使用 intel neural compute stick 需要用到 openvino 推理引擎，然而 raspberry pi 4 aarch64 系统并没有编译好的安装包，需要借助 yocto 环境来构建安装包。本文基于 yocto zeus 分支演示环境准备及构建安装过程。
操作系统： ubuntu 18.04 获取源代码：
git clone -b zeus https://git.yoctoproject.org/git/poky git clone -b zeus https://git.yoctoproject.org/git/meta-intel git clone -b zeus https://git.openembedded.org/meta-openembedded git clone -b zeus https://github.com/kraj/meta-clang.git准备烹饪的饼层
source poky/oe-init-build-env bitbake-layers add-layer ../meta-intel bitbake-layers add-layer ../meta-openembedded/meta-oe bitbake-layers add-layer ../meta-openembedded/meta-python bitbake-layers add-layer ../meta-clang修改烹饪菜谱, conf/local.conf
# Default to setting automatically based on cpu count BB_NUMBER_THREADS ?= "${@oe.utils.cpu_count()}" # Default to setting automatically based on cpu count PARALLEL_MAKE ?= "-j ${@oe.utils.cpu_count()}" MACHINE = "raspberrypi4-64" DISTRO = "poky" USER_CLASSES ?= "buildstats image-mklibs image-prelink" PATCHRESOLVE = "noop" CONF_VERSION = "1" ...开始烘烤
$ bitbake dldt-inference-engine Loading cache: 100% |######################################################################################################################| Time: 0:00:00 Loaded 6297 entries from dependency cache. Parsing recipes: 100% |####################################################################################################################| Time: 0:00:02 Parsing of 4786 .bb files complete (4770 cached, 16 parsed). 6299 targets, 219 skipped, 0 masked, 0 errors. NOTE: Resolving any missing task queue dependencies Build Configuration: BB_VERSION = "1.44.0" BUILD_SYS = "x86_64-linux" NATIVELSBSTRING = "universal" TARGET_SYS = "aarch64-poky-linux" MACHINE = "raspberrypi4-64" DISTRO = "poky" DISTRO_VERSION = "3.0.2" TUNE_FEATURES = "aarch64 cortexa72 crc crypto" TARGET_FPU = "" DISTRO_NAME = "Poky (Yocto Project Reference Distro)" ROS_DISTRO = "melodic" ROS_VERSION = "1" ROS_PYTHON_VERSION = "2" meta meta-poky meta-yocto-bsp = "zeus:74f229160c7f4037107c1dad8f0d02128c080a7e" meta-oe meta-python meta-multimedia meta-networking meta-perl = "zeus:9e60d30669a2ad0598e9abf0cd15ee06b523986b" meta-raspberrypi = "heads/origin/zeus:9e67048f3247762419decddec16d21482b7ba8e6" meta-security = "zeus:ecd8c30af6b1b957b4fa814d3f25c40f060b3ac1" meta-qt5 = "zeus:a582fd4c810529e9af0c81700407b1955d1391d2" meta-python2 = "zeus:4400f9155ec193d028208cf0c66aeed2ba2b00ab" meta-java = "zeus:e022ca616e161fdfce44f2956e3587894282a0ab" meta-tensorflow = "master:fcbe23b2a33ec9d8c900fb6be30bd9344742c2ab" meta-tensorflow-lite = "master:682d157f434348026fcd47e76b953988d80a8295" meta-intel = "zeus:32caa1705441615a07df01fe33ed12aee30e03d9" meta-ros1-melodic meta-ros1 meta-ros-common = "zeus-draft:840ee20985c501555d46308ac0c16b67ee4ae253" meta-rasp4 = "master:685ee71966ba84e00539f5be7a02692db887855a" meta-clang = "HEAD:711e593d5984aad3bf35c51b7ac4482982bc16c7" workspace = "<unknown>:<unknown>" Initialising tasks: 100% |#################################################################################################################| Time: 0:00:01 Sstate summary: Wanted 60 Found 3 Missed 57 Current 409 (5% match, 87% complete) NOTE: Executing Tasks NOTE: Setscene tasks completed Currently 1 running tasks (1873 of 1882) 99% |######################################################################################################## | 0: dldt-inference-engine-2019r3-r0 do_compile (pid 48314) 40% |#################################### |耐心等待烘烤结束
NOTE: Tasks Summary: Attempted 1882 tasks of which 1710 didn't need to be rerun and all succeeded.更新食谱菜单
