zhangyy12345 / mobilenet-ssd-snpe Goto Github PK

example of snpe with mobilenet-ssd trained by Caffe
caffe model MobileNet-Caffe

build

./build.sh

test on device

./run_example.sh

mobilenet-ssd model caffe to snpe
prepare snpe environment

nohup snpe-caffe-to-dlc --input_network ./model/deploy.prototxt \
--caffe_bin ./model/mobilenet_iter_73000.caffemodel \
--debug --o ./model/mobilenet_iter_73000.dlc \
> ./model/mobilenet_iter_73000.log 2>&1 &

prepare VOC raw data for snpe quantize
read from : data/VOC2007
generate resize image : data/VOC_resize
generate resize raw data : data/VOC_raw
generate raw list : data/VOC_raw_list.txt

python image_to_raw.py

quantize model to int8

nohup snpe-dlc-quantize --debug3 \
--input_dlc ./model/mobilenet_iter_73000.dlc \
--input_list ./data/VOC_raw_list.txt \
--output_dlc ./model/mobilenet_iter_73000_int8.dlc \
> ./model/mobilenet_iter_73000_int8.log 2>&1 &

** How to estimate runtime performance on dsp**

=> adb push snpe-net-run to path /data/local/tmp/test_demo/

=> cd /data/local/tmp/test_demo

=> chmod 777 snpe-net-run

=> settimg library path: export LD_LIBRARY_PATH=./${LD_LIBRARY_PATH} export ADSP_LIBRARY_PTAH=./:${ADSP_LIBRATY_PATH}

=> Then, need to do next step, Let's to show follow cmd: snpe-net-run --container *.dlc --input_list list.txt --use_dsp

=> if running is successful, we can have one output folder.

=> adb pull output folder to your local snpe develop path. follow cmd: snpe-diagview --input_log SNPEDiag.log --output LOGCSVFile > Log 2>&1
(Note: find file SNPEDiag.log from output folder)

*********** NOTE 1: runing run_example.sh************

On soc SA8155P, with problem about segmentation fault. Solution: copy library “libc++_shared” to "/system/lib"

********* NOTE 2: cross complie *********************

android-ndk-r21d with problem about lirary "libc++.so" and some undefined reference. Solution: delete library "libc++.so" from "./lib/snpe/armv7a-android"