Auto-Train-Once: Controller Network Guided Automatic Network Pruning from Scratch

TL;DR

Auto-Train-Once (ATO) introduces a controller-guided pruning method that trains and compresses neural networks from scratch in a single process, outperforming existing techniques in efficiency and accuracy.

Contribution

The paper presents a novel controller network guided pruning algorithm with a stochastic gradient method, enabling automatic network compression during training without fine-tuning.

Findings

Achieves state-of-the-art pruning performance on multiple architectures.

Effectively reduces computational and storage costs.

Demonstrates superior results on CIFAR and ImageNet datasets.

Abstract

Current techniques for deep neural network (DNN) pruning often involve intricate multi-step processes that require domain-specific expertise, making their widespread adoption challenging. To address the limitation, the Only-Train-Once (OTO) and OTOv2 are proposed to eliminate the need for additional fine-tuning steps by directly training and compressing a general DNN from scratch. Nevertheless, the static design of optimizers (in OTO) can lead to convergence issues of local optima. In this paper, we proposed the Auto-Train-Once (ATO), an innovative network pruning algorithm designed to automatically reduce the computational and storage costs of DNNs. During the model training phase, our approach not only trains the target model but also leverages a controller network as an architecture generator to guide the learning of target model weights. Furthermore, we developed a novel stochastic gradient algorithm that enhances the coordination between model training and controller network training, thereby improving pruning performance. We provide a comprehensive convergence analysis as well as extensive experiments, and the results show that our approach achieves state-of-the-art performance across various model architectures (including ResNet18, ResNet34, ResNet50, ResNet56, and MobileNetv2) on standard benchmark datasets (CIFAR-10, CIFAR-100, and ImageNet).