Frameworks and Distributed Training

How to choose a deep learning framework? How to enable distributed training for your models?
Frameworks and Distributed Training - Infrastructure and Tooling

Summary

  • Unless you have a good reason not to, you should use either TensorFlow or PyTorch.

  • Both frameworks are converging to a point where they are good for research and production.

  • fast.ai is a solid option for beginners who want to iterate quickly.

  • Distributed training of neural networks can be approached in 2 ways: (1) data parallelism and (2) model parallelism.

  • Practically, data parallelism is more popular and frequently employed in large organizations for executing production-level deep learning algorithms.

  • Model parallelism, on the other hand, is only necessary when a model does not fit on a single GPU.

  • Ray is an open-source project for effortless, stateful, parallel, and distributed computing in Python.

  • RaySGD is a library for distributed data parallel training that provides fault tolerance and seamless parallelization, built on top of Ray.

  • Horovod is Uber’s open-source distributed deep learning framework that uses a standard multi-process communication framework, so it can be an easier experience for multi-node training.