基于Intel OpenVINO的搭建及应用，包含分类，目标检测，及分割，超分辨-白红宇

基于Intel OpenVINO的搭建及应用，包含分类，目标检测，及分割，超分辨

阅读量：6150 次

发布时间：2019-06-21

本文共 20217 字，大约阅读时间需要 67 分钟。

PART I: 搭建环境OPENVINO+Tensorflow1.12.0

I: l_openvino_toolkit_p_2019.1.094

第一步常规安装参考链接：

第二步编译Inference Engine Samples:

cd /PATH/TO/deployment_tools/inference_engine/samples

run ./build_samples.sh

编译后的生成文件路径

/root/inference_engine_samples_build/intel64/Release

II. tensorflow编译

Bazel编译Tensorflow

参考链接：https://blog.csdn.net/chenyuping333/article/details/82108509

Bazel下载连接：

()

2. Tensorflow下载连接： (tensorflow-1.12.0)

Step 1: cd /PATH/TO/tensorflows

Step2: ./configure (全选No)

Step3: 编译freeze_graph

bazel build tensorflow/python/tools:freeze_graph

可能会遇到的问题：

error:Python.h:No such file or directory

solutions:yum install python34-devel

note:一定将默认python改为python3，缺少numpy文件，yum安装时报错，更改对应文件头部的/usr/bin/python地址为/usr/bin/python2，使用pip3 安装numpy

错误参考链接:

Step4: 编译transform_graph

bazel build tensorflow/tools/graph_transforms:transform_graph

Step5: 编译summarize_graph

bazel build tensorflow/tools/graph_transforms:summarize_graph

PART II: OPENVINO for Classification

数据集准备：ImageNet val

分别制作六个文件夹每个文件夹内的图片数量依次为1,8,16,32,64,96

形式如下

测试模型：VGG-19,resnet-50,resnet-101,resnet-152,inception-v3,inception-v4

参考链接:

StepI: 下载预训练的模型(6个)

mkdir pretrainedModels && cd pretrainedModels

wget

tar -xvf vgg_19_2016_08_28.tar.gz

wget

tar -xvf inception_v3_2016_08_28.tar.gz

wget

tar -xvf inception_v4_2016_09_09.tar.gz

wget

tar -xvf resnet_v1_50_2016_08_28.tar.gz

wget

tar -xvf resnet_v1_101_2016_08_28.tar.gz

wget

tar -xvf resnet_v1_152_2016_08_28.tar.gz

解压后的形式如下所示

StepII: 生成对应分类预训练模型的.pb文件

cd /PATH/TO/pretrainedModels

mkdir frozen && mkdir pb

(*不同网络的对应指令有些许不同)

1． python3.6 /PATH/TO/tensorflowModels/research/slim/export_inference_graph.py \

--alsologtostderr \

--model_name=vgg_19 \

--output_file=/PATH/TO/pb/vgg_19.pb \

--labels_offset=1

2. python3.6 /PATH/TO/tensorflowModels/research/slim/export_inference_graph.py \

--alsologtostderr \

--model_name=resnet_v1_50 \

--output_file=/PATH/TO/pb/resnet_v1_50.pb \

--labels_offset=1

3. python3.6 /PATH/TO/tensorflowModels/research/slim/export_inference_graph.py \

--alsologtostderr \

--model_name=resnet_v1_101 \

--output_file=/PATH/TO/pb/resnet_v1_101.pb \

--labels_offset=1

4. python3.6 /PATH/TO/tensorflowModels/research/slim/export_inference_graph.py \

--alsologtostderr \

--model_name=resnet_v1_152 \

--output_file=/PATH/TO/pb/resnet_v1_152.pb \

--labels_offset=1

5. python3.6 /PATH/TO/tensorflowModels/research/slim/export_inference_graph.py \

--alsologtostderr \

--model_name=inception_v3 \

--output_file=/PATH/TO/pb/inception_v3.pb

6. python3.6 /PATH/TO/tensorflowModels/research/slim/export_inference_graph.py \

--alsologtostderr \

--model_name=inception_v4 \

--output_file=/PATH/TO/pb/inception_v4.pb

StepIII: 对生成对应分类预训练模型的.pb文件进行freeze操作

(*若下面指令报错可以使用python3.6 运行对应的freeze_graph.py文件进行生成)

