v1.0.1: Config Refactor

This release patches the previous major release with updates on benchmark configurations.

pyg_spectral:

• Refactor model configuration with parameter list and ranges.
• Enhance loss configuration for more scenarios.
• Support importing some PyG models/layers for comparison purposes.

benchmark:

• Refactor data loading and logging configuration. Now the pipeline is more compact and extensible.
• Switch to stratified data splitting by default to reduce the effect of imbalanced label splits.
• Enhance parameter tuning scheme to reduce overfitting on tuning splits.

Note that the evaluation results may be different from those under v1.0.0 (and thus the released version of paper) due to configuration changes.

Full Changelog: v1.0.0-beta...v1.0.1-beta

v1.0.0: public beta

🎉 We are excited to mark this release as the first public version for our spectral GNN framework! This release solidifies the code arrangement of:

• benchmark/: codes for benchmark experiments.
• pyg_spectral/: core codes for spectral GNNs designs.
  • nn.conv: spectral spectral filters.
  • nn.models: common neural network architectures.

Currently, the framework supports two training schemes:

• Full-batch: available to Iterative and Decoupled models
• Mini-batch: available to Precomputed models

Milestone May 23

• Refactor model/conv layer division
• New filters
• Experiments to produce results

Milestone April 4

Basic pipeline, filter implementation, degree-specific evaluation