Installation¶

Installation from PyPI¶

LightlyTrain is available on PyPI and can be installed via pip or other package managers.

Warning

To successfully install LightlyTrain the Python version has to be >=3.8 and <=3.12 .

pip install lightly-train

To update to the latest version, run:

pip install --upgrade lightly-train

See Docker for Docker installation instructions.

Platform Compatibility¶

Platform	Supported Compute
Linux	CPU or CUDA
MacOS	CPU (MPS is planned)
Windows	CPU or CUDA (experimental)

Version Compatibility¶

`lightly-train`	`torch`	`torchvision`	`pytorch-lightning`	Python
`0.6`	`>=2.1`, `<2.6`	`>=0.16`	`>=2.1`, `<2.6`	`>=3.8`, `<3.13`

Warning

We recommend installing versions of the torch, torchvision, and pytorch-lightning packages that are compatible with each other. The latest compatible versions are torch==2.5 torchvision==0.21 pytorch-lightning==2.5.

See the Torchvision and PyTorch Lightning documentation for more information on version compatibility between different PyTorch packages.

Optional Dependencies¶

LightlyTrain has optional dependencies that are not installed by default. The following dependencies are available:

Logging¶

wandb: For logging to Weights & Biases

Model Support¶

rfdetr: For RF-DETR models
super-gradients: For SuperGradients models
timm: For TIMM models
ultralytics: For Ultralytics models

To install optional dependencies, run:

pip install "lightly-train[wandb]"

Or for multiple optional dependencies:

pip install "lightly-train[wandb,timm]"

Hardware Recommendations¶

An example hardware setup and its performance when using LightlyTrain is provided in Hardware Recommendations.