Pruning In Time (PIT): A Lightweight Network Architecture Optimizer for Temporal Convolutional Networks

TL;DR

This paper introduces Pruning In Time (PIT), an automatic dilation optimizer for Temporal Convolutional Networks that prunes weights on the time-axis, reducing model size and latency without accuracy loss.

Contribution

It presents a novel method that learns dilation factors and prunes weights simultaneously, producing efficient TCNs with Pareto-optimal trade-offs in size and accuracy.

Findings

Model size reduced by up to 7.4x on hardware

Inference latency decreased by up to 3x

Outperforms hand-designed solutions in size and accuracy

Abstract

Temporal Convolutional Networks (TCNs) are promising Deep Learning models for time-series processing tasks. One key feature of TCNs is time-dilated convolution, whose optimization requires extensive experimentation. We propose an automatic dilation optimizer, which tackles the problem as a weight pruning on the time-axis, and learns dilation factors together with weights, in a single training. Our method reduces the model size and inference latency on a real SoC hardware target by up to 7.4x and 3x, respectively with no accuracy drop compared to a network without dilation. It also yields a rich set of Pareto-optimal TCNs starting from a single model, outperforming hand-designed solutions in both size and accuracy.