MobileNet-Caffe-Classification

Introduction

Model prototxt files copy from https://github.com/shicai/MobileNet-Caffe and add some other files for training and inference conveniently. You can train mobilenet for classification task easily use this code.

[This is a Caffe implementation of Google's MobileNets (v1 and v2). For details, please read the following papers:

• [v1] MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
• [v2] Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation ]

Usage

1. Prepare data for classification task

1.1 Put all images into one dircectory (your_data_dir), use file name save your data's label

#before the separator "_" is your class index.
/data/1_xxx.jpg
/data/0_xxx.jpg
/data/2_xxx.png
...

1.2 Run this command then you will get a file named trainval.txt

sh get_train_data_list.sh your_data_dir

1.3 Data format for training

/imagepath class_index
# Example:
/data/1_pos.jpg 1
/data/0_neg.jpg 0

2. Train model

2.1 Modify solver.prototxt & train.prototxt

• solver.prototxt

  • net: your_train_prototxt_path

• train.prototxt

  • Notice Amend all "use_global_stats:false" when traning
  • source: your_data_trainval.txt_path
  • image shape
    • new_height: your_img_h
    • new_width: your_img_w
  • first conv layer
    • name: "conv1" for 3 channel image
    • name: "conv1_gray" for 1 channel image
  • fc7 layer
    • num_output : your_dataset_class_num

2.2 Run Command

sh train.sh mobilenet
# or
sh train.sh mobilenet_v2

3. Inference

3.1 Run Command

# Use the python that your caffe envrionment.
/xxx/python infer.py

File Structure

├── ckpt
│ ├── pretrain
│ │ ├──  mobilenet.caffemodel
│ │ ├──  mobilenet_v2.caffemodel
├── data
│ ├── get_train_data_list.sh  # util script for data
│ ├── your_data_dir
├── prototxt
│ ├── m1_solver.prototxt
│ ├── m2_solver.prototxt
│ ├── v1
│ │ ├── mobilenet_deplpoy.prototxt
│ │ ├── mobilenet_train.prototxt
│ ├── v2
│ │ ├── mobilenet_v2_deplpoy.prototxt
│ │ ├── mobilenet_v2_train.prototxt
├── README.md
├── LICENCE
├── train.sh
├── infer.py