* update ClassifierHead to allow different input format
* add output format support to patch embed
* fix some flatten issues for a few conv head models
* add Format enum and helpers for tensor format (layout) choices
based on is_exportable() set variable controlling behaviour of the block
CoreMLTools support im2col from 6.2 version, unfortunately col2im
is still not supported.
Tested with exporting to ONNX, Torchscript, CoreML, and TVM.
* Split CI tests to run them in parallel
The idea of this PR is to split tests into multiple sets that can be run
in parallel by GH. For this, all tests in test_models.py that would run
on GH get a pytest marker. The GH workflow matrix is factorized to run
only a single marker. That way, only a subset of tests should run per
worker, leading to quicker results.
There is also a worker that runs all the tests that are not inside
test_models.py.
* [skip ci] empty commit to abort ci
* Fix typo in marker name
* Split fx into forward and backward
* Comment out test coverage for now
Checking if it's responsible for regression in CI runtime.
* Remove pytest cov completely from requirements
* Remove cov call in pyproject.toml
Missed that one.
When tests finish, a report should be printed that shows the code
coverage of timm. This should give us a better idea where we should work
on test coverage.
I have tested this locally (on a subset of tests) and it worked.
Since the number of test dependencies was getting quite high, I created
a requirements-dev.txt and moved them there. GH action and
CONTRIBUTING.md are adjusted accordingly.
Furthermore, instead of extending the pytest invocation, I created a
pyproject.toml and added the coverage options there. For completeness, I
also added the black settings that come closest to the style of timm.
LMK if this is not desired.
For now, the coverage is only reported but not enforced. I.e. when a PR
is created that adds uncovered lines, CI will still succeed. We could
think about adding codecov or something like that, but it can be
annoying sometimes and the service was flaky for me in the past.
Description
Add type annotations to _registry.py so that they will pass mypy
--strict.
Comment
I was reading the code and felt that this module would be easier to
understand with type annotations. Therefore, I went ahead and added the
annotations.
The idea with this PR is to start small to see if we can align on _how_
to annotate types. I've seen people in the past disagree on how strictly
to annotate the code base, so before spending too much time on this, I
wanted to check if you agree, Ross.
Most of the added types should be straightforward. Some notes on the
non-trivial changes:
- I made no assumption about the fn passed to register_model, but maybe
the type could be stricter. Are all models nn.Modules?
- If I'm not mistaken, the type hint for get_arch_name was incorrect
- I had to add a # type: ignore to model.__all__ = ...
- I made some minor code changes to list_models to facilitate the
typing. I think the changes should not affect the logic of the function.
- I removed list from list(sorted(...)) because sorted returns always a
list.
* Some additions to the CONTRIBUTING guide
- how to run black if so desired
- install instructions for devs (following GH action)
- running tests
- minor fixups
If there is a guide on how to best add new models, it would be a good
idea to link it here, since I imagine this is what many contributors
would need most help with.
* [skip ci] empty commit to skip ci
Ross Wightman
Ross Wightman
SWHL
Piotr Sebastian Kluska
Benjamin Bossan
Benjamin Bossan