sage.factory.validation

Sage validation loop.

SageVanillaValidation: Mirrors SageVanillaTraining exactly except the model runs in eval + inference_mode, and per-epoch outputs (network predictions, targets, raw parameters) are saved to an HDF5 file for offline diagnostics (ROC curves, parameter recovery, etc.).

Classes

Standard Sage validation loop with full diagnostic output saving.

save_validation(nepoch, output, target, params, ...)

Append one epoch's validation results to an HDF5 file.

save_validation(nepoch, output, target, params, signal_idx, savepath)[source]

Append one epoch’s validation results to an HDF5 file.

Creates a group epoch_{nepoch:04d} inside savepath containing four gzip-compressed datasets.

Parameters:

nepoch (int) – Epoch index (used as the HDF5 group name).
output (torch.Tensor, shape (N, 1 + 2*num_pe)) – Network outputs: ranking statistic, predicted means, predicted sigmas (in physical units).
target (torch.Tensor, shape (N, num_pe + 1)) – Ground-truth targets (regression values + class label).
params (torch.Tensor, shape (N, num_params)) – Raw waveform parameters (theta) returned by the signal sampler.
signal_idx (torch.Tensor, shape (num_iter, S)) – Batch positions where signals were injected for each iteration.
savepath (str) – Path to the HDF5 output file.

class SageVanillaValidation(signal_sampler, noise_sampler, processor, model, loss_function, num_iterations, num_epochs)[source]

Standard Sage validation loop with full diagnostic output saving.

Runs the model in eval + inference_mode.
Calls signal_sampler(return_theta=True) to obtain raw waveform parameters for parameter-recovery diagnostics.
Unstandardises predicted means back to physical units.
Converts the raw sigma params to sigma via softplus (matching the loss’s _sigma), then un-standardises to physical units.
Writes per-epoch results to {export_dir}/validation_data.h5.

Auto-multibanding and GWBatch tracking work identically to the training loop — no extra configuration needed.

Parameters:

signal_sampler – Same objects passed to the corresponding training class.
noise_sampler – Same objects passed to the corresponding training class.
processor – Same objects passed to the corresponding training class.
model – Same objects passed to the corresponding training class.
loss_function – Same objects passed to the corresponding training class.
num_iterations (int) – Number of batches per validation epoch.
num_epochs (int) – Total epochs (pre-allocates the loss_components tracking tensor).
state (Initialize internal Module)
ScriptModule. (shared by both nn.Module and)