Skip to content

Install vision-synth

The base package requires Python 3.10 or newer. Direct generation through synth_datasets is torch-free. PyTorch is an optional torch extra required by the image-augmentation engine and by SyntheticIterableDataset for PyTorch DataLoader integration. Kornia, TorchVision, and Albumentations are optional on top of that because the native builder can create a useful augmentation pipeline without them.

Project maturity

The package is currently classified Beta. Pin versions in production or research environments and validate the exact pipeline after upgrades.

Base installation

python -m pip install vision-synth

This installs NumPy, Pillow, and the package itself — no torch. It is enough for synthetic dataset generation via import synth_datasets, which never imports torch. Generating COCO or YOLO data requires no optional extra, source images, or model download.

The SyntheticIterableDataset wrapper and its PyTorch DataLoader integration require the torch extra below; direct use of SyntheticGenerator and generate_dataset does not.

Image augmentation (needs torch)

Compose, FusedCompose, AugmentationSequential, and everything else at the fused_transforms package root need the torch extra:

python -m pip install "vision-synth[torch]"

This is enough for Compose.from_params with no optional adapter backend.

Optional backends

Install only the adapter ecosystems you use — each of these already includes the torch extra:

python -m pip install "vision-synth[kornia]"

python -m pip install "vision-synth[torchvision]"

python -m pip install "vision-synth[albumentations]"

python -m pip install "vision-synth[all]"

The extras enable adapter support; they do not make every upstream transform or parameter combination fusible. Check the capability tables.

Verify the installation

python -c "import synth_datasets"

This check uses only the base installation. For the image-augmentation features, verify the optional torch extra:

python -c "import torch, fused_transforms"

The base smoke check is executable in the generated documentation test suite:

import synth_datasets

assert callable(synth_datasets.generate_dataset)
assert synth_datasets.__version__

The augmentation engine lives under a single import namespace:

from fused_transforms import Compose

assert Compose.__name__ == "FusedCompose"

Build these docs locally

The repository uses uv for its locked development environment:

uv sync --group docs
uv run --group docs mkdocs serve

Run the release-style documentation gate with:

uv run --group docs mkdocs build --strict

The generated site is written to site/. The docs dependency group is not installed for package users.