Running evaluation on cpu - "AssertionError: Torch not compiled with CUDA enabled"

[GilLevi]

Hi,

Thank you for the great work in publishing this repository. I'm trying to evaluate CLIP in text to image retrieval on flickr30k by running evaluate.py with --cfg-path lavis/projects/clip/exp_flickr_ret_eval.yaml. However, I'm receiving the assertion error above, it occurs since PrefetchLoader defined in dataloader_utils.py uses the torch.cuda.Streamclass.

I removed the cuda.Stream from the PrefetchLoader class and managed to get the evaluation code to extract the text features (lines 729-739 of model.py, in the function compute_sim_matrix of the class Clip), but when reaching the image features computation, nothing seems to be running. I pinned it down and it occurs in line 742 of model.py when looping over the data loader (for samples in data_loader:). I even changed the batch size to 2 in case it just takes long and changed num_workers to 1 in case it's something there. I also tried to remove the use of PrefetchLoader by commenting out the line loader = PrefetchLoader(loader) in runner_base.py.

Is there support in the code for running on cpu? if not, what changes should I make to add support to it? I can share the changes that I made in PrefetchLoader if that's a good direction, but I figured there has to be a simpler solution.

Thanks in advance! Gil