Exploring use of kwargs for timm model and transforms creation

[rwightman]

What does this PR do?

Allow kwargs to be passed through to timm create_model and create_transforms fns via the TimmWrapperModel and TimmWrapperImageProcessor classes respectively.

Current exploring use and not ready to finalize. Would like to allow flexibility to configure model and transforms at runtime for added flexibility.

Examples:

Allow dynamic image size for vits
model = AutoModelForImageClassification.from_pretrained("timm/vit_base_patch16_224.augreg2_in21k_ft_in1k", dynamic_img_size=True)

Change image size of pre-processor

model = AutoModel.from_pretrained("timm/vit_large_patch14_dinov2.lvd142m", dynamic_img_size=True)
proc = transformers.AutoImageProcessor.from_pretrained("timm/vit_large_patch14_dinov2.lvd142m", img_size=448)

Customize image transforms, this will enable auto_augment and rand-erasing in train transforms and use the train resolution instead of test target size for eval transform

proc = transformers.AutoImageProcessor.from_pretrained("timm/resnet50.a1_in1k", use_train_size=True, auto_augment='rand-m9-inc1-mstd101', re_prob=0.3)

[rwightman]

I though this could also be neat to wire through in the image classification example script, though not quite sure how to add the args to HfArgumentParser, would be least obtrusive if a --model-kwargs and --preprocessor-kwargs could be added that'd pass through lists of key-value pairs as kwargs to the AutoModel and AutoImageProcessor...

[rwightman]

Not sure who else would have opinions re the example scripts, this could allow all the advanced timm augmentations, and also enabling of some model regularization like stochastic depth (e.g. drop_path=0.2) for training through transformers. @merveenoyan @NielsRogge

[qubvel]

Thanks for updating! Can you please add test cases?

[qubvel]

Would be great to have a variable for tuple(inspect.signature(timm.data.create_transform).parameters.keys()), easier to read and debug

[rwightman]

@qubvel changed that code a little in last commit

[rwightman]

I added properties to the image processor so use cases where people were read image_processor.size, image_processor.image_mean, etc will work. autotrain relied on this and now it works with timm models.

[qubvel]

Great! Can you please add some test cases to tests/models/timm_wrapper/? And would be great to add some code snippets to the docs timm_wrapper.md to make sure it's clear how to modify image processor/model parameters.

[rwightman]

Great! Can you please add some test cases to tests/models/timm_wrapper/? And would be great to add some code snippets to the docs timm_wrapper.md to make sure it's clear how to modify image processor/model parameters.

I'll get to this eventually, I wanted to explore the use cases a bit more with some notebooks, train runs and see if any other changes are worthwhile.