SwinTransformer' object has no attribute 'pos_drop'

[Sridhar-Ranganaboina]

getting error pos_drop attribute is not available

[0chandansharma]

Hi @Sridhar-Ranganaboina ,

This "SwinTransformer' is implemented in "timm" library, in the new release the 'pos_drop' might be removed
better to downgrade to some older version it should pass.

[poetaster]

Hi @Sridhar-Ranganaboina ,

This "SwinTransformer' is implemented in "timm" library, in the new release the 'pos_drop' might be removed better to downgrade to some older version it should pass.

EDIT: wrong repo, sorry!

I tried creating a custom env with conda, but it's not even possible to install the libraries as listed. any idea for a patch?

[rafikoham]

There's this forward function on the model.py of Donut ([/usr/local/lib/python3.10/dist-packages/donut/model.py])
Here's the code :
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Args:
x: (batch_size, num_channels, height, width)
"""
x = self.model.patch_embed(x)
x = self.model.pos_drop(x)
x = self.model.layers(x)
return x

Maybe we can comment the x = self.model.pos_drop(x) ?
Or we should notify some contributors on the SwinTransformer forward function?

[serser]

pos_drop is still in v0.6.13 of timm and removed after v0.9:

https://github.com/huggingface/pytorch-image-models/blob/2696eedd1b1763b9da06e68c1af3c5ed360fe687/timm/models/swin_transformer.py#L474

I downgraded to v0.6.13 stable and it passes.