Build Better Models Faster with Self-supervised Pre-training

LightlyTrain is a model training plug-in for self-supervised learning. It helps to pre-train models, generate embeddings, and backbones with 1 line of code

Get a demo

What are the benefits of self-supervised pre-training?

Models were pre-trained on the full COCO training set without using labels and then fine-tuned on 10% of the COCO train set using labels

36% higher mAP with pre-training

Better generalization

Reduce need of lots of labeled data

Higher accuracy

Get more insights

Self-supervised learning enables models to extract meaningful features from vast amounts of unlabeled images

Embeddings obtained from self-supervised pre-training are robust and generalizable, improving performance in various computer vision task

Reduce the reliance on large labeled datasets and extensive training

Case Study

10% of the data yields already >80% of total dataset accuracy

Which version is
right for you?

Lightly SSL

Open-source version for research and individuals with community support

Low level building blocks for research
SOTA Self-supervised methods
Compatible with Pytorch and Pytorch Lightning

GitHub

LightlyTrain

Enterprise version for embedding model endpoints and co-shaping the roadmap

Off-the-shelf modules for pre-training, optimized for downstream tasks such as object detection, classification, and segmentation.
Easy to use interface for training embedding models and generating embeddings with a single command
Automatic SSL method selection
Export multiple model formats
Available as Python or Docker Version
Tailoring & Hands-on Support