To develop a machine learning model for object detection (Pedestrians, Cars, Motorbike, Bicycle, Bus) which is supported by Snapdragon mobile platforms with the help of Neural Processing SDK.
The Machine learning model to detect the Objects uses Single Shot Detector(SSD) algorithm trained on Mobilenet network architecture. Why MobilenetSSD
- Gives realtime inference frame rates 8-13 FPS on HDK835
- Better perfomance and accuracy compared to other architecutures like YOLO ... Etc,
How to train the model Installation of MobilenetSSD and Caffe
- Clean the caffe source code from the below git repository using below commands.
$ git clone https://github.com/weiliu89/caffe.git
$ cd caffe
$ git checkout ssd
-
Depending on the Processor used either the CPU or GPU install the depending packages by executing respective instructions from http://caffe.berkeleyvision.org/install_apt.html
-
Build caffe using the below instructions
$ cp Makefile.config.example Makefile.config Please disable the option USE_OPENCV. USE_OPENCV := 0 ( Make necessary modification in the Makefile as per the device configuration) $ make -j8 $ make test $ make runtest $ make pycaffe ( Make sure you add the $CAFFE_ROOT/python to your PYTHON PATH once done with the make pycaffe command)
Getting the data ready
- Download the VOC2007 and VOC2012 dataset.
$ mkdir home/<username>/data
$ cd home/<username>/data
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
$ tar -xvf VOCtrainval_11-May-2012.tar
$ tar -xvf VOCtrainval_06-Nov-2007.tar
$ tar -xvf VOCtest_06-Nov-2007.tar
- Creating the Lightning Memory-Mapped Database (LMDB) file
$ cd $CAFFE_ROOT
$ ./data/VOC0712/create_list.sh
$ ./data/VOC0712/create_data.sh
Training the model with VOC
Now clone the MobileNetSSD implementation by chuanqi305 from github
$ git clone https://github.com/chuanqi305/MobileNet-SSD.git
$ cd MobileNet-SSD
- Creating the symbolic link to training and test data sets,
$ ln -s $CAFFE_DIR/PATH_TO_YOUR_TRAIN_LMDB trainval_lmdb
$ ln -s $CAFFE_DIR/PATH_TO_YOUR_TEST_LMDB test_lmdb
- Copy a labelmap_voc.prototxt file from ssd repo tree to MobileNetSSD Dir using below command,
$ cp $CAFFE_DIR/data/VOC0712/labelmap_voc.prototxt $MOBILENETSSD_DIR/labelmap.prototxt
Run gen_model.sh for generating the training & testing prototxt with the given number of classes. Please make sure that the number of classes while running gen_model.sh & number of classes mentioned in labelmap.prototxt should match.
$ ./gen_model.sh 21
Now train your model using train.sh, and keep on training unless the loss is in between 1.5 to 2.5.
$ ./train.sh
Test the trained model with test.sh script.
$ ./test.sh
Run merge_bn.py to generate your own no-bn caffemodel if necessary.
$ python merge_bn.py --model example/MobileNetSSD_deploy.prototxt --weights snapshot/mobilenet_iter_xxxxxx.caffemodel
once merge_bn.py has run successfully, You’ll get two files in the root directory of MobileNetSSD. These are our trained models,
- no_bn.prototxt
- no_bn.caffemodel.
Run demo.py for checking the detection.
How to Convert caffe into DLC?
Prerequisites Neural Processing SDK setup. Use the instructions from the below link to make the setup, https://developer.qualcomm.com/software/qualcomm-neural-processing-sdk/getting-started
Initialized environmental variables of Neural Processing SDK with caffe. For converting the model from caffe to dlc, you need 2 files, prototxt and caffemodel file Once you’ve both convert it into dlc using the following command:
$ snpe-caffe-to-dlc --caffe_txt MobileNetSSD_deploy.prototxt --caffe_bin MobileNetSSD_deploy.caffemodel --dlc caffe_mobilenet_ssd.dlc