API reference

import opendpd gives access to the functions and the trainer class below. They wrap the same training pipeline as python main.py; keyword arguments that are not listed are forwarded to the command-line configuration (arguments.py), so every main.py option is also available from Python. The opendpd-cli entry point is python main.py under another name: run opendpd-cli --help for the options.

OpenDPD API - User-friendly interface for training PA and DPD models

This module provides high-level functions for easy use of OpenDPD functionality.

OpenDPDTrainer

OpenDPDTrainer(dataset_name: Optional[str] = None, dataset_path: Optional[str] = None, **kwargs: Any)

Advanced trainer class for OpenDPD with more control over the training process.

This class provides a more object-oriented interface for advanced users who need fine-grained control over the training process.

Examples:

>>> import opendpd
>>> trainer = opendpd.OpenDPDTrainer(dataset_name='DPA_200MHz')
>>> trainer.train_pa(n_epochs=50)
>>> trainer.train_dpd(n_epochs=50)
>>> trainer.evaluate()

Initialize the OpenDPD trainer.

Parameters:

Name	Type	Description	Default
dataset_name	Optional[str]	Name of the dataset	None
dataset_path	Optional[str]	Path to custom dataset	None
**kwargs	Any	Additional configuration parameters	{}

train_pa

train_pa(**kwargs)

Train PA model

train_dpd

train_dpd(**kwargs)

Train DPD model

run

run(**kwargs)

Run DPD model

train_pa

train_pa(dataset_name: Optional[str] = None, dataset_path: Optional[str] = None, PA_backbone: str = 'gru', PA_hidden_size: int = 23, n_epochs: int = 300, batch_size: int = 64, lr: float = 0.005, accelerator: str = 'cpu', frame_length: int = 200, seed: int = 0, plot: bool = False, plot_every: int = 1, **kwargs: Any) -> Dict[str, Any]

Train a Power Amplifier (PA) behavioral model.

Parameters:

Name	Type	Description	Default
dataset_name	Optional[str]	Name of the dataset in the datasets/ folder (e.g., 'DPA_200MHz')	None
dataset_path	Optional[str]	Deprecated. Specify dataset_name after importing the dataset.	None
PA_backbone	str	Type of neural network backbone ('gru', 'lstm', 'dgru', 'deltagru', etc.)	'gru'
PA_hidden_size	int	Hidden size of the PA model	23
n_epochs	int	Number of training epochs	300
batch_size	int	Batch size for training	64
lr	float	Learning rate	0.005
accelerator	str	Device to use ('cpu', 'cuda', or 'mps')	'cpu'
frame_length	int	Length of signal frames	200
seed	int	Random seed for reproducibility	0
plot	bool	Enable plot generation during training	False
plot_every	int	Generate per-epoch plots every N epochs	1
**kwargs	Any	Additional arguments passed to the training configuration	{}

Returns:

Type	Description
Dict[str, Any]	Dictionary containing training results and model path

Examples:

>>> import opendpd
>>> results = opendpd.train_pa(dataset_name='DPA_200MHz', n_epochs=50)
>>> print(f"Model saved at: {results['model_path']}")

>>> # Train with plots
>>> results = opendpd.train_pa(dataset_name='DPA_200MHz', n_epochs=50, plot=True)

train_dpd

train_dpd(dataset_name: Optional[str] = None, dataset_path: Optional[str] = None, DPD_backbone: str = 'tres_deltagru', DPD_hidden_size: int = 15, PA_backbone: str = 'gru', PA_hidden_size: int = 23, n_epochs: int = 300, batch_size: int = 64, lr: float = 0.005, accelerator: str = 'cpu', frame_length: int = 200, seed: int = 0, thx: float = 0.0, thh: float = 0.0, plot: bool = False, plot_every: int = 1, collect_delta_stats: bool = False, cuda_graph_training: bool = False, **kwargs: Any) -> Dict[str, Any]

Train a Digital Pre-Distortion (DPD) model.

Parameters:

Name	Type	Description	Default
dataset_name	Optional[str]	Name of the dataset in the datasets/ folder	None
dataset_path	Optional[str]	Deprecated. Specify dataset_name after importing the dataset.	None
DPD_backbone	str	Type of DPD neural network backbone	'tres_deltagru'
DPD_hidden_size	int	Hidden size of the DPD model	15
PA_backbone	str	Type of PA backbone (must match the pre-trained PA model)	'gru'
PA_hidden_size	int	Hidden size of PA model (must match the pre-trained PA model)	23
n_epochs	int	Number of training epochs	300
batch_size	int	Batch size for training	64
lr	float	Learning rate	0.005
accelerator	str	Device to use ('cpu', 'cuda', or 'mps')	'cpu'
frame_length	int	Length of signal frames	200
seed	int	Random seed for reproducibility	0
thx	float	Threshold for input deltas (for delta-based models)	0.0
thh	float	Threshold for hidden state deltas (for delta-based models)	0.0
plot	bool	Enable plot generation during training	False
plot_every	int	Generate per-epoch plots every N epochs	1
collect_delta_stats	bool	Collect temporal sparsity diagnostics during training	False
cuda_graph_training	bool	Opt in to guarded CUDA-graph DPD training	False
**kwargs	Any	Additional arguments passed to the training configuration	{}

Returns:

Type	Description
Dict[str, Any]	Dictionary containing training results and model path

Examples:

>>> import opendpd
>>> dpd_results = opendpd.train_dpd(dataset_name='DPA_200MHz', n_epochs=50, plot=True)

run_dpd

run_dpd(dataset_name: Optional[str] = None, dataset_path: Optional[str] = None, DPD_backbone: str = 'tres_deltagru', DPD_hidden_size: int = 15, accelerator: str = 'cpu', plot: bool = False, **kwargs: Any) -> Dict[str, Any]

Run the trained DPD model to generate pre-distorted signals.

