Horovod is an open source software framework, designed for processing fast and efficient distributed deep learning models using TensorFlow, Keras, PyTorch, and Apache MXNet. It can scale up a single-GPU training script to run on multiple GPUs or hosts with minimal code changes.
This instructor-led, live training (online or onsite) is aimed at developers or data scientists who wish to use Horovod to run distributed deep learning trainings and scale it up to run across multiple GPUs in parallel.
By the end of this training, participants will be able to:
- Set up the necessary development environment to start running deep learning trainings.
- Install and configure Horovod to train models with TensorFlow, Keras, PyTorch, and Apache MXNet.
- Scale deep learning training with Horovod to run on multiple GPUs.
Format of the Course
- Interactive lecture and discussion.
- Lots of exercises and practice.
- Hands-on implementation in a live-lab environment.
Course Customization Options
- This course is focused on Horovod, but other software tools and frameworks such as TensorFlow, Keras, PyTorch, and Apache MXNet may be required. Please let us know if you have specific requirements or preferences.
- To request a customized training for this course, please contact us to arrange.
- An understanding of Machine Learning, specifically deep learning
- Familiarity with machine learning libraries (TensorFlow, Keras, PyTorch, Apache MXNet)
- Python programming experience
- Data scientists
- Overview of Horovod features and concepts
- Understanding the supported frameworks
Installing and Configuring Horovod
- Preparing the hosting environment
- Building Horovod for TensorFlow, Keras, PyTorch, and Apache MXNet
- Running Horovod
Running Distributed Training
- Modifying and running training examples with TensorFlow
- Modifying and running training examples with Keras
- Modifying and running training examples with PyTorch
- Modifying and running training examples with Apache MXNet
Optimizing Distributed Training Processes
- Running concurrent operations on multiple GPUs
- Tuning hyperparameters
- Enabling performance autotuning
Summary and Conclusion