Smater and Intelligent ways to Train Deep Learning Models.
AMAN PANDEY (~aman69) |
We all know Deep Learning, Machine Learning is doing wonders for us. It has its presence in every product we use today.
Data and Compute resource are two major driving factors behind this tremendous success. But have you ever wondered
- How to train deep learning model if we have less amount of data for a particular problem?
- How to achieve maximum accuracy from the model by simultaneously reducing the training time?
- What to do if the model is not converging beyond some loss value and it's stuck at some saddle point?
In this hands-on workshop, we take a look at how to train a deep learning model for state of the art result and learn several techniques for faster convergence which can be used both in Computer Vision and NLP task.
The talk will cover it in the following manner: -
Preparing Dataset using Google Images:- Here we will write MultiProcessing scripts to speed up the download and verification of images before using it as an input for the model. [10 min]
Understanding Theory Behind Transfer Learning and Implementing it on a ResNet50 Model. [20 min]
Understanding Different types of schedulers in Pytorch [15 min]
- Linear Scheduler
- Exponential Scheduler
- Cosine Scheduler
Understanding and Implementing different Super Convergence techniques [30 min]
- LRfinder Algorithm
- SGD with Warm Restarts
- One Cycle Policy
Model Training and Fine-Tuning using the discussed techniques. [60 min]
- Q&A [5-10 mins]
Two Things which will make this entire workshop fruitful for the attendees are:-
- All the Algorithms are in Pytorch from scratch.
- We are going to discuss a lot of things and understand their inner workings in a fun way and benchmark how these techniques boosted the model performance.
Articles and Research papers:-
Utkarsh Vardhan:- Currently working as a Data Scientist at Curl Analytics. Solving Computer Vision and NLP related problems using Machine Learning and Deep Learning.
Earlier worked at Worxogo Solutions as a Data Scientist. My role at Worxogo: Building and implementing the strategies behind data science solutions for the company with demonstrable ROI. Mostly focusing on applying machine learning, deep learning, reinforcement learning, and natural language processing in the context of modeling and explaining Human Behavior.
Aman Pandey:- Currently working as Research Intern at Instoried. Solving Computer Vision and NLP related problems using Machine Learning and Deep Learning.
Together we organize Meetup at Bengaluru on a weekly basis where we discuss the current trends in deep-learning and it's surprisingly effective role in solving real case scenarios. This meetup aims to educate, inspire, and enable you to rapidly prototype your next idea using ML and DL models. The strategy will be to focus more on Hands On experience first, and then, take you deeper into concepts. Artificial Intelligence and Machine Learning are evolving extremely fast which makes the concepts invented last year, obsolete this year. Therefore we will cover mostly the latest concepts used in the industry. This meetup focuses on "how to build and understand", not just "how to use".