Lukas Folle
Lukas Folle
Its probably this line https://github.com/lucidrains/vit-pytorch/blob/6c8dfc185ea41f4d2388e4d33bbb76f900ff8a0a/vit_pytorch/vit_pytorch.py#L63
I encountered this problem too. For a quick fix was changing https://github.com/mikgroup/sigpy/blob/master/sigpy/mri/samp.py#L157 to just `@jit`.
I havent benchmarked it. Probably depends also on the input size.
No worries. Unfortunately, I won't have time to benchmark it. ________________________________ Von: Siddharth Srinivasan ***@***.***> Gesendet: Mittwoch, 16. Februar 2022 21:27:20 An: mikgroup/sigpy Cc: Folle, Lukas (CS5); Comment Betreff: Re:...
@Qwlouse Any update on this? I still perceive the very same error.
To access the named blocks try this: ```python from torchvision.models.feature_extraction import create_feature_extractor, get_graph_node_names def get_model_attention(model, x, blocks=["blocks.11.attn.softmax"]): model_attention = create_feature_extractor(model, blocks) attention = model_attention(x) return list(attention.values())[0].detach() if __name__ == "__main__":...
Just click through the files and download "extended-cohn-kanade-images.zip". This file contains all the images.
@radarhere I believe this would enhance the Pillow package substantially. Especially using Pillow with 3-D support as the backend for torchvision would be awesome.
A public dataset would probably be a great choice, something like this for example: https://www.cancerimagingarchive.net/collection/tcga-lgg/
Good point! How about this sample data from Slicer? https://www.slicer.org/wiki/SampleData