Error(s) in loading state_dict for Generator: size mismatch for encoder.lstm.weight_ih_l0: copying a param with shape torch.Size([64, 512]) from checkpoint, the shape in current model is torch.Size([128, 512])

[ghost]

Helo, After trainig I try load model. I did not change any parameter.

Error(s) in loading state_dict for Generator:
    size mismatch for encoder.lstm.weight_ih_l0: copying a param with shape torch.Size([64, 512]) from checkpoint, the shape in current model is torch.Size([128, 512])

What make wrong?

[ghost]

I had to change the code to: G = Generator(**16**,256,512,32).eval().to(device) # (dim_neck, dim_emb, dim_pre, freq): Instead of: G = Generator(**32**,256,512,32).eval().to(device) # (dim_neck, dim_emb, dim_pre, freq):

[JohnHerry]

The same problem. What is the final solution?

[JohnHerry]

I am trying to retrain with python main.py --dim_neck=32 --freq=32 ; Maybe that is the same inputs shape with pretrained model.

[auspicious3000]

Yes. There might be some parameter discrepancies after the 2nd major repo update.

[ynalcakan]

So what is the solution? Will you make a correction on the repo? The code is still not reproducible with this current version.

[auspicious3000]

@ynalcakan To load a Pytorch checkpoint, the hyperparameters have to match. You need to set the corresponding hyperparameters before loading a Pytorch checkpoint. For example, you can try setting --dim_neck=32 --freq=32