Refactoring and visualization

[bilabbasi]

The main features added from this PR are as follows:

• a separate Metric class to simplify the FaceMetric, TangentMetric, etc.
• a pooling operation for meshes
• adds scripts to easily run experiments
• adds scripts for visualization of meshes
• the train scripts have been significantly refactors for ease of reuse through out the repo (and not to be used as just a stand alone script)

[trachelr]

Hi Bilal, thanks a lot for your PR.
Let me test it and do a quick review.

[trachelr]

Very nice improvement of the code!! I was able to include pytorch_geometric in the environment.yaml file to install it with conda. I installed the CPU version in Ubuntu18.04 (through WSL) but I wasnt able to try GPU support. I've train a model for few epochs in the segmentation task with COSEG and HumanSeg datasets but I wasn't able to get the FAUST dataset for correspondance tasks. The scripts for metric visualisation is working on COSEG but I wasnt able to run the script for pools visualisation. Is it supposed to work with autoencoders only? There's a few pytests errors as well, please see logs in test.txt.

Thanks again for your contribution!

[trachelr]

I think pytorch_geometric is missing, please see pyg package here
https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html

[bilabbasi]

The pytorch_geometric installation has separate instructions in the README, since it requires some additional stuff. There's a specific note in the README indicating that the environment.yaml doesn't contain installation for pytorch_geometric.

[trachelr]

my bad, I didn't read that note. I installed pyg version 2.0.4 and py38_torch_1.11.0_cu115
Let it like that if you'd like to keep the same installation version.

[trachelr]

There's a warning here, please use torch.linalg.eig

[bilabbasi]

That's strange; I'm not getting an error. Just to make sure, is your PyTorch up-to-date? I just ran it on my side, and nothing.

[trachelr]

I have version 1.11

[bilabbasi]

Hey Romain, thanks for the feedback. I've addressed a lot of the small things you've mentioned and submitted the changes. For the FAUST installation, can you tell me what your error is? If it persists, then I can just add details on how to download/set up the FAUST dataset.

I've removed the mesh pooling for now, since it's clear it requires a bit more testing before being implemented.

Thanks for your review so far :)

[trachelr]

Hello Bilal :)
Thanks for your changes and sorry for the mess-up with the installation... actually I'm running the code in a windows environment. For the FAUST dataset, it's a license issue, you know I'm not allowed to download the dataset... let's skip this part. Did you run the tests? I saw a few errors, don't know if your changes fixed it...
If you're confident with your changes, we can merge.