DLAP221上Yolov5s性能测试
2024-05-09 19:12:31
1. 导出yolov5 ONNX模型
选择一台x64的PC机，OS为Ubuntu18.04，在此台机器上完成Yolov5模型转ONNX模型。
1.1 下载原始yolov5代码、模型和权重文件
下载yolov5源代码
$ git clone -b v2.0 https://github.com/ultralytics/yolov5.git Cloning into 'yolov5'... remote: Enumerating objects: 6233, done. remote: Counting objects: 100% (27/27), done. remote: Compressing objects: 100% (20/20), done. remote: Total 6233 (delta 10), reused 14 (delta 7), pack-reused 6206 Receiving objects: 100% (6233/6233), 8.45 MiB | 19.00 KiB/s, done. Resolving deltas: 100% (4268/4268), done. Note: checking out '2a149a48d65bf4d358d4289058ef6066cda8c7c9'.You are in 'detached HEAD' state. You can look around, make experimental changes and commit them, and you can discard any commits you make in this state without impacting any branches by performing another checkout.If you want to create a new branch to retain commits you create, you may do so (now or later) by using -b with the checkout command again. Example:git checkout -b <new-branch-name>下载权重文件
$ cd yolov5/weights$ wget https://github.com/ultralytics/yolov5/releases/download/v2.0/yolov5s.pt如果不能在设备上直接下载权重文件，请在别的平台下载权重文件，然后拷贝到设备上。
1.2 Pytorch模型转ONNX模型
修改models/export.py脚本生成onnx文件，注意目前onnx版本应选择11
$ cd ~/yolov5/models $ vim export.py ... #省略部分# ONNX exporttry:import onnxprint('\nStarting ONNX export with onnx %s...' % onnx.__version__)f = opt.weights.replace('.pt', '.onnx') # filenamemodel.fuse() # only for ONNX####修改此处的opset_version为11######torch.onnx.export(model, img, f, verbose=False, opset_version=11, input_names=['images'],output_names=['classes', 'boxes'] if y is None else ['output'])# Checksonnx_model = onnx.load(f) # load onnx modelonnx.checker.check_model(onnx_model) # check onnx modelprint(onnx.helper.printable_graph(onnx_model.graph)) # print a human readable modelprint('ONNX export success, saved as %s' % f)except Exception as e:print('ONNX export failure: %s' % e) ...#省略部分安装docker，访问https://hub.docker.com/r/shinerchen/pytorch-1.8.1-py37-cpu下载docker 镜像
$sudo apt-get install docker.io $sudo docker version Client:Version: 20.10.2API version: 1.41Go version: go1.13.8Git commit: 20.10.2-0ubuntu1~18.04.2Built: Tue Mar 30 21:24:16 2021OS/Arch: linux/amd64Context: defaultExperimental: trueServer:Engine:Version: 20.10.2API version: 1.41 (minimum version 1.12)Go version: go1.13.8Git commit: 20.10.2-0ubuntu1~18.04.2Built: Mon Mar 29 19:27:41 2021OS/Arch: linux/amd64Experimental: falsecontainerd:Version: 1.3.3-0ubuntu1~18.04.4GitCommit:runc:Version: spec: 1.0.2-devGitCommit:docker-init:Version: 0.19.0GitCommit:$ sudo docker pull shinerchen/pytorch-1.8.1-py37-cpu $ sudo docker images REPOSITORY TAG IMAGE ID CREATED SIZE shinerchen/pytorch-1.8.1-py37-cpu latest 3f812182d768 2 hours ago 1.98GB运行docker镜像，导出ONNX模型文件