bitbake package-index Loading cache: 100% |######################################################################################################################| Time: 0:00:01 Loaded 6297 entries from dependency cache. Parsing recipes: 100% |####################################################################################################################| Time: 0:00:02 Parsing of 4786 .bb files complete (4784 cached, 2 parsed). 6299 targets, 219 skipped, 0 masked, 0 errors. NOTE: Resolving any missing task queue dependencies Build Configuration: BB_VERSION = "1.44.0" BUILD_SYS = "x86_64-linux" NATIVELSBSTRING = "universal" TARGET_SYS = "aarch64-poky-linux" MACHINE = "raspberrypi4-64" DISTRO = "poky" DISTRO_VERSION = "3.0.2" TUNE_FEATURES = "aarch64 cortexa72 crc crypto" TARGET_FPU = "" DISTRO_NAME = "Poky (Yocto Project Reference Distro)" ROS_DISTRO = "melodic" ROS_VERSION = "1" ROS_PYTHON_VERSION = "2" meta meta-poky meta-yocto-bsp = "zeus:74f229160c7f4037107c1dad8f0d02128c080a7e" meta-oe meta-python meta-multimedia meta-networking meta-perl = "zeus:9e60d30669a2ad0598e9abf0cd15ee06b523986b" meta-raspberrypi = "heads/origin/zeus:9e67048f3247762419decddec16d21482b7ba8e6" meta-security = "zeus:ecd8c30af6b1b957b4fa814d3f25c40f060b3ac1" meta-qt5 = "zeus:a582fd4c810529e9af0c81700407b1955d1391d2" meta-python2 = "zeus:4400f9155ec193d028208cf0c66aeed2ba2b00ab" meta-java = "zeus:e022ca616e161fdfce44f2956e3587894282a0ab" meta-tensorflow = "master:fcbe23b2a33ec9d8c900fb6be30bd9344742c2ab" meta-tensorflow-lite = "master:682d157f434348026fcd47e76b953988d80a8295" meta-intel = "zeus:32caa1705441615a07df01fe33ed12aee30e03d9" meta-ros1-melodic meta-ros1 meta-ros-common = "zeus-draft:840ee20985c501555d46308ac0c16b67ee4ae253" meta-rasp4 = "master:685ee71966ba84e00539f5be7a02692db887855a" meta-clang = "HEAD:711e593d5984aad3bf35c51b7ac4482982bc16c7" workspace = "<unknown>:<unknown>" Initialising tasks: 100% |#################################################################################################################| Time: 0:00:00 Sstate summary: Wanted 0 Found 0 Missed 0 Current 24 (0% match, 100% complete) NOTE: Executing Tasks NOTE: Setscene tasks completed Currently 1 running tasks (185 of 187) 98% |######################################################################################################### | 0: package-index-1.0-r0 do_package_index - 44s (pid 94131)然后在树莓派端查看最新的食谱菜单
rpi4:~/test$ sudo apt-get update Password: Ign:1 http://192.168.168.192:8000/all ./ InRelease Ign:2 http://192.168.168.192:8000/aarch64 ./ InRelease Ign:3 http://192.168.168.192:8000/raspberrypi4_64 ./ InRelease Get:4 http://192.168.168.192:8000/all ./ Release [1,213 B] Get:5 http://192.168.168.192:8000/aarch64 ./ Release [1,217 B] Get:6 http://192.168.168.192:8000/raspberrypi4_64 ./ Release [1,225 B] Ign:7 http://192.168.168.192:8000/all ./ Release.gpg Ign:8 http://192.168.168.192:8000/aarch64 ./ Release.gpg Ign:9 http://192.168.168.192:8000/raspberrypi4_64 ./ Release.gpg Get:10 http://192.168.168.192:8000/aarch64 ./ Packages [2,564 kB] Fetched 2,567 kB in 2s (1,280 kB/s) Reading package lists... Done查找食谱
rpi4:~/test$ acs dldt- dldt-inference-engine-dbg - OpenVINO(TM) Toolkit - Deep Learning Deployment Toolkit - Debugging files dldt-inference-engine-ptest - OpenVINO(TM) Toolkit - Deep Learning Deployment Toolkit - Package test files dldt-inference-engine - OpenVINO(TM) Toolkit - Deep Learning Deployment Toolkit dldt-inference-engine-vpu-firmware - OpenVINO(TM) Toolkit - Deep Learning Deployment Toolkit dldt-inference-engine-doc - OpenVINO(TM) Toolkit - Deep Learning Deployment Toolkit - Documentation files dldt-inference-engine-src - OpenVINO(TM) Toolkit - Deep Learning Deployment Toolkit - Source files dldt-inference-engine-samples - OpenVINO(TM) Toolkit - Deep Learning Deployment Toolkit dldt-inference-engine-dev - OpenVINO(TM) Toolkit - Deep Learning Deployment Toolkit - Development files引入食谱
