Documentation

Pyraug

This library provides a way to perform Data Augmentation using Variational Autoencoders in a reliable way even in challenging contexts such as high dimensional and low sample size data.

Installation

To install the library from pypi.org run the following using pip

$ pip install pyraug

or alternatively you can clone the github repo to access to tests, tutorials and scripts.

$ git clone https://github.com/clementchadebec/pyraug.git

and install the library

$ cd pyraug
$ pip install .

Augmenting your Data

In Pyraug, a typical augmentation process is divided into 2 distinct parts:

1. Train a model using the Pyraug's TrainingPipeline or using the provided scripts/training.py script
2. Generate new data from a trained model using Pyraug's GenerationPipeline or using the provided scripts/generation.py script

There exist two ways to augment your data pretty straightforwardly using Pyraug's built-in functions.

Using Pyraug's Pipelines

Pyraug provides two pipelines that may be used to either train a model on your own data or generate new data with a pretrained model.

note: These pipelines are independent of the choice of the model and sampler. Hence, they can be used even if you want to access to more advanced features such as defining your own autoencoding architecture.

Launching a model training

To launch a model training, you only need to call a TrainingPipeline instance. In its most basic version the TrainingPipeline can be built without any arguments. This will by default train a RHVAE model with default autoencoding architecture and parameters.

>>> from pyraug.pipelines import TrainingPipeline
>>> pipeline = TrainingPipeline()
>>> pipeline(train_data=dataset_to_augment)

where dataset_to_augment is either a numpy.ndarray, torch.Tensor or a path to a folder where each file is a data (handled data formats are .pt, .nii, .nii.gz, .bmp, .jpg, .jpeg, .png).

More generally, you can instantiate your own model and train it with the TrainingPipeline. For instance, if you want to instantiate a basic RHVAE run:

>>> from pyraug.models import RHVAE
>>> from pyraug.models.rhvae import RHVAEConfig
>>> model_config = RHVAEConfig(
...    input_dim=int(intput_dim)
... ) # input_dim is the shape of a flatten input data
...   # needed if you did not provide your own architectures
>>> model = RHVAE(model_config)

In case you instantiate yourself a model as shown above and you did not provide all the network architectures (encoder, decoder & metric if applicable), the ModelConfig instance will expect you to provide the input dimension of your data which equals to n_channels x height x width x .... Pyraug's VAE models' networks indeed default to Multi Layer Perceptron neural networks which automatically adapt to the input data shape.

note: In case you have different size of data, Pyraug will reshape it to the minimum size min_n_channels x min_height x min_width x ...

Then the TrainingPipeline can be launched by running:

>>> from pyraug.pipelines import TrainingPipeline
>>> pipe = TrainingPipeline(model=model)
>>> pipe(train_data=dataset_to_augment)

At the end of training, the model weights models.pt and model config model_config.json file will be saved in a folder outputs/my_model/training_YYYY-MM-DD_hh-mm-ss/final_model.

Important: For high dimensional data we advice you to provide you own network architectures and potentially adapt the training and model parameters see documentation for more details.

Launching data generation

To launch the data generation process from a trained model, run the following.

>>> from pyraug.pipelines import GenerationPipeline
>>> from pyraug.models import RHVAE
>>> model = RHVAE.load_from_folder('path/to/your/trained/model') # reload the model
>>> pipe = GenerationPipeline(model=model) # define pipeline
>>> pipe(samples_number=10) # This will generate 10 data points

The generated data is in .pt files in dummy_output_dir/generation_YYYY-MM-DD_hh-mm-ss. By default, it stores batch data of a maximum of 500 samples.

Retrieve generated data

Generated data can then be loaded pretty easily by running

>>> import torch
>>> data = torch.load('path/to/generated_data.pt')

Using the provided scripts

Pyraug provides two scripts allowing you to augment your data directly with commandlines.

note: To access to the predefined scripts you should first clone the Pyraug's repository. The following scripts are located in scripts folder. For the time being, only RHVAE model training and generation is handled by the provided scripts. Models will be added as they are implemented in pyraug.models

Launching a model training:

To launch a model training, run

$ python scripts/training.py --path_to_train_data "path/to/your/data/folder" 

The data must be located in path/to/your/data/folder where each input data is a file. Handled image types are .pt, .nii, .nii.gz, .bmp, .jpg, .jpeg, .png. Depending on the usage, other types will be progressively added.

At the end of training, the model weights models.pt and model config model_config.json file will be saved in a folder outputs/my_model_from_script/training_YYYY-MM-DD_hh-mm-ss/final_model.

Launching data generation

Then, to launch the data generation process from a trained model, you only need to run

$ python scripts/generation.py --num_samples 10 --path_to_model_folder 'path/to/your/trained/model/folder' 

The generated data is stored in several .pt files in outputs/my_generated_data_from_script/generation_YYYY-MM-DD_hh_mm_ss. By default, it stores batch data of 500 samples.

Important: In the simplest configuration, default configurations are used in the scripts. You can easily override as explained in documentation. See tutorials for a more in depth example.

Retrieve generated data

Generated data can then be loaded pretty easily by running

>>> import torch
>>> data = torch.load('path/to/generated_data.pt')

Getting your hands on the code

To help you to understand the way Pyraug works and how you can augment your data with this library we also provide tutorials that can be found in examples folder:

• getting_started.ipynb explains you how to train a model and generate new data using Pyraug's Pipelines
• playing_with_configs.ipynb shows you how to amend the predefined configuration to adapt them to your data
• making_your_own_autoencoder.ipynb shows you how to pass your own networks to the models implemented in Pyraug

Dealing with issues

If you are experiencing any issues while running the code or request new features please open an issue on github

Citing

If you use this library please consider citing us:

@article{chadebec2022data,
  title={Data augmentation in high dimensional low sample size setting using a geometry-based variational autoencoder},
  author={Chadebec, Cl{\'e}ment and Thibeau-Sutre, Elina and Burgos, Ninon and Allassonni{\`e}re, St{\'e}phanie},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2022},
  publisher={IEEE}
}

Credits

Logo: SaulLu