MonkeyKing-KK / Huaguoshan

Lightweight model reimplementation

This is a pytorch implementation of MobileNetV2, ShuffleNetV2 and GhostNet on CIFAR-10 datasets.

Models are modified as the inputs are all 32*32 RGB images. We only conduct three downsample operations and you can find the modified parts in .py files.

Reference

MobileNetV2 ShuffleNetV2 ghostnet

Network

-MobileNetV2
-ShuffleNetV2
-GhostNet

Accuracy After Training:

Model				Accuracy
MobileNetV2			93.92%
ShuffleNetV2			92.60%
GhostNet			92.57%

Requirements:

-requirements.txt

Quick Start:

default settings: batch_size = 128, max_epoch = 190, lr = 0.1.

python train.py

More Training Options

-bs			batch_size, type=int, default=128
-max_epoch		num of training epochs, type=int, default=190
-lr			learning_rate, default=0.1
-arc			model to use['MobileNetV2 shufflenetv2 ghost_net'] [default=MobileNetV2]

To Do

Add more lightweight models
Clean up the code