AutoSlim: Towards One-Shot Architecture Search for Channel Numbers

TL;DR

AutoSlim introduces a one-shot method for optimizing per-layer channel counts under resource constraints by training a slimmable network to approximate the performance of many widths, followed by a greedy slimming pass to meet budgets. The approach delivers superior speed-accuracy trade-offs on ImageNet across MobileNet, MNasNet, and ResNet-50, including 74.2% top-1 for AutoSlim-MobileNet-v2 at 305M FLOPs and gains over RL-searched baselines. It leverages universally slimmable training techniques and evaluates configurations without retraining, providing a practical, scalable alternative to exhaustive search or iterative pruning. The results highlight nonuniform channel allocations favoring deeper layers and reveal dataset-dependent transferability of optimized configurations, with code and models to be released.

Abstract

We study how to set channel numbers in a neural network to achieve better accuracy under constrained resources (e.g., FLOPs, latency, memory footprint or model size). A simple and one-shot solution, named AutoSlim, is presented. Instead of training many network samples and searching with reinforcement learning, we train a single slimmable network to approximate the network accuracy of different channel configurations. We then iteratively evaluate the trained slimmable model and greedily slim the layer with minimal accuracy drop. By this single pass, we can obtain the optimized channel configurations under different resource constraints. We present experiments with MobileNet v1, MobileNet v2, ResNet-50 and RL-searched MNasNet on ImageNet classification. We show significant improvements over their default channel configurations. We also achieve better accuracy than recent channel pruning methods and neural architecture search methods. Notably, by setting optimized channel numbers, our AutoSlim-MobileNet-v2 at 305M FLOPs achieves 74.2% top-1 accuracy, 2.4% better than default MobileNet-v2 (301M FLOPs), and even 0.2% better than RL-searched MNasNet (317M FLOPs). Our AutoSlim-ResNet-50 at 570M FLOPs, without depthwise convolutions, achieves 1.3% better accuracy than MobileNet-v1 (569M FLOPs). Code and models will be available at: https://github.com/JiahuiYu/slimmable_networks

Figures (7)

• Figure 1: Various heuristics for setting channel numbers across entire network ($(A) - (B)$) simonyan2014veryhan2017deepzhang2017pyramidal, and inside network building blocks ($(a) - (f)$) sandler2018invertedhe2016deephan2017deepzhang2017pyramidaltan2018mnasnetcai2018proxylessnas.
• Figure 2: The flow diagram of our proposed approach AutoSlim.
• Figure 3: The flow diagram of network pruning methods liu2017learning.
• Figure 4: The flow diagram of network architecture search methods tan2018mnasnetcai2018proxylessnaszoph2018learningzoph2016neural.
• Figure 5: The optimized number of channels (left) and the percentage compared with default channels (right) of MobileNet v2. The channels of depthwise convolutions are ignored in the figure, since its output channels are always equal to the previous $1\times1$ convolution outputs.