How dbis-uibk/srgnn-emo works
SRGNN-EMO predicts nine music-emotion targets from track features and a multi-relational track graph. The workflow prepares session-, tag-, and genre-based graph relations, loads graph/features/labels, then trains and cross-validates a semi-supervised graph model and reports metrics. The implementation separates graph construction, orchestration, and model layers. Some trainer and graph-builder excerpts are partial; edges are limited to shown calls or README statements, so unsampled internals are not inferred.
Main entry points: main.py Experiment entrypoint · trainer.py Fold trainer · shgr.py Semi-supervised SHGR · encoder.py Weighted graph encoder
Components 15
Graph preparation
- Relation graph builder
preprocessing/build_graphs.py - Multi-relation graph
Experiment runtime
- Experiment entrypoint
main.py - Representation settings
params.py - Feature and label loader
preprocessing/utils.py - Seed and device utilities
model/utils.py
Model and learning
- Semi-supervised SHGR
model/shgr.py - Weighted graph encoder
model/encoder.py - Relational convolution
model/layer.py - Emotion regressor
model/regressor.py
Training and evaluation
- Fold trainer
model/trainer.py - Graph augmentations
model/transforms.py
Other components
- Researcher
- Music metadata
- Prepared features and labels
Connections 19
- Researcher → Experiment entrypoint: runs experiment
README.md - Music metadata → Relation graph builder: builds relations
README.md - Relation graph builder → Multi-relation graph: writes graph
README.md - Prepared features and labels → Feature and label loader: provides inputs
README.md - Experiment entrypoint → Representation settings: gets settings
main.py - Experiment entrypoint → Seed and device utilities: seeds runtime
main.py - Experiment entrypoint → Feature and label loader: loads features
main.py - Experiment entrypoint → Multi-relation graph: loads graph
main.py - Experiment entrypoint → Weighted graph encoder: constructs encoder
main.py - Experiment entrypoint → Emotion regressor: constructs regressor
main.py - Experiment entrypoint → Semi-supervised SHGR: assembles model
main.py - Experiment entrypoint → Fold trainer: starts folds
main.py - Weighted graph encoder → Relational convolution: uses convolution
model/encoder.py - Fold trainer → Semi-supervised SHGR: trains and evaluates
model/trainer.py - Fold trainer → Weighted graph encoder: computes representations
model/trainer.py - Fold trainer → Emotion regressor: predicts targets
model/trainer.py - Fold trainer → Graph augmentations: configures graph drops
model/trainer.py - Semi-supervised SHGR → Weighted graph encoder: owns encoder
model/shgr.py - Semi-supervised SHGR → Emotion regressor: owns regressor
model/shgr.py