ShuffleNet v2 in Pytorch

• The paper about shufflenetv2: shufflenet v2
• I implement the shufflenetv2, and test the performance on classification and detection tasks. You can use these codes to train model on your dataset.

Version

• Python 3.6
• torch 1.1.0
• torchvision 0.3.0

To do

• Network structure
• Used for Classification
• Used for Objection detection

Usage

For classification

• If you want to test classification demo.

1. Download the project https://github.com/ZhuYun97/shufflenetv2.git
2. You should download the dataset.
3. Then, put the dataset into the corresponding location.

Training

• Training from a fresh start: run the command python main.py
• Training from a trained model: python main.py -t 1 -pre "/save/1563186508.pkl"

Testing

• python main.py -t 0 -pre "./save/1563186508.pkl"

Explanation for some obscure arguments

• t means you decide to train(1) or test(0) the model. If you assign 0 to t, it's better to assign a trained model to pre
• e means how many epochs
• bs means batch size

Comparision with other models

As for classification task batch_size=1, CPU

Type	Acc	Time	MFLOPs
EfficientNet-B3	94.8	9.9 FPS	1800
ShuffleNet v2	94.7	48.3 FPS	146
MobileNet v2	94.3	30.5 FPS	300
MobileNet v3-Large	89.8	29.7 FPS	219
MobileNet v3-Small	90.9	45.0 FPS	66

For detection

In another repo

Training own your dataset

For classification

TBD

For detection

In another repo

Experiments

I train the model about 5 eopchs, and in each eopch, I test the performance of trained model.

Phase train loss: 0.6354673637662616, acc: 0.6564571428571429
Phase val loss: 0.5708242939949035, acc: 0.7146666666666667
Phase train loss: 0.493809922170639, acc: 0.7606285714285714
Phase val loss: 0.5668963393211365, acc: 0.724
Phase train loss: 0.4324655994551522, acc: 0.7994857142857142
Phase val loss: 0.4208303438186646, acc: 0.8048
Phase train loss: 0.38515312327657425, acc: 0.8273714285714285
Phase val loss: 0.37815397882064183, acc: 0.8298666666666666
Phase train loss: 0.3477836193084717, acc: 0.8467428571428571
Phase val loss: 0.34451772966384886, acc: 0.8441333333333333

And I didn't adjust any hyper parameters. After 15 epochs, the accuracy can reach 92.7% Phase val loss: 0.18356857439478239, acc: 0.9269333333333334

Dataset

For classification dataset: I use a dataset from kaggle which contains two classes(cats, dogs). For detection dataset: In another repo

References

• For structure: shufflenet-v2
• For data processing: process data