MobileNetV2_pytorch_cifar

This is a complete implementation of MobileNetv2 in PyTorch which can be trained on CIFAR10, CIFAR100 or your own dataset. This network comes from the paper below

Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation https://arxiv.org/abs/1801.04381v3

In this network, an inverted residual structure and depthwise convolution is used. Please refer to that paper for more detailed information

Usage

This project is compiled and run on Python 2.7 and PyTorch 0.4.0 Here are some necessaries dependencies:

torch 0.4.0
torchvision 0.2.1
numpy 1.14.3
tensorboardX 1.2

use pip to install them first

Train and Test

1. Download CIFAR10 or CIFAR100 dataset or prepare your own dataset like a dataloader defined in PyTorch

2. Modify config.py into your own configuration, eg. change image_size or something

3. Run python main.py --dataset cifar10 --root /your/path/to/dataset/

Tensorboard is also available, just use

tensorboard --logdir=./exp_dir/summaries

logdir is also changable in config.py

I compared my implementation with the implementation of MG2033, for he has many stars, is quite confusing that my implementation on cifar100 has about 8% higher accuracy-74% vs 66%, my training policy is the same with his.

my implementation accuracy on cifar100: