Determined GAN

• 

  Determined GAN Architecture Proposal

• 

  Classifier accuracy between using only original MNIST VS 59K MNIST data + 1K Generated Image

• 

  Classifier accuracy 30k MNIST data + 30k Generated Image

Inspiration

In many fields there is a growing issue in data analysis, like data unavailability due to ethical and legislative restrictions, biases in data, etc. To address data unavailability, people are leveraging machine learning algorithms such as Generative Adversarial Networks (GAN) to augment data from existing samples. With many GAN algorithm it is practically impossible to predict which algorithm best fits the problem in question, unless many algorithms are tested.

What it does

In this hackathon we used Determined AI to accelerate GAN exploration by increasing the training speed, distributed training and help customer to pick the best GAN Architecture. We are proposing to support multiple GAN model templates on top of the determined AI stack, so as to accelerate the model training, and comparative analysis, find best model using advanced hyperparameter tuning.

How we built it

We ported CGAN (Conditional GAN) and CNN classifier in Determined AI template format.

Challenges we ran into

Multiple inputs were not accepted due to the TF API conformance. Due to which we had to some reworks.

Accomplishments that we're proud of

Successful run of CNN classifier using CGAN generated images.

What we learned

1. Learning and understanding of determined AI platform.
2. Replacing part of original data with generated image, this helps train the model with different structures or scenario of data than already available. Helping us to train model across different types of data and biased data etc.

What's next for Determined GAN

1. Encourage our HPE business like HPC and GreenLake to explore the benefits of Determined AI.
2. Also Implement other GAN models in Determined AI

Built With

• determinedai
• keras
• python
• tensorflow
• yaml