How lightly-ai/lightly works
LightlySSL is a PyTorch-oriented framework for self-supervised computer-vision training and image embeddings. A user can start training through the Python API or CLI: configuration reaches the training entry point, which assembles data, transforms, model and loss, then trains an embedding module. Embedding inference produces vectors, labels and filenames; benchmarking utilities support downstream evaluation. The repository also provides video datasets, cropping tools, and modular model, loss and transform building blocks. Several feature links are documented or implied by the tree rather than shown in sampled call sites, so uncertain links are omitted or marked dashed.
Main entry points: train_cli.py Training CLI · embed_cli.py Embedding CLI · embedding.py Embedding trainer · dataset.py Image dataset
Components 16
User entry points
- Python API
lightly/core.py - Training CLI
lightly/cli/train_cli.py - Embedding CLI
lightly/cli/embed_cli.py
Data and transforms
- Image dataset
lightly/data/dataset.py - Video dataset
lightly/data/_video.py - View transforms
lightly/transforms
Self-supervised learning
- Model heads
lightly/models/modules/heads.py - Backbones
lightly/models/resnet.py - Batch normalization
lightly/models/batchnorm.py - SSL losses
lightly/loss
Embedding and evaluation
- Embedding trainer
lightly/embedding/embedding.py - Benchmarking
lightly/utils/benchmarking/knn_classifier.py
Data utilities
- Crop command
lightly/cli/crop_cli.py - Bounding-box cropping
lightly/utils/cropping/crop_image_by_bounding_boxes.py - Bounding boxes
lightly/utils/bounding_box.py
Other components
- User
Connections 21
- User → Python API: calls
- User → Training CLI: runs
- User → Embedding CLI: runs
- Python API → Training CLI: invokes training
lightly/core.py - Python API → Embedding CLI: invokes embedding
lightly/core.py - Training CLI → Image dataset: configures data
lightly/cli/train_cli.py - Training CLI → View transforms: uses
lightly/cli/train_cli.py - Training CLI → Model heads: configures model
- Training CLI → Backbones: configures model
lightly/cli/train_cli.py - Training CLI → Batch normalization: selects normalization
lightly/cli/train_cli.py - Training CLI → SSL losses: configures criterion
lightly/cli/train_cli.py - Training CLI → Embedding trainer: trains through
- Image dataset → Video dataset: loads video
lightly/data/dataset.py - Embedding CLI → Image dataset: loads data
lightly/cli/embed_cli.py - Embedding CLI → View transforms: applies
lightly/cli/embed_cli.py - Embedding CLI → Embedding trainer: computes embeddings
- Embedding trainer → Image dataset: reads samples
lightly/embedding/embedding.py - Embedding trainer → Benchmarking: uses utilities
lightly/embedding/embedding.py - User → Crop command: runs
- Crop command → Bounding-box cropping: calls
lightly/cli/crop_cli.py - Bounding-box cropping → Bounding boxes: uses
lightly/utils/cropping/crop_image_by_bounding_boxes.py