Missing key(s) in state_dict: "blocks.1.synapse.weight_g", "blocks.1.synapse.weight_v", "blocks.3.synapse.weight_g", "blocks.3.synapse.weight_v", "blocks.4.synapse.weight_g", "blocks.4.synapse.weight_v"

[HAMBUK]

Describe the bug
After updating lava-dl to 0.6.0, the naming way of weight has been changed.
So when loading the pt file(which is trained in lava-dl 0.6.0 environment) for inference, I get this error

Missing key(s) in state_dict: "blocks.1.synapse.weight_g", "blocks.1.synapse.weight_v", "blocks.3.synapse.weight_g", "blocks.3.synapse.weight_v", "blocks.4.synapse.weight_g", "blocks.4.synapse.weight_v"

Problem cause:
Maybe the naming way of weights when making .pt file has been changed since the update, but loading .pt file stays same. So that is why we got the mismatch issue.

Solution:
So I downgrade my lava-dl to 0.5.0 and trained it again and get the new .pt file and check it works fine now.

Screenshots
If applicable, add screenshots to help explain your problem. Remove section otherwise.

Environment (please complete the following information):
Ubuntu 22.04 LTS
Lava-dl 0.6.0
Pytorch 2.1.0
Pyhton 3.10.0

• Device: [e.g. Laptop, Intel cloud]
• OS: [e.g. Linux]
• Lava version [e.g. 0.6.1]

Additional context
Add any other context about the problem here. Remove section otherwise.

[HAMBUK]

I have noticed that this is not the lava-dl version issue.
During training, after saving the best.pt file, I also load pt file and export it with .net file like below

    if stats.testing.best_accuracy:
        epochs_without_improvement = 0
        print("Current Process Saved.")
        torch.save(network.state_dict(), trained_folder + "/network.pt")
        network.load_state_dict(torch.load(trained_folder + '/network.pt'))
        network.export_hdf5(trained_folder + '/network.net')

This was the probelm, when exporting pt to net, the weight architectures are changed after this epoch.

My export_hdf5 function is below:

def export_hdf5(self, filename):
    # network export to hdf5 format
    h = h5py.File(filename, "w")
    layer = h.create_group("layer")
    for i, b in enumerate(self.blocks):
        b.export_hdf5(layer.create_group(f"{i}"))

So, it works fine after I just save only pt files and do the conversion later or separately