csiphon

csiphon is an online-first processing library for WiFi Channel State Information. Compose reusable DSP steps, compile them against your capture setup, then pour a complete recording or stream live chunks through the same validated pipeline.

from csiphon import AcquisitionProfile, Pipeline
from csiphon.steps import GainNormalize, Magnitude, WindowedVariance

# Define the hardware setup
profile = AcquisitionProfile(
    n_rx_antennas=3,
    subcarrier_indices=tuple(range(52)),
    sampling_rate_hz=1000.0,
)

# Define the pipeline steps
siphon = (
    Pipeline()
    .then(Magnitude())
    .then(GainNormalize())
    .then(WindowedVariance(win_size_s=0.1))
    .compile(profile)
)

# Generate a summary of the pipeline
print(siphon.describe())

# Push some data through it to run
features = siphon.pour(profile.raw_signal(csi, timestamps)).single()

What it does

• Catches structural mistakes before execution. Compilation checks axes, shapes, value semantics, and physical representations before data starts flowing.
• Runs the same recipe in batch or live. Every step states whether its streaming computation matches batch, differs intentionally, or requires the complete recording.
• Supports branching pipelines. Split into parallel feature paths, merge them again, and expose named intermediate outlets.
• Keeps real timestamps. Non-uniform CSI sampling is expected; resampling is explicit rather than silently assumed.
• Describes itself. Steps and compiled siphons expose their contracts, layouts, parameters, and streaming behavior, and can save that information alongside results for reproducibility.
• Measures itself. siphon.measure(signal) pours once and reports each step's time and output shape (memory on request), per step or for a named group of consecutive steps.

The built-in steps cover calibration, cleaning, filtering, delay and time-frequency transforms, temporal features, pooling, statistics, reduction, and restructuring. Custom steps use the same contracts and inspection tools.

Install

pip install -e .

The core only depends on NumPy. Install every optional transform with:

pip install -e ".[all]"

Runnable recipes live in examples/.