Partial freeze and partial train

[mannyv]

Hi @yiwc,

If you are using torch, you could write a function to freeze the layers of interest by setting requires_grad=False on those layers parameters.

If you are using tf and keras you can set the layer.trainable=False

With either framework, could then create a new trainer that apply this function similar to the method described in this post:

How to get mode summary if I use tune.run()?

@bug404, You could also do something like this. I don’t know if it counts as “a better way” but it would work. I do not think that updating after_init on the Trainer will step on any other callbacks because I think those are defined in the policies rather than the trainer but @sven1977 would know better. from ray.tune.registry import register_trainable from ray.rllib.agents.ppo import PPOTrainer ModelPrintPPOTrainer = PPOTrainer.with_updates(after_init=lambda trainer: trainer.get_p…