$ sudo docker run -it --ipc=host -v <你的yolov5的代码目录>:/data pytorch-1.8.1-py37-cpu:latest $ cd /data $ python3.7 models/export.py --weights ./weights/yolov5s.pt --img 640 --batch 1 Namespace(batch_size=1, img_size=[640, 640], weights='./weights/yolov5s.pt') [W NNPACK.cpp:80] Could not initialize NNPACK! Reason: Unsupported hardware.Starting TorchScript export with torch 1.8.1+cpu... /usr/local/lib/python3.7/dist-packages/torch/jit/_trace.py:940: TracerWarning: Encountering a list at the output of the tracer might cause the trace to be incorrect, this is only valid if the container structure does not change based on the module's inputs. Consider using a constant container instead (e.g. for `list`, use a `tuple` instead. for `dict`, use a `NamedTuple` instead). If you absolutely need this and know the side effects, pass strict=False to trace() to allow this behavior._force_outplace, TorchScript export success, saved as ./weights/yolov5s.torchscript.ptStarting ONNX export with onnx 1.9.0... ... ONNX export success, saved as ./weights/yolov5s.onnxStarting CoreML export with coremltools 4.1... WARNING:root:Tuple detected at graph output. This will be flattened in the converted model. Converting Frontend ==> MIL Ops: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████▋| 730/732 [00:00<00:00, 779.53 ops/s] Running MIL optimization passes: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 18/18 [00:00<00:00, 55.74 passes/s] Translating MIL ==> MLModel Ops: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████| 944/944 [00:01<00:00, 768.94 ops/s] CoreML export success, saved as ./weights/yolov5s.mlmodelExport complete. Visualize with https://github.com/lutzroeder/netron.简化模型
对导出的onnx模型使用onnx-simplifer工具进行简化
$ pip3.7 install onnx-simplifier $ cd yolov5/weights $ python3.7 -m onnxsim --skip-optimization yolov5s.onnx yolov5s_sim.onnx Simplifying... Checking 0/3... Checking 1/3... Checking 2/3... Ok!创建包含以下代码的modify_yolov5.py脚本
# Copyright 2020 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import sys import onnxINT_MAX = sys.maxsizeif len(sys.argv) == 1:print("usage: python3.7 modify_yolov5.py model.onnx [n,c,h,w]")exit(0)model_path = sys.argv[1] model = onnx.load(model_path)b, c, h, w = 1, 3, 640, 640 if len(sys.argv) == 3:b, c, h, w = list(map(int, sys.argv[2].split(',')))print(f"input shape: {b, c, h, w}")def get_node_by_name(nodes, name):for n in nodes:if n.name == name:return nreturn -1""" before: after:input input/ / \ \ / \ slice4 slice14 slice24 slice34 slice4 slice14| | | | | | slice9 slice19 slice29 slice39 t t\ \ / / / \ / \concat slice9 slice29 slice19 slice39| | | |t t t t\ \ / /concat """model.graph.node.remove(get_node_by_name(model.graph.node, "Slice_24")) model.graph.node.remove(get_node_by_name(model.graph.node, "Slice_34"))prob_info1 = onnx.helper.make_tensor_value_info('to_slice9', onnx.TensorProto.FLOAT, [b, c, h, w//2]) prob_info3 = onnx.helper.make_tensor_value_info('to_slice19', onnx.TensorProto.FLOAT, [b, c, h, w//2]) prob_info5 = onnx.helper.make_tensor_value_info('from_slice9', onnx.TensorProto.FLOAT, [b, c, h//2, w//2]) prob_info6 = onnx.helper.make_tensor_value_info('from_slice19', onnx.TensorProto.FLOAT, [b, c, h//2, w//2]) prob_info7 = onnx.helper.make_tensor_value_info('from_slice29', onnx.TensorProto.FLOAT, [b, c, h//2, w//2]) prob_info8 = onnx.helper.make_tensor_value_info('from_slice39', onnx.TensorProto.FLOAT, [b, c, h//2, w//2])# slice4 