Fix Pretrained ResNet/ViT not working when features_only=True

[lccol]

Fix the AssertionError raised when pretrained ResNet/ViT models (from timm) were created with features_only = True.

ResNet18 and Swin models were not affected by the issue (Swin models are created with torchvision instead of timm).

Closes #2640.

[isaaccorley]

We will probably have to bump the min version of timm we support because looks like FeatureGetterNet and FeatureListNet aren't available in the min version. Will have to search to see when they were first added

[lccol]

We will probably have to bump the min version of timm we support because looks like FeatureGetterNet and FeatureListNet aren't available in the min version. Will have to search to see when they were first added

Yes. Unfortunately, in the current min version of timm, ViT model does not support the features_only param:

RuntimeError: features_only not implemented for Vision Transformer models.

EDIT: it seems that features_only was introduced for ViT in version 1.0.3 of timm. The previous one (0.9.16) still raises the error

[adamjstewart]

Should we skip the tests for older versions of timm? We don't necessarily need to bump the minimum version if it's very new.

[lccol]

I checked and timm 1.0.3 was released in April/May 2024, so the introduction of features_only on ViT is quite recent.

[isaaccorley]

You can just skip the tests using pytest.mark.skip and check what version of timm is installed.

import pytest
import timm
    
@pytest.mark.skipif(int(timm.__version__.split(".")[0]) < 1, reason="Requires timm 1.0.3 or higher")
def test_features_only():
   ...

[adamjstewart]

Or better:

import pytest

def test_features_only():
    pytest.importorskip('timm', minversion='1.0.3')

[adamjstewart]

Just a minor request to make the tests simpler, I can do this myself if you want.

[adamjstewart]

Tests look much cleaner now, thanks!

[adamjstewart]

I think this looks good now, thanks for putting up with my tedious change requests!

Note that this PR does increase the time our unit tests take to run from 11m 34s to 12m 52s (+11%). For now that's tolerable. If we add more models in the future and this becomes too much I may only run the features_only=True tests on the release branch.

Thanks again for the PR!