avoid initializing with Y if X is set

[svlandeg]

Inspired by this issue. The "trouble" is being caused by the shape inference code in chain, but it can't hurt to add some additional robustness into layernorm to avoid using Y if not necessary.

[codecov]

Codecov Report

Merging #466 (c7cae55) into master (ca42188) will decrease coverage by 0.19%.
The diff coverage is 61.86%.

@@            Coverage Diff             @@
##           master     #466      +/-   ##
==========================================
- Coverage   59.66%   59.47%   -0.20%     
==========================================
  Files         100      101       +1     
  Lines        7079     7285     +206     
==========================================
+ Hits         4224     4333     +109     
- Misses       2855     2952      +97     

Impacted Files	Coverage Δ
thinc/layers/logistic.py	100.00% <ø> (ø)
thinc/layers/reduce_first.py	100.00% <ø> (ø)
thinc/layers/reduce_last.py	100.00% <ø> (ø)
thinc/tests/layers/test_layers_api.py	0.00% <ø> (ø)
thinc/tests/layers/test_with_transforms.py	0.00% <0.00%> (ø)
thinc/layers/layernorm.py	90.90% <50.00%> (-9.10%)	⬇️
thinc/backends/ops.py	83.58% <76.92%> (+0.07%)	⬆️
thinc/about.py	100.00% <100.00%> (ø)
thinc/api.py	100.00% <100.00%> (ø)
thinc/config.py	97.05% <100.00%> (+0.98%)	⬆️
... and 14 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 3d5ab74...c7cae55. Read the comment docs.

[svlandeg]

How is this little PR suddenly triggering mypy issues? mypy 0.8 was out already 15 days ago, and master was still building correctly 13 days ago (with the same mypy version), so that can't be it?

[adrianeboyd]

This is probably also related to the newest version of numpy, but I'm not sure why it looks different than in #465, which didn't fail until the test suite.

[svlandeg]

Hm, I guess we'll see when the other PR is merged whether it also fixes this one 🤞

[justindujardin]

I'm late to this conversation, but you're right, the failures were related to the latest numpy release.

I added some # type: ignore comments in the test suite to fix them

The latest numpy release has its own type hints, so our usages in thinc don't completely line up.

At some point we could probably benefit from using the built-in numpy types in thinc, but for now a few comments did the trick. 😅