On-the-fly Operation Batching in Dynamic Computation Graphs

TL;DR

This paper introduces an algorithm for automatic operation batching in dynamic neural network frameworks, enabling efficient on-the-fly batching that simplifies development and achieves high performance comparable to manual batching.

Contribution

The paper presents a novel batching algorithm implemented in DyNet that automatically batches operations during execution, reducing developer effort and improving efficiency in dynamic neural networks.

Findings

Achieves throughput similar to manual batching.

Provides speedups over single-instance learning.

Effective on complex architectures that are hard to batch manually.

Abstract

Dynamic neural network toolkits such as PyTorch, DyNet, and Chainer offer more flexibility for implementing models that cope with data of varying dimensions and structure, relative to toolkits that operate on statically declared computations (e.g., TensorFlow, CNTK, and Theano). However, existing toolkits - both static and dynamic - require that the developer organize the computations into the batches necessary for exploiting high-performance algorithms and hardware. This batching task is generally difficult, but it becomes a major hurdle as architectures become complex. In this paper, we present an algorithm, and its implementation in the DyNet toolkit, for automatically batching operations. Developers simply write minibatch computations as aggregations of single instance computations, and the batching algorithm seamlessly executes them, on the fly, using computationally efficient batched operations. On a variety of tasks, we obtain throughput similar to that obtained with manual batches, as well as comparable speedups over single-instance learning on architectures that are impractical to batch manually.