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.

[0.11.4] - 2025-10-08¶

Added¶

Add support for saving the best semantic segmentation checkpoints and model weights during training.
Expose more arguments for the checkpointing callback in pretraining.
Add LT-DETR inference support with DINOv2 and DINOv3 ConvNext backbones.

Changed¶

Change default precision to bf16-mixed for pretraining on GPUs that support it.

Deprecated¶

Removed¶

Fixed¶

Fix warning about too few epochs for DINOv2 which occurs with the default epoch calculation.

[0.11.3] - 2025-09-25¶

Added¶

Add EoMT semantic segmentation benchmark results and model weights trained on ADE20k, COCO-Stuff, and Cityscapes datasets.
Add support for exporting the semantic segmentation model weights to exported_models/exported_last.pt.
Add support for allow loading semantic segmentation model weights for training.
Add simplify flag to ONNX export_task.
Add support for using DINOv3 models as teacher in distillationv1.

Fixed¶

Fix a bug in model.predict() with ignore_index.
Speed up listing of filenames in large datasets.

[0.11.2] - 2025-09-08¶

Added¶

Add support for using multi-channel masks for the inputs in semantic segmentation.
Add support for training models on multi-channel images with transform_args={"num_channels": 4}.
Add support for using custom mask names for the inputs in semantic segmentation.
Add precision flag to ONNX export task to specify if we export with float16 or float32 precision.

Fixed¶

Fix issue where segmentation fine-tuning could fail when encountering masks containing only unknown classes.
Fix issue with mmap cache when multiple runs use the same dataset on the same machine.
Speed up logging of datasets with many files.

[0.11.1] - 2025-08-28¶

Added¶

Add support for DINOv2 linear semantic segmentation models. You can train them with model="dinov2/vits14-linear" in the train_semantic_segmentation command. Those models are trained with a linear head on top of a frozen backbone and are useful to evaluate the quality of pretrained DINOv2 models.
Make fine-tune transform arguments configurable in the train_semantic_segmentation command. You can now use the transform_args argument like this
```
transform_args={
  "image_size": (448, 448), # (height, width)
  "normalize": {"mean": (0.0, 0.0, 0.0), "std": (0.5, 0.5, 0.5)}, # (r, g, b) channels
}
```
to customize the image augmentations used during training and validation. See the semantic segmentation documentation for more information.
Add support for the channel drop transform in the train_semantic_segmentation command.

Add support for mapping multiple classes into a single class for semantic segmentation datasets. You can now use a dictionary in the classes entry of the data argument in the train_semantic_segmentation command like this:

data={
  "classes": {
    0: {"name": "background", "values": [0, 255]}, # Map classes 0 and 255 to class 0
    1: {"name": "class 1", "values": [1]},
    2: "class 2",  # Still supported. Equivalent to {"name": "class 2", "values": [2]}
  },
}

Fixed¶

Models loaded with load_model_from_checkpoint are now automatically moved to the correct device.
Loading EoMT models with load_model_from_checkpoint no longer raises a missing key error.
Fix MLFlow logging on AzureML.

[0.11.0] - 2025-08-15¶

🚀 New DINOv3 Support: Pretrain your own model with distillation from DINOv3 weights. Or fine-tune our SOTA EoMT semantic segmentation model with a DINOv3 backbone! 🚀

Added¶

Distillation now supports DINOv3 pretrained weights as teacher.
Semantic Segmentation now supports DINOv3 pretrained weights as EoMT backbone.

Changed¶

LightlyTrain now infers the best number of epochs based on the chosen method, dataset size and batch size.

[0.10.0] - 2025-08-04¶

🔥 New: Train state-of-the-art semantic segmentation models with our new DINOv2 semantic segmentation fine-tuning method! 🔥

Added¶

DINOv2 semantic segmentation fine-tuning with the train_semantic_segmentation command. See the semantic segmentation documentation for more information.
Support for image resolutions that are not a multiple of the patch size in DINOv2.

Changed¶

DINOv2 model names to be more consistent with the new naming scheme. The model name scheme changed from dinov2_vit/vits14_pretrain to dinov2/vits14. The pretrained weights are now always loaded by default, making the pretrain suffix redundant.
DINOv2 models are now using registers by default which increases model performance. You can continue using models without registers with the -noreg suffix: dinov2/vits14-noreg.
DINOv2 and distillation to work with image resolutions that are not a multiple of the patch size.
Temporary files are now stored in the ~/.cache/lightly-train directory be default. The location can be changed with the LIGHTLY_TRAIN_CACHE_DIR environment variable.

Deprecated¶

The dinov2_vit/vits14_pretrain model name is deprecated and will be removed in a future release. Use dinov2/vits14 instead.

