Load pretrained hybrid vit with my custom image size, patch size?

[juanigp]

Hi, I wanna fine tune the model vit_small_r26_s32_224 (or vit_small_r26_s32_224_in21k) to my own dataset.

If I call:

timm.create_model('vit_small_r26_s32_224', pretrained=True, patch_size = 7, img_size = 224*5, num_classes = 2)

will I get the model with the pretrained weights + the patch embedding layer adapted (somehow) for the patch_size, img_size that I desire, or are some layers going to be randomly initialized?

Thank you!