search

dmlc / dgl

Python package built to ease deep learning on graph, on top of existing DL frameworks.
http://dgl.ai
Apache License 2.0
13.58k stars 3.02k forks source link

Functionize NN modules to allow dynamic weight matrices #3237

Open BarclayII opened 3 years ago

BarclayII commented 3 years ago

🚀 Feature

We should implement NN modules as functions to allow building models with dynamically generated weights.

Motivation

As discussed in #3190, EvolveGCN is a typical example where GCN's weight is generated by an LSTM. While dgl.nn.GraphConv does allow the user specifying the weights in the forward pass, other modules do not.

This is also an issue for PyTorch modules in general, and PyTorch only functionizes some of the modules like conv2d or prelu that are most likely needed for dynamic weights. So I guess the priority to implement this is reasonably low.

This issue bookkeeps the feature request.

Alternatives

One could copy the implementation and change the weight initialization to use the generated weights.

github-actions[bot] commented 2 years ago

This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you