[0.9.0] - 2025-07-21¶

Added¶

Add an extra teacher_weights argument in method_args to allow loading pretrained DINOv2 teacher weights for distillation methods.
Add support for allowing images with different number of channels in the channel drop transform.
Add documentation for the RT-DETRv2 models.
Add warning for situations where the number of steps is below the recommendation for DINOv2.

Changed¶

The callback DeviceStatsMonitor is now disabled by default.
Replace epoch-based warmup hyperparameters with iteration-based ones.

Removed¶

Remove a note and specific recommendation for using DINOv2 with pretrained weights in the documentation.

[0.8.1] - 2025-06-23¶

Added¶

Add student_freeze_backbone_epochs option to DINOv2 method to control how many epochs the student backbone is frozen during training. We suggest setting it to 1 when starting from DINOv2 pretrained weights. See the DINOv2 documentation for more information.
Add channel drop transform.
Add option to load multi-channel images with LIGHTLY_TRAIN_IMAGE_MODE="UNCHANGED".
Add option to reuse memmap dataset file via environment variable: LIGHTLY_TRAIN_MMAP_REUSE_FILE=True.

[0.8.0] - 2025-06-10¶

Added¶

DINOv2 pretraining is now available with the method="dinov2" argument. The method is in beta and further improvements will be released in the coming weeks. See the DINOv2 documentation for more information.
Support for Torchvision ShuffleNetV2 models.
RT-DETR has now an integrated model wrapper.

Changed¶

The Ultralytics YOLO tutorial now highlights better how to use YOLO with LightlyTrain.

Deprecated¶

The resume parameter in the train command is deprecated in favor of resume_interrupted and will be removed in a future release. The new parameter behaves the same as the old one but is more explicit about its purpose. See the documentation for more information.

[0.7.0] - 2025-05-26¶

Added¶

Add Distillation v2 method that achieves higher accuracy and trains up to 3x faster than Distillation v1. The new method is selected as default by LightlyTrain with method="distillation".
Add MLflow logger to enable system and model metric logging.
Add support for lists of files and folders as input to the embed and train commands.
Add faster dataset initialization (SLURM and Windows).
Add configurable periodic model export.
Add training precision option float32_matmul_precision.
Add tutorial: Train embedding models with LightlyTrain.

Changed¶

Distillation v1 is now selected with method="distillationv1".
All commands (embed, export, and train) now require keyword arguments as input.
Custom models now require the get_model method to be implemented.
Distillation methods now use the teacher model from the official DINOv2 implementation.
The RT-DETR example uses RT-DETRv2, imposing fewer constraints on package versions.

Removed¶

Dependency on the transformers library.

Fixed¶

num_workers="auto" now limits the number of workers to a maximum of 8 workers/GPU to avoid overloading systems with many CPU cores.

[0.6.3] - 2025-04-23¶

Added¶

Transforms and methods are now documented on dedicated pages.
Add version compatibility table to the documentation.

Fixed¶

Fix image size mismatch issue when using TIMM models and DINO.

[0.6.2] - 2025-04-09¶

Added¶

Document RF-DETR models.
Add frequently asked questions page.
Add Torchvision classification tutorial.
Add depth estimation tutorial.

Changed¶

Increase minimum Wandb version to 0.17.2 which contains fixes for numpy>=2.0 support.
Limit PyTorch version to <2.6. We’ll remove this limitation once PyTorch Lightning 2.6 is released.
Limit Python version to <3.13. We’ll remove this limitation once PyTorch supports Python 3.13.

Removed¶

Remove Albumentations versions 1.4.18-1.4.22 support.

[0.6.1] - 2025-03-31¶

Added¶

Add support for RFDETR models.
Document platform compatibility.

Changed¶

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

Removed¶

Remove DenseCL from the documentation.

[0.6.0] - 2025-03-24¶

Added¶

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.

Changed¶

The default pretraining method is now "distillation".
The default embedding format is now "torch".
The log messages in the train.log file are now more concise.

Fixed¶

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

[0.5.0] - 2025-03-04¶

Added¶

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.

Changed¶

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.

Fixed¶

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.

[0.4.0] - 2024-12-05¶

Added¶

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

Changed¶

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

Removed¶

Remove DenseCLDINO method
Remove DINO teacher_freeze_last_layer_epochs argument

[0.3.2] - 2024-11-06¶

Added¶

Log data loading and forward/backward pass time as data_time and batch_time
Batch size is now more uniformly handled

Changed¶

The custom model feature_dim property is now a method
Replace FeatureExtractor base class by the set of Protocols

Fixed¶

Datasets support symlinks again

[0.3.1] - 2024-10-29¶

Added¶

The documentation is now available at https://docs.lightly.ai/train
Support loading checkpoint weights with the checkpoint argument
Log resolved training config to tensorboard and WandB

Fixed¶

Support single-channel images by converting them to RGB
Log config instead of locals
Skip pooling in DenseCLDino

[0.3.0] - 2024-10-22¶

Added¶

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