ShuffleNet and ShuffleNetV2

Note

This is unofficial implement repo

• Reimplement ShuffleNet and ShuffleNetV2 for timm style
• Mapping original model weight for timm style model

Original Pretrain Weight

• OneDrive download: Link
• BaiduYun download: Link (extract code: mc24)

Quick Start

import torch
from shufflenet_original import create_shufflenet_original
from shufflenet_timm import create_shufflenet_timm

x = torch.rand(1, 3, 224, 224)

model_original = create_shufflenet_original('v2', '1.0x', None).eval()
model_timm = create_shufflenet_timm('shufflenetv2_100').eval()

y_original = model_original(x)
y_timm = model_timm(x)

print('All close: ', torch.allclose(y_timm, y_original))
>> All close:  True

print('MAE: ', torch.mean(torch.abs(y_timm - y_original)).item())
>> MAE:  0.0

print('Original Model')
flops_params(model_original)
>> | FLOPs:297.368 MFLOPS   MACs:145.814 MMACs   Params:2.2788 M 

print('Timm Model')
flops_params(model_timm)
>> | FLOPs:297.368 MFLOPS   MACs:145.814 MMACs   Params:2.2788 M 

Usage

python valid.py

python mapping_weight.py

Results

• requirement package calflops

Model	FLOPs	MACs	Params	Top-1	Top-5
ShuffleNetV1 0.5x (group=3)	79.82M	37.59M	0.72M	57.3	80.0
ShuffleNetV1 0.5x (group=8)	88.20M	40.47M	1.01M	58.8	81.0
ShuffleNetV2 0.5x	84.17M	40.48M	1.37M	61.1	82.6
ShuffleNetV1 1.0x (group=3)	284.20M	137.46M	1.86M	67.8	87.7
ShuffleNetV1 1.0x (group=8)	290.84M	138.36M	2.43M	68.0	86.4
ShuffleNetV2 1.0x	297.37M	145.81M	2.28M	69.4	88.9
ShuffleNetV1 1.5x (group=3)	598.19M	292.44M	3.44M	71.6	90.2
ShuffleNetV1 1.5x (group=8)	600.93M	290.19M	4.27M	71.0	89.6
ShuffleNetV2 1.5x	605.90M	298.97M	3.50M	72.6	90.6
ShuffleNetV1 2.0x (group=3)	1.07G	524.05M	5.45M	74.1	91.4
ShuffleNetV1 2.0x (group=8)	1.07G	521.94M	6.52M	72.9	90.8
ShuffleNetV2 2.0x	1.19G	590.78M	7.40M	75.0	92.4

Acknowledgement

This repository is built using the timm and Megvii Research

Citation

@inproceedings{zhang2018shufflenet,
    title={Shufflenet: An extremely efficient convolutional neural network for mobile devices},
    author={Zhang, Xiangyu and Zhou, Xinyu and Lin, Mengxiao and Sun, Jian},
    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
    pages={6848--6856},
    year={2018}
}

@inproceedings{ma2018shufflenet, 
    title={Shufflenet v2: Practical guidelines for efficient cnn architecture design},  
    author={Ma, Ningning and Zhang, Xiangyu and Zheng, Hai-Tao and Sun, Jian},  
    booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},  
    pages={116--131}, 
    year={2018} 
}