Parameters:

Name	Type	Description	Default
dataset_name	Optional[str]	Name of the dataset in the datasets/ folder	None
dataset_path	Optional[str]	Deprecated. Specify dataset_name after importing the dataset.	None
DPD_backbone	str	Type of DPD backbone (must match trained model)	'tres_deltagru'
DPD_hidden_size	int	Hidden size of DPD model (must match trained model)	15
accelerator	str	Device to use ('cpu', 'cuda', or 'mps')	'cpu'
plot	bool	Enable plot generation	False
**kwargs	Any	Additional arguments	{}

Returns:

Type	Description
Dict[str, Any]	Dictionary containing output paths and results

Examples:

>>> import opendpd
>>> results = opendpd.run_dpd(dataset_name='DPA_200MHz', plot=True)

plot_dpd

plot_dpd(dataset_name: Optional[str] = None, PA_backbone: str = 'gru', PA_hidden_size: int = 23, DPD_backbone: str = 'tres_deltagru', DPD_hidden_size: int = 15, accelerator: str = 'cpu', **kwargs: Any) -> Dict[str, Any]

Generate comparison plots: PA output without DPD vs with DPD.

This function loads trained PA and DPD models, runs inference on the test set for both scenarios, and generates side-by-side comparison plots including PSD, AM/AM, AM/PM, constellation diagrams, waveforms, and a metrics summary.

Parameters:

Name	Type	Description	Default
dataset_name	Optional[str]	Name of the dataset in the datasets/ folder	None
PA_backbone	str	Type of PA backbone (must match the pre-trained PA model)	'gru'
PA_hidden_size	int	Hidden size of PA model (must match the pre-trained PA model)	23
DPD_backbone	str	Type of DPD backbone (must match trained model)	'tres_deltagru'
DPD_hidden_size	int	Hidden size of DPD model (must match trained model)	15
accelerator	str	Device to use ('cpu', 'cuda', or 'mps')	'cpu'
**kwargs	Any	Additional arguments	{}

Returns:

Type	Description
Dict[str, Any]	Dictionary containing plot directory path

Examples:

>>> import opendpd
>>> results = opendpd.plot_dpd(dataset_name='DPA_200MHz')
>>> print(f"Plots saved at: {results['plot_dir']}")

load_dataset

load_dataset(dataset_path: str) -> Dict[str, Any]

Load a dataset from a CSV file or directory.

This function supports both formats: 1. Split CSV files (train_input.csv, train_output.csv, etc.) 2. Single CSV file with all data (I_in, Q_in, I_out, Q_out columns)

Parameters:

Name	Type	Description	Default
dataset_path	str	Path to the dataset directory or single CSV file	required

Returns:

Type	Description
Dict[str, Any]	Dictionary containing loaded data arrays

Examples:

>>> import opendpd
>>> data = opendpd.load_dataset('datasets/DPA_200MHz')
>>> print(data.keys())
dict_keys(['X_train', 'y_train', 'X_val', 'y_val', 'X_test', 'y_test'])

>>> # Load from single CSV
>>> data = opendpd.load_dataset('my_data.csv')

create_dataset

create_dataset(csv_path: str, output_dir: str, dataset_name: str, train_ratio: float = 0.6, val_ratio: float = 0.2, test_ratio: float = 0.2, dataset_format: str = 'single_csv', csv_filename: Optional[str] = None, **spec_kwargs: Any) -> str

Create a dataset in OpenDPD format from a single CSV file.

The input CSV should have 4 columns: I_in, Q_in, I_out, Q_out

Parameters:

Name	Type	Description	Default
csv_path	str	Path to the input CSV file	required
output_dir	str	Directory where the dataset will be created	required
dataset_name	str	Name of the dataset	required
train_ratio	float	Ratio of data for training (default: 0.6)	0.6
val_ratio	float	Ratio of data for validation (default: 0.2)	0.2
test_ratio	float	Ratio of data for testing (default: 0.2)	0.2
dataset_format	str	'single_csv' (default) to keep a single CSV or 'split_csv' to generate the classic six-file layout	'single_csv'
csv_filename	Optional[str]	Optional filename for the generated single CSV (defaults to data.csv)	None
**spec_kwargs	Any	Additional fields for spec.json (e.g., input_signal_fs, bw_main_ch)	{}

Returns:

Type	Description
str	Path to the created dataset directory

Examples:

>>> import opendpd
>>> dataset_path = opendpd.create_dataset(
...     csv_path='my_measurements.csv',
...     output_dir='datasets',
...     dataset_name='MyPA_Data',
...     input_signal_fs=800e6,
...     bw_main_ch=200e6
... )
>>> print(f"Dataset created at: {dataset_path}")