huggingface
make --dataset.episodes work
What this does
There was an "index out of range error" when --dataset.episodes is specified with train.py. This change fixes it.
Details: The original code shortens the length of episode_data_index array (computed by get_episode_data_index()) to the number of episodes specified. But the episode no is used to access the episode-data-index. For example, specifying --dataset.episodes='[10]' makes episode_data_index contain only one element. But later in code episode-data-index[10] is called to access data for episode 10, causing index-out-of-range.
How it was tested
Tested with training a model with the change.
Examples:
python lerobot/scripts/train.py
...
--dataset.episodes='[40,41,42,43]'
...
How to checkout & try? (for the reviewer)
Try training with --dataset.episodes specified.
The previous cl broke some tests. I made a fix. Can you take another look and also prove the workflows to see if it's fine now.
I was facing this issue., Thanks for the patch. Any updates on merging this to main?
This was a lifesaver when I was filtering out some bad episodes during training - thank you and I hope some one resolves/approves this PR!
