Sample-Augment: Tool for Synthetic Data Augmentation

This is a Master's Thesis Project exploring Sampling Strategies for Synthetic Data Augmentation utilizing Generative Models, such as Generative Adversarial Networks (GANs). The objective is to augment a dataset synthetically to improve performance on downstream tasks such as image classifcation.

As of June 2023, the foundation of the project is established, including the basic pipeline and modules for data processing and model training. The next steps involve training a Generator in the form of StyleGAN2-ADA and implementing different sampling strategies based on it.

Project Structure

The main Python package for this project is located in sample_augment/. It contains two main components:

• core/: A general framework for constructing reproducible and serializable pipelines tailored to Machine Learning needs. Its main features are an interface for pipeline steps with dependency resolution, centralized Config and intermediary Data management.

• The other modules are specific to the thesis:

  • data/: Scripts for data processing steps, such as preprocessing, train-test split. It also contains an adaptor for the GC10-DET dataset which gets covered in this thesis.
  • models/: Scripts for training and evaluating the classifier and generator models.

Installation

First, initialize an Anaconda environment. You can use one of the environment.yml files to replicate a working environment.

• environment_cpu.yml
  • Python 3.10 Environment for local development without CUDA installed.
  • Can use this for anything but training/evaluating models
  • This environment was tested on a local Ubuntu 22 installation.
• environment_gpu.yml
  • Python 3.10 environment with CUDA installed
  • This environment was used for running experiments (Classifier training, etc.) on the IPT DS machine ipt-d-0432
  • This can NOT run StyleGAN
  • TODO provide this env file
• environment_stylegan.yml
  • StyleGAN depends on a specific version of PyTorch and on Python 3.7
  • not every method in sampling_aug is compatible with this configuration, so only use this when training/evaluating StyleGAN
  • TODO provide this env file Enter this directory and call pip install -e .. Pytorch installation instructions: TODO

License

The project is licensed under MIT.

Note that this repository contains a submodule pointing to a fork of StyleGAN2-ADA, which is released under a license for non-commercial use.