Try OpenDPD Studio in your browser →

No installation. Explore example data or upload your own CSV, train PA/DPD models on shared CUDA compute, and download your checkpoints. Temporary data and results are deleted within 24 hours.

OpenDPD

Model a power amplifier. Train a digital predistorter. Understand the result.

OpenDPD is a PyTorch framework for power amplifier (PA) modeling and digital predistortion (DPD), developed by the Efficient Machine Intelligence Lab at TU Delft. Use OpenDPD Studio in the browser or locally for a guided workflow, or automate experiments with the CLI and Python API. All three use the original OpenDPD training core.

Documentation · Studio walkthrough · Examples & Colab · Papers & citation

What's new

OpenDPD 2.2.4 separates PSD plots by signal-chain position: DPD Input, DPD Output / PA Input, and PA Output. Output references and with/without-DPD comparisons share only the PA Output plot, with compact legends and independent chart controls.

Signal Generator creates a PA Input Dataset with separate CSV and metadata downloads. The new PA Library offers nine mathematical Virtual PAs with editable formulas and linked parameter controls. Explicitly simulate the output, create a paired synthetic dataset, and continue to PA/DPD training and testing. An expandable workflow diagram follows your progress.

See the 2.2.4 release notes and Signal Generator guide. Standard presets are uncoded engineering stimuli; Wi-Fi 8 is experimental. Local Studio also adds research comparisons, publication figures and reproduction, measurement sessions, Sweep Board, hardware cost evidence and optional dataset contribution PRs for human review.

• CUDA replay for supported native models reduces dispatch overhead while retaining the existing optimizer, precision, batches and scheduler.
• Quick/full training defaults are 10/150 epochs; plots update once per epoch; DPD previews capture the intermediate signal in a bounded shadow-model forward pass.
• See the 2.2.1 performance measurements and release notes.
• Guided experiments: explore built-in I/Q data, train and test PA/DPD models, and choose from the original backbone registry.
• Live feedback: separate epoch and batch progress bars, NMSE and other task metrics, live signal plots, reconnectable experiments and a Stop control.
• Download models while training: save the best checkpoint so far; after training, download the selected final model. Compare compatible runs and export reports.
• Browser and local workbench: nine interface languages, English by default, CUDA when available, touch-friendly plots and system light/dark themes.

Bring your own CSV: upload UTF-8 CSV with two complex columns or four real I/Q columns, up to 25 MiB and 1,000,000 paired samples. Every row is validated in quarantine before preview; rejected uploads are deleted. Code, package and checkpoint uploads are unavailable in the public app.

For a hosted installation, the public Studio deployment guide covers GitHub Pages, a Cloudflare Tunnel and isolated local VM compute, with temporary sessions and automatic file deletion within 24 hours.

Feature history · Verified platform status

Get started with Studio

Open the hosted Studio now, or install the packaged local app with Python 3.10–3.13:

python -m pip install "opendpd[gui]==2.2.4"
opendpd gui

The wheel includes the frontend; Node.js is not needed. For development from source, also install Git and Node.js 22.22+:

git clone https://github.com/lab-emi/OpenDPD.git
cd OpenDPD
python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[gui]"
npm --prefix frontend ci
npm --prefix frontend run build
opendpd gui

This opens Studio locally in your browser. For Windows, a native desktop window, GPU setup, or a core-only installation, see Installation.

Click Get Started → Signal Generator to create a waveform, then Choose Virtual PA to simulate a paired dataset. Or choose Use an existing dataset → DPA_200MHz to go directly to PA Training. Use Starting settings → Quick trial to check the pipeline, then choose Full training for a longer experiment. Quick trial defaults to 10 epochs; full training defaults to 150 epochs. Plots update once per epoch; DPD previews show x, u and PA(u) from the same bounded validation probe. Advanced settings offer an optional batch preview interval with a red warning because extra previews can severely slow training.

The PA → DPD workflow

A DPD result evaluated through a PA model is a simulation. Exported u = DPD(x) is the PA input; a physical PA measurement is needed to establish measured linearization performance. See the Studio walkthrough and measured DPD guide.

PA Library guide · Reading signal-chain PSD plots

Choose your next step

Contribute & cite

Contributions of models, tests and documentation are welcome. Start with CONTRIBUTING.md; see testing and how we maintain the docs.

If you use OpenDPD in research, cite the OpenDPD paper. BibTeX and related papers · CITATION.cff

Chang Gao — Project Leader · Yizhuo Wu — Leading Developer. Meet the team · EMI Lab