Multi-GPU distributed training with TensorFlow

Author: fchollet Date created: 2020/04/28 Last modified: 2023/06/29 Description: Guide to multi-GPU training for Keras models with TensorFlow.
View in Colab • GitHub source
There are generally two ways to distribute computation across multiple devices:
Data parallelism, where a single model gets replicated on multiple devices or multiple machines. Each of them processes different batches of data, then they merge their results. There exist many variants of this setup, that differ in how the different model replicas merge results, in whether they stay in sync at every batch or whether they are more loosely coupled, etc.
Model parallelism, where different parts of a single model run on different devices, processing a single batch of data together. This works best with models that have a naturally-parallel architecture, such as models that feature multiple branches.
This guide focuses on data parallelism, in particular synchronous data parallelism, where the different replicas of the model stay in sync after each batch they process. Synchronicity keeps the model convergence behavior identical to what you would see for single-device training.
Specifically, this guide teaches you how to use the tf.distribute API to train Keras models on multiple GPUs, with minimal changes to your code, on multiple GPUs (typically 2 to 16) installed on a single machine (single host, multi-device training). This is the most common setup for researchers and small-scale industry workflows.
In this setup, you have one machine with several GPUs on it (typically 2 to 16). Each device will run a copy of your model (called a replica). For simplicity, in what follows, we'll assume we're dealing with 8 GPUs, at no loss of generality.
At each step of training:
In practice, the process of synchronously updating the weights of the model replicas is handled at the level of each individual weight variable. This is done through a mirrored variable object.
To do single-host, multi-device synchronous training with a Keras model, you would use the tf.distribute.MirroredStrategy API. Here's how it works:
Importantly, we recommend that you use tf.data.Dataset objects to load data in a multi-device or distributed workflow.
Schematically, it looks like this:
Here's a simple end-to-end runnable example:
When using distributed training, you should always make sure you have a strategy to recover from failure (fault tolerance). The simplest way to handle this is to pass ModelCheckpoint callback to fit(), to save your model at regular intervals (e.g. every 100 batches or every epoch). You can then restart training from your saved model.
Here's a simple example:
When doing distributed training, the efficiency with which you load data can often become critical. Here are a few tips to make sure your tf.data pipelines run as fast as possible.
Note about dataset batching
When creating your dataset, make sure it is batched with the global batch size. For instance, if each of your 8 GPUs is capable of running a batch of 64 samples, you call use a global batch size of 512.
Calling dataset.cache()
If you call .cache() on a dataset, its data will be cached after running through the first iteration over the data. Every subsequent iteration will use the cached data. The cache can be in memory (default) or to a local file you specify.
This can improve performance when:
Calling dataset.prefetch(buffer_size)
You should almost always call .prefetch(buffer_size) after creating a dataset. It means your data pipeline will run asynchronously from your model, with new samples being preprocessed and stored in a buffer while the current batch samples are used to train the model. The next batch will be prefetched in GPU memory by the time the current batch is over.