Managing Tensorflow training and inference with a simple RESTful framework
Training models for self-driving cars is daunting task – there are continuous improvement in the field of deep learning which creates a need to constantly train on newer architectures.
For this, a framework approach is highly essential as it helps to conduct multiple experiments, keep track of the sessions and easily run and extract the results.
The same approach works well for any deep learning task – training speech models, financial estimation models etc. I will talk more about the simple RESTFul framework for conducting Tensorflow training and inference. This architecture makes training and evaluation highly scalable and manage-able by:
- Plugging in your model for training
- Centralizing logging framework
- Storing inputs and outputs at a convenient location
- Conducting a large number of experiments
What the talk will cover:
- Small intro to Deep Learning with Tensorflow -
- Why is it different from other python libraries?
- Conducting an image segmentation task in Tensorflow
- How do you make it run on REAL data?
- ( Train + explore ) x N
- How to setup the an experiment for the best results in the least time
- Knowledge of RESTFul APIs
- Basic understanding of machine learning
I am currently working on computer vision solutions like object detection, semantic segmentation for self-driving cars at Intel. I have worked across a wide range of platforms for computer vision – from hardware based night-vision equipment at Tonbo Imaging and DRDO to cloud machine learning pipelines. I have a deep passion to solve the next generation of problems in technology by building computers that “see” and interpret the world as humans do.