test.py model_dict = torch.load(pretrained_pth)["state_dict"] gives error about research-contributions HOT 3 OPEN

from research-contributions.

Hi @davidkvcs , Thanks for reporting this issue. The pre-trained model checkpoints should contain three keys {'epochs': , 'best_acc', 'state_dict': }

Key "state_dict" should be the model weights, could you direct me to the model checkpoints you've tried, I will double check whether the saved model has consistent key-value items.
Thank you.

from research-contributions.

Hi @tangy5

I trained the model my self, using this repo.
Do you have some way I can share this .pth file with you?

Since the next line, model.load_state_dict(model_dict) works fine and the prediction look pretty good I assume, the weights are loaded correctly with the changes I made?

It seems my trained model does not have the three keys {'epochs': , 'best_acc', 'state_dict': }, but holds the state_dict only. When I run model_dict = torch.load(pretrained_pth), model_dict is a OrderedDict of length 159, holding what could look like the trained model weights. Examples of field names:

...
swinViT.patch_embed.proj.weight
swinViT.patch_embed.proj.bias
swinViT.layers1.0.blocka.0.norm1.weight
and so on...

Does that make sense?

from research-contributions.