Training_System

This is a tutorial of how tp train the MobileNet-SSD detection network. The implementation of MobileNet-SSD detection network used was taken from here.

Run

1. Download the caffe source code from BVLC or here and compile.
2. Put the MobileNet-SSD directory in caffe/examples.
3. Run demo.py.

Create lmdb for your own dataset

1. Prepare the caffe-ssd VOC training datasets.
2. Put MyDataSet directory in caffe/data.
3. Replace the content in MyDataSet/Images with the VOC training datasets Images (should be original images), replace and content in MyDataSet/Labels with the VOC training datasets Labels (should be xml files). Each image in Images folder should have a unique label file in Labels folder with same name.
4. Modify MyDataSet/create_data.py file if necessary.
5. Create MyDataSet/ImageSets/Main.
6. Run MyDataSet/create_data.py, which will generate four txt files in ImageSets/Main.
7. Modify the paths and directories in create_list.sh and create_data.sh as specified in same file in comments (The scripts was taken from here).
8. Modify labelmap_voc.prototxt according to your classes defined in the caffe-ssd VOC training datasets.
9. Run create_list.sh and create_data.sh.

Train MobileNet-SSD network

1. Go to caffe/examples/MobileNet-SSD
2. Run gen_model.sh with one argument n where n is the number of your classes defined in MyDataSet/labelmap_voc.prototxt (e.g. bash gen_model.sh 2 and there are 2 classed including background defined in MyDataSet/labelmap_voc.prototxt).
3. Create symlinks.

ln -s PATH_TO_YOUR_TRAIN_LMDB trainval_lmdb
ln -s PATH_TO_YOUR_TEST_LMDB test_lmdb

4. Modify train.sh if necessary (Currently the training mode is CUP-ONLY).
5. Run train.sh which will generate snapshots in caffe/examples/MobileNet-SSD/snapshot directory.
6. Run merge_bn.py to obtain the caffemodel and prototxt based on snapshots generated from the previous step. One example:

python merge_bn.py --model ./example/MobileNetSSD_deploy.prototxt --weights ./snapshot/mobilenet_iter_50000.caffemodel