AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

NetAdapt: Platform-Aware Neural Network Adaptation for Mobile Applications

Tien-Ju Yang, Andrew Howard, Bo Chen, Xiao Zhang, Alec Go, Mark Sandler, Vivienne Sze, Hartwig Adam

ECCV 2018 · Architectures and efficiency › RNNs and normalization · approach: Efficient network compression and training · 607 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
adapt with NetAdapt instead of MorphNet, small MobileNetV1 on mobile CPU (latency-guided)small MobileNetV1 · ImageNet, Pixel 1 CPUspeed and top-1 accuracy: 1.6x faster with 0.3% higher accuracyImproved
adapt with NetAdapt instead of width/resolution multipliers, small MobileNetV1small MobileNetV1 · ImageNet, Pixel 1 CPUspeed at same accuracy: up to 1.7x faster with the same or higher accuracyImproved
adapt with NetAdapt instead of multipliers and ADC, large MobileNetV1 on mobile CPUlarge MobileNetV1 · ImageNet, Pixel 1 CPUspeed: 1.4x and 1.2x speed-upImproved
adapt with NetAdapt on mobile GPU instead of benchmark algorithms, large MobileNetV1large MobileNetV1 · ImageNet, Galaxy S8 GPUspeed: up to 1.2x speed-up with higher accuracyImproved
add dropout and label smoothing to the training configuration of NetAdapt networksNetAdapt-adapted large MobileNetV1 · ImageNet, Pixel 1 CPUtop-1 accuracy: increased by 1.3%Improved
guide NetAdapt by latency (direct metric) rather than MACs (ablation)NetAdapt · ImageNet, Pixel 1 CPU, MACs-targeted (Table 1)latency: MorphNet 6.52 ms (140%) vs NetAdapt 6.01 ms (129%); fewer MACs (81%) did not give lower latency than baseline 4.65 msImproved
add short-term fine-tuning (ten thousand iterations) to each network proposal (ablation)NetAdapt on small MobileNetV1 · ImageNet, no long-term fine-tunetop-1 accuracy: accuracy drops to nearly zero without it; above 20% with ten thousand iterationsImproved
increase short-term fine-tuning from ten thousand to more iterations (ablation)NetAdapt on small MobileNetV1 · ImageNettop-1 accuracyImproved
add long-term fine-tuning at the end (ablation)NetAdapt on small MobileNetV1 · ImageNet, Pixel 1 CPUtop-1 accuracy: up to 4.5% or 3.4% on average (range 1.8% to 4.5%)Improved
use resource reduction decay 0.96 instead of 1.0 (initial 0.5 ms) (ablation)NetAdapt on small MobileNetV1 · ImageNet, Table 2top-1 accuracy: 47.4 -> 47.7 (28 vs 20 iterations)Improved
use larger initial resource reduction 0.8 ms (decay 0.95) instead of 0.5 ms (decay 0.96) (ablation)NetAdapt on small MobileNetV1 · ImageNet, Table 2top-1 accuracy: 47.7 -> 46.7Hurt
adapt MobileNetV2 with NetAdapt instead of multipliers (similar latency)75% MobileNetV2 (224) · ImageNet, Pixel 1 CPUtop-1 accuracy: 69.8 -> 70.9Improved

Later papers that compared with this method

5 later accepted papers report NetAdapt as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy