You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I am training my own text and motion embedding models for evaluation. I noticed in the TextMotionMatchTrainer class, there is a potential bug in the shift applied to create negative examples for the contrastive loss.
If we shift 0, then the "negative" examples with be compared to itself. This is especially problematic when training with low batch sizes, like in the README (batch size 8). The correction is
shift=np.random.randint(1, batch_size-1)
After doing this, I see improved training curves. Below, the grey curve is with the bug fix and the purple curve is with the original code. Batch size is 8.
The text was updated successfully, but these errors were encountered:
I am training my own text and motion embedding models for evaluation. I noticed in the TextMotionMatchTrainer class, there is a potential bug in the shift applied to create negative examples for the contrastive loss.
If we shift 0, then the "negative" examples with be compared to itself. This is especially problematic when training with low batch sizes, like in the README (batch size 8). The correction is
After doing this, I see improved training curves. Below, the grey curve is with the bug fix and the purple curve is with the original code. Batch size is 8.
The text was updated successfully, but these errors were encountered: