TokEye

TokEye is an open-source Python-based application for automatic classification and localization of fluctuating signals. It is designed to be used in the context of plasma physics, but can be used for any type of fluctuating signal.

Check out this preprint for more information.

Example Demonstration

Expected processing time:

• V100: < 0.5 seconds on any size spectrogram after warmup.
• CPU: ~5-10 seconds.

Quickstart

pip install "tokeye[app]"   # web app + CLI   (or: uv tool install "tokeye[app]")
tokeye app                  # opens web app on http://localhost:7860

• The default model downloads automatically from Hugging Face on first use (~30 MB).
• No data on hand? Click "Load Example Signal" in the app, or generate one from the shell with tokeye example.
• Runs on Linux, macOS (Apple silicon) and Windows with Python >= 3.11; a laptop CPU is enough. uvx/uv tool install fetch a compatible Python automatically.
• Something off? tokeye info shows the versions, the device TokEye will use, and which models are cached; it exits 1 if a check fails.

Zero-install trial: uvx "tokeye[app]" app runs the app without installing anything into your environment (on Linux without a GPU, add --torch-backend=cpu; see below).

Install variants

Heavy dependencies are split into extras so you download only what you use:

Training is research code that runs from a clone (uv sync --group train); it is not part of the installed package.

GPU vs CPU PyTorch. A plain install pulls the default PyTorch wheel. On macOS and Windows that is already small; on Linux it is the CUDA (GPU) build (~2.5 GB). On a Linux machine without a GPU, install the CPU wheel (~200 MB) instead:

# pip: install CPU torch first, then tokeye
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install "tokeye[app]"

# uv
uv pip install "tokeye[app]" --torch-backend=cpu
uv tool install "tokeye[app]" --torch-backend=cpu
uvx --torch-backend=cpu "tokeye[app]" app

Python API

To use TokEye inside your own program, import the TokEye class:

import numpy as np
from tokeye import TokEye

eye = TokEye()  # loads the default model (auto-downloads on first use)

mask = eye(signal)             # 1D time series → STFT → inference
mask = eye(spectrogram)        # 2D spectrogram → inference directly
coherent, transient = mask     # (2, H, W) sigmoid scores in [0, 1]

seg = eye.segment(signal, fs=200_000)   # mask + spectrogram + axes
seg["coherent"], seg.freqs, seg.times   # Hz and seconds when fs is known
seg.save("shot.npz")                    # reload with tokeye.Segmentation.load

From a file to a labelled plot, tokeye.io.load_signal reads the same formats as tokeye run (.h5 needs tokeye[hdf5]) and returns the array with the sampling rate the file records (or None):

from tokeye import TokEye
from tokeye.io import load_signal

x, fs = load_signal("shot.h5", key="data")  # key only when the file holds several arrays
seg = TokEye().segment(x, fs=fs)
ax = seg.plot()                             # matplotlib Axes: spectrogram + mask, in Hz and s
ax.figure.savefig("shot.png")

Input is auto-detected by shape: a 1D array is treated as a raw time series (TokEye computes the spectrogram), a 2D array as a ready spectrogram. eye(...) and eye.predict(...) return (C, H, W), or (H, W) for a single-channel model (as before 1.0); eye.segment(...).mask is always (C, H, W). Standardization happens internally, and long inputs run in tiles so memory stays bounded. With big_tf_unet, tiled output matches an untiled run to float32 rounding; other models may differ slightly, and TokEye warns when it can tell.

If your 2D spectrogram is stored in linear scale (raw STFT magnitude/power), pass log=True so TokEye applies log1p first since the model expects log-scaled input:

mask = eye(linear_spectrogram, log=True)      # per call
eye = TokEye(log=True)                        # or for every call

log is off by default and ignored for 1D inputs (the STFT already log-scales). Everything is configurable through the constructor, but the defaults just work:

eye = TokEye(
    model="big_tf_unet",   # registry name or path to a local .pt/.pt2
    device="auto",         # "cpu", "cuda", "mps", or "auto" (CUDA, then MPS, then CPU)
    n_fft=1024, hop=128,   # STFT settings (1D inputs only)
    clip_dc=True, clip_low=1.0, clip_high=99.0,
    log=False,             # log1p for linear-scale 2D spectrograms
)

from tokeye import SpectrogramConfig
config = SpectrogramConfig(n_fft=1024, hop=128, clip_dc=True,
                           clip_low=1.0, clip_high=99.0, log=False)
eye = TokEye(config=config)  # the same settings, as one object

Batch processing (CLI)

For headless / scripted use (no browser needed), run inference directly (tokeye, or python -m tokeye when the script is not on your PATH). For example:

tokeye run "files/*.npy" --output-dir results

INPUT arguments can be files (.npy, .npz, .wav, .flac, .ogg, .csv, .txt, .mat, .h5), directories (every supported file inside except .csv/.txt), or quoted glob patterns. Glob patterns match supported files only. Outputs are named after each input's file stem, so inputs that share a stem, ignoring case (shot.npy and shot.wav, or Shot.npy and shot.npy), are rejected; run them into separate --output-dirs. The sampling rate is read from the file when it records one (or from a _sr<fs> filename suffix); --fs sets it explicitly. In .npz, .mat and .h5/.hdf5 files, TokEye reads the array named data, signal, x, spectrogram, values or y, or else the only numeric array. When several qualify, it stops and lists them; --key NAME picks one. A suffix of whole HDF5 path components, such as tree/pointname/data, selects the same array in every file. Quote the key in the shell when it contains a backslash. A 2-column .csv/.txt is read as (time, value), with time in seconds unless its header says ms or us. Each input is interpreted by its shape:

• 1D array — a raw time series. TokEye computes its STFT spectrogram using the flags below before running inference. Frames are centred on samples 0, hop, 2·hop, … (the model's training framing), so N samples give 1 + N // hop columns (for the default, even, n_fft) and column j is at j·hop/fs s.
• 2D array — a precomputed spectrogram, fed to the model directly. Its columns are assumed to follow the same rule.

For each input file, tokeye run writes:

• <stem>_mask.npy — float32 sigmoid scores per pixel, shape (C, H, W); (2, H, W) for the default model (channel 0 = coherent, channel 1 = transient).
• <stem>_preview.png — a grayscale spectrogram with the mask overlaid (green = coherent, red = transient), unless --no-png is passed.
• <stem>_params.json — the model, version, device, input, key, fs, preprocessing settings and tiling used (tile as requested; tile_shape as [h, w], or null when the run was untiled), and the weights used (weights: repo, file name, revision and sha256 for a registry model; file name and sha256 for a local checkpoint, which model names by its file name only).
• With --format npz, <stem>_tokeye.npz (mask + spectrogram + axes, loadable with tokeye.Segmentation.load) replaces the .npy mask.

The run ends with one line on stdout, such as wrote 3 inputs -> tokeye_output/ (mask, preview, params), and shows a progress bar only on a terminal. Exit codes: 0 = every input succeeded, 1 = at least one input failed, 2 = usage or configuration error (bad flag, unknown model, no inputs, model cannot be loaded), 130 = interrupted. Errors and warnings are one line each; add -v for details.

Flags:

The defaults (n_fft=1024, hop=128) match the released model's training configuration. A larger hop (e.g. --hop 256) halves the columns for faster, lighter runs at some fidelity cost.

On HPC clusters where compute nodes have no internet access, pre-fetch the weights on the login node, then run the batch job on the compute node. Without cached weights, the CLI says so in one line and exits 2.

tokeye download big_tf_unet   # on the login node; prints the cached path
tokeye run ... --model big_tf_unet   # on the compute node — model is already cached

Mode-analysis suite

Beyond segmentation, tokeye bundles the analyses DIII-D researchers usually reach for separate tools to get. Each is a subcommand; --help on any of them shows the full flags.

elmspec and alfvenspec take --key as tokeye run does.

Web app guide

tokeye app (or python -m tokeye.app) launches a Gradio interface with three tabs:

• Analyze — load a signal, compute its spectrogram, run a model, and visualize the result. Guided for first-time use: the model dropdown defaults to the bundled big_tf_unet model, the STFT transform has working defaults, and "Load Example Signal" generates a synthetic demo signal so a brand-new user needs zero files. "Analyze" runs the whole load-model → infer → visualize pipeline in one click. View modes: Original, Enhanced (percentile-clipped amplitude), Mask (thresholded model output), Amplitude.
• Annotate — manually draw and save mask annotations over a read-only backdrop image.
• Utilities — audio-format conversion and .npy file inspection.

