Pytorch hub models don't download to correct directory

[onetonfoot]

Currently the code that download a models weights doesn't use torch.hub.hub_dir, so the weights are downloaded to the current working directory instead of that specified by default or torch.hub.set_dir. It also means that the weights are re-downloaded each time the working directory is changed e.g

ssd_model = torch.hub.load(
            'NVIDIA/DeepLearningExamples', 'nvidia_ssd')

os.chdir("../")

ssd_model = torch.hub.load('NVIDIA/DeepLearningExamples', 'nvidia_ssd')

#Downloads twice rather than using the cache

One possible fix could be to use torch.hub.load_state_dict_from_url to download instead however it requires the filenames to follow the naming convention of model-<sha256>.pth, so the weight files would have to be renamed.

Or alternatively something along the lines of

ckpt_file = os.path.join(
            torch.hub._get_torch_home(), "checkpoints", os.path.basename(checkpoint))

Thanks

[nv-kkudrynski]

Thanks for suggestions. This is a known issue. We have work on checkpoint conventions further in our roadmap, but PR with temporary fix should be merged very soon.

[onetonfoot]

Alright cool since you guys are already on it I'll close this then