sudo apt-get install dldt-inference-engine dldt-inference-engine-samples查看可用推理设备
rpi4:~/test$ hello_query_device Available devices: Device: CPU Metrics: AVAILABLE_DEVICES : [ ] SUPPORTED_METRICS : [ AVAILABLE_DEVICES SUPPORTED_METRICS FULL_DEVICE_NAME OPTIMIZATION_CAPABILITIES SUPPORTED_CONFIG_KEYS RANGE_FOR_ASYNC_INFER_REQUESTS RANGE_FOR_STREAMS ] FULL_DEVICE_NAME : Non Intel Architecture OPTIMIZATION_CAPABILITIES : [ FP32 INT8 BIN ] SUPPORTED_CONFIG_KEYS : [ CPU_BIND_THREAD CPU_THREADS_NUM CPU_THROUGHPUT_STREAMS DUMP_EXEC_GRAPH_AS_DOT DYN_BATCH_ENABLED DYN_BATCH_LIMIT EXCLUSIVE_ASYNC_REQUESTS PERF_COUNT ] RANGE_FOR_ASYNC_INFER_REQUESTS : { 1, 1, 1 } RANGE_FOR_STREAMS : { 1, 4 } Default values for device configuration keys: CPU_BIND_THREAD : YES CPU_THREADS_NUM : 0 CPU_THROUGHPUT_STREAMS : 1 DUMP_EXEC_GRAPH_AS_DOT : "" DYN_BATCH_ENABLED : NO DYN_BATCH_LIMIT : 0 EXCLUSIVE_ASYNC_REQUESTS : NO PERF_COUNT : NO Device: MYRIAD Metrics: AVAILABLE_DEVICES : [ 1.2-ma2480 ] FULL_DEVICE_NAME : Intel Movidius Myriad X VPU SUPPORTED_METRICS : [ AVAILABLE_DEVICES FULL_DEVICE_NAME SUPPORTED_METRICS SUPPORTED_CONFIG_KEYS OPTIMIZATION_CAPABILITIES RANGE_FOR_ASYNC_INFER_REQUESTS ] SUPPORTED_CONFIG_KEYS : [ VPU_HW_STAGES_OPTIMIZATION VPU_LOG_LEVEL VPU_PRINT_RECEIVE_TENSOR_TIME VPU_NETWORK_CONFIG VPU_COMPUTE_LAYOUT VPU_CUSTOM_LAYERS VPU_IGNORE_IR_STATISTIC VPU_MYRIAD_FORCE_RESET VPU_MYRIAD_PLATFORM EXCLUSIVE_ASYNC_REQUESTS LOG_LEVEL PERF_COUNT CONFIG_FILE DEVICE_ID ] OPTIMIZATION_CAPABILITIES : [ FP16 ] RANGE_FOR_ASYNC_INFER_REQUESTS : { 3, 6, 1 } Default values for device configuration keys: VPU_HW_STAGES_OPTIMIZATION : UNSUPPORTED TYPE VPU_LOG_LEVEL : UNSUPPORTED TYPE VPU_PRINT_RECEIVE_TENSOR_TIME : UNSUPPORTED TYPE VPU_NETWORK_CONFIG : UNSUPPORTED TYPE VPU_COMPUTE_LAYOUT : UNSUPPORTED TYPE VPU_CUSTOM_LAYERS : UNSUPPORTED TYPE VPU_IGNORE_IR_STATISTIC : UNSUPPORTED TYPE VPU_MYRIAD_FORCE_RESET : UNSUPPORTED TYPE VPU_MYRIAD_PLATFORM : UNSUPPORTED TYPE EXCLUSIVE_ASYNC_REQUESTS : UNSUPPORTED TYPE LOG_LEVEL : UNSUPPORTED TYPE PERF_COUNT : UNSUPPORTED TYPE CONFIG_FILE : UNSUPPORTED TYPE DEVICE_ID : UNSUPPORTED TYPE我们来写一个简单的opencv dnn 推理测试，直接使用 caffe 模型，利用神经棒加速 test-opencv-dnn-vpu.py
#copyright dxxd, #东写西Do @ import numpy as np import time import cv2 caffe_root = '/home/pi/ml/' image = cv2.imread('/home/pi/test/beagle.jpeg') labels_file = caffe_root + 'models/ilsvrc12/classification/synset_words.txt' prototxt = caffe_root + 'models/bvlc_googlenet/bvlc_googlenet.prototxt' model = caffe_root + 'models/bvlc_googlenet/bvlc_googlenet.caffemodel' # load the labels file rows = open(labels_file).read().strip().split("\n") classes = [r[r.find(" ") + 1:].split(",")[0] for r in rows] blob = cv2.dnn.blobFromImage(image,1,(224,224),(104,117,123)) print("[INFO] loading model...") net = cv2.dnn.readNetFromCaffe(prototxt,model) net.setPreferableBackend(cv2.dnn.DNN_BACKEND_INFERENCE_ENGINE) print("use dnn target myriad...") net.setPreferableTarget(cv2.dnn.DNN_TARGET_MYRIAD) # set the blob as input to the network and perform a forward-pass to # obtain our output classification net.setInput(blob) start = time.time() preds = net.forward() end = time.time() print("[INFO] classification took " + str((end-start)*1000) + " ms")执行测试
rpi4:~/test$ python3 test-opencv-dnn-vpu.py [INFO] loading model... use dnn target myriad... [INFO] classification took 3804.2752742767334 ms完
欢迎评论交流，收藏备用，点赞转发分享！谢谢！
