More Related Content
Similar to Exploring GAN Architecture using Determined AI (20)
More from madhucharis (20)
Recently uploaded (20)
Exploring GAN Architecture using Determined AI
- 1. Determined GAN RECOMMENDATION SYSTEM FOR MULTIPLE GAN ARCHITECTURE USING DETERMINED AI TEAM MEMBERS: • G S, ARCHANA • RR, SOUMYA • JOLLY, JINU MARIAM • DHIMATE, VIKRANT MAH • NAMBIAR, DIVYA C • S, MADHUSOODHANA CHARI
- 2. 2 PROJECT OBJECTIVE • In this hackathon we used Determined AI to accelerate customer work by increasing the training speed, distributed training and in turn help customer to pick the right GAN Architecture.
- 3. SCOPE OF WORK • We propose a solution having support for multiple GAN models using determined AI • This will help user to bring in their use cases and run it across multiple GAN model with the following benefits o Helps to train model faster and obtain comparative analysis of the solution across model with distributed training. o Find better models with advanced hyperparameter tuning. o This helps customer to take a well informed and efficient solution for their problem.
- 4. HACKATHON IMPLEMENTATION • Dataset used : • MNIST dataset of handwritten digits (training data – 60000 images and test data – 10000 images). • CGAN generated MNIST data (1k images)
- 5. CONFIDENTIAL LESSONS LEARNED • We started our GAN exploration using goggle Collab. With Determined AI, we were able to run parallel experiments and able to get better visualisation of metrics data. • A single configuration file to control the inputs like hyperparameter helps in better control of the model executions. • During our GAN exploration, there was existing DCGAN example, we found that DCGAN wouldn't let us choose the class of digits we were trying to generate. To be able to control what we want, we need to condition the GAN output on a semantic input, such as the class of an image. • Those related changes were identified and added which helped us to achieve the conditional GAN algorithm. • The DCGAN has implementation for model training. To support augmentation, exporting and storing images using CGAN, we referred the documentation to make it work with project requirements. • Thanks to DetAI team for helping us identify and resolve few of the issues faced during our exploration journey. • We showcased our exploration of GAN in determined AI platform in various events like HPE sustainability Hackathon, University Relationship Program etc. 5
- 7. OBSERVATIONS Data Validation Accuracy Original MNIST 60K 0.9912 59K MNIST + 1K CGAN 0.9919 30K MNIST + 30K CGAN 0.9860 Fig. Accuracy graph of classifier run for data with 30k MNIST + 30k CGAN data
- 8. CONFIDENTIAL FUTURE SCOPE • Explore the implementation of other GAN models in Determined AI • Experiment with various types of data on the GAN stack to identify the best model for the use case • Explore work on hyperparameter tunning • Explore the distributed training option for the GAN flow 8
- 9. ACKNOWLEDGEMENTS DETERMINED AI TEAM FOR VALUABLE SUPPORT, • GHODGAONKAR, ISHA • ILIA • RYAN • HIROKI ASAKURA
- 10. THANK YOU