Pavel C

icon indicating a website link czemp.in

icon company University of Southern California icon location PhD student @USC-Lira, advisor @ebiyik

Results 4 comments of Pavel C

Weird behavior training CNN policies with train_rl

I think either of these two CnnPolicies would mesh better with imitation. Maybe imitation could just provide its own CnnPolicy? Don't allow changing feature extractor (not super elegant, parameter still...

Weird behavior training CNN policies with train_rl

Yes, that would work!

Weird behavior training CNN policies with train_rl

@AdamGleave Just tested with the current master. Yes the same issues exists. `python -m imitation.scripts.train_rl with common.env_name=AsteroidsNoFrameskip-v4 ` gives you a FF policy, as it should. Using named config a)...

Remove normalize_input_layer default

One could also change the default [here](https://github.com/HumanCompatibleAI/imitation/blob/40a2a559706e50bf60d7cc388a2c36dd0d4e8619/src/imitation/util/networks.py#L213), though that wouldn't be visible in the sacred config if we remove the hook. Maybe non-trivial defaults like this should always be visible...