How jkalogero/scalegmn works
ScaleGMN is a research codebase for learning over neural-network parameters, with experiments for INR classification, INR editing, and prediction of trained CNN generalization. The principal workflow is to configure an experiment, load parameter-space datasets, represent networks as graphs, and run ScaleGMN for task-specific training and evaluation; INR editing additionally applies predicted parameter updates to render images. Phase canonicalization is a separate documented preprocessing workflow. Edges are limited to excerpted imports/calls and README statements; unsampled implementation details are not inferred.
Main entry points: inr_editing.py INR editing · inr_classification.py INR classification · models.py ScaleGMN models · predicting_generalization.py Generalization prediction
Components 22
Experiment workflows
- Experiment configs
configs - INR classification
inr_classification.py - INR editing
inr_editing.py - Generalization prediction
predicting_generalization.py
Data processing
- Phase canonicalization
src/phase_canonicalization/canonicalization.py - Phase shift parameters
src/phase_canonicalization/shift_params.py - Dataset factory
src/data/__init__.py - INR datasets
src/data - CNN parameter datasets
src/data/cifar10_dataset.py - Dataset graph encoding
src/data/base_datasets.py - Graph data utilities
src/data/data_utils.py
ScaleGMN model
- ScaleGMN models
src/scalegmn/models.py - Graph initialization
src/scalegmn/graph_init.py - GNN message passing
src/scalegmn/gnn_layers.py - Equivariant layers
src/scalegmn/layers.py - MLP blocks
src/scalegmn/mlp.py - SIREN runtime
src/scalegmn/inr.py
Training components
- Loss functions
src/utils/loss.py - Optimizer setup
src/utils/optim.py - CLI arguments
src/utils/setup_arg_parser.py - Experiment helpers
src/utils/helpers.py
Other components
- Experimenter
Connections 39
- Experimenter → Experiment configs: selects
README.md - Experiment configs → INR classification: configures
README.md - Experiment configs → INR editing: configures
README.md - Experiment configs → Generalization prediction: configures
README.md - Experimenter → Phase canonicalization: runs preprocessing
README.md - Phase canonicalization → Phase shift parameters: uses
src/phase_canonicalization/canonicalization.py - Phase canonicalization → Dataset factory: loads data
src/phase_canonicalization/canonicalization.py - Phase canonicalization → Loss functions: uses
src/phase_canonicalization/canonicalization.py - INR classification → Dataset factory: loads splits
inr_classification.py - INR editing → Dataset factory: loads splits
inr_editing.py - INR classification → ScaleGMN models: runs model
inr_classification.py - INR editing → ScaleGMN models: runs model
inr_editing.py - Generalization prediction → ScaleGMN models: uses model
predicting_generalization.py - INR classification → Loss functions: selects criterion
inr_classification.py - INR editing → Loss functions: selects criterion
inr_editing.py - Generalization prediction → Loss functions: uses loss
predicting_generalization.py - INR classification → Optimizer setup: sets optimization
inr_classification.py - INR editing → Optimizer setup: sets optimization
inr_editing.py - Generalization prediction → Optimizer setup: sets optimization
predicting_generalization.py - INR classification → Experiment helpers: uses helpers
inr_classification.py - INR editing → Experiment helpers: uses helpers
inr_editing.py - Generalization prediction → Experiment helpers: uses helpers
predicting_generalization.py - INR classification → CLI arguments: parses arguments
inr_classification.py - INR editing → CLI arguments: parses arguments
inr_editing.py - Generalization prediction → CLI arguments: parses arguments
predicting_generalization.py - Dataset factory → INR datasets: dispatches
src/data/__init__.py - Dataset factory → CNN parameter datasets: dispatches
src/data/__init__.py - INR datasets → Dataset graph encoding: inherits
src/data/mnist_inr_dataset.py - CNN parameter datasets → Graph data utilities: uses
src/data/cifar10_dataset.py - Dataset graph encoding → Graph data utilities: encodes graphs
src/data/base_datasets.py - ScaleGMN models → Graph initialization: initializes graph
src/scalegmn/models.py - ScaleGMN models → GNN message passing: constructs GNN
src/scalegmn/models.py - ScaleGMN models → Equivariant layers: constructs layers
src/scalegmn/models.py - ScaleGMN models → Graph data utilities: uses graph utilities
src/scalegmn/models.py - GNN message passing → Equivariant layers: uses blocks
src/scalegmn/gnn_layers.py - Equivariant layers → MLP blocks: builds MLPs
src/scalegmn/layers.py - Equivariant layers → SIREN runtime: uses Sine
src/scalegmn/layers.py - INR editing → SIREN runtime: renders INRs
inr_editing.py - Experiment helpers → CLI arguments: uses argument helper
src/utils/helpers.py