Training on sparse tensors

data
1

I want to train my neural network on sparse tensors that I’ve made by using:
pytorch sparse tensors
But while trying to propagate my sparse tensors into the CNN I get this error:
RuntimeError: Input type (torch.cuda.sparse.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same

What should I change in my network in order to be able to train on sparse tensors?

Thank you

1 Like
2

I read and learn a lot from your comments here @ptrblck, maybe you could help me here?

Thank you!

3

Hello @nirzaa,

We currently don’t have support for convolutions for sparse inputs. Could I kindly ask you to open a feature request on https://github.com/pytorch/pytorch with your intended use case (model, application area, dataset)? Also, please include details on the intended semantics if you want the input to be treated differently if it has a sparse layout vs. a regular strided layout.

Thanks,
Christian

5

the link to the issue I opened on github is here:
link to the issue on github

Thank you very much

6

This library from facebookresearch at least at first glance seems to implement the desired features for basic CNNs. GitHub - facebookresearch/SparseConvNet: Submanifold sparse convolutional networks