search

LoSealL / VideoSuperResolution

A collection of state-of-the-art video or single-image super-resolution architectures, reimplemented in tensorflow.
MIT License
1.61k stars 295 forks source link

tecagon model defination vs the pretrained TecoPSNR.pth mismatched keys #118

Closed atrah22 closed 4 years ago

atrah22 commented 4 years ago

Hello, The pretrained TecoPSNR.pth (https://github.com/LoSealL/Model/releases/download/tecogan/tecogan.zip) model mismatches with the model definition in TecoGAN.py. The missing keys while loading the state dict that is, model.load_state_dict(torch.load(TecoPSNR.pth)) are shown below:

Missing key(s) in state_dict: "gnet.body.0.body.0.weight", "gnet.body.0.body.0.bias", "gnet.body.1.body.0.weight", "gnet.body.1.body.0.bias", "gnet.body.1.body.2.weight", "gnet.body.1.body.2.bias", "gnet.body.17.body.0.1.body.0.weight", "gnet.body.17.body.0.1.body.0.bias", "gnet.body.17.body.1.1.body.0.weight", "gnet.body.17.body.1.1.body.0.bias", "gnet.body.18.body.0.weight", "gnet.body.18.body.0.bias".

Unexpected key(s) in state_dict: "gnet.body.19.weight", "gnet.body.19.bias", "gnet.body.0.weight", "gnet.body.0.bias", "gnet.body.17.body.2.weight", "gnet.body.17.body.2.bias", "gnet.body.17.body.0.weight", "gnet.body.17.body.0.bias", "gnet.body.18.body.1.0.weight", "gnet.body.18.body.1.0.bias", "gnet.body.18.body.0.0.weight", "gnet.body.18.body.0.0.bias".

BRs, Atul

LoSealL commented 4 years ago

The training of the tecogan is not finished yet. The weights are not converged. Even though, you can try with the latest code, the unmatched keys shall not block the inference. While the results are ugly :(