cd /PATH/TO/TensorflowModels

1. bazel-bin/tensorflow/python/tools/freeze_graph \

--input_graph=/PATH/TO/pb/vgg_19.pb \

--input_checkpoint=/PATH/TO/vgg_19.ckpt \

--input_binary=true \

--output_graph=/PATH/TO/frozen/frozen_vgg_19.pb \

--output_node_names=vgg_19/fc8/squeezed

2. bazel-bin/tensorflow/python/tools/freeze_graph \

--input_graph=/PATH/TO/pb/resnet_v1_50.pb \

--input_checkpoint=/PATH/TO/resnet_v1_50.ckpt \

--input_binary=true \

--output_graph=/PATH/TO/frozen/frozen_resnet_v1_50.pb \

--output_node_names=resnet_v1_50/predictions/Reshape_1

3. bazel-bin/tensorflow/python/tools/freeze_graph \

--input_graph=/PATH/TO/pb/resnet_v1_101.pb \

--input_checkpoint=/PATH/TO/resnet_v1_101.ckpt \

--input_binary=true \

--output_graph=/PATH/TO/frozen/frozen_resnet_v1_101.pb \

--output_node_names=resnet_v1_101/predictions/Reshape_1

4. bazel-bin/tensorflow/python/tools/freeze_graph \

--input_graph=/PATH/TO/pb/resnet_v1_152.pb \

--input_checkpoint=/PATH/TO/resnet_v1_152.ckpt \

--input_binary=true \

--output_graph=/PATH/TO/frozen/frozen_resnet_v1_152.pb \

--output_node_names=resnet_v1_152/predictions/Reshape_1

5. bazel-bin/tensorflow/python/tools/freeze_graph \

--input_graph=/PATH/TO/pb/inception_v3.pb \

--input_checkpoint=/PATH/TO/inception_v3.ckpt \

--input_binary=true \

--output_graph=/PATH/TO/frozen/frozen_inception_v3.pb \

--output_node_names=InceptionV3/Predictions/Reshape_1

6. bazel-bin/tensorflow/python/tools/freeze_graph \

--input_graph=/PATH/TO/pb/inception_v4.pb \

--input_checkpoint=/PATH/TO/inception_v4.ckpt \

--input_binary=true \

--output_graph=/PATH/TO/frozen/frozen_inception_v4.pb \

--output_node_names=InceptionV4/Logits/Predictions

StepIV: 生成IR文件

cd /PATH/TO /deployment_tools/model_optimizer

1. python3.6 mo.py --framework tf \

--input_model /PATH/TO/frozen/frozen_vgg_19.pb \

--data_type FP32 \

--output_dir /PATH/TO/frozen/ \

--reverse_input_channels \

--batch 64

(batch大小可选1,8,16,32,64,96)

2. python3.6 mo.py --framework tf \

--input_model /PATH/TO/frozen/frozen_inception_v3.pb \

--data_type FP32 \

--scale 255 \

--reverse_input_channels \

--output_dir /PATH/TO/frozen/ \

--batch 16

3. python3.6 mo.py --framework tf \

--input_model /PATH/TO/frozen/frozen_inception_v4.pb \

--data_type FP32 \

--scale 255 \

--reverse_input_channels \

--output_dir /PATH/TO/frozen/ \

--batch 16

4. python3.6 mo.py --framework tf \

--input_model /PATH/TO/frozen/frozen_resnet_v1_50.pb \

--data_type FP32 \

--output_dir /PATH/TO/frozen/ \

--reverse_input_channels \

--batch 16

5. python3.6 mo.py --framework tf \

--input_model /PATH/TO/frozen/frozen_resnet_v1_101.pb \

--data_type FP32 \

--output_dir /PATH/TO/frozen/ \

--reverse_input_channels \

--batch 16

6. python3.6 mo.py --framework tf \

--input_model /PATH/TO/frozen/frozen_resnet_v1_152.pb \

--data_type FP32 \

--output_dir /PATH/TO/frozen/ \

--reverse_input_channels \

--batch 16

StepV:测试(-ni -niter 迭代100次)

