Stage1 training crashes on eval about motiongpt HOT 6 CLOSED

I have the same problem, how do I know if processing the HumanML3D dataset have a problem?
it looks like it passed correctly

from motiongpt.

Hi, @gaoxiao
can you tell me how to resolve the dataset corruption problem?

Replacing the test data with validation data fixed this issue, confirming that the test data is indeed corrupted.

from motiongpt.

Hi, I was facing the same issue.

• I found a couple files in the test set had non-finite values ( Files '007975', 'M007975').
• I removed them and their corresponding motion file counterparts.
• Yet, the same issue continues.

However, when I change the SPLIT in, # Evaluating Configuration

EVAL:
  BATCH_SIZE: 32 # Evaluating Batch size
  SPLIT: test -->  val

The training works fine. Confirming the experience from @gaoxiao above.
I would like to know why the Evaluation Configuration uses 'test' SPLIT by default? is this recomennded?

from motiongpt.

Hi, @gaoxiao

This error is most likely caused by the incorrectly processed dataset. We recommend re-processing the HumanML3D dataset following the instructions in their official repo. Hope this can help you solve the problem.

from motiongpt.

Hi, @gaoxiao

This error is most likely caused by the incorrectly processed dataset. We recommend re-processing the HumanML3D dataset following the instructions in their official repo. Hope this can help you solve the problem.

Replacing the test data with validation data fixed this issue, confirming that the test data is indeed corrupted.

from motiongpt.

hi,
I get the same error,
these are the results of the HumanML3D data preparation,
the error looks ok, what else should I check?

from motiongpt.