About load pretrained models error

[Cxz-dev]

Hi! Thanks for your code release! when I use moganet_base, I set pretrained=True, the error as follows:

RuntimeError: Error(s) in loading state_dict for MogaNet: Unexpected key(s) in state_dict: "head.weight", "head.bias" Can u give me some advice?

[Cxz-dev]

I dowload the pretrained model from https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_base_sz224_8xbs128_ep300.pth.tar

[Lupin1998]

Hi, @Cxz-dev, thanks for using MogaNet. Could you please describe your usage in detail? The RuntimeError you met is caused by no classification head in the built model. I provide two situations of loading pre-trained models and hope these will be helpful to you.

• In image classification tasks, you can build MogaNet-Base and load pre-training without any errors:

  import models
  from timm.models import create_model
  cnn = create_model('moganet_base', pretrained=True)  # build by timm
  cnn = models.moganet.moganet_base(pretrained=True)  # or call the model directly

• For downstream tasks, you will get RuntimeError: Error(s) in loading state_dict for MogaNet if you build the model like this. Since fork_feat=True will remove the classification head in MogaNet and the pre-trained models in moganet-in1k-weights contain parameters of the head, it will raise the error using model.load_state_dict(checkpoint["state_dict"]), where strict defaults to True.

  import models
  from timm.models import create_model
  cnn = create_model('moganet_base', pretrained=True, fork_feat=True)  # fork_feat=True for dense predictions

  Therefore, I suggest you load the model as follows or directly perform dense prediction tasks by our implementations in MogaNet/detection, MogaNet/segmentation, and MogaNet/pose_estimation.

  import torch
  import models
  from timm.models import create_model
  cnn = create_model('moganet_base', fork_feat=True)
  checkpoint = torch.hub.load_state_dict_from_url(url='https://github.com/Westlake-AI/MogaNet/releases/download/moganet-in1k-weights/moganet_base_sz224_8xbs128_ep300.pth.tar', map_location="cpu", check_hash=True)
  cnn.load_state_dict(checkpoint["state_dict"], strict=False)  # allow unexpected keys

[Lupin1998]

Thanks for your question. I will close this issue if you have no more questions. Please feel free to open a new issue when new problems occur.