Quick Start¶

Installation¶

pip install lightly-train[tensorboard]

Prepare Data¶

You can use any image dataset for training. No labels are required, and the dataset can be structured in any way, including subdirectories. If you don’t have a dataset at hand, you can download one like this:

git clone https://github.com/lightly-ai/dataset_clothing_images.git my_data_dir

See the data guide for more information on supported data formats.

Train¶

Once the data is ready, you can train the model like this:

Python

import lightly_train

lightly_train.train(
    out="out/my_experiment",            # Output directory
    data="my_data_dir",                 # Directory with images
    model="torchvision/resnet18",       # Model to train
    method="dino",                      # Self-supervised learning method
    epochs=100,                         # Number of epochs to train
    batch_size=128,                     # Batch size
)

Command Line

lightly-train train out="out/my_experiment" data="my_data_dir" model="torchvision/resnet18" method="dino" epochs=100 batch_size=128

Tip

Decrease the number of epochs and batch size for faster training.

This will pretrain a Torchvision ResNet-18 model using images from my_data_dir and the DINO self-supervised learning method. All training logs and checkpoints are saved to the output directory at out/my_experiment.

Once the training is complete, the out/my_experiment directory should contain the following files:

out/my_experiment
├── checkpoints
│   ├── epoch=99-step=123.ckpt          # Intermediate checkpoint
│   └── last.ckpt                       # Last checkpoint
├── events.out.tfevents.123.0           # Tensorboard logs
├── metrics.jsonl                       # Training metrics
└── train.log                           # Training logs

The most important file is out/my_experiment/checkpoints/last.ckpt, which contains the final model checkpoint.

While the trained model has already learned good representations of the images, it cannot yet make any predictions for tasks such as classification, detection, or segmentation. To solve these tasks, the model needs to be fine-tuned on a labeled dataset.

Export¶

Before the model can be fine-tuned, it needs to be exported. The export command makes sure that only the model weights needed for fine-tuning are exported from the LightlyTrain checkpoint:

Python

import lightly_train

lightly_train.export(
    out="my_exported_model.pth",                            # Exported model
    checkpoint="out/my_experiment/checkpoints/last.ckpt",   # LightlyTrain checkpoint
    part="model",                                           # Export ResNet-18 model with classification head
    format="torch_state_dict",                              # Export format
)

Command Line

lightly-train export out="my_exported_model.pth" checkpoint="out/my_experiment/checkpoints/last.ckpt" part="model" format="torch_state_dict"

Fine-Tune¶

Now the model is ready for fine-tuning! You can use your favorite library for this step. Below is a simple example using PyTorch:

import torch
from torch import nn, optim
from torch.utils.data import DataLoader
from torchvision import datasets, transforms, models

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor()
])
dataset = datasets.ImageFolder(root="my_data_dir", transform=transform)
dataloader = DataLoader(dataset, batch_size=16, shuffle=True, drop_last=True)

# Load the exported model
model = models.resnet18()
model.load_state_dict(torch.load("my_exported_model.pth", weights_only=True))

# Update the classification head with the correct number of classes
model.fc = nn.Linear(model.fc.in_features, len(dataset.classes))

criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)

print("Starting fine-tuning...")
num_epochs = 10
for epoch in range(num_epochs):
    for inputs, labels in dataloader:
        optimizer.zero_grad()
        outputs = model(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()
    print(f"Epoch [{epoch+1}/{num_epochs}], Loss: {loss.item():.4f}")

The output should show the loss decreasing over time:

Starting fine-tuning...
Epoch [1/10], Loss: 2.1686
Epoch [2/10], Loss: 2.1290
Epoch [3/10], Loss: 2.1854
Epoch [4/10], Loss: 2.2936
Epoch [5/10], Loss: 1.9303
Epoch [6/10], Loss: 1.9949
Epoch [7/10], Loss: 1.8429
Epoch [8/10], Loss: 1.9873
Epoch [9/10], Loss: 1.8179
Epoch [10/10], Loss: 1.5360

Congratulations! You just trained and fine-tuned a model using LightlyTrain!

Tip

LightlyTrain has integrated support for popular libraries such as Ultralytics and SuperGradients, which allow you to fine-tune the exported models directly from the command line.

Embed¶

Instead of fine-tuning the model, you can also use it to generate image embeddings. This is useful for clustering, retrieval, or visualization tasks. The embed command generates embeddings for all images in a directory:

Python

import lightly_train

lightly_train.embed(
    out="my_embeddings.pth",                                # Exported embeddings
    checkpoint="out/my_experiment/checkpoints/last.ckpt",   # LightlyTrain checkpoint
    data="my_data_dir",                                     # Directory with images
    format="torch",                                         # Embedding format
)

Command Line

lightly-train embed out="my_embeddings.pth" checkpoint="out/my_experiment/checkpoints/last.ckpt" data="my_data_dir" format="torch"

The embeddings are saved to my_embeddings.pth and can be loaded like this:

import torch

embeddings = torch.load("my_embeddings.pth")
embeddings["filenames"]     # List of filenames
embeddings["embeddings"]    # Tensor with embeddings with shape (num_images, embedding_dim)