Freeze Layers for Training

[akshay-bapat-magna]

Have you implemented a way to freeze training for certain layers of the network? I have the backbone and geometric feature portion of the network trained, I would like to fine-tune the Patch-PnP and the head.

If you have not implemented this, would you recommend editing the model itself, or do it some other way?

[shanice-l]

You can set the parameter "FREEZE" to True.

[akshay-bapat-magna]

But this does not work if I need to train up to a checkpoint with all layers trainable, then freezing some and resuming the training. I think the optimizer finds a mismatch in parameters from the checkpoint.