search

HuguesTHOMAS / KPConv-PyTorch

Kernel Point Convolution implemented in PyTorch
MIT License
762 stars 152 forks source link

Batch clone #223

Open WesamOM opened 1 year ago

WesamOM commented 1 year ago

First of all thank you so much for this amazing program! I am trying to run DALES dataset using the code from train_S3DIS.py, I was following steps from #72 but I reached an issue. while running the program, I am having trouble with function def forward(self, batch, config): Specifically the line: x = batch.features.clone().detach() The batch created from the code using DALES dataset creates a batch with these info:

Screenshot 2023-06-27 173419

where features is not formatted the same way in S3DIS dataset. Then I get an error that is originating from blocks.py line 361:

x = torch.cat((x, torch.zeros_like(x[:1, :])), 0)

I get this error:

x = torch.cat((x, torch.zeros_like(x[:1, :])), 0)
IndexError: too many indices for tensor of dimension 1

Now I am not sure why I am getting features to be a different tensor than when I run it on S3DIS dataset. If you can help me @HuguesTHOMAS I'd really appreciate it.

HuguesTHOMAS commented 1 year ago

Can you make sure your feature matrix batch.features has 2 dimension (N, D) where N is the number of points and D the dimension of the features.

In your case D should be equal to 1, but the feature dimension should still be (N,1) and not (N,)