Overview / Usage

Deep learning researchers from across the world have produced amazing results using GANs, since they were introduced back in 2014 and I have also had interest in them.

So, to test my knowledge of GANs and also to learn more about them I decided to participate in this competition on Kaggle. The task was to train GANs on the Standford Dogs Dataset (which consisted of 20000 images of dogs from 120 breeds) and submit the generated dog images for evaluation.

Methodology / Approach

The dataset was also provided with the information about the bounding boxes so I used it to crop the images and keep only the information about dogs and not the surroundings.

Then I trained and tested various GAN models starting with DCGAN, WGAN, GAN with RaLS loss and CGAN. I used Intel DevCloud to train my model since it comes with Intel optimized Python and Tensorflow and is very easy to setup. I gained the most accurate results from CGAN so I decided to go with it and fine tun it further. There are various tips and tricks for training GANs that I learnt while working on this project, some of which are:

• Using BatchNorm and Dropout layers
• Using small batch size
• Trying out different custom loss functions
• Tuning the learning rate of both Discriminator and Generator so that our model does not suffer from Mode Collapse.
• Using LeakyRELu

When I was satisfied with my model's output is submitted my results to Kaggle.

Technologies Used

Language: Python

Deep learning Frameworks: Keras, Tensorflow

Platform: Intel DevCloud and Kaggle Kernels

Visualization: Matplotlib

​

Repository

https://www.kaggle.com/techytushar/dog-dcgan