Does the compute status norm could be faster?

[lucasjinreal]

Current for everysingle frame in every single episode, it reads episode.jsonl again and again.
Causing a very huge computing norm stratus time.

Does it really needed? The overall framework is not very engineered.

[yuzhms]

Hi, thanks for pointing this out.

For the norm computation, we only sample a subset of frames (500 in total) instead of iterating over every frame in every episode, so the actual computation time should be limited. You can check the implementation here:

dexbotic/dexbotic/exp/base_exp.py

Lines 489 to 493 in 932192c

	for batch_idx, batch in tqdm.tqdm(
	enumerate(dataloader), desc='Computing norm stats'):
	# only use the first 500 * 128 samples
	if batch_idx > 500:
	break

In addition, we have a caching mechanism for the norm values. The norm statistics are computed only once and then reused, so this process will not be repeatedly triggered during training.

If you are still observing significant overhead, could you share more details about your setup? That would help us further diagnose the issue.

[lucasjinreal]

actually it computes 128 * 2500 about 320000 frames, and the root reason caused time cost as I stated above didn't get alleviated.

Also, if my subset too much, ( i have about 32 subsets) it will cause a very painful time.