Tensorboard with HeteroGraph?

[archids]

Hi:

I have a Heterograph set up for GCN and trying to follow the example to write to Tensorboard:

model = HeteroGraphGCN(data.metadata(), hidden_channels=64, out_channels = num_room_labels,
                  num_layers=2)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
data, model = data.to(device), model.to(device)

with torch.no_grad():  
    out = model(data.x_dict, data.edge_index_dict)

optimizer = torch.optim.Adam(model.parameters(), lr=0.004, weight_decay=0.001)


transform = T.RandomNodeSplit('train_rest', num_val=200, num_test=500)
data = transform(data)

from torch.utils.tensorboard import SummaryWriter
from torch.utils.tensorboard import SummaryWriter
model(data.x_dict, data.edge_index_dict)
writer = SummaryWriter()
writer.add_graph(model, [data.x_dict, data.edge_index_dict])

And i get the following error:

Cannot create dict for key type '(str, str, str)', only int, float, complex, Tensor, device and string keys are supported
Error occurs, No graph saved

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
f:\condos\220904_condos_heterograph_node_classification_final.ipynb Cell 67 in <cell line: 3>()
      1 from torch.utils.tensorboard import SummaryWriter
      2 writer = SummaryWriter()
----> 3 writer.add_graph(model, [data.x_dict, data.edge_index_dict])

Please help!

Thanks!

[rusty1s]

I think there exists a limitation in PyTorch+Tensorboard that model inputs cannot take dictionaries with tuples as input. One workaround would be to convert all edge types (a 3-tuple) into a string first, use this as model input, and split them immediately at the start of your forward pass. Let me know if this works.

[archids]

Hello Matthew:

Thanks for the tip. For now I have discarded incorporating Tensorboard due to deadline. But i will try your possible solution.

Thanks!