slice24 后的Transpose,若自定义模型，inputs中的编号需根据onnx图中实际编号修改 node1 = onnx.helper.make_node('Transpose',inputs=[get_node_by_name(model.graph.node, "Slice_4").output[0]],outputs=['to_slice9'],perm=[0, 1, 3, 2] ) node3 = onnx.helper.make_node('Transpose',inputs=[get_node_by_name(model.graph.node, "Slice_14").output[0]],outputs=['to_slice19'],perm=[0, 1, 3, 2] ) # slice9 slice19 slice29 slice39后的Transpose node5 = onnx.helper.make_node('Transpose',inputs=['from_slice9'],outputs=[get_node_by_name(model.graph.node, "Slice_9").output[0]],perm=[0, 1, 3, 2] ) node6 = onnx.helper.make_node('Transpose',inputs=['from_slice19'],outputs=[get_node_by_name(model.graph.node, "Slice_19").output[0]],perm=[0, 1, 3, 2] ) node7 = onnx.helper.make_node('Transpose',inputs=['from_slice29'],outputs=[get_node_by_name(model.graph.node, "Slice_29").output[0]],perm=[0, 1, 3, 2] ) node8 = onnx.helper.make_node('Transpose',inputs=['from_slice39'],outputs=[get_node_by_name(model.graph.node, "Slice_39").output[0]],perm=[0, 1, 3, 2] ) model.graph.node.append(node1) model.graph.node.append(node3) model.graph.node.append(node5) model.graph.node.append(node6) model.graph.node.append(node7) model.graph.node.append(node8)# slice9 slice19 换轴 model.graph.initializer.append(onnx.helper.make_tensor('starts_9', onnx.TensorProto.INT64, [1], [0])) model.graph.initializer.append(onnx.helper.make_tensor('ends_9', onnx.TensorProto.INT64, [1], [INT_MAX])) model.graph.initializer.append(onnx.helper.make_tensor('axes_9', onnx.TensorProto.INT64, [1], [2])) model.graph.initializer.append(onnx.helper.make_tensor('steps_9', onnx.TensorProto.INT64, [1], [2])) newnode1 = onnx.helper.make_node('Slice',name='Slice_9',inputs=['to_slice9', 'starts_9', 'ends_9', 'axes_9', 'steps_9'],outputs=['from_slice9'] ) model.graph.node.remove(get_node_by_name(model.graph.node, "Slice_9")) model.graph.node.insert(9, newnode1) newnode2 = onnx.helper.make_node('Slice',name='Slice_19',inputs=['to_slice19', 'starts_9', 'ends_9', 'axes_9', 'steps_9'],outputs=['from_slice19'] ) model.graph.node.remove(get_node_by_name(model.graph.node, "Slice_19")) model.graph.node.insert(19, newnode2)# slice29 slice39 换轴 model.graph.initializer.append(onnx.helper.make_tensor('starts_29', onnx.TensorProto.INT64, [1], [1])) model.graph.initializer.append(onnx.helper.make_tensor('ends_29', onnx.TensorProto.INT64, [1], [INT_MAX])) model.graph.initializer.append(onnx.helper.make_tensor('axes_29', onnx.TensorProto.INT64, [1], [2])) model.graph.initializer.append(onnx.helper.make_tensor('steps_29', onnx.TensorProto.INT64, [1], [2])) newnode3 = onnx.helper.make_node('Slice',name='Slice_29',inputs=['to_slice9', 'starts_29', 'ends_29', 'axes_29', 'steps_29'],outputs=['from_slice29'] ) model.graph.node.remove(get_node_by_name(model.graph.node, "Slice_29")) model.graph.node.insert(29, newnode3) newnode4 = onnx.helper.make_node('Slice',name='Slice_39',inputs=['to_slice19', 'starts_29', 'ends_29', 'axes_29', 'steps_29'],outputs=['from_slice39'] ) model.graph.node.remove(get_node_by_name(model.graph.node, "Slice_39")) model.graph.node.insert(39, newnode4)onnx.save(model, model_path.split('.')[0] + "_t.onnx") print("success")修改生成的onnx模型文件
$ cd yolov5/weights $ python3.7 modify_yolov5.py yolov5s_sim.onnx #生成的模型名为：yolov5s_sim_t.onnx
1.3 升级DLAP221的系统
由于华为在CANN3.3.0版本及配套的1.0.9固件版本中针对yolov5做了优化，如果DLAP221的系统版本不是此版本，建议直接将DLAP221的SD系统整体升级到此版本。Adlink提供一键制卡工具包，方便客户制作该版本的系统。制卡工具包可以从以下的网盘地址下载。阅读工具包中README文件进行制卡。
制卡
###注意以下步骤，不要在Docker中执行，在PC机的Host环境(Ubuntu18.04)中执行!!! ###通过如下命令进行安装制卡工具包的系统依赖。 $ sudo apt-get install -y qemu-user-static python3-yaml binfmt-support gcc-aarch64-linux-gnu g++-aarch64-linux-gnu expect unzip squashfs-tools #注意Ubuntu 18.04系统：“gcc-aarch64-linux-gnu”与“g++-aarch64-linux-gnu”版本要求为7.4.0，其他依赖软件包无版本要求。默认安装的gcc版本为7.4.0。 $ tar xvf sd_maker_1.0.9_can3.3.0alpha006.tgz $ cd sd_maker #注意将micro SD卡插入PC机，如果PC机没有Micro SD卡槽，请用Micro SD读卡器进行转接。 $ sudo python3.7.5 make_sd_card.py local /dev/sdX #这里的sdX需要根据Micro SD卡插入Linux系统后生成的磁盘名来修改！！启动系统
将制好的卡插入DLAP221设备，上电启动，第一次启动系统会自动安装一些的平台依赖包和开发工具包，整个过程大概10多分钟，完成后系统会自动重启。建议用户接串口线来监视整个过程。串口波特率为115200 8N1，具体可以参考DLAP221的使用手册。
2. 模型转换
在DLAP221上完成ONNX模型到OM模型的转换
2.1 安装昇腾软件包
| 软件 | 版本 | 说明 | 获取方式 |
|---|---|---|---|
| mxManufacture | 2.0.1 | mxManufacture软件包 | mxManufacture下载页面 选择下载ARM安装包 |
| Ascend-CANN-toolkit | 3.3.0.alpha006 | Ascend-cann-toolkit开发套件包，社区版 | 社区版官网获取链接选择历史版本中的3.3.0.alpha006 |
安装Ascend Toolkit
从上表中的链接下载toolkit，保存到/home/HwHiAiUser目录中并安装
$ chmod +x Ascend-cann-toolkit_3.3.0.alpha006_linux-aarch64.run $ ./Ascend-cann-toolkit_3.3.0.alpha006_linux-aarch64.run --install --chip=Ascend310-minirc安装mxManufacture
从上表中的链接下载mxManufacture，保存到/home/HwHiAiUser目录中并安装
$ chmod +x Ascend-mindxsdk-mxmanufacture_2.0.1_linux-aarch64.run $ ./Ascend-mindxsdk-mxmanufacture_2.0.1_linux-aarch64.run --install $ source ~/.bashrc # 激活mxManufacture环境变量运行结果：生成mxManufacture-2.0.1目录和mxManufacuture软链接。
2.2 模型转换
将Yolov5的ONNX模型拷贝到DLAP221上
$ scp yolov5s_sim_t.onnx HwHiAiUser@<DLAP221的IP>:~/yolov5创建aipp配置文件aipp_yolov5s.cfg, 并输入以下内容
aipp_op { aipp_mode : static related_input_rank : 0 input_format : YUV420SP_U8 src_image_size_w : 640 src_image_size_h : 640 crop : false csc_switch : true rbuv_swap_switch : false matrix_r0c0 : 256 matrix_r0c1 : 0 matrix_r0c2 : 359 matrix_r1c0 : 256 matrix_r1c1 : -88 matrix_r1c2 : -183 matrix_r2c0 : 256 matrix_r2c1 : 454 matrix_r2c2 : 0 input_bias_0 : 0 input_bias_1 : 128 input_bias_2 : 128 var_reci_chn_0 : 0.0039216 var_reci_chn_1 : 0.0039216 var_reci_chn_2 : 0.0039216 }模型转换
$ source env.sh ##此步可选，默认制卡时已将环境变量设置好了 $ atc --model=./yolov5s_sim_t.onnx \--framework=5 \--output=./yolov5s_sim_t \--input_format=NCHW \--input_shape="images:1,3,640,640" \--enable_small_channel=1 \--insert_op_conf=./aipp_yolov5s.cfg \--soc_version=Ascend310 \--log=info
3. 推理测试
3.1 准备
编写pipeline文件
已为您编排好使用yolov5s进行目标检测的pipeline文件如下，该pipeline中只包括一条名为"classification+detection"的业务流，该业务流运行在0号芯片（DLAP221也只有一颗atlas200），由appsrc、mxpi_imagedecoder、mxpi_imageresize、mxpi_modelinfer、mxpi_dataserialize、appsink六个插件依次串连而成。注意，mxpi_imageresize插件设置缩放后的图片宽高为640*640且为等比缩放（“resizeType”: “Resizer_KeepAspectRatio_Fit”）；mxpi_modelinfer插件中使用了与yolov5s_sim_t.om相适应的配置文件yolov5s.cfg和标签文件coco.names，均在此文档中提供。
