cDCGAN on MNIST & Fashion MNIST w/ Tensorflow & Keras

This is the implementation of conditional deep convolutional generative adversarial network (cDCGAN) on the MNIST and Fashion MNIST Dataset using Tensorflow and Keras. The implementation is based on this code by @gaborvescei. The example code by him runs on the MNIST Dataset. I've added the implementation for Fashion MNIST dataset on his code.

Here is an abstract model of this cDCGAN implementation:

Setup

Install Tensorflow, Keras, Numpy, Matplotlib, os, cv2, math, subprocess.
All of which can be installed using Anaconda
You also need to install Python and an editor to run the program, in my case, I use PyCharm

Installing "Swiss-Army-Tensorboard" is also necessary. It's a modification of the usual Tensorboard by @gaborvescei
Run this on the editor to install it: pip install git+https://github.com/gaborvecsei/Swiss-Army-Tensorboard.git
This is used to log the model's loss during training.

Run

For the MNIST experiment, open the files in CDCGAN-Keras-MNIST/cdcgan
For the Fashion MNIST experiment, open the files in CDCGAN-Keras-FMNIST/cdcgan
There are 4 important files in each of those folders:

• cdcgan_models.py this contains the code for the model definition (The Generator & Discriminator)
• cdcgan_utils.py this contains the code for helper functions, such as saving the generated images
• cdcgan_train.py this contains the code for training the model and saving that trained model
• test_model.ipynb this contains the code for testing the generative capability of the trained model

I've included the model that I've trained in CDCGAN-Keras-MNIST/models/weights, for the MNIST experiment, and
CDCGAN-Keras-FMNIST/models/weights, for the Fashion MNIST experiment. In those folders, you'll find these files:

• generator.h5, it's the generator model for the MNIST experiment.
• discriminator.h5, it's the discriminator model for the MNIST experiment.
• generator_fashion_mnist.h5, it's the generator model for the Fashion MNIST experiment.
• discriminator_fashion_mnist.h5, it's the discriminator model for the Fashion MNIST experiment.

The model is too big to upload on Github, so I uploaded it into Google Drive, here are the links:

• generator.h5 & discriminator.h5
  https://drive.google.com/open?id=1OO1GBK7V5oAWK8b9p0BeRM0KET12OUNl
• generator_fashion_mnist.h5 & discriminator_fashion_mnist.h5
  https://drive.google.com/open?id=1t_K4kuL14qmqfGfim7J7xxp_A-VkJVd7

Test on those models can be done by running the test_model.ipynb file on Jupyter Notebook or Google Collab.
Creating and training model from scratch is also possible by running the cdcgan_train.py file. It'll train a new model and overwrite the previously saved model.

Experiment Results

The model was run for 100 epochs for both the MNIST and Fashion MNIST Dataset.
The gifs below show the generated images from the model from the first epoch until the 100th epoch.

MNIST

The Loss (MNIST)

The loss is visualized using the "Swiss-Army-Tensorboard".
Running the "Swiss-Army-Tensorboard" is the same with running the usual Tensorboard.
Follow these steps to run the Tensorboard:

• Open Anaconda Command Prompt.
• Go to CDCGAN-Keras-MNIST/models.
• Type in tensorboard --logdir=logs/ --host=127.0.0.1.
• Open localhost:6006 on your browser, and it'll open up Tensorboard.

On the Tensorboard, there will be graphs showing the Generator and Discriminator loss when training on the MNIST Dataset. This graph is created based on the log files. The log files is created automatically by the program when the model is being trained. I've included the log of my model training in the CDCGAN-Keras-MNIST/models/logs/loss folder. Here are the graphs created from the training that I've done on my models:

Generator Loss

Discriminator Loss

Fashion MNIST

The loss (Fashion MNIST)

The loss is visualized using the "Swiss-Army-Tensorboard" as well. To run it follow the same step as before, but change the directory to CDCGAN-Keras-FMNIST/models instead of CDCGAN-Keras-MNIST/models.

Generator Loss

Discriminator Loss