Updated parts of the network during fine-tuning? about finetune-detr HOT 1 CLOSED

Thank you for the kind words!

I mention this aspect in a paragraph at the start of the "Fine-tuning" section of my notebook:

NB: There is a --frozen_weights argument. However,
i) I have yet to figure out how it is used,
ii) it is of no use for box detection. Indeed, "frozen training is meant for segmentation only" (as mentioned at this line).

In summary, for box detection, which is my use-case, nothing has been frozen.

However, for segmentation:

• it is possible to freeze parts of the network,
• only the mask head is trained, based on the comments found in main.py and quoted below.

    # Model parameters
    parser.add_argument('--frozen_weights', type=str, default=None,
                        help="Path to the pretrained model. If set, only the mask head will be trained")

https://github.com/facebookresearch/detr/blob/a54b77800eb8e64e3ad0d8237789fcbf2f8350c5/main.py#L31-L33

    if args.frozen_weights is not None:
        assert args.masks, "Frozen training is meant for segmentation only"
    print(args)

https://github.com/facebookresearch/detr/blob/a54b77800eb8e64e3ad0d8237789fcbf2f8350c5/main.py#L109-L111

    if args.frozen_weights is not None:
        checkpoint = torch.load(args.frozen_weights, map_location='cpu')
        model_without_ddp.detr.load_state_dict(checkpoint['model'])

https://github.com/facebookresearch/detr/blob/a54b77800eb8e64e3ad0d8237789fcbf2f8350c5/main.py#L167-L169

from finetune-detr.