DeepPeak 0.1.0

Key Features

• Classical peak detection: Height, sigma, prominence, zero-crossing, and non-maximum-suppression methods.

• Optional neural deconvolution: CNN, WaveNet, and 1D U-Net models can reconstruct a cleaner pulse signal before peak detection.

• Synthetic data generation: Gaussian, Lorentzian, square, Dirac, custom, and two-lobe kernels with configurable noise and peak-count models.

• Trace analysis: Arrival-time, amplitude, width, pulse-shape, noise, and dead-time analysis.

• Direct-versus-deconvolved evaluation: Compare event counts, time-of-arrival, amplitude, and width distributions on the same traces.

• Dilution-series workflows: Standard and flash dilution-series analysis with detector-specific metrics and plots.

• Plotting and diagnostics: Figures are returned as Matplotlib objects so they can be customized, saved, or embedded in notebooks.

Installation

Install the released package from PyPI:

pip install DeepPeak

The classical signal-processing and generation APIs do not require TensorFlow. Install the optional neural-network stack when using the models package:

pip install "DeepPeak[ml]"

For development, install the repository and its test/documentation tools in your preferred virtual environment.

Quickstart: generate a signal

The generation API can create reproducible training and evaluation data:

from DeepPeak import Gaussian, SignalGenerator, UniformCount

generator = SignalGenerator(sequence_length=1_000)
dataset = generator.generate(
    n_samples=128,
    kernel=Gaussian(
        amplitude=(1.0, 10.0),
        position=(0.1, 0.9),
        width=(0.02, 0.05),
    ),
    peak_count=UniformCount(bounds=(1, 4)),
    seed=42,
    noise_std=0.05,
)

signals = dataset.signals

Analysis quickstart

For standard dilution-series analysis, provide trace files as (filename, dilution) pairs and run the configured workflow:

from DeepPeak.analysis import HeightPeakTrigger, StandardDilutionSeries

series = StandardDilutionSeries(
    folder="path/to/traces",
    files=[
        ("path/to/traces/trace_1.csv", 1.0),
        ("path/to/traces/trace_2.csv", 10.0),
    ],
    trigger=HeightPeakTrigger(height=0.05, hysteresis=0.03),
    initial_concentration=1.0,
    nrows=100_000,
)

result = series.run()
series.plot.standard_detection(index=0)

series.poisson.plot.expected_histogram(
    index=0,
    base_index=0,
    detector="standard",
    x_axis="time",
)

series.amplitude.plot.histogram(index=0, detector="standard")
series.width.plot.histogram(index=0, detector="standard", x_axis="time")

For reusable settings, prefer the typed configuration objects exposed by DeepPeak.core. They validate values at construction time and can be passed to analyzers, dilution-series workflows, and plotting helpers:

import numpy as np

from DeepPeak.core import DetectionConfig, PlotConfig, Trace
from DeepPeak.detection import HeightPeakTrigger
from DeepPeak.analysis import StandardTraceAnalyzer

detection_config = DetectionConfig(
    sequence_length=1_000,
    normalization="zscore",
    trigger=HeightPeakTrigger(height=0.05),
)
analyzer = StandardTraceAnalyzer(config=detection_config)
signal = np.zeros(1_000)
signal[500] = 1.0
detection = analyzer.detect(Trace(signal=signal, dx=1.0))
record = analyzer.analyze_processed_signal(signal, dx=1.0)

figure = record.plot_standard_detection(
    config=PlotConfig(show=False, close=False, dpi=150),
)

The plotting API always returns figures. Set show=False for scripts and tests, and set close=True when a figure should be closed automatically after it is created.

The detector-specific classes StandardDilutionSeries and FlashDilutionSeries are also available when a workflow should expose only one detector mode. Use FlashDilutionSeries with a trained neural model for CNN-based workflows.

Direct versus deconvolved comparison

The neural model is an optional deconvolution stage. The same peak-detection concept can therefore be evaluated directly on the raw trace and after neural deconvolution:

raw trace ────────────────> detector ──> direct result
    │
    └─ optional CNN/WaveNet/U-Net ──> detector ──> deconvolved result

Use TraceComparisonAnalyzer to compare the two branches. The result provides arrival-time, amplitude, and width distributions, together with summary statistics and distribution differences. Set deconvolver=None to run only the direct branch.

from DeepPeak.analysis import TraceComparisonAnalyzer
from DeepPeak.core import Trace
from DeepPeak.detection import HeightPeakTrigger

comparison = TraceComparisonAnalyzer(
    standard_trigger=HeightPeakTrigger(height=0.05),
    deconvolver=trained_model,  # optional CNN, WaveNet, or U-Net wrapper
).compare(Trace(signal=signal, dx=1e-9))

arrival_comparison = comparison.compare_distribution("arrival")
amplitude_comparison = comparison.compare_distribution("amplitude")
width_comparison = comparison.compare_distribution("width")

The same comparison can be run over multiple traces with compare_many. This makes it possible to quantify changes in event counts, time-of-arrival distributions, retrieved amplitudes, widths, and distribution distances.

For this workflow, the neural model must be trained as a reconstruction model: its target should be a clean or deconvolved pulse trace, not only a binary ROI mask. For example, a WaveNet model should use a linear output head and a regression loss when it is trained to predict pulse amplitudes. The existing ROI-classification examples demonstrate a related model use case, but are not deconvolution models by themselves.

Architecture

DeepPeak is being organized around clear domain boundaries:

DeepPeak/
├── core/          shared types, protocols, configuration, exceptions
├── generation/    synthetic signals, kernels, noise, datasets
├── detection/     classical and neural detection algorithms
├── models/        trainable neural-network architectures and losses
├── analysis/      trace and dilution-series workflows
├── metrics/       numerical diagnostics and distribution summaries
├── plotting/      visualization of traces, detections, and metrics
└── io/            trace loading and result serialization

The intended dependency direction is:

core
  ↓
generation / detection / models
  ↓
analysis
  ↓
metrics / plotting / io

This separation keeps numerical analysis independent from plotting and keeps TensorFlow-specific code isolated from the core signal-processing API. The domain namespaces are the supported public API for new code.

Public API guide

Use these namespaces when writing new code:

DeepPeak.analysis

Trace analyzers, dilution-series workflows, triggers, and analysis results.

DeepPeak.generation

Synthetic datasets, kernels, noise models, and peak-count models.

DeepPeak.detection

Detection algorithms, triggers, and the common detection-result type.

DeepPeak.models

DenseNet, WaveNet, UNet1D, neural losses, and model utilities.

DeepPeak.metrics

Detection, amplitude, width, arrival-time, and series metrics.

DeepPeak.plotting and DeepPeak.io

Figure helpers and trace/file loading utilities.

DeepPeak.core

Stable Trace, DetectionResult, MetricResult, and SeriesResult objects, plus typed TraceConfig, DetectionConfig, SeriesConfig, and PlotConfig settings.

The root DeepPeak namespace exposes the most common user-facing types for interactive work and notebooks.

Documentation

The full API reference, theory notes, and executable examples are available at the DeepPeak documentation.

Contact

For questions or contributions, contact martin.poinsinet.de.sivry@gmail.com.