cd /root/inference_engine_samples_build/intel64/Release

./classification_sample \

-i /PATH/TO/frozen/frozen_inception_v3.xml \

-d CPU -ni 100 \

-l /PATH/TO/deployment_tools/inference_engine/samples/intel64/Release/lib/libcpu_extension.so \

-nt 1 \

-i /PATH/TO/val1

./benchmark_app \

-m /PATH/TO /frozen/ frozen_inception_v3.xml \

-d CPU -api async -niter 100 \

-l /PATH/TO/deployment_tools/inference_engine/samples/intel64/Release/lib/libcpu_extension.so -nireq 32 \

-i /PATH/TO/val1

(-nireq单CPU核的数量，通过lscpu指令查看)

(上述输入图片的数量与batch的大小相同，对应前面的val数据集，通过更改不同模型对应的.xml文件来测试不同的模型)

PART III: OPENVINO for Object Detection Test

数据集准备：COCO val2017

分别制作六个文件夹每个文件夹内的图片数量依次为1,8,16,32,64,96

形式如下

Note: Tensorflow Object_Detection API参考：

Step I:预训练模型的下载

mkdir object_detection && cd object_detection && mkdir test_models

下载对应的预训练的目标检测模型

模型链接：

解压后的文件格式如下图

Step II: 生成IR文件

参考链接：

（更改SSD,Faster R-CNN,RFCN,Mask R-CNN对应的 .json及pipeline.config文件

对应的.json文件如下

ssd_v2_support.json

faster_rcnn_support.json

rfcn_support.json

mask_rcnn_support.json）

python3.6 mo_tf.py \

--input_model=/PATH/TO/frozen_inference_graph.pb \

--tensorflow_use_custom_operations_config /PATH/TO/deployment_tools/model_optimizer/extensions/front/tf/ssd_v2_support.json \

--tensorflow_object_detection_api_pipeline_config /PATH/TO/pipeline.config \

--reverse_input_channels --batch 16

(batch只针对SSD的大小可调,Faster RCNN 及RFCN只能测试batch为1的情形)

StepIII:测试

For SSD/Faster RCNN/RFCN series

(benchmark_app只适用于SSD)

./benchmark_app \

-m /PATH/TO/frozen_inference_graph.xml \

-d CPU -api async -niter 100 \

-l /PATH/TO/intel64/Debug/lib/libcpu_extension.so \

-nireq 32 -i /PATH/TO /val1/

(-nireq代表单CPU的核数，通过lscpu可以查看，val与batch的大小相对应)

(RFCN, SSD 及 Faster R-CNNs, 测试指令都为object_detection_sample_ssd

***Faster R-CNN/RFCN的只适用于batch=1

)

./object_detection_sample_ssd \

-m /PATH/TO/frozen_inference_graph.xml -d CPU -ni 100 \

-l /PATH/TO/intel64/Debug/lib/libcpu_extension.so -i /PATH/TO /val1/

For Mask R-CNN

参考链接：

./mask_rcnn_demo \

-m /PATH/TO/frozen_inference_graph.xml -d CPU -ni 100 \

-l /root/inference_engine_samples_build/intel64/Debug/lib/libcpu_extension.so \

-i /PATH/TO/coco_val/val64/

在Mask R-CNN进行batch 值大于16以上时，会出现错误

“segmentation fault”是因为在测试时会生成在当前路径下生成图片，随着batch的增加，内存爆掉。此部分代码可以注释，并不影响正常时间的测量。

注释部分代码如下图。

main.cpp 路径：/opt/intel/openvino_2019.1.0394/deployment_tools/inference_engine/samples/mask_rcnn_demo

