Fix metrics not being torch-scriptable due to new is_differentiable property

[maximsch2]

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

What does this PR do?

Metrics were made scriptable in Lightning-AI/pytorch-lightning#4428 but this was broken with addition of is_differentiable property which is triggering TorchScript issue (reported to PyTorch). For now, I'm suggesting we just return None there instead of raising an exception.

[pep8speaks]

Hello @maximsch2! Thanks for updating this PR.

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-04-15 02:51:02 UTC

[codecov]

Codecov Report

Merging #172 (16a642e) into master (da00174) will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #172      +/-   ##
==========================================
+ Coverage   96.14%   96.16%   +0.01%     
==========================================
  Files         180       90      -90     
  Lines        5580     2791    -2789     
==========================================
- Hits         5365     2684    -2681     
+ Misses        215      107     -108     

Flag	Coverage Δ
Linux	79.79% <50.00%> (+<0.01%)	⬆️
Windows	79.79% <50.00%> (+<0.01%)	⬆️
cpu	96.16% <100.00%> (+0.03%)	⬆️
gpu	?
macOS	96.16% <100.00%> (+0.03%)	⬆️
pytest	96.16% <100.00%> (+0.01%)	⬆️
python3.6	96.15% <100.00%> (+0.03%)	⬆️
python3.8	96.16% <100.00%> (+0.03%)	⬆️
python3.9	96.05% <100.00%> (+0.03%)	⬆️
torch1.3.1	95.07% <100.00%> (+0.03%)	⬆️
torch1.4.0	95.19% <100.00%> (+0.03%)	⬆️
torch1.8.1	96.05% <100.00%> (+0.03%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
torchmetrics/metric.py	95.14% <100.00%> (+0.39%)	⬆️
__w/2/s/torchmetrics/regression/spearman.py
__w/2/s/torchmetrics/functional/__init__.py
.../s/torchmetrics/classification/precision_recall.py
...rics/functional/classification/confusion_matrix.py
__w/2/s/torchmetrics/utilities/__init__.py
__w/2/s/torchmetrics/functional/regression/psnr.py
__w/2/s/torchmetrics/functional/regression/ssim.py
...chmetrics/classification/precision_recall_curve.py
__w/2/s/torchmetrics/retrieval/retrieval_metric.py
... and 81 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update da00174...16a642e. Read the comment docs.

[Borda]

but this was broken with addition of is_differentiable property which is triggering TorchScript issue

heh can we also add a test for this so prevent it in the future... :]

[Borda]

that was my concern earlier... #154 (comment)

[maximsch2]

It should hopefully be OK to throw once PyTorch fixes the issue (it's not respecting the ignore on properties!), but that would be a while till it gets to the released version, etc.

[maximsch2]

Actually there is a fix on PyTorch side already, but it's not part of the released version yet and given the desire for backwards-compatibility I don't think we'll be able to use it for a while: pytorch/pytorch#52367

We need to use __jit_ignored_attributes__ there essentially which will be supported in new pytorch, I've updated the PR.