ViTModel custom image sizes

[pmin20]

Hi! I'm looking to use ViTModel, but want to input images with standardized size of 1920 x 1080. I see that most Huggingface examples with ViT utilize ViTModel.from_pretrained("google/vit-base-patch16-224-in21k"), but in my understanding this doesn't allow me to input images that were preprocessed with parameter size = {"height": 1920, "width": 1080} in ViTImageProcessor. How can I use ViTModel with a custom image size?

[rwightman]

@pmin20 this is timm, not transformers, in timm you could

mm = timm.create_model('vit_base_patch16_224.augreg_in21k', pretrained=True, img_size=(1920, 1088))

to statically resize the model weights to this img size at creation time and will now accept mm(torch.randn(1, 3, 1088, 1920))

or

mm = timm.create_model('vit_base_patch16_224.augreg_in21k', pretrained=True, dynamic_img_size=True)

to allow dynamic resize per batch accepts mm(torch.randn(1, 3, 1088, 1920)) and other sizes as long as divisible by patch size

or

mm = timm.create_model('vit_base_patch16_224.augreg_in21k', pretrained=True, dynamic_img_size=True, dynamic_img_pad=True)

to allow mm(torch.randn(1, 3, 1080, 1920)) (1080 is not divisible by patch size 16 so needs padding).

I am not familiar with transformers equivalents.