Command-line tool

The LightlySSL framework provides you with a command-line interface (CLI) to train self-supervised models and create embeddings without having to write a single line of code.

You can also have a look at this video to get an overview of how to work with the CLI.

Check the installation of LightlySSL

To see if the LightlySSL command-line tool was installed correctly, you can run the following command which will print the version of the installed LightlySSL package:

lightly-version

If LightlySSL was installed correctly, you should see something like this:

lightly version 1.1.4

Crop Images using Labels or Predictions

For some tasks, self-supervised learning on an image level has disadvantages. For example, when training an object detection model we care about local features describing the objects rather than features describing the full image.

One simple trick to overcome this limitation, is to use labels or to use a pre-trained model to get bounding boxes around the objects and then cropping the objects out of the image.

We can do this using the lightly-crop CLI command. The CLI command crops objects out of the input images based on labels and copies them into an output folder. The new folder consists now of the cropped images.

# Crop images and set the crop to be 20% around the bounding box
lightly-crop input_dir=images label_dir=labels output_dir=cropped_images crop_padding=0.2

# Crop images and use the class names in the filename
lightly-crop input_dir=images label_dir=labels output_dir=cropped_images \
             label_names_file=data.yaml

The labels should be in the yolo format. For each image you should have a corresponding .txt file. Each row in the .txt file has the following format:

  • class x_center y_center width height

0 0.23 0.14 0.05 0.04
1 0.43 0.13 0.12 0.08

An example for the label names .yaml file:

names: [cat, dog]

You can use the output of the lightly-crop command as the input_dir for your lightly-train command.

Training and Embedding in a Go – Magic

Lightly-magic is a singular command for training a self-supervised model and use it to compute embeddings

  • To start with, we need to input the directory of the dataset, pass it to input_dir.

  • It requires information on the number of epochs to perform, set trainer.max_epochs.

  • To use a pre-trained model, simply set trainer.max_epochs=0.

  • The embedding model is used to embed all images in the input directory and saves the embeddings in a CSV file.

  • To set a custom batch size just set the value to loader.batch_size for the same.

# Embed images from an input directory
# Setting trainer.max_epochs=10 trains a model for 10 epochs.
# loader.num_workers=8 specifies the number of cpu cores used for loading images.
lightly-magic input_dir=data_dir trainer.max_epochs=10 loader.num_workers=8


# To use a custom batch size, pass the batch size to loader.batch_size parameter
# updating the previous example by passing value for loader.batch_size
lightly-magic input_dir=data_dir trainer.max_epochs=10 loader.batch_size=128 \
loader.num_workers=8

Train a model using the CLI

Training a model using default parameters can be done with just one command. Let’s assume you have a folder of cat images named cat and want to train a model on it. You can use the following command to train a model and save the checkpoint:

# train a model using default parameters
lightly-train input_dir=cat

# train a model for 5 epochs
lightly-train input_dir=cat trainer.max_epochs=5

# continue training from a checkpoint for another 10 epochs
lightly-train input_dir=cat trainer.max_epochs=10 checkpoint=mycheckpoint.ckpt

# continue training from the last checkpoint
lightly-train input_dir=cat trainer.max_epochs=10 \
              checkpoint=$LIGHTLY_LAST_CHECKPOINT_PATH

# train with multiple gpus
# the total batch size will be trainer.gpus * loader.batch_size
lightly-train input_dir=data_dir trainer.gpus=2

The path to the latest checkpoint you created using the lightly-train command will be saved under an environment variable named LIGHTLY_LAST_CHECKPOINT_PATH. This can be useful for continuing training or for creating embeddings from a checkpoint.

For a full list of supported arguments run

lightly-train --help

You can get an overview of the various CLI parameters you can set in Default Settings.

Create embeddings using the CLI

Once you have a trained model checkpoint, you can create an embedding of a dataset.

# use pre-trained models provided by Lighly
lightly-embed input_dir=cat

# use custom checkpoint
lightly-embed input_dir=cat checkpoint=mycheckpoint.ckpt

# use the last checkpoint you created
lightly-embed input_dir=cat checkpoint=$LIGHTLY_LAST_CHECKPOINT_PATH

The path to the latest embeddings you created using the lightly-embed command will be saved under an environment variable named LIGHTLY_LAST_EMBEDDING_PATH.

The embeddings.csv file should look like the following:

embeddings_example.csv

filenames

embedding_0

embedding_1

embedding_2

embedding_3

labels

101053-1.jpg

-51.535

-2.325

-21.750

78.265

0

101101-1.jpg

-67.958

-2.800

-28.861

103.812

0

101146-1.jpg

-59.831

-2.719

-25.413

90.945

0

Download data using the CLI

You can download a dataset with a given tag from the Lightly Platform using the following CLI command. The CLI provides you with three options:

  • Download the list of filenames for a given tag in the dataset.

  • Download the images for a given tag in the dataset.

  • Copy the images for a given tag from an input directory to a target directory.

The last option allows you to very quickly extract only the images in a given tag without the need to download them explicitly.

# download a list of files
lightly-download tag_name=my_tag_name dataset_id=your_dataset_id token=your_token

# download the images and store them in an output directory
lightly-download tag_name=my_tag_name dataset_id=your_dataset_id token=your_token \
                 output_dir=path/to/output/dir

# copy images from an input directory to an output directory
lightly-download tag_name=my_tag_name dataset_id=your_dataset_id token=your_token \
                 input_dir=path/to/input/dir output_dir=path/to/output/dir

Breakdown of lightly-magic

If you want to break the lightly-magic command into separate steps, you can use the following:

# lightly-magic command
lightly-magic input_dir=data_dir
# equivalent breakdown into single commands

# train the embedding model
lightly-train input_dir=data_dir
# embed the images with the embedding model just trained
lightly-embed input_dir=data_dir checkpoint=$LIGHTLY_LAST_CHECKPOINT_PATH