Inspiration
The project started from the question of how to integrate actual human feedback into the process of model refinement in the background of responsible AI.
What it does
Hero Academy is a StyleGAN based superhero generator that is integrated into a well built web platform with many features: easy authentication, automated superhero generator, ML model improvement based on user feedback, My Gallery - where users can save their favorite superheroes and check them back at any time, fast response given its caching mechanisms.
Challenges we ran into
Ensuring stability during training. The training instance suffered a lot of interruptions. For this we developed a component called Training Manager that restarts training from the moment of an interruption and continues training from checkpoints of progress.
Another challenge was integrating the old data used for training with the new data for model improvement and updating the model behind the scenes in a way subtle to users.
What we're proud of
Quality of generated images
Training Manager able to ensure that training will be finalized in case of instance malfunctions
Feedback loop to periodically retrain the model with new data to improve the quality of the images and increase diversity among superheroes
Caching mechanisms to ensure great user experience in terms of fast response times.
What we learned
New cloud services: Azure services for ML, CI/CD and storage.
New technologies: StyleGAN2-ADA insights
What's next
Opening our product to the public and integrate it into our website for users to provide feedback related to improvements so that we can further refine it.
Built with #Python #JavaScript #TypeScript #CSS #SQ #Django #ReactJS #TensorFlow #Azure #NodeJS #PostgreSQL #AzureBlobStorage #AzureCacheForRedis #AzureML #AzureDevOps
For more information check this: https://responsibleai.devpost.com/
Comments