SciTeX NN (scitex-nn)

PyTorch neural-network building blocks for signal processing — BNet, Hilbert, PAC, Wavelet, Filters, AxiswiseDropout, and more.

Full Documentation · uv pip install scitex-nn[all]

Problem and Solution

Installation

Requires Python >= 3.9.

pip install scitex-nn

2 Interfaces

import scitex_nn as nn

# Differentiable Hilbert transform
hilbert = nn.Hilbert(seq_len=512, dim=-1)

# Bandpass filter bank
filt = nn.Filters(...)

# Axis-wise dropout (drop entire channels/features)
drop = nn.AxiswiseDropout(dropout_prob=0.5, dim=1)

# Phase-amplitude coupling
pac = nn.PAC(...)

# Reference architectures
model = nn.BNet(...)

Gallery

Each notebook is self-contained — clone, open, run cell-by-cell. Cell outputs are baked in (jupyter nbconvert --execute --inplace) so every figure renders inline on GitHub. Ordered simple → complex.

scitex-dev skills export --package scitex-nn  # Export to Claude Code

Demo

The shortest end-to-end demo: differentiable Hilbert envelope on a multi-channel signal, axis-wise dropout for SSL pre-training.

import torch
import scitex_nn

x = torch.randn(8, 19, 1024)              # (batch, channels, samples)

env = scitex_nn.Hilbert(seq_len=1024)(x)  # analytic signal: (..., 2)
phase, amplitude = env[..., 0], env[..., 1]

drop = scitex_nn.AxiswiseDropout(dropout_prob=0.5, dim=1).train()
y = drop(x)                               # whole channels zeroed

flowchart LR
    raw["raw signal<br/>(B, C, T)"]
    aug["aug<br/>(DropoutChannels,<br/>FreqGainChanger,<br/>...)"]
    filt["filter<br/>(BandPassFilter,<br/>GaussianFilter,<br/>...)"]
    spec["spectral<br/>(Hilbert,<br/>Spectrogram,<br/>Wavelet, PSD)"]
    coup["coupling<br/>(ModulationIndex,<br/>PAC)"]
    arch["architecture<br/>(ResNet1D,<br/>MNet1000, BNet)"]
    raw --> aug --> filt --> spec --> coup
    raw --> arch
    spec --> arch

For tutorial-style runnable examples covering every public class, see the Gallery below — each is a self-contained examples/<NN>_*.ipynb whose cell outputs render inline on GitHub. examples/00_run_all.sh re-executes every notebook in place.

Architecture

scitex-nn is a flat collection of nn.Modules grouped by what they do to a (batch, channels, samples) tensor:

scitex_nn/
├── _Filters.py            # FIR-init bandpass / lowpass / highpass / bandstop
├── _GaussianFilter.py     # Gaussian smoothing (kernel = 6·sigma)
├── _Hilbert.py            # analytic-signal extraction (FFT-based)
├── _PSD.py                # power spectral density
├── _Spectrogram.py        # STFT magnitude per channel
├── _Wavelet.py            # Morlet CWT
├── _ModulationIndex.py    # Tort 2010 KL-MI
├── _PAC.py                # phase-amplitude coupling pipeline
├── _AxiswiseDropout.py    # axis-wise dropout (channel / time / feature)
├── _DropoutChannels.py    # whole-channel dropout
├── _ChannelGainChanger.py # softmax-weighted per-channel gain
├── _FreqGainChanger.py    # softmax-weighted per-band gain (julius)
├── _SwapChannels.py       # random channel permutation
├── _SpatialAttention.py   # 1×1 conv channel attention
├── _ResNet1D.py           # 1D ResNet backbone
├── _MNet_1000.py          # 4-stage Conv2d EEG/MEG classifier
├── _BNet.py / _BNet_Res.py# B-shaped multi-modality wrapper
└── _vendor_dsp_utils/     # vendored helpers (no scitex-dsp dep)

Modules compose as ordinary nn.Sequential. The signal-processing layers operate on the last (time) axis; channel-aware augmentations operate on dim=1.

Available Modules

Part of SciTeX

scitex-nn is part of SciTeX. Install via the umbrella with pip install scitex[nn] to use as scitex.nn (Python).

import scitex
import scitex_nn as nn

@scitex.session
def main(CONFIG=scitex.INJECTED):
    signal = scitex.io.load("signal.npy")
    hilbert = nn.Hilbert(seq_len=signal.shape[-1], dim=-1)
    out = hilbert(signal)
    scitex.io.save(out, "analytic.npy")
    return 0

The SciTeX system follows the Four Freedoms for Research below, inspired by the Free Software Definition:

Four Freedoms for Research

0. The freedom to run your research anywhere — your machine, your terms.
1. The freedom to study how every step works — from raw data to final manuscript.
2. The freedom to redistribute your workflows, not just your papers.
3. The freedom to modify any module and share improvements with the community.

AGPL-3.0 — because we believe research infrastructure deserves the same freedoms as the software it runs on.