yolov5s_example.pipelines
{"classification+detection": {"stream_config": {"deviceId": "0"},"mxpi_imagedecoder0": {"factory": "mxpi_imagedecoder","next": "mxpi_imageresize0"},"mxpi_imageresize0": {"props": {"parentName": "mxpi_imagedecoder0","resizeHeight": "640","resizeWidth": "640","resizeType": "Resizer_KeepAspectRatio_Fit"},"factory": "mxpi_imageresize","next": "mxpi_modelinfer0"},s"mxpi_modelinfer0": {"props": {s"parentName": "mxpi_imageresize0","modelPath": "../models/yolov5s/yolov5s_sim_t.om","postProcessConfigPath": "../models/yolov5s/yolov5s.cfg","labelPath": "../models/yolov5s/coco.names","postProcessLibPath": "../../../lib/libMpYOLOv5PostProcessor.so"},"factory": "mxpi_modelinfer","next": "mxpi_dataserialize0"},"mxpi_dataserialize0": {"props": {"outputDataKeys": "mxpi_modelinfer0"},"factory": "mxpi_dataserialize","next": "appsink0"},"appsrc0": {"props": {"blocksize": "409600"},"factory": "appsrc","next": "mxpi_imagedecoder0"},"appsink0": {"props": {"blocksize": "4096000"},"factory": "appsink"}} }将如上文件命名为yolov5s_example.pipeline，并保存至/home/HwHiAiUser/mxManufacture-2.0.1/samples/mxManufacture/pipeline目录。
编写yolov5s配置文件yolov5s.cfg
CLASS_NUM=80 BIASES_NUM=18 BIASES=10,13,16,30,33,23,30,61,62,45,59,119,116,90,156,198,373,326 SCORE_THRESH=0.6 OBJECTNESS_THRESH=0.6 IOU_THRESH=0.5 YOLO_TYPE=3 ANCHOR_DIM=3 MODEL_TYPE=1 RESIZE_FLAG=0将如上文件命名为yolov5s.cfg，并保存至/home/HwHiAiUser/mxManufacture-2.0.1/samples/mxManufacture/models/yolov5s目录。
标签文件coco.names
# This file is originally from https://github.com/pjreddie/darknet/blob/master/data/coco.names person bicycle car motorbike aeroplane bus train truck boat traffic light fire hydrant stop sign parking meter bench bird cat dog horse sheep cow elephant bear zebra giraffe backpack umbrella handbag tie suitcase frisbee skis snowboard sports ball kite baseball bat baseball glove skateboard surfboard tennis racket bottle wine glass cup fork knife spoon bowl banana apple sandwich orange broccoli carrot hot dog pizza donut cake chair sofa pottedplant bed diningtable toilet tvmonitor laptop mouse remote keyboard cell phone microwave oven toaster sink refrigerator book clock vase scissors teddy bear hair drier toothbrush将如上文件命名为coco.names，并保存至/home/HwHiAiUser/mxManufacture-2.0.1/samples/mxManufacture/models/yolov5s目录。
离线模型文件
拷贝yolov5s_sim_t.om文件到/home/HwHiAiUser/mxManufacture-2.0.1/samples/mxManufacture/models/yolov5s目录
至此，你已准备好运行yolov5s sample需要的所有文件。
4. 推理应用调度代码开发
4.1 理解并运行sample code
mxManufacture已为你准备好sample code，在/home/HwHiAiUser/mxManufacture-2.0.1/samples/mxManufacture/C++和/home/HwHiAiUser/mxManufacture-2.0.1/samples/mxManufacture/Python目录下。我们以C++代码为例，为了不影响默认提供的demo代码，我们另外在/home/HwHiAiUser/mxManufacture-2.0.1/samples/mxManufacture目录中创建一个文件夹：yolov5s_C++, 同时copy C++目录中的main.cpp和run.sh到我们新创建的目录。
$ cd /home/HwHiAiUser/mxManufacture-2.0.1/samples/mxManufacture $ mkdir yolov5s_C++ $ cp C++/main.cpp yolov5s_C++/ $ cp C++/run.sh yolov5s_C++/ $ ls main.cpp run.sh编辑main.cpp
