Ross Wightman

@abcamiletto I have thought about this one as it would be useful, unfortunately I don't have a good solution at present. The features only capability for a given model isn't...

I have thought about this a bit more, I don't see a clean way to do this automatically ... * module local _create__* helpers have the needed info but it's...

```python def _create_swin_transformer(variant, pretrained=False, **kwargs): #CHANGE move default out indices in to build_mode_cfg as it can be resolved with pretrained_cfg there? default_out_indices = tuple(i for i, _ in enumerate(kwargs.get('depths', (1,...

@abcamiletto FYI, there are now only 62 of 1045 model arch that don't support features_only=True after this PR is merged https://github.com/huggingface/pytorch-image-models/pull/2165 The remaining can be supported with the new forward_intermediates()...

These are the remaining: ``` ['coat_lite_medium', 'coat_lite_medium_384', 'coat_lite_mini', 'coat_lite_small', 'coat_lite_tiny', 'coat_mini', 'coat_small', 'coat_tiny', 'convit_base', 'convit_small', 'convit_tiny', 'convmixer_768_32', 'convmixer_1024_20_ks9_p14', 'convmixer_1536_20', 'crossvit_9_240', 'crossvit_9_dagger_240', 'crossvit_15_240', 'crossvit_15_dagger_240', 'crossvit_15_dagger_408', 'crossvit_18_240', 'crossvit_18_dagger_240', 'crossvit_18_dagger_408', 'crossvit_base_240', 'crossvit_small_240', 'crossvit_tiny_240',...

at 34 of of 1053 now that will throw/assert with features_only=True... so instead of spending time on this idea, will close... people can file issue if there is a high...

@Adenialzz @b5y sorry for delay, been trying to get some other things out the door. So, where I got with this, I verified the modelling aspect works. The masking /...

The batch size in inference and train results tables is the ~min~ max batch size (with some reasonable step granularity) they run at (from a 1024 or 512 starting point...

Hmm, I'm 99% sure the issue is in this block (https://github.com/huggingface/pytorch-image-models/blob/394e8145551191ae60f672556936314a20232a35/timm/models/volo.py#L70-L91), something to do with the unfold/fold (col2im is fold) and handling of the shapes Fold support was just added...

@WildChlamydia hmm, needing aten fallback is going to make it not very pratical for many users wanting ONNX export to use outside of the torch ecosystem no?