felixopolka / stgcn-pytorch Goto Github PK

Hi,

I am confused with the Adjacency Matrix in your adj_mat.npy. It doesn't seem to be an adjacency matrix.

Best

The whole validation dataset (batch_size = 6,854) was passed in after a training epoch. Then it caused 'CUDA out of memory' ERROR.
This is how to fix the problem in main.py. ( The original lines are annotated. )

with torch.no_grad():

    net.eval()
    val_input = val_input.to(device=args.device)
    val_target = val_target.to(device=args.device)

    # out = net(A_wave, val_input)
    # --------------------------------
    tmp_val_losses = []
    tmp_maes = []
    for i in range(0, val_input.shape[0], batch_size):
        out = net(A_wave, val_input[i:i+batch_size, ...])
        loss = loss_criterion(out, val_target[i:i+batch_size, ...]).to(device="cpu")
        tmp_val_losses.append(np.ndarray.item(loss.detach().numpy()))
        out_unnormalized = out.detach().cpu().numpy() * stds[0] + means[0]
        target_unnormalized = val_target[i:i+batch_size, ...].detach().cpu().numpy() * stds[0] + means[0]
        mae = np.mean(np.absolute(out_unnormalized - target_unnormalized))
        tmp_maes.append(mae)
    val_loss = sum(tmp_val_losses) / len(tmp_val_losses)
    validation_losses.append(val_loss)
    mae = sum(tmp_maes) / len(tmp_maes)
    validation_maes.append(mae)
    # --------------------------------
    # val_loss = loss_criterion(out, val_target).to(device="cpu")
    # validation_losses.append(np.ndarray.item(val_loss.detach().numpy()))
    # out_unnormalized = out.detach().cpu().numpy() * stds[0] + means[0]
    # target_unnormalized = val_target.detach().cpu().numpy() * stds[0] + means[0]
    # mae = np.mean(np.absolute(out_unnormalized - target_unnormalized))
    # validation_maes.append(mae)
    out = None
    val_input = val_input.to(device="cpu")
    val_target = val_target.to(device="cpu")

print("Training loss: {}".format(training_losses[-1]))
print("Validation loss: {}".format(validation_losses[-1]))
print("Validation MAE: {}".format(validation_maes[-1]))

Is there someone who also has this problem?

Why we set number of input and output channel to 64?
Why we set number of spatial channel to 16?

We predict the traffic speed at each node, we can say it's (207 * 1). So what's the number I should use for the input and output channel if I want to predict the traffic flow from each node to each node which means (207 * 207)?

Thanks a lot :-)

X = X.transpose((1, 2, 0))
Why the number of features (observation) is 2 at a specific time for each station?

When running main.py, a runtime error occurs: one of the variables needed for gradient computation has been modified by an inplace operation. What are the possible lines that cause this issue? Is it line 83 (torch.einsum) in stgcn.py?

Hi,

Thanks for your work,
BTW, could I use the dataset from https://github.com/Davidham3/STGCN on your work?

Hi,

I found no related codes about 1st-order approximation in your product. Does it mean you only provide the Chebyshev approx in your code?

Best

What features mean in node_values.npy?

In the tensorflow implementation, the temporal_conv_layer result is the product of conv and sigmoid.
“ return (x_conv[:, :, :, 0:c_out] + x_input) * tf.nn.sigmoid(x_conv[:, :, :, -c_out:]) ”

could you explain why the add operation is used here?

temp = self.conv1(X) + torch.sigmoid(self.conv2(X))
out = F.relu(temp + self.conv3(X))

In the original paper, the function of temporal block is P*σ(Q), but in this implementation in PyTorch, I can only find 3 summation, is there anything wrong or it's just my mis-understanding in this paper?

Hi author,
Thank you for your STGCN of PyTorch version. I found that the matrix in adj_mat.npy is not symmetric, but the matrix should be Wij=exp(-dij^2/sigma^2) (if Wij≠0). So why is that not symmetric?
Best wishes