scitex-nn

name

scitex-nn

description

PyTorch neural-network building blocks for neuroscience and signal processing. Differentiable filters (`BandPassFilter`, `BandStopFilter`, `HighPassFilter`, `LowPassFilter`, `GaussianFilter`, `DifferentiableBandPassFilter`) operate on 1D channels-first tensors. `Hilbert` provides differentiable Hilbert transform for analytic-signal extraction. Architecture blocks: `BNet` / `BNet_Res` (B-shaped backbone with optional residual connections), `BHead` (decoder head). Augmentation/regularization: `AxiswiseDropout`, `DropoutChannels`, `ChannelGainChanger`, `FreqGainChanger` (gain perturbation in time and frequency domains for SSL training). Drop-in replacement for hand-rolled `torch.nn.Conv1d` butterworth-init wrappers, scattered Hilbert implementations using `torch.fft`, and bespoke channel-dropout layers. Use whenever a model needs trainable filter banks, differentiable spectral features, or SSL-style time/frequency augmentation.

Interfaces: Python ⭐⭐⭐ (primary) · CLI — · MCP — · Skills ⭐⭐ · Hook — · HTTP —

scitex-nn

PyTorch building blocks specialized for neuroscience / signal-processing models — differentiable filters, Hilbert, B-shaped backbones, and spectral augmentation.

Differentiable filters

from scitex_nn import BandPassFilter, BandStopFilter, GaussianFilter

bp = BandPassFilter(low=4.0, high=8.0, fs=1000.0, order=4)
y = bp(x)                       # x: (batch, channels, samples)

DifferentiableBandPassFilter learns low/high end-to-end (use when the band of interest is itself a hyperparameter).

Hilbert transform

from scitex_nn import Hilbert
analytic = Hilbert()(x)          # x: (..., samples) → complex tensor

Backbones

from scitex_nn import BNet, BNet_Res, BNet_config_v1

cfg = BNet_config_v1(in_chans=64, out_chans=2, ...)
model = BNet(cfg)

BNet_Res adds residual connections; the same config dataclass works.

Augmentation

• AxiswiseDropout(p, axis) — drops along a chosen axis (channel, time, frequency)
• DropoutChannels(p) — convenience wrapper
• ChannelGainChanger(min_gain, max_gain) — random per-channel gain
• FreqGainChanger(...) — same in frequency domain via FFT

These compose as plain nn.Modules — drop into a nn.Sequential.

When to use

• ✅ Trainable / differentiable signal processing inside a model
• ✅ SSL-style time / frequency augmentation pipelines
• ✅ Replacing tens of lines of hand-rolled FIR/Butterworth init
• ❌ Classical (non-trainable) filtering — use scipy.signal or scitex-dsp

See also

• scitex-dsp — non-trainable counterparts (numpy/scipy-backed)
• General skill 01_arch_06_local-state-directories.md if model checkpoints / cache directories need a canonical location