Thomas Kipf
Thomas Kipf
The block-diagonal representation suggested in https://github.com/tkipf/gcn/issues/4 works with arbitrarily sized graphs (each instance can be of different size). There's a PyTorch framework that implemented this strategy for GCNs (they call...
The closest to this (in TensorFlow) is the Graph Nets library: https://github.com/deepmind/graph_nets -- but even this library doesn't provide an out-of-the-box solution for your case and will require some tinkering....
Sounds like underfitting to me if the training loss is ~5 times higher than with using the MLP? Maybe try some other GNN architectures, like in this paper: https://arxiv.org/abs/1802.04687 On...
Yes if your graph is fully-connected then you obtain the same feature representation for each node, which will not work if you want to classify nodes differently. On Wed, Aug...
If you have a fully-connected graph, then you can simply use the Transformer architecture from https://arxiv.org/abs/1706.03762 . No need to use Graph Attention Networks (which are just Transformers with masked-out...
You can find a TensorFlow implementation of Transformer-style self-attention here: https://github.com/deepmind/graph_nets/blob/master/graph_nets/modules.py#L593 On Wed, Aug 14, 2019 at 12:01 PM Thomas Kipf wrote: > If you have a fully-connected graph, then...
Hard to say. Have you tried increasing model capacity (more hidden units) and disable regularizers to see if you can at least overfit on the training set? If you can’t,...
No changes required w.r.t. multiplication with the weight matrix: the weight matrix acts independently on each node feature vector. On Thu, Nov 28, 2019 at 2:05 PM JanoschMenke wrote: >...
This implementation does not use mini-batches and hence one batch constitutes the full dataset. Hence 1 iteration = 1 epoch. On Fri, Dec 6, 2019 at 10:23 AM YSN1011 wrote:...
Have a look here: https://github.com/tkipf/gcn/issues/12 On Tue 30. Oct 2018 at 16:50 Riccardo Bucco wrote: > Hi Thomas, > I'm trying to reproduce the karate club example of your paper...