search

octo-models / octo

Octo is a transformer-based robot policy trained on a diverse mix of 800k robot trajectories.
https://octo-models.github.io/
MIT License
885 stars 166 forks source link

Problem of frozen_keys in ```02_finetune_new_observation_action.py``` #107

Open BUAAZhangHaonan opened 5 months ago

BUAAZhangHaonan commented 5 months ago

I found a small problem in examples/02_finetune_new_observation_action.py:

frozen_keys.append("BlockTransformer_0")

By printing param_partitions in freeze_weights, I found that it cannot effectively freeze the parameters of BlockTransformer_0. I think it should be expressed as follows: 

frozen_keys.append("octo_transformer.BlockTransformer_0.*")

In addition, I found that all frozen modes are not as good as training from scratch (of course, the default *hf_model* needs to be frozen)