MNN_MSSD

1. Training MSSD model

trainning your own MSSD model by Google Object Detection API Tensorflow Object Detection API. Here we suppose that you have already finished the training phase.

2. Exporting tflite model

For more details of exporting tflite model, please refer to export tflite. And we List two important steps as follows:

object_detection/export_tflite_ssd_graph.py  \
--pipeline_config_path=$CONFIG_FILE          \
--trained_checkpoint_prefix=$CHECKPOINT_PATH \
--output_directory=$OUTPUT_DIR               \
--add_postprocessing_op=false

After this step, you got a pb file. If you would like to use tensorflow pb model for this MNN project, you can straightly jump the /tensorflow/jni, there are codes and details for it.

If you would like to use tflite for this MNN project, please be patient. You still need to go throught the rest of this document.

bazel run --config=opt tensorflow/lite/toco:toco -- \
--input_file=$OUTPUT_DIR/tflite_graph.pb            \
--output_file=$OUTPUT_DIR/detect.tflite             \
--input_shapes=1,224,224,3                          \
--input_arrays=normalized_input_image_tensor        \
--output_arrays='concat','concat_1'                 \
--inference_type=FLOAT                              \
--change_concat_input_ranges=false

3. Compile MNN for android

MNN is a lightweight deep neural network inference engine MNN. You can follow this link for compilation and benckmark.

4. Convert tflite model to mnn model

For details, please follow link. By only simple command line as follow you can get MNN model for MSSD, if you do it right. The following command is only for example, you need to replace items in --modelFile and --MNNModel with yours respectively.

./MNNConvert -f TFLITE --modelFile ~/Desktop/DMS-data/facedet/trained_model/detect.tflite --MNNModel ~/Desktop/face_det.mnn --bizCode MNN

And for tensorflow pb files, type

./MNNConvert -f TF --modelFile ~/Desktop/DMS-data/facedet/trained_model/detect.tflite --MNNModel ~/Desktop/face_det.mnn --bizCode MNN

5. Compile and execute demo

This demo depends on OpenCV on Android and MNN, please download them respectively. First, you need to revise OpenCV_BASE and MNN_BASE in jni/Android.mk acording to your desktop environment. And remember to install NDK, I use android-ndk-r17c for compilation in my desktop.

Then, by following simple command, you should get libs and executable file

if you have done things right.

cd jni
ndk-build

In order to test it, we push them to the android devices. For example, here we suppose you are in jni folder.

adb push ../libs/arm64-v8a/* /data/local/tmp
adb push image.jpg /data/local/tmp
adb push face_det.mnn /data/local/tmp

Then we need to set up the enviroment for testing in android devices, like

adb shell   
cd /data/local/tmp
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/local/tmp

then, in /data/local/tmp, run the executable file mssd

./mssd

The last step, close the adb shell session, and pull the output result back to your desktop.

adb pull /data/local/tmp/output.jpg ./

6. Discussion

For more discussion, you can contact me via Wechat: zxd675816777