此处开始注释

此处结束注释

文件修改后重新进行编译

run ./build_samples.sh

PART IV: OpenVINO for DeepLabV3+

Reference：

Tensorflow运行指令(wwen.sh)：

echo 1 > /proc/sys/vm/compact_memory

echo 3 > /proc/sys/vm/drop_caches

echo 100 > /sys/devices/system/cpu/intel_pstate/min_perf_pct

echo 0 > /sys/devices/system/cpu/intel_pstate/no_turbo

echo 0 > /proc/sys/kernel/numa_balancing

cpupower frequency-set -g performance

export KMP_BLOCKTIME=0

export KMP_SETTINGS=1

export KMP_AFFINITY=granularity=fine,compact,1,0

export OMP_NUM_THREADS=16

numactl --physcpubind=0-15,32-47 --membind=0 python3.6 demo_multi.py --input_folder ./img --output_folder ./output --logdir ./model > node_2_1.log 2>&1 &

numactl --physcpubind=16-31,48-63 --membind=1 python3.6 demo_multi.py --input_folder ./img --output_folder ./output --logdir ./model > node_2_2.log 2>&1 &

Step1: IR文件的生成

python3.6 mo_tf.py --input_model /home/gsj/deeplab/research/deeplab/model/frozen_inference_graph.pb --data_type FP32 --output_dir /home/gsj/super-resolution/tf_estimator_barebone/models/ --input 0:xception_65/Pad --output aspp0/Relu,aspp1_pointwise/Relu,aspp2_pointwise/Relu,aspp3_pointwise/Relu,ResizeBilinear_1 --input_shape [1,1953,2593,3](根据下一步inference过程，预处理图片对应输出尺寸进行调整，如下图)

Step2: Inference(infer_IE_TF.py位于intel64/Release下)

echo 1 > /proc/sys/vm/compact_memory

echo 3 > /proc/sys/vm/drop_caches

echo 100 > /sys/devices/system/cpu/intel_pstate/min_perf_pct

echo 0 > /sys/devices/system/cpu/intel_pstate/no_turbo

echo 0 > /proc/sys/kernel/numa_balancing

cpupower frequency-set -g performance

export KMP_BLOCKTIME=0

export KMP_SETTINGS=1

export KMP_AFFINITY=granularity=fine,compact,1,0

export OMP_NUM_THREADS=16

python3.6 infer_IE_TF.py -m /home/gsj/super-resolution/tf_estimator_barebone/models/frozen_inference_graph.xml -i 1.jpg -l lib/libcpu_extension .so

