Implement Transfer Learning API / Instructions

[lapp0]

It would be helpful if we could piggy back on your library of pre-trained model for transfer learning. Perhaps this may be accomplished by freezing the first 6 (2L) of 9 layers of Mask2Formers transformer decoder.

Usage may look like this:

export DETECTRON2_DATASETS=/path/to/dir/containing/new/dataset

python train_net.py \
  --config-file <pretrained model config> \
  --pretrained-model /path/to/checkpoint_file
  --transfer-learning-dataset <new dataset>

[bowenc0221]

I did not get the question. You can always finetune the model on your own dataset by registering it in Detectron2.

[lapp0]

Sorry I wasn't clear.

What I'm requesting is a feature similar to keras' trainable attribute which freezes the weights and biases of a layer if false. https://keras.io/guides/transfer_learning/#the-typical-transferlearning-workflow

I would like to use one of Mask2Formers pre-trained COCO models to learn features of smaller datasets.

[bowenc0221]

In PyTorch, you can freeze any parameter by setting its require_grad to False.

[lapp0]

I will experiment with freezing different layers. If I have any success I'll make a PR.

[jkim50104]

@lapp0 Where did you add the require_grad False to freeze the layers? And just from curiosity any results about freezing different layers?