MobileNetv2-SSDLite

Caffe implementation of SSD detection on MobileNetv2, converted from tensorflow.

This repo. provides deploy.caffemodel and deploy.prototxt trained via COCO dataset that work fine for object detection task. If you want to remake your caffemodel and prototxt via own dataset then read "how to make caffemodel and prototxt in this repo." section.

Depends

• Tensorflow detection model zoo

Reference original

• Caffe implementation of SSD and SSDLite detection on MobileNetv2, converted from tensorflow.

Prerequisites

• Tensorflow and Caffe version SSD is properly installed on your computer.
• Ubuntu16.04 on PC, update and upgrade

Usage

COCO Detection Demo bellow,

    $ python demo_caffe.py

How to make ./deploy.caffemodel and ./deploy.prototxt in this repo.

Notice!: As of now ssd/ version is working fine, but ssdlite/ version bellow is not working correctly.

0. Firstly you should download the original model from Tensorflow detection model zoo and install tensorflow framework.

    $ pip install tensorflow==1.5

1. Use gen_model.py to generate the train.prototxt and deploy.prototxt (or use the default prototxt).
   coco/labelmap_coco.prototxt includes 10 name: 'N/A' and 10 display_name: 'N/A' and background, so number of classes is 91(=80+1+10).

    // -c 91 means category number of coco included 'N/A'(==10 items) and background.
    
    // for ssd
    $ python ssd/gen_model.py -s deploy -c 91 >deploy.prototxt
    
    // for ssdlite
    $python ssdlite/gen_model.py -s deploy -c 91 >deploy.prototxt

Check generated .prototxt directory.

2. Use dump_tensorflow_weights.py to dump the weights of conv layer and batchnorm layer.

    // for ssd
    $ wget  http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v2_coco_2018_03_29.tar.gz
    $ tar xzf ssd_mobilenet_v2_coco_2018_03_29.tar.gz
    $ python ssd/dump_tensorflow_weights.py
    
    // for ssdlite
    $ wget http://download.tensorflow.org/models/object_detection/ssdlite_mobilenet_v2_coco_2018_05_09.tar.gz
    $ tar xzf ssdlite_mobilenet_v2_coco_2018_05_09.tar.gz
    $ python ssdlite/dump_tensorflow_weights.py

Check generated contents of ./output/ directory.

3. Use load_caffe_weights.py to load the dumped weights to deploy.caffemodel.

    // for ssd
    $ python ssd/load_caffe_weigts.py

    // for ssdlite
    $ python ssdlite/load_caffe_weigts.py

Check generated deploy.caffemodel.

4. Use the code in src to accelerate your training if you have a cudnn7, or add "engine: CAFFE" to your depthwise convolution layer to solve the memory issue.
5. The original tensorflow model is trained on MSCOCO dataset, maybe you need deploy.caffemodel for VOC dataset, use coco2voc.py to get deploy_voc.caffemodel.

Attempt Demo Scripts

Try Object Detection Demo to check generated deploy.prototxt and deploy.caffemodel.

// modify demo script
$ vi demo_caffe.py
net_file    = 'deploy.prototxt'
caffe_model = 'deploy.caffemodel'
$ python demo_caffe.py

Train your own dataset

1. Generate the trainval_lmdb and test_lmdb from your dataset.
2. Write a labelmap.prototxt
3. Use gen_model.py to generate some prototxt files, replace the "CLASS_NUM" with class number of your own dataset such as coco class number 91(=80class + 10 stuff + 1 background)

python gen_model.py -s train  -c CLASS_NUM >train.prototxt
python gen_model.py -s test   -c CLASS_NUM >test.prototxt
python gen_model.py -s deploy -c CLASS_NUM >deploy.prototxt

4. Copy coco/solver_train.prototxt and coco/train.sh to your project and start training.
   replace path to lmdb, prototxt file names and caffe root path, etc.

Note

There are some differences between caffe and tensorflow implementation:

1. The padding method 'SAME' in tensorflow sometimes use the [0, 0, 1, 1] paddings, means that top=0, left=0, bottom=1, right=1 padding. In caffe, there is no parameters can be used to do that kind of padding.
2. MobileNet on Tensorflow use ReLU6 layer y = min(max(x, 0), 6), but caffe has no ReLU6 layer. Replace ReLU6 with ReLU cause a bit accuracy drop in ssd-mobilenetv2, but very large drop in ssdlite-mobilenetv2. There is a ReLU6 layer implementation in my fork of ssd.