gsnn.optim

Classes

TrainingDiagnostics(model[, track_every, ...])

Comprehensive training diagnostics for monitoring model optimization.

class gsnn.optim.TrainingDiagnostics(model: torch.nn.Module, track_every: int = 10, window_size: int = 100, track_activations: bool = True, track_weights: bool = True, track_gradients: bool = True, track_curvature: bool = False, verbose: bool = True)

Bases: object

Comprehensive training diagnostics for monitoring model optimization.

Tracks gradient flow, activation patterns, weight spectral properties, and basic curvature metrics to help diagnose training issues and guide optimization decisions.

Usage:

diagnostics = TrainingDiagnostics(model, track_every=10)

# During training loop: loss.backward() diagnostics.update(model, loss.item(), batch_idx) optimizer.step()

# Generate reports: summary = diagnostics.get_summary() diagnostics.plot_diagnostics()

get_gradient_histogram_data() → Dict[str, numpy.ndarray]

Get gradient histograms for current step.

Returns:

Dictionary mapping layer names to gradient histograms

get_summary() → Dict[str, Any]

Generate a comprehensive diagnostic summary.

Returns:

Dictionary containing current diagnostic state and recommendations

plot_diagnostics(save_path: Optional[str] = None, figsize: Tuple[int, int] = (15, 12))

Generate comprehensive diagnostic plots.

Parameters:

• save_path – Optional path to save the plot

• figsize – Figure size tuple

reset()

Reset all diagnostic tracking.

update(model: torch.nn.Module, loss: float, step: Optional[int] = None)

Update diagnostics with current training state.

Parameters:

• model – Current model state

• loss – Current loss value

• step – Optional step counter (uses internal if None)

Modules

gsnn.optim.EarlyStopper
gsnn.optim.Environment
gsnn.optim.FunctionEdgeInferer	Post-hoc inference of latent function -> function edges in a trained GSNN.
gsnn.optim.GradDiagnostics
gsnn.optim.MagnitudeEdgeInferer	Tier-0 edge inference via magnitude correlation between node activations and node gradients.
gsnn.optim.MagnitudeEdgeKGE	Post-hoc Tier-0 edge inference via node2vec on the augmented function graph.
gsnn.optim.MagnitudeEdgeRegressor	Online Tier-0 edge inference via auxiliary linear regression during GSNN training.
gsnn.optim.OutputEdgeInferer	Lightweight optimizer to infer output edges from intermediate GSNN node activations.
gsnn.optim.REINFORCE
gsnn.optim.RewardScaler
gsnn.optim.TrainingDiagnostics(model[, ...])	Comprehensive training diagnostics for monitoring model optimization.
gsnn.optim.utils