Skip to content

anthonychen1251/BudgetNet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

BudgetNet

Overall Structure

image In this work, we focus on how to dynamically regulate the computational cost of a given model based on target budget constraints during the inference phase. Therefore, we adopt existing CNNs as our pre-trained model rather than design a new network architecture. ResNeXt, a variant of ResNet, exploits a split-transform-merge strategy in building blocks.This strategy introduces a hyper-parameter called cardinality which is the size of the set of transformations, or in other words, the number of independent paths. These independent paths can be considered as a kind of branch structure and is suitable for our requirements.

System Environment

Python3.6

PyTorch1.1+

Pre-trained model

cifar10_ResNext-29_4x16d

*To achieve dynamic paths selecting, we implement RexNeXt bottleneck in type A, instead of the common implementation of grouped convolutions version (type C).

Results

image To get this result, download model weights in here and run

python test.py --model ResNext_C10 --depth 29 --base_width 16 --cardinality 4 --load ./ckpt_E_1280.t7 --output ResNeXt_29_C10.txt

Reference

https://github.com/Tushar-N/blockdrop

https://github.com/prlz77/ResNeXt.pytorch

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages