Missing aggregate_dataset_length in train.py

[ustcwhy]

System Info

I setup the env by the README.

Reproduction

Thanks for your wonderful work!

I found that in the latest commit, aggregate_dataset_length function in scripts/train.py has been removed, which introduces some bugs while running fine-tuning in simpler-env dataset.

Expected behavior

Can I safely run the fine-tuning after removing these codes that uses aggregate_dataset_length?

[DelinQu]

Hi ustcwhy! Thanks so much for the issues. I have updated the imports in train and squashed the commits. Please sync the latest commits. Thanks!

git fetch origin
git switch main
git reset --hard origin/main

[ustcwhy]

Thanks for your quick response, Delin!

I met some bugs when trying to conduct evaluation on simpler-env dataset. I am trying to evaluate a fine-tuned model on fractal20220817_data_lerobot.

Following the README, I run the evaluation on SimplerEnv using bridge.sh. However, it seems that in LINE 20 obs["agent"] does not have a key of "eef_pos", only include odict_keys(['qpos', 'qvel', 'controller', 'base_pose']).

File "/home/hyw/eo1/experiments/3_simpler/SimplerEnv-OpenVLA/simpler_env/evaluation/maniskill2_evaluator.py", line 120, in run_maniskill2_eval_single_episode
eef_pos=obs["agent"]["eef_pos"]
~~~~~~~~~~~~^^^^^^^^^^^
KeyError: 'eef_pos'

[DelinQu]

The eef_pos is a new feature in allenzren/ManiSkill2_real2sim. To support state input, you should use the submodule ManiSkill2_real2sim from https://github.com/allenzren/ManiSkill2_real2sim

[ustcwhy]

Thanks for your help! I found that exec_horizon is not claimed in init, but are refered in LINE 178.

Set the self.exec_horizon = exec_horizon in init could fix this issue.