fit current timm import, and fix pos_embed shape when using 384

[leondgarse]

Fit current timm register_model and to_2tuple import, and fix pos_embed shape when using 384. When using:

sys.path.append('../pytorch-image-models/')
import torch
from models import inception_transformer
tt = inception_transformer.iformer_small_384(pretrained=True)

met 3 errors:

ImportError: cannot import name 'register_model' from 'timm.models.registry'

ModuleNotFoundError: No module named 'timm.models.layers.helpers'

RuntimeError: Error(s) in loading state_dict for InceptionTransformer:
size mismatch for pos_embed1: copying a param with shape torch.Size([1, 56, 56, 96]) from checkpoint, the shape in current model is torch.Size([1, 96, 96, 96]).
size mismatch for pos_embed2: copying a param with shape torch.Size([1, 28, 28, 192]) from checkpoint, the shape in current model is torch.Size([1, 48, 48, 192]).
size mismatch for pos_embed3: copying a param with shape torch.Size([1, 14, 14, 320]) from checkpoint, the shape in current model is torch.Size([1, 24, 24, 320]).
size mismatch for pos_embed4: copying a param with shape torch.Size([1, 7, 7, 384]) from checkpoint, the shape in current model is torch.Size([1, 12, 12, 384]).

• For the first 2, adds a try except block fitting current newest timm imports.
• For the 3rd one, I think it should use 224 calculating the first num_patches for any input image size.

[shangshang0912]

您的邮件已经收到。我会及时做出回复。