#infer_IE_TF.py代码#author:fourmi_gsjfrom __future__ import print_functionimport sysimport os from argparse import ArgumentParserimport numpy as npimport cv2import time import tensorflow as tffrom tensorflow.python.platform import gfilefrom openvino.inference_engine import IENetwork,IEPlugindef build_argparser():    parser = ArgumentParser()    parser.add_argument("-m", "--model" ,help="Path to an .xml file with a trained model.",required=True,type=str)    parser.add_argument("-i", "--input", help="Path to a folder with images or path to an image files", required=True,                        type=str)    parser.add_argument("-l", "--cpu_extension",                        help="MKLDNN (CPU)-targeted custom layers.Absolute path to a shared library with the kernels "                             "impl.", type=str, default=None)    parser.add_argument("-pp", "--plugin_dir", help="Path to a plugin folder", type=str, default=None)    parser.add_argument("-d", "--device",                        help="Specify the target device to infer on; CPU, GPU, FPGA or MYRIAD is acceptable. Sample "                             "will look for a suitable plugin for device specified (CPU by default)", default="CPU",                        type=str)    parser.add_argument("-nt", "--number_top", help="Number of top results", default=10, type=int)    parser.add_argument("-pc", "--performance", help="Enables per-layer performance report", action='store_true')    return parserdef resoze_for_concat(a0,a1,a2,a3,RB):    iimg_ir = []    '''    print('a0:',a0.shape)    print('a1:',a1.shape)    print('a2:',a2.shape)    print('a3:',a3.shape)    print('RB:',RB.shape)    resize_aspp0 =np.float32(np.zeros((1,256,123,163)))    resize_ResizeBilinear_1=np.float32(np.zeros((1,256,123,163)))    resize_aspp1 =np.float32(np.zeros((1,256,123,163)))    resize_aspp2 =np.float32(np.zeros((1,256,123,163)))    resize_aspp3 =np.float32(np.zeros((1,256,123,163)))    for i in range(256):        resize_aspp0[0,i]=cv2.resize(a0[0,i],(163,123), interpolation=cv2.INTER_LINEAR)        resize_ResizeBilinear_1[0,i]=cv2.resize(RB[0,i],(163,123), interpolation=cv2.INTER_LINEAR)        resize_aspp1[0,i]=cv2.resize(a1[0,i],(163,123), interpolation=cv2.INTER_LINEAR)        resize_aspp2[0,i]=cv2.resize(a2[0,i],(163,123), interpolation=cv2.INTER_LINEAR)        resize_aspp3[0,i]=cv2.resize(a3[0,i],(163,123), interpolation=cv2.INTER_LINEAR)    ResizeBilinear_1=resize_ResizeBilinear_1.transpose((0,2,3,1))    aspp0=resize_aspp0.transpose((0,2,3,1))    aspp1=resize_aspp1.transpose((0,2,3,1))    aspp2=resize_aspp2.transpose((0,2,3,1))    aspp3=resize_aspp3.transpose((0,2,3,1))    '''    ResizeBilinear_1=RB.transpose((0,2,3,1))    aspp0=a0.transpose((0,2,3,1))    aspp1=a1.transpose((0,2,3,1))    aspp2=a2.transpose((0,2,3,1))    aspp3=a3.transpose((0,2,3,1))    '''    print(aspp0.shape)    print(aspp1.shape)    print(aspp2.shape)    print(aspp3.shape)    print(ResizeBilinear_1.shape)    '''    iimg_ir.append(ResizeBilinear_1)    iimg_ir.append(aspp0)    iimg_ir.append(aspp1)    iimg_ir.append(aspp2)    iimg_ir.append(aspp3)    return iimg_irclass _model_preprocess():    def __init__(self):        graph = tf.Graph()        f_handle = gfile.FastGFile("/home/gsj/deeplab/research/deeplab/model/frozen_inference_graph.pb",'rb')        graph_def = tf.GraphDef.FromString(f_handle.read())        with graph.as_default():            tf.import_graph_def(graph_def,name='')        self.sess = tf.Session(graph=graph)    def _pre_process(self,image):        seg_map = self.sess.run('sub_7:0',feed_dict={
