How to freeze layers and models multiple times during training and and unfreeze them

implementation help
1

Hi, how can I freeze layers and activate certain custom functions multiple times while using Pytorch lighting? I can freeze and unfreeze models multiple times in Pytorch easily, however, not sure how to do it in Pytorch Lighting and utilizing EarlyStop and BASEFINETUNING CallBacks.

So basically, I need to freeze and unfreeze the model/layers in the following steps:

  1. freeze backbone and train the classifier only.
    1.1 stop training once validation loss is not decreasing for certain epochs (Early stop callbacks in PL)?

  2. unfreeze backbone and train the whole model for another 200 epochs but stop if validation loss is not decreasing? (how to do two early stop callbacks here?)

  3. activate multi-inputs (my model need to take multi inputs in certain layers in this step) and continue training for another 100 epochs but stop if validation loss is not decreasing

  4. activate the attention mechanism in some blocks and train for some epochs and stop if validation loss is not decreasing.

Is it possible to do all these in a single Trainer? or do I have to save/load checkpoints multiple times and reinitialize a new Trainer object?

Thanks in advance