How to fine-tune the pre-trained on Imagenet with 384x384 resolution? about uniformer

[Andy1621]

You can simply comment these codes since I never use it. Similar problem can be found in here. I forget to delete it when merging codes.

UniFormer/image_classification/token_labeling/tlt/utils/utils.py

Lines 92 to 101 in e802470

	old_posemb = state_dict['pos_embed']
	if model.pos_embed.shape != old_posemb.shape: # need resize the position embedding by interpolate
	if len(old_posemb.shape)==3:
	if int(math.sqrt(old_posemb.shape[1]))**2==old_posemb.shape[1]:
	new_posemb = resize_pos_embed_without_cls(old_posemb, model.pos_embed)
	else:
	new_posemb = resize_pos_embed(old_posemb, model.pos_embed)
	elif len(old_posemb.shape)==4:
	new_posemb = resize_pos_embed_4d(old_posemb, model.pos_embed)
	state_dict['pos_embed'] = new_posemb

[yunh-w]

@Andy1621 Thanks!

[Andy1621]

As there is no more activity, I am closing the issue, don't hesitate to reopen it if necessary.