temporal-filters

Learnable, parameterized temporal filters (1D/2D wavelet, 1D/2D sinc) designed as drop-in preprocessing layers for EEG deep-learning architectures such as EEGNet, SpatialNet, and FBCSP.

Overview

Fixed band-pass filter banks are a common first stage in EEG deep-learning pipelines, but their cutoffs are hand-tuned rather than learned. TemporalFilter implements wavelet- and sinc-based filters as differentiable layers with a small, interpretable parameter set — a low-frequency parameter for all filter types, plus a scale parameter for the wavelet and sinc variants — so the filter bank can be optimized end-to-end with the rest of the network.

What's inside

• A universal TemporalFilter layer implementation covering all four filter types (wavelet 1D/2D, sinc 1D/2D).
• Example integrations comparing the filters inside EEGNet-, SpatialNet-, and FBCSP-style architectures, to validate that the layer behaves as expected inside a real training loop.

Repository contents

Path	Purpose
src/temporal_filters.py	The filter layers themselves (sinc, wavelet, Hilbert, complex-wavelet — 1D and 2D)
src/reference_models.py	Reference EEG architectures (EEGNet family, DeepConvNet, ShallowConvNet) used to sanity-check the filters inside real models
EEGML_WithinSession/	Within-session EEG model experiments using the filters (carry this folder over from the old repo as-is)

Setup

pip install -r requirements.txt   # torch + numpy for src/temporal_filters.py
pip install tensorflow            # only needed for src/reference_models.py

Tech stack

Python, PyTorch, NumPy (TensorFlow/Keras additionally for the reference models)