This workshop features three main topics:
- how to select the best hardware configurations for your workload with minimal effort.
- how to optimize your PyTorch training and inference pipelines with system level optimizations.
- how to visualize and analyze the training performance of your model with an interactive profiling tool embedded in your code editor.
We cover these topics by optimizing a real-life image classification example. We show that by using a combination of tools and techniques, we can improve the performance by almost 10x with little to no convergence impact. For the best experience, participants are encouraged to bring their own laptop with Visual Studio Code server hosted on Vector's cluster, or simply bring a laptop and use VM's provided by CentML.