Graph Laplacian 🌐

[KristinaUlicna]

PR contribution summary

Why is this PR useful / good for? Please describe the problem(s) you're trying to address.
2 intertwined problems:

1. No node features to train the classifier on: ▶️ feature_extractor.py

• After re-consideration, compute the node features on the spot at runtime:
  • If config.extractor_fn is None, uses a simple image descriptor (mean, sum, max, min) of the underlaying patch to generate features. Super fast, preferable at the moment.
  • If config.extractor_fn is provided (must be called /path/to/resnetXX.pt), use it to extract image patch embeddings
  • E.g. ResNet18 uses as little as 6-8 seconds per image to compute the features, speeding up the processing.

2. Neighbourhood-level representation for each node: ▶️ graph_laplacian.py

• Learning performed on a vector twice as long as feature vector:
  • Appends another node feature representation to the graph to make node embeddings aware of their neighbourhood.
  • Calculates Laplacian of a graph, and multiplies the NODE_FEATURES matrix by the Laplacian matrix, generating NODE_EMBEDDINGS (see grace/base.py)
  • At dataset creation, both of these vectors are concatenated to create Data.x attribute (see datasets.py)

Note: No need to check the rest of the code.

List of proposed changes / linked issues & discussions

• ✅ Resolves [DEVELOPMENT] Allow feature / property extraction to be computed at run-time #297
• ✅ Resolves [DEVELOPMENT] Concatenate Laplacian-multiplied node features #330

What should a reviewer concentrate their feedback on?

• 🏃 Scripts / notebooks to run
  • run.py with config.extractor_fn as None
  • and/or config.extractor_fn called /path/to/resnet18.pt (or any resnet)
• 💻 Code quality
• No need to focus on the rest of the code.

What type of PR is this? (check all applicable)

• 🪄 Feature
• 🐛 Bug fix
• 🧑‍💻 Code refactor / style
• 🔥 Performance Improvements
• ✅ Test (no need to check)

Added tests?

• 🙋 no, because I need some help

[marjanfamili]

@KristinaUlicna explained all the changes to me. and I have inspected the code. Also ran this with the attached config file (from @KristinaUlicna ) both using the resent feature extractor and without which would be using the default.

The default extracts the mean and min and max and standard deviation

config_hyperparams.txt

(format has to be .yaml which can't be attached here)