sage.data.noise.recolour

Filename : recolour.py Description : Short description of the file

Created on 2026-02-09 23:39:57

__author__ = Narenraju Nagarajan __copyright__ = Copyright 2026, ProjectName __license__ = MIT Licence __version__ = 0.0.1 __maintainer__ = Narenraju Nagarajan __affiliation__ = N/A __email__ = N/A __status__ = [‘inProgress’, ‘Archived’, ‘inUsage’, ‘Debugging’]

GitHub Repository: NULL

Documentation: NULL

Classes

RecolourPostprocess

GPU postprocessing step: stochastic PSD recolouring from one noise epoch

Module Contents

class RecolourPostprocess(*, p_recolour, recolour_dataset_dir, eps=1e-38)[source]

Bases: torch.nn.Module

GPU postprocessing step: stochastic PSD recolouring from one noise epoch to another, operating entirely in the frequency domain.

Motivation

Sage trains on O3b noise but evaluates on O3a noise (different GPS epoch, different spectral shape). Simply whitening with O3b PSDs and testing on O3a produces a distribution shift. With p_recolour probability, each training sample is:

Whitened using the segment’s own O3b ASD (removing O3b colour).
Recoloured by multiplying with a randomly chosen O3a ASD (adding O3a colour).

The remaining 1 - p_recolour fraction of the batch passes through unchanged (plain FD conversion only).

This bridges the spectral gap between training and evaluation epochs without using any actual O3a time-domain data during training. Note that glitch morphology is not altered — only the spectral amplitude envelope changes.

param p_recolour:: Per-sample probability of applying the whiten + recolour transform. Typical value: 0.37.
type p_recolour:: float in [0, 1]
param recolour_dataset_dir:: Root directory of the target noise epoch dataset (e.g. the O3a data release directory). Must contain a data_dir/recolour_psds/ sub-directory with pre-computed per-detector ASD banks.
type recolour_dataset_dir:: str
param eps:: Small value added to ASDs before division/multiplication to prevent division by zero in very quiet frequency bins.
type eps:: float
param Inputs / Outputs:
param —————-:
param forward(batch_td:: batch_td : (B, D, T) float32 — time-domain noise windows. segment_ids: (B, D) int64 — index into the segment ASD bank (used to select the correct per-segment whitening ASD).
param segment_ids):: batch_td : (B, D, T) float32 — time-domain noise windows. segment_ids: (B, D) int64 — index into the segment ASD bank (used to select the correct per-segment whitening ASD).
param Returns ``(B:
param D:
param F)`` complex64 — frequency-domain (recoloured) strain.:
param Initialize internal Module state:
param shared by both nn.Module and ScriptModule.: