Add Create_model's Image Size setting

[LReion]

Is your feature request related to a problem? Please describe.
I hope to use create_model to create a model, such as the vit_tiny_224 model, to be able to use other sizes of image input without error.

Describe the solution you'd like
use like create_model('vit_tiny_224',img_size=args.img_size)
Describe alternatives you've considered
Through such a modification, the image size of the input model can be changed in the model determined using the specified input image size, so that the created vit_tiny_224 can be a model suitable for an image size of 448
Additional context
I think this avoids the confusion of the model input size with the model name, making the input size of the model a separate parameter to create the model, thus making it possible to get the initial model for each image size

[rwightman]

@LReion the issue / concern here is that only models that need a static image size up front have the img_size arg, convnets don't need it and do n't use it. You just create a preprocessing pipeline with any image size.

For vit as an example, you can create a vit with a different image size by passing the img_size= arg, BUT it is not the native resolution so it forces a resize of the position embedding and some models don't have the resize so they just error out if you pass a different size...

All that said, for your needs, I did actually add a way of doing this that works with the scripts as is, I added a --model-kwargs that will take any number of arg_name=value pairs and pass it through to kwargs of the model. You do need a bit of duplication, the args.img_size is passed to the preprocessing pipepline setup and the model-kwargs is forwarded to model only. Note you can even override args like patch_size, etc.

python validate.py /data/imagenet  -j 8 --model vit_tiny_patch16_224 --model-kwargs img_size=320 --img-size 320
Validating in float32. AMP not enabled.
Loading pretrained weights from Hugging Face hub (timm/vit_tiny_patch16_224.augreg_in21k_ft_in1k)
Resized position embedding: (14, 14) to (20, 20).
torch.Size([1, 1, 192]) torch.Size([1, 400, 192])

python validate.py /data/f/imagenet  -j 8 --model vit_base_patch16_clip_224 --model-kwargs patch_size=12 img_size=288 --img-size 288
Validating in float32. AMP not enabled.
Loading pretrained weights from Hugging Face hub (timm/vit_base_patch16_clip_224.laion2b_ft_in12k_in1k)
Resized position embedding: (14, 14) to (24, 24).
torch.Size([1, 1, 768]) torch.Size([1, 576, 768])
Resize patch embedding torch.Size([768, 3, 16, 16]) to (12, 12), w/ bicubic interpolation.
Model vit_base_patch16_clip_224 created, param count: 86602216

[rwightman]

Those args work for the train and validation scripts.

[LReion]

Thank you for your reply. I was a little unclear about the parameters that can be modified in --model-kwargs.