search

akensert / molgraph

Graph neural networks for molecular machine learning. Implemented and compatible with TensorFlow and Keras.
https://molgraph.readthedocs.io/en/latest/
MIT License
48 stars 5 forks source link

Training based on the output of GradCAM? #27

Closed vcanogil closed 5 months ago

vcanogil commented 5 months ago

Is it possible using molgraph to write a model that could be trained based on the output of it's GradCAM map? I.e. a GNN that takes a molecule and adjusts its weight based on the heatmap it produces?

akensert commented 5 months ago

Hi @vcanogil, not sure how this would be done, and also why?

The saliency maps are basically gradients of the output with respect to the data (the node features). I don't see how you would then be able to obtain gradients of that with respect to the weights. Perhaps I've misunderstood?

vcanogil commented 5 months ago

Never mind, after thinking about it some, I don't think it'll make sense either