PA modeling and DPD training

For a first experiment, follow the Studio walkthrough or the workspace CLI guide. This page explains the learning pipeline and the original research commands, which remain available as python main.py and opendpd-cli.

How the pipeline works

1. Prepare paired I/Q data. Each complex input sample has a time-aligned measured PA output. Training, validation and test data have distinct roles; framing happens within their boundaries. Use the dataset's recorded splits rather than assuming one ratio for every dataset. See dataset formats and the split protocol.
2. Learn the PA response. A behavioral model predicts the measured output from the input. Recurrent models learn from sequences using backpropagation through time.
3. Learn a predistorter. Put a trainable DPD model before the fitted PA model. Gradients pass through the frozen PA model to optimize the DPD against the linear reference.
4. Evaluate and export. Score held-out data and inspect the waveforms and spectra. Exported u = DPD(x) is an input for the physical PA; the surrogate's PA(u) is a simulated output.

A good surrogate result does not establish measured PA performance. For hardware validation, follow the measured DPD guide. Metric definitions and comparability rules live in metric profiles.

Original CLI: train, apply, plot

Run these commands in order from the repository root after installation. This example uses the built-in DPA_200MHz dataset on CPU and the original default models:

python main.py --dataset_name DPA_200MHz --step train_pa --accelerator cpu
python main.py --dataset_name DPA_200MHz --step train_dpd --accelerator cpu
python main.py --dataset_name DPA_200MHz --step run_dpd --accelerator cpu
python main.py --dataset_name DPA_200MHz --step plot --accelerator cpu

For an installed core package, substitute opendpd-cli for python main.py. Use --accelerator cuda or --accelerator mps only with a working backend on your machine. The defaults are research runs; use the workspace CLI's smoke recipes for a shorter first check.

The original pipeline writes checkpoints under save/, logs under log/ and predistorted I/Q CSVs under dpd_out/. The workspace service runs this same pipeline inside each run directory and records the resolved configuration and provenance. See the CLI artifact table.

When changing a model, keep the PA checkpoint, dataset, seed and framing settings consistent between stages. Studio and opendpd run validate these references for you. Query opendpd models for the current registry and opendpd recipes for preset experiments; the model guide explains how backbones are registered.

Classical baselines

The workspace recipes pa-mp-ls-v1, pa-gmp-ls-v1, dpd-mp-ila-v1 and dpd-gmp-ila-v1 fit memory-polynomial baselines on CPU. PA fits use direct least squares on the training split; DPD fits use indirect learning (ILA) and are evaluated through a gradient-trained PA surrogate.

These fits have no training epochs or random seed. Their results record retained rank, condition number, singular-value cutoff (rcond) and train residual in fit.json. The result names its training path so comparisons remain explicit. A least-squares PA is a modeling reference and cannot be selected as the DPD simulation surrogate. See the benchmark protocol for comparisons under a shared budget and reference.

Quantization-aware training

Quantization-aware training fine-tunes a compatible DPD model with fixed-point weight and activation constraints. First train the float model of the same quantization-capable backbone:

python main.py --dataset_name DPA_200MHz --step train_dpd --accelerator cpu --DPD_backbone qgru

Then set FLOAT_CHECKPOINT to that run's actual DPD_*_M_QGRU_*.pt checkpoint under save/. The following is a shell template; replace the path before running it:

FLOAT_CHECKPOINT="/path/to/float-qgru-checkpoint.pt"
python main.py --dataset_name DPA_200MHz --step train_dpd --accelerator cpu --DPD_backbone qgru --quant --n_bits_w 16 --n_bits_a 16 --pretrained_model "$FLOAT_CHECKPOINT" --quant_dir_label w16a16
python main.py --dataset_name DPA_200MHz --step run_dpd --accelerator cpu --DPD_backbone qgru --quant --n_bits_w 16 --n_bits_a 16 --quant_dir_label w16a16

Keep the backbone, bit widths and label consistent when applying the quantized checkpoint. For paper-specific settings use the reproduction scripts; for a checked fixed-point C99 package see deployment export.

Inspect training

Studio shows live task metrics, signal comparisons and worker output; see visualization. In the original CLI, Rich tables display run information and train/validation/test metrics. Set their precision with --log_precision:

python main.py --dataset_name DPA_200MHz --step train_pa --log_precision 4

Python users can call train_pa, train_dpd, run_dpd and plot_dpd; see the examples and API reference.

Starting in Studio 2.2.4

A generated signal is a PA Input Dataset, not a training pair. Use the Virtual PA Library to explicitly synthesize y, or supply measured input/output pairs. The Virtual PA is the mathematical simulation device; PA Model Training subsequently learns a separate surrogate from its dataset. PA and DPD each combine Training and Testing tabs. Live DPD previews include x, u and PA(u) from the same bounded validation forward pass, with separate PSD panels.

ILC linearization reference

Studio 2.2.5 adds dpd-ilc-ila-v1 (ilc_dpd): bounded gain-inverse waveform learning through a trained PA surrogate, followed by a training-only ILA memory-polynomial fit. PA and DPD pages expose this workflow, including conservative default hyperparameters and a separate Ideal test-waveform reference. See ILC and ILA DPD for equations, convergence, artifacts and evidence limits.