bkj / pytorch-graphsage

Representation learning on large graphs using stochastic graph convolutions.

graphsage

pytorch version of https://github.com/williamleif/GraphSAGE

There are also a handful of new features, including:

• scripts for preprocessing data
• attention-based aggregator
• sparse edge sampler (eg, don't use the dense 2D edgelist)
• richer, pluggable preprocessing classes

This version is missing some features, as well. Namely:

• we don't support unsupervised models (yet)

A simplified version of the original Tensorflow version can be found on branch tf-simple.

LICENSE

MIT

pytorch-graphsage's Issues

Running

How do you run train.py for ppi? What file is needed for the argument "--problem-path" so I can run it?
Also for reddit because the h5 files are not working for running as well.

typo?

https://github.com/bkj/pytorch-graphsage/blob/master/models.py#L99

self.forward()?

What does the problem path mean?

I note that in your code, there is a problem path parameter. What does this mean?
And how to get this worked?

Thanks.

Initialize with links to a dummy node?

Hi,

I notice that when you create the adjacency matrix, you use the trick "Initialize w/ links to a dummy node" (see utils/convert.py function make_adjacency), however, even if I change it to a normal adjacency matrix, the code still works, could you explain why to use this trick?

Thanks!

Could you upload your requirements.txt

implement unsupervised version

Need to implement:

• method to generate context pairs
• edge sampler
• triplet loss model