepare the dataset in coco format (image + JSON)
2. Download the pre-learning model

• Resnet50
  https://drive.google.com/file/d/1Jy3yCdbatgXa5YYIdTCRrSV0S9V5g1rn/view?usp=sharing
• Resnet101
  https://drive.google.com/file/d/1tvqFPd4bJtakOlmn-uIA492g2qurRChj/view?usp=sharing
• Darknet53
  https://drive.google.com/file/d/17Y431j4sagFpSReuPNoFcj9h7azDTZFf/view?usp=sharing

1. Modify the config file (2 places) First fill in the database definition,

       test_dataset = dataset_base.copy({
           'name': 'Test Dataset',
   
           'train_images': 'path_to_training_images',
           'train_info':   'path_to_training_annotation',
   
           'valid_images': 'path_to_validation_images',
           'valid_info':   'path_to_validation_annotation',
   
           'has_gt': True,
           'class_names': ('my_class_id_1', 'my_class_id_2', 'my_class_id_3', ...)
       })
   

Then change the default config file and exit.

```py:data/config.py
    yolact_base_config = coco_base_config.copy({
        'name': 'yolact_base',

        # Dataset stuff
        # 'dataset': coco2017_dataset,    # default
        'dataset': test_dataset,    # Original Dataset

        # 'num_classes': len(coco2017_dataset.class_names) + 1,
        'num_classes': len(test_dataset.class_names) + 1,
    })
```

1. Learning

• Change config depending on model

• Store the initial model (resnet50-19c8e357.pth) in save_folder

  python train.py --config=yolact_plus_resnet50_config --save_folder=/path/to/workspace/ --save_interval=1000
  

1. Model accuracy evaluation

• Can be evaluated with the following code.

• If you add --output_coco_json to the end, you can generate json files under ~ / results /.

  python eval.py --trained_model=weights/yolact_base_54_800000.pth