scitex-nn 0.1.14

Neural network building blocks (BNet, Hilbert, PAC, Wavelet, Filters, …) — standalone module from the SciTeX ecosystem

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

#	Problem	Solution
1	Signal-processing layers are scattered across research codebases — Hilbert, PAC, Wavelet, bandpass filters	Drop-in PyTorch modules — differentiable, batched, and composable into any nn.Module
2	Standard nn.Dropout operates element-wise — no axis-wise option for channel/feature drop	AxiswiseDropout, DropoutChannels — zero out entire features along a chosen axis
3	Custom blocks (BNet, MNet, ResNet1D) must be re-implemented for every project	Vetted reference implementations with consistent APIs and shape conventions

Installation

Requires Python >= 3.9.

pip install scitex-nn

2 Interfaces

Python API

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(...)

Full API reference

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.

#	Notebook	Topic
01	examples/01_axiswise_dropout.ipynb	AxiswiseDropout — drop entire slices along an axis
02	examples/02_channel_aug.ipynb	DropoutChannels / SwapChannels / ChannelGainChanger
03	examples/03_gaussian_filter.ipynb	GaussianFilter — temporal smoothing at three sigmas
04	examples/04_filter_bank.ipynb	LowPass / HighPass / BandPass / BandStop frequency response
05	examples/05_psd.ipynb	PSD vs scipy.signal.welch on sine, two-tone, 1/f noise
06	examples/06_freq_gain_changer.ipynb	FreqGainChanger — softmax-weighted random per-band gain
07	examples/07_hilbert.ipynb	Hilbert vs scipy.signal.hilbert on a chirp + AM signal
08	examples/08_spectrogram.ipynb	Spectrogram STFT magnitude on a 5→60 Hz chirp
09	examples/09_wavelet.ipynb	Wavelet Morlet CWT — adaptive time-frequency resolution
10	examples/10_modulation_index.ipynb	ModulationIndex (Tort 2010) on coupled vs uncoupled theta-gamma
11	examples/11_pac.ipynb	PAC end-to-end comodulogram on synthetic theta-gamma
12	examples/12_differentiable_bandpass.ipynb	DifferentiableBandPassFilter — learnable band centres
13	examples/13_spatial_attention.ipynb	SpatialAttention — per-channel gain from a 1×1 conv
14	examples/14_resnet1d.ipynb	ResNet1D — tiny train loop on synthetic 1D data
15	examples/15_mnet1000.ipynb	MNet1000 forward + backward + per-parameter gradient norms
16	examples/16_bnet.ipynb	BNet_v1 2-modality forward + per-submodule parameter distribution

Skills — for AI Agent Discovery

Skills provide workflow-oriented guides that AI agents query to discover capabilities and usage patterns.

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

Category	Modules
Signal transforms	Hilbert, Wavelet, Spectrogram, PSD, Filters, GaussianFilter
Coupling / features	PAC, ModulationIndex
Dropout variants	AxiswiseDropout, DropoutChannels
Augmentation	ChannelGainChanger, FreqGainChanger, SwapChannels
Architectures	BNet, BNet_Res, MNet_1000, ResNet1D
Utilities	SpatialAttention, TransposeLayer

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.

---