Flags: tokeye app [--host 127.0.0.1] [--port 7860] [--open | --no-browser] [--workspace DIR] [--share]. The app binds to 127.0.0.1 (--host defaults to $GRADIO_SERVER_NAME if set), moves to the next free port if 7860 is busy and says which port it chose, and opens a browser tab when you run it locally. --share creates a public Gradio link that anyone with the URL can use.

If you're on a remote server (e.g. an HPC login node), forward the port over SSH instead of using --share:

ssh -L 7860:localhost:7860 user@remote

Then open http://localhost:7860 in your local browser.

Verified Datatypes

• DIII-D Fast Magnetics (cite)
• DIII-D CO2 Interferometer (cite)
• DIII-D Electron Cyclotron Emission (cite)
• DIII-D Beam Emission Spectroscopy (cite)

Evaluation

Recall Scores:

• TJII2021: 0.8254
• DCLDE2011 (Delphinus capensis): 0.7708
• DCLDE2011 (Delphinus delphis): 0.7953

With more data, comes better models. Please contribute to the project!

Installation (from source / development)

uv is the dev tool for this repo:

git clone git@github.com:PlasmaControl/TokEye.git
cd TokEye
uv sync --dev --extra all                # dev tools (pytest, ruff, etc.) + every extra (web app, ae model, HDF5)
uv sync --dev --extra all --group train  # the same + training deps (lightning, h5py, etc.)

Each uv sync makes .venv match its flags exactly, removing any extra or group you leave out, so name everything you need in one command. PyTorch comes as the default build (GPU/CUDA on Linux).

Additionally, run these the first time

uv run pre-commit install
uv run ruff check .
uv run pytest

This creates a .venv/; activate it with source .venv/bin/activate, or prefix commands with uv run.

Models

Weights download automatically the first time a registry name is used (cached in ~/.cache/huggingface). Override the default repo with the TOKEYE_HF_REPO environment variable (per-model repos are fixed in the registry).

To use a local checkpoint instead, put .pt/.pt2 files in a model/ directory (picked up by the app's model dropdown) or pass a path directly via --model PATH.

Input should be a tensor that has shape (B, 1, H, W) where B, H, and W can vary Output will be a tensor of shape (B, 2, H, W)

Best performance when spectrograms are oriented so that when they are plotted with matplotlib, the lowest frequency bin is oriented with the bottom when origin='lower'. Spectrograms should be standardized (mean = 0, std = 1). If baseline activity is very strong, clipping the input may help, but is generally not needed.

The first channel of the output will return preferential measurements of coherent activity (useful for most tasks) The second channel of the output will return preferential measurements of transient activity

Data

Keep signals as 1D float arrays (raw time series) in any supported format (see Batch processing). No need to normalize or preprocess them. The CLI also accepts 2D arrays (precomputed spectrograms) directly. The app scans a signal directory for .npy files (default data/input, configurable in the Analyze tab).

Bringing your own data takes two lines:

import numpy as np

signal = ...  # any 1D float array: tokamak diagnostic, hydrophone, etc.
np.save("shots/myshot.npy", signal)

tokeye run shots/myshot.npy --output-dir results

No data yet? tokeye example writes a synthetic demo signal (tokeye_example_sr200000.npy) you can run immediately, and the web app has a matching "Load Example Signal" button.

Citation

If you use this code in your research, please cite:

@article{chen_TokEye_2026,
  title={TokEye: Fast Signal Extraction for Fluctuating Time Series via Offline Self-Supervised Learning From Fusion Diagnostics to Bioacoustics},
  author={Chen, Nathaniel},
  year={2026},
  publisher={ArXiv},
  doi={10.48550/arXiv.2602.20317},
  url={https://www.arxiv.org/abs/2602.20317}
}

Contact

Nathaniel Chen — nathaniel [at] princeton [dot] edu — https://nathanielchen.net