AttributeError: 'GCNConv' object has no attribute '_explain'

[Pantelispanka]

Had that error in from 2.0.3 to 2.0.4 when loading a model. Are there breaking changes in the minor versions? Is there a way to ensure that?

[rusty1s]

We sadly cannot prevent any renaming of internal attributes even for minor version releases. Your best bet to prevent such kind of failures is to not store the Python module, but only its state_dict:

torch.save(model.state_dict())

This should fix your issues. You can read more about it here.

[Pantelispanka]

We store also the model_dict and that solved the issue! That way the "receiver" of the model requires the class / module of the model. We also deliver a "wrapper" class through pickle in our " runtime " ( a python module ). Would dynamic importing of the pyg layers solve the issue ; ( haven't tried that yet )

Also a question regarding the convergence of pyg models. We have seen that the model convergence is different from cpu to gpu and by a lot. Is there a thread for that I should follow ?
Thanks a lot !!

[rusty1s]

Not sure what that means, but ideally you store the hyperparams of the model to initialize it, and then use the state_dict to fill in parameter values.

Can you share some more information about the divergence between CPU and GPU runs? Performance should be equivalent besides any floating point precision errors.

[Debottam]

getting error while inferring on a trained model the model details
EdgeClassifier(
(node_embedding): Linear(in_features=8, out_features=256, bias=True)
(conv1): GCNConv(256, 128)
(conv2): GCNConv(128, 64)
(fc): Linear(in_features=128, out_features=1, bias=True)
)
AttributeError: 'GCNConv' object has no attribute '_user_args'
for pytorch-2.0.1

[akihironitta]

@Debottam Sorry for the delay. Have you had a chance to try this?

 path = "./GNNCalo_net_cluster_500evs_1.pth"
-torch.save(model, path)
+torch.save(model.state_dict(), path)

[Debottam]

no let me try this. Thanks @akihironitta

[Debottam]

Hi @akihironitta ,
How do I run inference now? I did saved as stat_dict. Now I did

input_size = 8
hidden_dim = 256
output_dim = 1
path = "./GNNCalo_net_cluster_500evs_1.pth"
model = EdgeClassifier(input_size, hidden_dim, output_dim )
model = torch.load(path)

but getting following error while running the inference

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Cell In[65], line 3
      1 # Evaluate the model on the test set
      2 with torch.no_grad():
----> 3     model.eval()
      4     test_edge_scores = model(x_1, edge_index_test_loop, edge_index_out_ch_1)

AttributeError: 'collections.OrderedDict' object has no attribute 'eval'

[Debottam]

ok I did

path = "./GNNCalo_net_cluster.pth"
model = EdgeClassifier(input_size, hidden_dim, output_dim )
#model.load(torch.load(path))
model = torch.load(path)

now its working

[akihironitta]

Glad to hear it's working now :)