Training

Data preparation

Download the training dataset (DeepExtractor reconstructions of seven LIGO O3 glitch classes: Blip, Fast Scattering, Koi Fish, Low Frequency Burst, Scattered Light, Tomte, and Whistle) directly from HuggingFace using the built-in helper:

from glitchgan import download_data

paths = download_data("data/")
# paths["samples"]     → data/glitch_GAN_samples_scaled_balanced.npy
# paths["labels"]      → data/glitch_GAN_labels_balanced.npy
# paths["label_order"] → data/glitch_GAN_label_order.npy

To also download the time-derivative array required for cDVGAN training (~2.1 GB extra):

paths = download_data("data/", include_derivatives=True)
# paths["derivatives"] → data/glitch_GAN_deriv_samples_balanced.npy

The dataset is hosted at tomdooney/deepextractor-glitch-reconstructions on HuggingFace (35,000 samples, 7 classes, 8192 samples at 4096 Hz).

Expected directory layout after download:

data/
├── glitch_GAN_samples_scaled_balanced.npy   # (35000, 8192) whitened waveforms
├── glitch_GAN_labels_balanced.npy            # (35000, 7)   one-hot class labels
├── glitch_GAN_label_order.npy                # (7,)         class name order
└── glitch_GAN_deriv_samples_balanced.npy     # (35000, 8191) derivatives (optional)

Training a model

glitchgan-train \
    --variant cDVGAN \
    --data-dir data/ \
    --epochs 500 \
    --output-dir outputs/

Available model variants

Variant	Description
cWGAN	Conditional Wasserstein GAN with gradient penalty (single discriminator)
cDVGAN	Dual-discriminator cWGAN with derivative discriminator (recommended)
cDVGAN2	Extended cDVGAN with additional second-derivative discriminator

Python API

from glitchgan.tf import build_gan, train_gan, GlitchDataset
import numpy as np

X = np.load("data/glitch_GAN_samples_scaled_balanced.npy")
y = np.load("data/glitch_GAN_labels_balanced.npy")

dataset = GlitchDataset(X, y, batch_size=64)
gan     = build_gan("cDVGAN", noise_dim=100, num_classes=7, signal_length=8192)

train_gan(gan, dataset, epochs=500, checkpoint_dir="checkpoints/")

Checkpointing

Weights are saved every 10 epochs to checkpoint_dir/. Training can be resumed by pointing --output-dir at an existing checkpoint directory.

Run glitchgan-train --help for the full list of arguments.