VictorAtPL / CIFAR-10_GAN_Tensorflow

Generative Adversarial Networks for CIFAR-10 dataset written as part of my MSc in Data Science degree.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Generative Adversarial Network for CIFAR-10 dataset

  • project implemented as part of the Deep Learning Methods course attended on MSc degree of Data Science at the Faculty of Mathematics and Information Sciences in Warsaw University of Technology,
  • the implemented architectures are DC-GAN and MLP-GAN,
  • written in TensorFlow r1.13.

Prerequisites to run

  • Google Colab access,
  • Google Storage Bucket with public access.

How to run via Google Colab

There are three files for Google Colab, which are described below:

  • Google_Colab_Train_eval_predict.ipynb - main notebook which allows to train GAN model, make small evaluation and generate images based on noise,
  • Google_Colab_Most_similar.ipynb - after training model you can use this notebook to find most similar images in training set to that which your model can generate,
  • Google_Colab_Latent_space_interpolation.ipynb - this notebook can ensure you that the GAN model has learnt some deep knowledge about the training set examples.

In order to run any of script please do following:

  1. Open Google_Colab_*.ipynb in Google Colab by clicking any of these links: train_eval_predict, most_similar, latent_space_interpolation,
  2. set Google Colab environment to support TPU computation,
  3. modify variables placed in the first cell of each notebook,
  4. run all cells,
  5. after all cells execution, please take a look either on the cells output or the Google Storage Bucket, which you provided to store model's checkpoints and generated images in.

Results

Animation showing generated images during training of MLP-GAN and DC-GAN architectures:

  1. MLP-GAN:

    Generated images during training for MLP-GAN

  2. DC-GAN:

    Generated images during training for DC-GAN

Image generated for 1000-epoch trained DC-GAN using Google_Colab_Latent_space_interpolation.ipynb:

Latent space showcase for 1000-epoch trained DC-GAN

Image generated for 1000-epoch trained DC-GAN using Google_Colab_Most_similar.ipynb:

Most similar images showcase for 1000-epoch trained DC-GAN

For more results and diagrams of architectures please read Project_Final_Report_PL.pdf file.

About

Generative Adversarial Networks for CIFAR-10 dataset written as part of my MSc in Data Science degree.

License:MIT License


Languages

Language:Jupyter Notebook 100.0%