   'ImageTensor:0':[image]})        #print('The shape of the seg_map is :',seg_map.shape)        return seg_mapclass _model_postprocess():    def __init__(self):        graph = tf.Graph()        f_handle = gfile.FastGFile("/home/gsj/deeplab/research/deeplab/model/frozen_inference_graph.pb",'rb')        graph_def = tf.GraphDef.FromString(f_handle.read())                with graph.as_default():                        new_input0=tf.placeholder(tf.float32,shape=(1,123,163,256),name='new_input0')            new_input1=tf.placeholder(tf.float32,shape=(1,123,163,256),name='new_input1')            new_input2=tf.placeholder(tf.float32,shape=(1,123,163,256),name='new_input2')            new_input3=tf.placeholder(tf.float32,shape=(1,123,163,256),name='new_input3')            new_input4=tf.placeholder(tf.float32,shape=(1,123,163,256),name='new_input4')            tf.import_graph_def(graph_def,input_map={
   'ResizeBilinear_1:0':new_input0,'aspp0/Relu:0':new_input1,'aspp1_pointwise/Relu:0':new_input2,'aspp2_pointwise/Relu:0':new_input3,'aspp3_pointwise/Relu:0':new_input4},name='')                    self.sess = tf.Session(graph=graph)    def _post_process(self,image_ir,image):        seg_map = self.sess.run('SemanticPredictions:0', feed_dict={
   'ImageTensor:0': [image], 'new_input0:0': image_ir[0],                                    'new_input1:0': image_ir[1],'new_input2:0': image_ir[2],'new_input3:0': image_ir[3],                                    'new_input4:0': image_ir[4]})        return seg_map_pre = _model_preprocess()_post = _model_postprocess()def main_IE_infer():    args = build_argparser().parse_args()    model_xml = args.model    model_bin = os.path.splitext(model_xml)[0] + ".bin"    image = cv2.imread(args.input)    print("The size of the orig image is:",image.shape[0],image.shape[1])        h_input_size=1360 #the height of the output    w_input_size=1020 #the width of the output        h_ratio = 1.0 * h_input_size / image.shape[0]     w_ratio = 1.0 * w_input_size / image.shape[1]     shrink_size = (int(w_ratio * image.shape[1]),int(h_ratio*image.shape[0]))    image = cv2.resize(image,shrink_size, interpolation=cv2.INTER_LINEAR)    print("The shape of the resized Image is:",image.shape)            # Plugin initialization for specified device and load extensions library if specified    plugin = IEPlugin(device=args.device, plugin_dirs=args.plugin_dir)    if args.cpu_extension and 'CPU' in args.device:        plugin.add_cpu_extension(args.cpu_extension)    if args.performance:        plugin.set_config({
    "PERF_COUNT": "YES"})    # Read IR    net = IENetwork.from_ir(model=model_xml, weights=model_bin)    #print("the output Info of the net is :",net.outputs)    input_blob = next(iter(net.inputs))    print('input_blob is :',input_blob)    exec_net = plugin.load(network=net)    img_ir = []          for itr in range(1):        now = time.time()        image_ = _pre._pre_process(image)        image_ = image_.transpose((0,3,1,2))        #print("the shape of the Front Net'output:",image_.shape)                res =exec_net.infer(inputs={input_blob:image_})        #print(res.keys())        aspp0 = res['aspp0/Relu']        aspp1 = res['aspp1_pointwise/Relu']        aspp2 = res['aspp2_pointwise/Relu']        aspp3 = res['aspp3_pointwise/Relu']        ResizeBilinear_1=res['ResizeBilinear_1']        img_ir = resoze_for_concat(aspp0,aspp1,aspp2,aspp3,ResizeBilinear_1)        result = _post._post_process(img_ir,image)[0]    print('time cost:',time.time()-now)        #print(result)    result[result!=0]=255            cv2.imwrite('./result_deeplabv3.jpg', result)    del net    del exec_net    del plugin    if __name__=='__main__':    sys.exit(main_IE_infer() or 0)

PART V: OpenVINO for Super Resolution

Step 1: tensorflow进行测试

Github:

运行inference程序，查找该模型的输出节点，查找到的节点名称为”clip_by_value”

指令执行路径：/home/gsj/super-resolution/tf_estimator_barebone/

运行相关指令如下：

export KMP_BLOCKTIME=1

export KMP_AFFINITY=granularity=fine,compact,1,0

export OMP_NUM_THREADS=16

numactl -C 0-15,32-47 -m 0 python3.6 -m datasets.div2k --mode wdsr --model-dir /home/gsj/super-resolution/tf_estimator_barebone/models/ --input-dir /home/gsj/super-resolution/tf_estimator_barebone/data/DIV2K_valid_HR/ --output-dir ../output

Step 2:

将文件夹models下的模型相关文件(saved_model.pb, variabels文件夹)进行处理，freeze saved_model.pb文件，生成pruned_saved_model_or_whatever.pb文件

指令执行路径：/home/gsj/super-resolution/tf_inference_demo/tensorflow-1.12.0