mlpack  blog
Implementing Essential Deep Learning Modules - Week 5

Implementing Essential Deep Learning Modules - Week 5

Shikhar Jaiswal, 18 June 2018

Phase I evaluations came in and I'm glad that Marcus is happy with what we have been able to achieve.

The DCGAN code is now completely functional, and has been debugged for the CelebA dataset as well. However, I have been unable to prepare the dataset for CelebA in our preferable format, because it is memory intensive, and always crashes my system.

This week we got some promising results with the DCGAN code as well on the 10,000 image subset and the full 70,000 image set of MNIST:

10,000 Images

70,000 Images

After discussing with Marcus, we decided to merge the DCGAN code in its present state, which would be done shortly. Also, this week I implemented the support for mini-batches in the convolutional layers, and the GAN implementation. The PR is also completely debugged and ready to get merged. I really couldn't thank Marcus enough for his instant reviews and debugging help.

The only thing remaining to be done now for optimization of our code is the implementation of the GANOptimizer class, which would allow us to separate our optimizers for the generator and the discriminator. Also, this week, I'll start working on the Wasserstein GAN implementation. We are right on track to wrap up all of the planned work for the summer by the end of Phase II, after which, I can take up the work pertaining to the additional goals planned.