Errors from different version of PyG

[hkim716]

Hi Matt,

I have a GNN model with SplineConv made by old version of torch_geometric. (1.7.2) And I tried to load this model on another machine, 2.0.2 torch_geometric was installed on.

I got an error when I load the model on new machine like this.

RuntimeError: Error(s) in loading state_dict for autoencoder:
	Missing key(s) in state_dict: "encoder.down_path.0.lin.weight", "encoder.down_path.1.lin.weight", "encoder.down_path.2.lin.weight", "encoder.down_path.3.lin.weight", "encoder.down_path.4.lin.weight", "encoder.down_path.5.lin.weight", "encoder.down_path.6.lin.weight", "decoder.up_path.0.lin.weight", "decoder.up_path.1.lin.weight", "decoder.up_path.2.lin.weight", "decoder.up_path.3.lin.weight", "decoder.up_path.4.lin.weight", "decoder.up_path.5.lin.weight", "decoder.up_path.6.lin.weight". 
	Unexpected key(s) in state_dict: "encoder.down_path.0.root", "encoder.down_path.1.root", "encoder.down_path.2.root", "encoder.down_path.3.root", "encoder.down_path.4.root", "encoder.down_path.5.root", "encoder.down_path.6.root", "decoder.up_path.0.root", "decoder.up_path.1.root", "decoder.up_path.2.root", "decoder.up_path.3.root", "decoder.up_path.4.root", "decoder.up_path.5.root", "decoder.up_path.6.root". 

When I print name, param in model.named_parameters():, the two machines gave me different names.
In the old version, it prints like this :

1000
encoder.down_path.0.root torch.Size([1, 8])
1008
encoder.down_path.0.bias torch.Size([8])
1016
...

In the new version, it prints like this :

encoder.down_path.0.weight torch.Size([125, 1, 8])
1000
encoder.down_path.0.bias torch.Size([8])
1008
encoder.down_path.0.lin.weight torch.Size([8, 1])
1016

How can I solve this issue, and why were their names changed?
Should I re-install the whole packages with specific versions installed on the old machine?

[zcaicaros]

I have had the same problem before. The way I solved it is to install the legacy version of PyG 1.7.2. Suppose you are using pytorch 1.7.0 + cuda 11.0, This is simple:

pip install --user --upgrade pip
pip install --user torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html
pip install --user torch-sparse -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html
# For additional but optional functionality, can be skipped:
pip install --user torch-cluster -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html
# For additional but optional functionality, can be skipped:
pip install --user torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.7.0+cu110.html  
pip install --user torch-geometric==1.7.2

I guess in 2.0.2, they made some changes to the architectures (keys) of the network so when loading an old version of the model, the new interpreter cannot recognize the module's name.

[rusty1s]

Yes, one needs to sadly modify the names of the state_dict, see my answer below.

[rusty1s]

This is the result of our efforts to bring lazy initialization of model parameters to PyG. A simple fix is to replace all root occurrences with lin.weight and transpose the tensor, e.g.:

state_dict["encoder.down_path.0.lin.weight"] = state_dict["encoder.down_path.0.root"].t()