打开main.cpp文件，在main函数中指定待推理的图片（88行）和待创建的业务流编排文件（94行），然后依次调用mxManufacture初化接口（102行，InitManager，初化始日志级别等）、业务流创建接口（108行，CreateMultipleStreams，根据pipeline文件创建业务流）、发送推理请求接口（116行，SendData，发送推理请求，参数中指定业务流名称和待推理图片）、获取推理结果接口（122行，GetResult，获取推理结果）和业务流销毁接口（132行，DestroyAllStreams，程序退出前调用）。
1 /*2 * Copyright (c) 2020.Huawei Technologies Co., Ltd. All rights reserved.3 *4 * Licensed under the Apache License, Version 2.0 (the "License");5 * you may not use this file except in compliance with the License.6 * You may obtain a copy of the License at7 *8 * http://www.apache.org/licenses/LICENSE-2.09 *10 * Unless required by applicable law or agreed to in writing, software11 * distributed under the License is distributed on an "AS IS" BASIS,12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13 * See the License for the specific language governing permissions and14 * limitations under the License.15 */1617 #include <cstring>18 #include "MxBase/Log/Log.h"19 #include "MxStream/StreamManager/MxStreamManager.h"20 namespace {21 APP_ERROR ReadFile(const std::string& filePath, MxStream::MxstDataInput& dataBuffer)22 {23 char c[PATH_MAX + 1] = { 0x00 };24 size_t count = filePath.copy(c, PATH_MAX + 1);25 if (count != filePath.length()) {26 LogError << "Failed to copy file path(" << c << ").";27 return APP_ERR_COMM_FAILURE;28 }29 // Get the absolute path of input file30 char path[PATH_MAX + 1] = { 0x00 };31 if ((strlen(c) > PATH_MAX) || (realpath(c, path) == nullptr)) {32 LogError << "Failed to get image, the image path is (" << filePath << ").";33 return APP_ERR_COMM_NO_EXIST;34 }35 // Open file with reading mode36 FILE *fp = fopen(path, "rb");37 if (fp == nullptr) {38 LogError << "Failed to open file (" << path << ").";39 return APP_ERR_COMM_OPEN_FAIL;40 }41 // Get the length of input file42 fseek(fp, 0, SEEK_END);43 long fileSize = ftell(fp);44 fseek(fp, 0, SEEK_SET);45 // If file not empty, read it into FileInfo and return it46 if (fileSize > 0) {47 dataBuffer.dataSize = fileSize;48 dataBuffer.dataPtr = new (std::nothrow) uint32_t[fileSize];49 if (dataBuffer.dataPtr == nullptr) {50 LogError << "allocate memory with \"new uint32_t\" failed.";51 return APP_ERR_COMM_FAILURE;52 }5354 uint32_t readRet = fread(dataBuffer.dataPtr, 1, fileSize, fp);55 if (readRet <= 0) {56 fclose(fp);57 return APP_ERR_COMM_READ_FAIL;58 }59 fclose(fp);60 return APP_ERR_OK;61 }62 fclose(fp);63 return APP_ERR_COMM_FAILURE;64 }6566 std::string ReadPipelineConfig(const std::string& pipelineConfigPath)67 {68 std::ifstream file(pipelineConfigPath.c_str(), std::ifstream::binary);69 if (!file) {70 LogError << pipelineConfigPath <<" file dose not exist.";71 return "";72 }73 file.seekg(0, std::ifstream::end);74 uint32_t fileSize = file.tellg();75 file.seekg(0);76 std::unique_ptr<char[]> data(new char[fileSize]);77 file.read(data.get(), fileSize);78 file.close();79 std::string pipelineConfig(data.get(), fileSize);80 return pipelineConfig;81 }82 }8384 int main(int argc, char* argv[])85 {86 // read image file and build stream input87 MxStream::MxstDataInput dataBuffer;88 APP_ERROR ret = ReadFile("./test.jpg", dataBuffer);89 if (ret != APP_ERR_OK) {90 LogError << GetError(ret) << "Failed to read image file.";91 return ret;92 }93 // read pipeline config file94 std::string pipelineConfigPath = "../pipeline/yolov5s_example.pipeline";95 std::string pipelineConfig = ReadPipelineConfig(pipelineConfigPath);96 if (pipelineConfig == "") {97 LogError << "Read pipeline failed.";98 return APP_ERR_COMM_INIT_FAIL;99 } 100 // init stream manager 101 MxStream::MxStreamManager mxStreamManager; 102 ret = mxStreamManager.InitManager(); 103 if (ret != APP_ERR_OK) { 104 LogError << GetError(ret) << "Failed to init Stream manager."; 105 return ret; 106 } 107 // create stream by pipeline config file 108 ret = mxStreamManager.CreateMultipleStreams(pipelineConfig); 109 if (ret != APP_ERR_OK) { 110 LogError << GetError(ret) << "Failed to create Stream."; 111 return ret; 112 } 113 std::string streamName = "classification+detection"; 114 int inPluginId = 0; 115 // send data into stream 116 ret = mxStreamManager.SendData(streamName, inPluginId, dataBuffer); 117 if (ret != APP_ERR_OK) { 118 LogError << GetError(ret) << "Failed to send data to stream."; 119 return ret; 120 } 121 // get stream output 122 MxStream::MxstDataOutput* output = mxStreamManager.GetResult(streamName, inPluginId); 123 if (output == nullptr) { 124 LogError << "Failed to get pipeline output."; 125 return ret; 126 } 127 128 std::string result = std::string((char *)output->dataPtr, output->dataSize); 129 LogInfo << "Results:" << result; 130 131 // destroy streams 132 mxStreamManager.DestroyAllStreams(); 133 delete dataBuffer.dataPtr; 134 dataBuffer.dataPtr = nullptr; 135 136 delete output; 137 return 0; 138 }实际开发时，用户可以参考如上代码调用mxManufacture的五个接口:InitManager, CreateMultipleStreams, SendData, GetResult, DestroyAllStreams, 即可在自己的推理应用中使用mxManufacture进行推理了。这里我们指定业务流编排文件为3.1中编写好的yolov5s_example.pipeline文件，图片使用C++目录下的默认图片（test.jpg，是一张背景纯白的狗），无需修改。在C++目录下运行如下命令，可在日志中看到推理结果。
$ bash run.sh 省略部分输出.... I0518 04:23:20.026540 7934 MxGstBase.cpp:764] element(mxpi_imageresize0) gst_change_state NULL_TO_READY. I0518 04:23:20.026688 7934 MxGstBase.cpp:632] element(mxpi_imageresize0) dynamicImageSize=640,640; I0518 04:23:20.026892 7934 MxGstBase.cpp:764] element(mxpi_imagedecoder0) gst_change_state NULL_TO_READY. I0518 04:23:20.027192 7934 MxStreamManager.cpp:384] Creates stream(classification+detection) successfully. I0518 04:23:20.027390 7934 MxStreamManager.cpp:434] Creates streams successfully. I0518 04:23:20.075621 7934 main.cpp:129] Results:{"MxpiObject":[{"classVec":[{"classId":16,"className":"dog","confidence":0.67863529899999997,"headerVec": []}],"x0":72.779052699999994,"x1":941.77044699999999,"y0":132.82678200000001,"y1":604.33477800000003}]} ##此行为推理结果！！ I0518 04:23:20.075755 7934 MxsmStream.cpp:688] Begin to destroy stream(classification+detection). I0518 04:23:20.075904 7934 MxsmStream.cpp:743] Send custom eos to the Stream successfully. I0518 04:23:20.076239 7934 MxsmStream.cpp:749] Send eos to the Stream successfully. I0518 04:23:20.076385 7934 MxsmStream.cpp:755] Flushes the Stream data successfully. 省略部分输出....
