Add Native Support for Fastai Tensor Types

fastai extends PyTorch’s tensors to have custom data types like TensorCategory, etc. I’m not sure how exactly the WandbCallback for fastai works but I assume it converts them to PyTorch tensors when passing it to Wandb’s JSONEncoders.

Do you think it makes sense to add native support for this? It would require modifying wandb/util.py like so:

def is_fastai_tensor_typename(typename):
    # restrict to TensorCategory and TensorMultiCategory?
    return typename.startswith('fastai') and 'Tensor' in typename

def json_friendly(obj):
    ...
    elif is_pytorch_tensor_typename(typename) or is_fastai_tensor_typename(typename):
    ...

[rsomani95]

@borisdayma I was actually only trying to use the standard W&B callback with the PyTorch 1.8 nightly version. Not trying to log anything extra.

After making the changes mentioned in the related fastai issue – specifically commenting out _patch_tb() here, I ran into the tensor type conversion issue which I was able to bypass making the changes I proposed above.

However, using the fastai W&B callback now leads to a python multiprocessing error that I’m not sure how to debug, so I decided to stop hacking around and let the official versions release.

I do not have a colab notebook for you to try out. Is the description above specific enough? It’s tricky to reproduce because I’ve made modifications to both fastai and w&b source. LMK if you plan on tackling this and how I can be of better help.

Also should we try to close the related fastai issue first?

Sure! I’m just not sure how to proceed

EDIT: I can reproduce my steps in an isolated environment and post the specific error messages here

[issue-label-bot[bot]]

Issue-Label Bot is automatically applying the label feature_request to this issue, with a confidence of 0.94. Please mark this comment with 👍 or 👎 to give our bot feedback!

Links: app homepage, dashboard and code for this bot.