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
Thank you for building Mammoth, which has helped me a lot. However, while using Mammoth to compare baseline methods, I accidentally discovered an error in dualprompt, specifically in lines 117-122 of ./mammoth/models/dualprompt_utils/prompt.py.
While transforming the dimensions of a tensor, reshape should be used with caution, especially since the goal here is to swap two dimensions.
Furthermore, I checked the referenced code in Mammoth and found that this issue has already been raised by someone else: JH-LEE-KR/dualprompt-pytorch#8 (comment)
For the sake of reproducibility of the methods, I suggest making the necessary corrections.
Many thanks.
The text was updated successfully, but these errors were encountered:
Hi @AmeliaBartxn, yes I agree in general the permute should have been used instead of the reshape to swap the dimensions. However, since the operation is performed directly on a nn.Parameter it should not make a difference (and indeed I verified this by changing the operation with batched_prompt_raw = batched_prompt_raw.permute(0, 2, 1, 3, 4, 5, 6) and got the same results).
I will leave the code with the reshape for the moment and maybe add an hyperparameter to conditionally select the permute operation, since the original implementation of dualprompt also uses the reshape instead of the permute.
Thank you for building Mammoth, which has helped me a lot. However, while using Mammoth to compare baseline methods, I accidentally discovered an error in dualprompt, specifically in lines 117-122 of
./mammoth/models/dualprompt_utils/prompt.py
.While transforming the dimensions of a tensor, reshape should be used with caution, especially since the goal here is to swap two dimensions.
Furthermore, I checked the referenced code in Mammoth and found that this issue has already been raised by someone else: JH-LEE-KR/dualprompt-pytorch#8 (comment)
For the sake of reproducibility of the methods, I suggest making the necessary corrections.
Many thanks.
The text was updated successfully, but these errors were encountered: