search

cedricvincentcuaz / TFGW

Code for Template based Graph Neural Network with Optimal Transport Distances (NeurIPS 2022)
MIT License
9 stars 0 forks source link

This is a very interesting job, is it possible to make the code public. #1

Open guangmingjian opened 1 year ago

guangmingjian commented 1 year ago

I'm curious and unsure about the specific implementation of the thesis part. For example: how to connect the POT library with pytorch without affecting backpropagation.

cedricvincentcuaz commented 1 year ago

Hello @guangmingjian

I updated the readme to mention that we are currently working on a factorised implementation with a pipeline compatible with the user-defined GNN pre-processing architecture, which will be released in a few days. You can find an exemple with GIN pre-processing, as set by default in the paper, in the supplementary material shared during the reviewing process. Link: vanilla implementation. The code to reproduce experiments is there too, except those with GAT architectures.

We are also working on speeding up FGW solvers implemented in POT. This will allow a 20% to 50% gain in speed over graphs from 10 to 1000 nodes, compared to POT 0.8.2. These new solvers will be released soon in POT 0.8.3 . Basically, the FGW2 function of POT manually sets the gradients to allow back-propagation through the FGW loss.

Best regards :)