Deep Convolutional GAN (DCGAN)

This project is part of the course IDATT2502 - Applied Machine Learning at the Norwegian University of Science and Technology (NTNU) by:

• Carl Johan Gützkow - GitHub
• Eirik Elvestad - GitHub
• Tomas Beranek - GitHub

Overview

In this project, we explore various applications and metrics for Generative Adversary Networks.

Installation and Execution

1. Clone the repository:

   git clone https://github.com/CJGutz/DCGAN-IDATT2502.git

   OR

   git clone [email protected]:CJGutz/DCGAN-IDATT2502.git

2. Download the required packages:

   pip install -r requirements.txt

3. Add necessary folders
   
   The structure beneath should be followed for the different datasets

    DCGAB-IDATT2502
   ├── datasets
   │   ├── figures
   │   └── model
   
   

4. To train a DCGAN model with default values:

      python3 Entrypoint.py dcgan MNIST --channels 1

   Models are saved automatically(if you don't want to save the models use no-model-save):

    python3 Entrypoint.py dcgan MNIST --channels 1 --no-model-save

   To load models use:

   python3 Entrypoint.py dcgan MNIST --channels 1 --load-model

Datasets

Specify your preferred dataset directory or zip path as an argument to use it. For torchvision datasets, refer to DatasetLoader.py for supported datasets.

Example Script:

python3 Entrypoint.py dcgan datasets/celeba-dataset -c 3 -i 64 -l 3 -b 128 -e 5 -lr 0.0002 -b1 0.5 --ndf 64 --ngf 64 --nz 100

Tested Datasets

These datasets are already a part of the program and can be directly accessed from the command line using the specific titles listed in the table below.

Datasets	Description
MNIST	Handwritten digits (0-9) in 28x28 grayscale images
celeba-dataset	200,000+ celebrity images with 40 attribute labels each
FashionMNIST	28x28 grayscale images of 10 fashion categories
CIFAR10	60,000 32x32 color images across 10 classes

If the celeba dataset is to be used follow the steps beneath

1. Download the zipfile from "https://drive.google.com/drive/folders/0B7EVK8r0v71pTUZsaXdaSnZBZzg?resourcekey=0-rJlzl934LzC-Xp28GeIBzQ"

   
   

2. Place the zipfile into your dataset folder or directory The structure beneath should be followed for the different datasets

    DCGAB-IDATT2502
   ├── datasets
   │   ├── figures
   │   ├── model
   │   └── celeba.zip
   
   
   

(back to top)