Hangar, PyTorch & RedisAI; missing pieces of a complete deep learning workflow

Sherin Thomas (~hhsecond)


Managing DL workflow is always a nightmare. Problems include handling the scale, efficient resource utilization, version controlling the data. With the highly optimized RedisAI, super flexible PyTorch and heavily organized Hangar, all the sleepless nights are stories of the past. Even with the advancement, the community had made in the past couple of years, problems of a DL engineer starts right from the beginning when they think about version controlling the data and model. None of the toolset available right now could make a platform "git" had provided for programmers years ago. With the entry of Hangar and it's python APIs, we are now moving ahead of the game. Having a version controlling system like hangar in place, DL folks are still struggling with production deployment with their framework of choice. PyTorch is the most flexible and easy framework that all deep learning developers had loved. But without a plug n play deployment toolkit, PyTorch always suffered to attract people who want to deploy their model to production. Meanwhile, RedisAI is trying to solve the problem every deep learning engineer faces especially when they try to scale. Making sure the production environment is highly available is probably the most daunting task DevOps experts worried especially when they have a deep learning service in production. I'll be presenting the killer combo of PyTorch and RedisAI and explain how can we develop super optimized DL model with LibTorch but without losing the flexibility provided by PyTorch and how to ship it to production without even worried about writing a wrapper service for the model. Not just that, I'll be showing how to put this deployment into a multi-node cluster and make sure you have 100% availability always. In a nutshell, the talk covers a deep learning workflow by introducing three toolkits for the user that makes the whole pipeline seamless.

Talk outline

  • Introduction to modern world deep learning workflow
  • Problems with the existing toolkits in the ecosystem
  • Solving the problems
    • Hangar: Building reliable data pipeline
      • Saving data as tensors
      • Managing conflicts, branching, merging
      • CLI and Python APIs
    • PyTorch: Ideating, designing and building the model architecture
      • First steps with PyTorch
      • Data loader for loading data from the hangar repository
      • Training and exporting the model using redisai-py
    • RedisAI: Trained models can be pushed to a highly scalable Redis environment
      • Loading exported pytorch model using redisai-py
      • Serving the model at scale
      • High availability using sentinel


The audience should have basic python knowledge and a brief understanding of deep learning.

Speaker Info:

I am working as a part of the development team of [Tensor]werk, an infrastructure development company focusing on deep learning deployment problems. I and my team focus on building open source tools for setting up a seamless deep learning workflow this includes RedisAI & hangar. I have been programming since 2012 and started using python since 2014 and moved to deep learning in 2015. I am an open source enthusiast and have contributed to the core of several widely used projects like PyTorch. I spend most of my research time on improving the interpretability of AI models using TuringNetwork. I have authored a deep learning book. I go by hhsecond on internet

Speaker Links:

  • https://github.com/hhsecond
  • https://medium.com/@hhsecond
  • https://www.amazon.in/dp/B078TLWD3F

Id: 1137
Section: Data Science, Machine Learning and AI
Type: Talks
Target Audience: Intermediate
Last Updated: