Examples

After installing the package, you can directly use the generator and datasets in your project. See the following examples for guidance on how to instantiate the generator and datasets.

Recommended: Using Configuration Files

For reproducibility and greater flexibility, we recommend using the configuration files provided in the package. These files define the parameters for the generator and datasets, allowing for quick and easy setup with hydra.

Steps to Use Configuration Files

1. Copy the configuration files (entire configs/dataset folder) to your project root directory into configs/dataset.
2. Use the following methods to instantiate the generator and dataset classes:

For examples below, set up cfg_path = "configs/dataset" as the path to the configuration files and cfg_name as the name of the configuration file you want to use (cfg_name='synth_train' for example for the synthetic training dataset).

See the Configs page for detailed information on configuration files and available generation modes.

Using the Imperative API

import hydra

with hydra.initialize(config_path=cfg_path, version_base="1.2"):
    cfg = hydra.compose(config_name=cfg_name)
    print(f"Composed config: {cfg}")
    dataset = hydra.utils.instantiate(cfg)

Using the Declarative API

import hydra
from omegaconf import DictConfig

@hydra.main(config_path=cfg_path, config_name=cfg_name)
def my_app(cfg: DictConfig) -> None:
    print(cfg)

Note

Ensure that the bids and seeds paths in the configuration files are updated to the absolute paths for your data.

Using Direct Instantiation

You can manually instantiate required classes from the FetalSynthGen in your project as needed. For example, to instantiate the FetalSynthDataset class and generator components follow the example below:

# Import necessary classes
from fetalsyngen.data.datasets import FetalSynthDataset

from fetalsyngen.generator.model import FetalSynthGen
from fetalsyngen.generator.augmentation.synthseg import (
    RandBiasField,
    RandGamma,
    RandNoise,
    RandResample,
)
from fetalsyngen.generator.deformation.affine_nonrigid import SpatialDeformation
from fetalsyngen.generator.intensity.rand_gmm import ImageFromSeeds

# Instantiate the generator components
intensity_generator = ImageFromSeeds(
    min_subclusters=1,
    max_subclusters=3,
    seed_labels=[1, 2, 3, 4, 5],
    generation_classes=[1, 2, 3, 4, 5],
    meta_labels=4,
)
spatial_deform = SpatialDeformation(
    max_rotation=10,
    max_shear=1,
    max_scaling=1,
    size=(256, 256, 256),
    nonlinear_transform=1,
    nonlin_scale_min=1,
    nonlin_scale_max=1,
    nonlin_std_max=1,
    flip_prb=1,
    device="cuda",
)
resampler = RandResample(prob=0.5, max_resolution=1.5, min_resolution=0.5)
bias_field = RandBiasField(
    prob=0.5, scale_min=0.5, scale_max=1.5, std_min=0.5, std_max=1.5
)
noise = RandNoise(prob=0.5, std_min=0.5, std_max=1.5)
gamma = RandGamma(prob=0.5, gamma_std=0.5)

# Instantiate the generator
generator = FetalSynthGen(
    shape=(256, 256, 256),
    resolution=(0.5, 0.5, 0.5),
    device="cuda",
    intensity_generator=intensity_generator,
    spatial_deform=spatial_deform,
    resampler=resampler,
    bias_field=bias_field,
    noise=noise,
    gamma=gamma,
)

# Instantiate the dataset
dataset = FetalSynthDataset(
    bids_path="./../../data",
    generator=generator,
    seed_path="./../../data/derivatives/seeds",
    sub_list=None,
)

Additional Resources

For more examples of generator instantiation with hydra, see the Generator Instantiation notebook.