CustomDatasetTutorial.md

Tutorial for finetuning on a custom dataset

Step 1. Convert the custom dataset to coco format.

• The coco format can be found in Link. Specifically, the coco annotation file (.json) includes three necessary feilds, i.e., image, annotation, categories. A toy sample (named toy_sample.json) is provided below:

{
  "categories": 
  [{
      "supercategory": "person", 
      "id": 1, 
      "name": "person"
  }], 
 "images": 
  [{
      "license": 1, 
      "file_name": "000000425226.jpg",        
      "coco_url": "http://images.cocodataset.org/val2017/000000425226.jpg", 
      "height": 640, 
      "width": 480, 
      "date_captured": 
      "2013-11-14 21:48:51", 
      "flickr_url": 
      "http://farm5.staticflickr.com/4055/4546463824_bc40e0752b_z.jpg", 
      "id": 1
  }], 
 "annotations": 
  [{
      "image_id": 1, 
      "category_id": 1, 
      "segmentation": [], 
      "area": 47803.279549999985, 
      "iscrowd": 0, 
      "bbox": [73.35, 206.02, 300.58, 372.5], 
      "id": 1
  }]
}

• Then we can organize the custom dataset (including images and annotations) as follows:

├── Custom_coco
│   ├── annotations
│   │   └── toy_sample.json
│   ├── images
│   │   └── 000000425226.jpg

Step2. Link custom dataset into DAMO-YOLO

• Link your dataset into datasets.

ln -s path/to/Custom_coco datasets/toy_sample

• Add the custom dataset into damo/config/paths_catalog.py. Note, the added dataset should contain coco in their names to declare the dataset format, e.g., here we use sample_train_coco and sample_test_coco.

'sample_train_coco': {
    'img_dir': 'toy_sample/images',
    'ann_file': 'toy_sample/annotations/toy_sample.json'
},
'sample_test_coco': {
    'img_dir': 'toy_sample/images',
    'ann_file': 'toy_sample/annotations/toy_sample.json'
},

Step3. Modify the config file.

In this tutorial, we finetune on DAMO-YOLO-Tiny as example.

• Download the DAMO-YOLO-Tiny torch model from Model Zoo
• Add the following pretrained model path into damoyolo_tinynasL20_T.py.

self.train.finetune_path='path/to/damoyolo_tinynasL20_T.pth'

• Modify the custom dataset in config file. Change coco_2017_train and coco_2017_test in damoyolo_tinynasL20_T.py to sample_train_coco and sample_test_coco respectively. https://github.com/tinyvision/DAMO-YOLO/blob/6e38813220900955d0f6138429c91a33a79c922f/configs/damoyolo_tinynasL20_T.py#L33-L34

• Modify the category number in config file. Change 'num_classes': 80 in damoyolo_tinynasL20_T.py to 'num_classes': 1. Because in our toy sample, there is only one category, so we set num_classes to 1. https://github.com/tinyvision/DAMO-YOLO/blob/6e38813220900955d0f6138429c91a33a79c922f/configs/damoyolo_tinynasL20_T.py#L64-L66

Step4. Finetune on custom dataset

You can run the finetuning with the following code:

python -m torch.distributed.launch --nproc_per_node=8 tools/train.py -f configs/damoyolo_tinynasL20_T.py