Porting Deep‐Learning Phase Pickers From TensorFlow to PyTorch for SeisBench Without Retraining
- 1Bureau of Economic Geology The University of Texas at Austin Austin TX USA
Abstract
SeisBench has become the premier open‐source library for accessing pretrained deep‐learning seismic phase pickers through a unified Python interface. Because SeisBench distributes models exclusively in PyTorch, phase pickers originally developed in TensorFlow must be converted to PyTorch before integration into the library. Although retraining is possible, doing so produces new learned weights that can alter the numerical behavior and picking characteristics of the original model while demanding substantial time and compute. We present a four‐step workflow for converting static TensorFlow/Keras phase pickers to PyTorch that preserves original model weights without retraining: (a) architectural alignment, (b) weight extraction and layout transformation, (c) explicit parameter mapping, and (d) staged validation from intermediate activations through waveform traces and phase picks. We demonstrate the workflow by converting the Earthquake Compact Convolutional Transformer (EQCCT), a production phase picker used by the Texas Seismological Network, from TensorFlow to PyTorch. Across 100,000 STEAD and TXED waveform windows, CPU and GPU mean absolute errors remained between and with TF32 disabled. Detection decisions at thresholds of 0.1, 0.3, and 0.5 were identical on both devices, while argmax locations agreed for all but one CPU P‐wave window, which differed by one sample. The converted EQCCT models are now available in the main SeisBench repository as EQCCTP and EQCCTS. This workflow serves as a reference for integrating TensorFlow‐based seismic phase pickers into SeisBench by converting their original weights into a PyTorch‐compatible format while preserving their operational behavior. Validation in this study is limited to EQCCT.
Related
- SeisBench DAS: A machine learning framework for Distributed Acoustic Sensing
- Machine-learning prediction of freezing temperature in Pb-contaminated soils remediated with biochar
- Atmospheric Predictability beyond 30 Days with Machine Learning
- Machine-learning modelling of regional landslide susceptibility in eThekwini, South Africa
- Hybrid stacked bidirectional deep learning with attention based air pollution prediction for environmental sustainability
BibTeX
@article{skevofilax2026porting,
title = {{Porting Deep‐Learning Phase Pickers From TensorFlow to PyTorch for SeisBench Without Retraining}},
author = {C. G. Skevofilax and Y. Chen and A. Savvaidis},
journal = {Earth and Space Science},
year = {2026},
month = oct,
doi = {10.1029/2026ea005507},
url = {https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EA005507},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Porting Deep‐Learning Phase Pickers From TensorFlow to PyTorch for SeisBench Without Retraining
AU - C. G. Skevofilax
AU - Y. Chen
AU - A. Savvaidis
PY - 2026
DA - 2026/10/30
JO - Earth and Space Science
DO - 10.1029/2026ea005507
UR - https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EA005507
AB - SeisBench has become the premier open‐source library for accessing pretrained deep‐learning seismic phase pickers through a unified Python interface. Because SeisBench distributes models exclusively in PyTorch, phase pickers originally developed in TensorFlow must be converted to PyTorch before integration into the library. Although retraining is possible, doing so produces new learned weights that can alter the numerical behavior and picking characteristics of the original model while demanding substantial time and compute. We present a four‐step workflow for converting static TensorFlow/Keras phase pickers to PyTorch that preserves original model weights without retraining: (a) architectural alignment, (b) weight extraction and layout transformation, (c) explicit parameter mapping, and (d) staged validation from intermediate activations through waveform traces and phase picks. We demonstrate the workflow by converting the Earthquake Compact Convolutional Transformer (EQCCT), a production phase picker used by the Texas Seismological Network, from TensorFlow to PyTorch. Across 100,000 STEAD and TXED waveform windows, CPU and GPU mean absolute errors remained between and with TF32 disabled. Detection decisions at thresholds of 0.1, 0.3, and 0.5 were identical on both devices, while argmax locations agreed for all but one CPU P‐wave window, which differed by one sample. The converted EQCCT models are now available in the main SeisBench repository as EQCCTP and EQCCTS. This workflow serves as a reference for integrating TensorFlow‐based seismic phase pickers into SeisBench by converting their original weights into a PyTorch‐compatible format while preserving their operational behavior. Validation in this study is limited to EQCCT.
SN - 2333-5084
SN - 2333-5084
ER -