TabNet explainability on custom data

[Alexander Watanabe]

Hello, thank you for this repo. I am trying to run the TabNet notebook on a custom data set, have got everything working up to the explainability decorator which fails with this error:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-37-bfedb2755251> in <module>
----> 1 learn.explain(dl)

<ipython-input-36-c2a2fc0e0447> in explain(x, dl)
      6   for batch_nb, data in enumerate(dl):
      7     with torch.no_grad():
----> 8       out, M_loss, M_explain, masks = x.model(data[0], data[1], True)
      9     for key, value in masks.items():
     10       masks[key] = csc_matrix.dot(value.numpy(), matrix)

~/dev/Practical-Deep-Learning-for-Coders-2.0/.venv-nix/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
    725             result = self._slow_forward(*input, **kwargs)
    726         else:
--> 727             result = self.forward(*input, **kwargs)
    728         for hook in itertools.chain(
    729                 _global_forward_hooks.values(),

TypeError: forward() takes 3 positional arguments but 4 were given

I am reading the docs to better understand how to fix it, if you have any insights/pointers they would be appreciated.

[Zach Mueller]

What version of tabnet do you have installed?

…

On Wed, Dec 2, 2020 at 8:28 PM alexanderwatanabe ***@***.***> wrote: Hello, thank you for this repo. I am trying to run the TabNet notebook on a custom data set, have got everything working up to the explainability decorator which fails with this error: `--------------------------------------------------------------------------- TypeError Traceback (most recent call last) in ----> 1 learn.explain(dl) in explain(x, dl) 6 for batch_nb, data in enumerate(dl): 7 with torch.no_grad(): ----> 8 out, M_loss, M_explain, masks = x.model(data[0], data[1], True) 9 for key, value in masks.items(): 10 masks[key] = csc_matrix.dot(value.numpy(), matrix) ~/dev/Practical-Deep-Learning-for-Coders-2.0/.venv-nix/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs) 725 result = self._slow_forward(*input, **kwargs) 726 else: --> 727 result = self.forward(*input, **kwargs) 728 for hook in itertools.chain( 729 _global_forward_hooks.values(), TypeError: forward() takes 3 positional arguments but 4 were given` I am reading the docs to better understand how to fix it, if you have any insights/pointers they would be appreciated. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#38>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB3YCVZRE3CYHEDIVOMIPMTSS3SVLANCNFSM4ULJC6AQ> .

[Alexander Watanabe]

here are what I think the relevant libraries i have in this environment are:
fastai 2.1.5
fastcore 1.3.6
fast_tabnet 0.2.0
pytorch-tabnet 2.0.1
fastinference 0.0.30
pytorch 1.7.0

also my model is setup to solve for single-variable regression

[Zach Mueller]

If you look at my TabNet notebook, I changed the install dependencies as things were breaking and I didn’t have the time to adjust things.

…

On Wed, Dec 2, 2020 at 8:48 PM alexanderwatanabe ***@***.***> wrote: here are what I think the relevant libraries i have in this environment are: fastai 2.1.5 fastcore 1.3.6 fast_tabnet 0.2.0 pytorch-tabnet 2.0.1 fastinference 0.0.30 pytorch 1.7.0 — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB3YCV6RQFICM5X5J4M7M4DSS3VAJANCNFSM4ULJC6AQ> .

[Alexander Watanabe]

Got it thanks!

[Zach Mueller]

There may be some issues with the new fastai or torch version though (wouldn’t be surprised) so let me know if you run into trouble!

…

On Wed, Dec 2, 2020 at 8:53 PM alexanderwatanabe ***@***.***> wrote: Got it thanks! — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB3YCV5FBVZDJ665W3USFIDSS3VQZANCNFSM4ULJC6AQ> .

[Alexander Watanabe]

Got it working with your pinned versions and the new verions of fastai/torch. If you have any notes or outline for how you might approach fixing it for the new versions I'd love to get involved with contributing. Understand you are probably busy so it's a standing offer for later if necessary!