Downloading model does not work

[thomasloux]

Running this line torch.hub.download_url_to_file("https://figshare.com/ndownloader/files/59143058", "matris_10m_oam.pt") does not actually download the model, although if I manually go on this url with my browser, the file is automatically downloaded on my computer.
So that MatRIS.load() fails with: EOFError: Ran out of input.

[daniisler]

I have also ran across this issue. You can solve it most easily by replacing the download URLs by the api URLs from figshare:

https://api.figshare.com/v2/file/download/<file_id>

Would be nice to see this in a future version. Also, could you upload the weights for the model pretrained on omat24? I would be delighted to see that!

[bjthorpe]

FYI for anyone in the future still having this issue I wrote the attached python script to get the files and put them in the right place i.e. ~/.cache/matris. Just run it as python Download_matris_model.py "MODEL_NAME" before running MatRIS.

Download_matris_model.py

For future reference to the devs it would be nice not have it hard-coded to a specific directory, maybe allow users to override it it with an environment variable like MATRIS_CACHE_DIR. Or even allowing you to provide a path to the file instead of downloading it.

I say this having dealt with the, admittedly unique situation of using a container which has a different home directory "/root" at build time compared to run time. Short version is I had to pre-download the checkpoints at build time since the system I'm working on does not have internet access at run time. Hence the python script.