Closed ardasahiner closed 2 years ago
Hello @ardasahiner. Timm augmentations are not numba augmentations and therefore won't be able to benefit from FFCV. There is a #154 that might be of some help (maybe you can even contribute if you feel like it). Unfortunately, our team is too small to add support for all existing augmentations. We hope to get some help from the community for this.
Hi,
I was attempting to use FFCV with timm, using the fact that
torch.nn.Module
s should be compatible with thepipelines
argument of FFCV'sLoader
. However, I am getting some strange errors and would like some clarification on what is going wrong here.Please see my simple reproducible implementation below. I use CIFAR100 images and use timm's
create_transform
function. While each transform is not an instance ofnn.Module
, I attempted to wrap it in a simple module with theCustomClass
. However, I get the following issue as documented below.Would you have any suggestions what is causing this issue, or any ideas for a simpler integration with timm? Any help is appreciated.
Error:
Implementation: