[go: up one dir, main page]

Skip to content

A regularly updated comparison of CNN architectures in terms of accuracy, operations and model size

Notifications You must be signed in to change notification settings

BertMoons/Comparing-CNN-Architectures

Repository files navigation

A Comparison of Network Architectures

This is a comparison of Network Architectures in the accuracy - operations - model_size space.
The repo will be regularly updated if new network architectures come out. Let me know if you have any suggestions for architectures to add, or if any relevant data is missing.

This comparison is built on the work done by Canziani et al.

Network Efficiency

In the graphs below, 'MSD' are Multi-Scale DenseNets, 'DN' are DenseNets and 'MobNets' are MobileNets. Checkout interactive_plot.py or interactive_plot.html for a more readable visualization of all data points.

MobileNets stand out through their high accuracy using tiny models. DenseNets and Multi-Scale DenseNets achieve higher accuracy at the same computational budget, but require larger models.

Figure 1 IMAGENET top-1 accuracy vs #flops, blob size is the model size

Figure 2 IMAGENET top-1 accuracy vs #weights, blob size is the #flops

Remarks

A higher amount of operations or a larger model-size does not necessarily translate in a higher energy consumption on an embedded platform. The real relevant comparison point would hence be top-1 accuracy vs energy/classification, which is heavily platform dependent.

If you like/use this comparison, please cite Canziani et al and:

(to appear)
@phdthesis{moons2018Embedded,
    title={Embedded Deep Learning - Algorithms, Architectures and Circuits for Always-On Neural Network Processing},
    author = {Bert Moons},
    advisor = {prof. Marian Verhelst}
    school=[KU Leuven},
    year={2018}
    }

About

A regularly updated comparison of CNN architectures in terms of accuracy, operations and model size

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published