Training¶

Glitch reconstruction dataset¶

DeepExtractor reconstructions of seven LIGO O3 glitch classes (35,000 samples) are available for download from HuggingFace: tomdooney/deepextractor-glitch-reconstructions

These reconstructions are suitable for training generative models such as GlitchGAN and for any downstream task requiring high-quality time-domain glitch waveforms.

from deepextractor.data import download_glitch_data

paths = download_glitch_data("data/glitches/")
# paths["samples"]     → (35000, 8192) whitened waveforms
# paths["labels"]      → (35000, 7)   one-hot class labels
# paths["label_order"] → ['Blip', 'Fast_Scattering', 'Koi_Fish',
#                          'Low_Frequency_Burst', 'Scattered_Light', 'Tomte', 'Whistle']

To also download the first-order time-derivative array (required for cDVGAN training):

paths = download_glitch_data("data/glitches/", include_derivatives=True)
# paths["derivatives"] → (35000, 8191)

Data preparation¶

Generate time-domain data:

deepextractor-generate --output-dir data/ --num-train 250000 --num-val 25000

# Or with bilby noise:
deepextractor-generate --output-dir data/ --num-train 250000 --bilby-noise

Convert to spectrograms (for 2D models):

deepextractor-specgen \
    --input-dir data/pycbc_noise/time_domain/ \
    --output-dir data/pycbc_noise/spectrogram_domain/

Expected directory layout after data generation:

data/
└── pycbc_noise/
    ├── time_domain/
    │   ├── glitch_train_scaled_pycbc.npy
    │   ├── background_train_scaled_pycbc.npy
    │   ├── glitch_val_scaled_pycbc.npy
    │   └── background_val_scaled_pycbc.npy
    └── spectrogram_domain/
        ├── glitch_train_scaled_mag_phase.npy
        ├── background_train_scaled_mag_phase.npy
        ├── glitch_val_scaled_mag_phase.npy
        └── background_val_scaled_mag_phase.npy

Training a model¶

deepextractor-train \
    --model DeepExtractor_257 \
    --data-dir data/pycbc_noise/spectrogram_domain/ \
    --checkpoint-dir checkpoints/ \
    --batch-size 32 \
    --epochs 150

Available models¶

Model name	Architecture	Domain
`DeepExtractor_257`	UNET2D (257×257 spectrograms)	Spectrogram
`DeepExtractor_129`	UNET2D (129×129 spectrograms)	Spectrogram
`UNET1D`	1D U-Net	Time-domain
`DnCNN1D`	1D DnCNN	Time-domain
`Autoencoder1D`	1D Autoencoder	Time-domain

Hyperparameter options¶

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