What is Horovod?

Review of distributed training framework for TensorFlow, developed by Uber
20 October 2017   3064

What is Horovod?

Horovod is a distributed training framework for TensorFlow. The goal of Horovod is to make distributed Deep Learning fast and easy to use.

Why not traditional Distributed TensorFlow?

The primary motivation for this project is to make it easy to take a single-GPU TensorFlow program and successfully train it on many GPUs faster. This has two aspects:

  1. How much modifications does one have to make to a program to make it distributed, and how easy is it to run it.
  2. How much faster would it run in distributed mode?

Internally at Uber we found that it's much easier for people to understand an MPI model that requires minimal changes to source code than to understand how to set up regular Distributed TensorFlow.

To give some perspective on that, this commit into our fork of TF Benchmarks shows how much code can be removed if one doesn't need to worry about towers and manually averaging gradients across them, tf.Server()tf.ClusterSpec()tf.train.SyncReplicasOptimizer()tf.train.replicas_device_setter() and so on. If none of these things makes sense to you - don't worry, you don't have to learn them if you use Horovod.

In addition to being easy to use, Horovod is fast. Below is a chart representing the benchmark that was done on 32 servers with 4 Pascal GPUs each connected by RoCE-capable 25 Gbit/s network:

Horovod Benchmark
Horovod Benchmark

Horovod achieves 90% scaling efficiency for both Inception V3 and ResNet-101, and 79% scaling efficiency for VGG-16.

While installing MPI and NCCL itself may seem like an extra hassle, it only needs to be done once by the team dealing with infrastructure, while everyone else in the company who builds the models can enjoy the simplicity of training them at scale.

Learn more at GitHub.

Nvidia to Open SPADE Source Code

SPADE machine learning system creates realistic landscapes based on rough human sketches
15 April 2019   679

NVIDIA has released the source code for the SPADE machine learning system (GauGAN), which allows for the synthesis of realistic landscapes based on rough sketches, as well as training models associated with the project. The system was demonstrated in March at the GTC 2019 conference, but the code was published only yesterday. The developments are open under the non-free license CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0), allowing use only for non-commercial purposes. The code is written in Python using the PyTorch framework.

Sketches are drawn up in the form of a segmented map that determines the placement of exemplary objects on the scene. The nature of the generated objects is set using color labels. For example, a blue fill turns into sky, blue into water, dark green into trees, light green into grass, light brown into stones, dark brown into mountains, gray into snow, a brown line into a road, and a blue line into the river. Additionally, based on the choice of reference images, the overall style of the composition and the time of day are determined. The proposed tool for creating virtual worlds can be useful to a wide range of specialists, from architects and urban planners to game developers and landscape designers.

Objects are synthesized by a generative-adversarial neural network (GAN), which, based on a schematic segmented map, creates realistic images by borrowing parts from a model previously trained on several million photographs. In contrast to the previously developed systems of image synthesis, the proposed method is based on the use of adaptive spatial transformation followed by transformation based on machine learning. Processing a segmented map instead of semantic markup allows you to achieve an exact match of the result and control the style.

To achieve realism, two competing neural networks are used: the generator and the discriminator (Discriminator). The generator generates images based on mixing elements of real photos, and the discriminator identifies possible deviations from real images. As a result, a feedback is formed, on the basis of which the generator begins to assemble more and more qualitative samples, until the discriminator ceases to distinguish them from the real ones.