Timm doesnt work well with AutoModelForImageClassification, num_labels doesn't work as expected.

[lucian-student]

System Info

'4.57.1'

RuntimeError: Error(s) in loading state_dict for Linear:
	size mismatch for bias: copying a param with shape torch.Size([1000]) from checkpoint, the shape in current model is torch.Size([6]).

Who can help?

Practically anybody, probably an easy fix.

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

model = AutoModelForImageClassification.from_pretrained("timm/efficientvit_b3.r256_in1k",num_labels=6)

Expected behavior

Whenever i pass num_labels i get this since effiicient_net was trained using 1000 labels.

timm.create_model works fine, so I assume timm integration isn't refined.

Expected behaviour that chechkpoint will be loaded using timm. Not other means.

[Rocketknight1]

You probably want to set ignore_mismatched_sizes=True!