Changelog¶

All notable changes to LightlyTrain will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.6.1] - 2025-03-31¶

TensorBoard is now automatically installed and no longer an optional dependency.
Update the Models documentation.
Update the YOLO tutorial

Add support for DINOv2 distillation pretraining with the "distillation" method.
Add support for YOLO11 and YOLO12 models.
Add support for RT-DETR models.
Add support for YOLOv12 models by the original authors.
The Git info (branch name, commit, uncommited changes) for the LightlyTrain package and the directory from where the code runs are now logged in the train.log file.

Ensures proper usage of the blur_limit parameter in the GaussianBlur transforms.

Add tutorial on how to use LightlyTrain with YOLO.
Show the data_wait percentage in the progress bar to better monitor performance bottlenecks.
Add auto format export with example logging, which automatically determines the best export option for your model based on the used model library.
Add support for configuring the random rotation transform via transform_args.random_rotation.
Add support for configuring the color jitter transform via transform_args.color_jitter.
When using the DINO method and configuring the transforms: Removes local_view_size, local_view_resize and n_local_views from DINOTransformArgs in favor of local_view.view_size, local_view.random_resize and local_view.num_views. When using the CLI, replace transform_args.local_view_size with transform_args.local_view.view_size, … respectively.
Allow specifying the precision when using the embed command. The loaded checkpoint will be casted to that precision if necessary.

Increase default DenseCL SGD learning rate to 0.1.
Dataset initialization is now faster when using multiple GPUs.
Models are now automatically exported at the end of a training.
Update the docker image to PyTorch 2.5.1, CUDA 11.8, and cuDNN 9.
Switched from using PIL+torchvision to albumentations for the image transformations. This gives a performance boost and allows for more advanced augmentations.
The metrics batch_time and data_time are grouped under profiling in the logs.

Fix Ultralytics model export for Ultralytics v8.1 and v8.2
Fix that the export command may fail when called in the same script as a train command using DDP.
Fix the logging of the train_loss to report the batch_size correctly.

Log system information during training
Add Performance Tuning guide with documentation for multi-GPU and multi-node training
Add Pillow-SIMD support for faster data processing
- The docker image now has Pillow-SIMD installed by default
Add ultralytics export format
Add support for DINO weight decay schedule
Add support for SGD optimizer with optim="sgd"
Report final accelerator, num_devices, and strategy in the resolved config
Add Changelog to the documentation

Various improvements for the DenseCL method
- Increase default memory bank size
- Update local loss calculation
Custom models have a new interface
The number of warmup epochs is now set to 10% of the training epochs for runs with less than 100 epochs
Update default optimizer settings
- SGD is now the default optimizer
- Improve default learning rate and weight decay values
Improve automatic num_workers calculation
The SPPF layer of Ultralytics YOLO models is no longer trained

Add Ultralytics model support
Add SuperGradients PP-LiteSeg model support
Save normalization transform arguments in checkpoints and automatically use them in the embed command
Better argument validation
Automatically configure num_workers based on available CPU cores
Add faster and more memory efficient image dataset
Log more image augmentations
Log resolved config for CallbackArgs, LoggerArgs, MethodArgs, MethodTransformArgs, and OptimizerArgs