4.2 推理结果可视化&性能打点
为将推理结果更加直观地显示并简单测试推理性能，对run.sh脚本和main.cpp文件作简单修改：
run.sh脚本包含opencv的头文件和引入opencv_world动态库。
#!/bin/bash# Copyright 2020 Huawei Technologies Co., Ltd # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License.set -eCUR_PATH=$(cd "$(dirname "$0")" || { warn "Failed to check path/to/run.sh" ; exit ; } ; pwd)# Simple log helper functions info() { echo -e "\033[1;34m[INFO ][MxStream] $1\033[1;37m" ; } warn() { echo >&2 -e "\033[1;31m[WARN ][MxStream] $1\033[1;37m" ; }export MX_SDK_HOME="${CUR_PATH}/../../.." export LD_LIBRARY_PATH="${MX_SDK_HOME}/lib":"${MX_SDK_HOME}/opensource/lib":"${MX_SDK_HOME}/opensource/lib64":"/usr/local/Ascend/ascend-toolkit/latest/acllib/lib64":${LD_LIBRARY_PATH} export GST_PLUGIN_SCANNER="${MX_SDK_HOME}/opensource/libexec/gstreamer-1.0/gst-plugin-scanner" export GST_PLUGIN_PATH="${MX_SDK_HOME}/opensource/lib/gstreamer-1.0":"${MX_SDK_HOME}/lib/plugins"# complie ###包含opencv的头文件和引入opencv_world动态库 g++ main.cpp -I "${MX_SDK_HOME}/include/" -I "${MX_SDK_HOME}/opensource/include/" -I "${MX_SDK_HOME}/opensource/include/opencv4" -L "${MX_SDK_HOME}/lib/" -L "${MX_SDK_HOME}/opensource/lib/" -L "${MX_SDK_HOME}/opensource/lib64/" -std=c++11 -D_GLIBCXX_USE_CXX11_ABI=0 -Dgoogle=mindxsdk_private -fPIC -fstack-protector-all -g -Wl,-z,relro,-z,now,-z,noexecstack -pie -Wall -lglog -lmxbase -lmxpidatatype -lplugintoolkit -lstreammanager -lcpprest -lmindxsdk_protobuf -lopencv_world -o main# run ./mainmain.cpp添加利用opencv显示坐标框和性能打点代码。
第20行插入如下代码
#include "opencv4/opencv2/opencv.hpp"第116行插入如下代码
auto startTime = std::chrono::high_resolution_clock::now();第125-128行插入如下代码
auto endTime = std::chrono::high_resolution_clock::now(); double costMs = std::chrono::duration<double, std::milli>(endTime - startTime).count(); LogInfo << "[SendData-GetResult] cost: " << costMs << "ms. "; LogInfo << "[SendData-GetResult] fps: " << 1000/costMs << "fps";第137-173行插入如下代码
web::json::value jsonText = web::json::value::parse(result); if (jsonText.is_object()) { web::json::object textObject = jsonText.as_object(); auto itInferObject = textObject.find("MxpiObject"); if (itInferObject == textObject.end() || (!itInferObject->second.is_array())) { return 0; } auto iter = itInferObject->second.as_array().begin(); cv::Mat src = cv::imread("./test.jpg"); for (; iter != itInferObject->second.as_array().end(); iter++) { if (iter->is_object()) { auto modelInferObject = iter->as_object(); float x0 = 0; float x1 = 0; float y0 = 0; float y1 = 0; auto it = modelInferObject.find("x0"); if (it != modelInferObject.end()) { x0 = float(it->second.as_double()); } it = modelInferObject.find("x1"); if (it != modelInferObject.end()) { x1 = float(it->second.as_double()); } it = modelInferObject.find("y0"); if (it != modelInferObject.end()) { y0 = float(it->second.as_double()); } it = modelInferObject.find("y1"); if (it != modelInferObject.end()) { y1 = float(it->second.as_double()); } cv::Rect rect(x0, y0, x1 - x0, y1 - y0); cv::rectangle(src, rect, cv::Scalar(0, 255, 0),5, cv::LINE_8,0); } } cv::imwrite("./result.jpg", src); }
修改完后的代码如下：
1 /*2 * Copyright (c) 2020.Huawei Technologies Co., Ltd. All rights reserved.3 *4 * Licensed under the Apache License, Version 2.0 (the "License");5 * you may not use this file except in compliance with the License.6 * You may obtain a copy of the License at7 *8 * http://www.apache.org/licenses/LICENSE-2.09 *10 * Unless required by applicable law or agreed to in writing, software11 * distributed under the License is distributed on an "AS IS" BASIS,12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13 * See the License for the specific language governing permissions and14 * limitations under the License.15 */1617 #include <cstring>18 #include "MxBase/Log/Log.h"19 #include "MxStream/StreamManager/MxStreamManager.h"20 #include "opencv4/opencv2/opencv.hpp"21 namespace {22 APP_ERROR ReadFile(const std::string& filePath, MxStream::MxstDataInput& dataBuffer)23 {24 char c[PATH_MAX + 1] = { 0x00 };25 size_t count = filePath.copy(c, PATH_MAX + 1);26 if (count != filePath.length()) {27 LogError << "Failed to copy file path(" << c << ").";28 return APP_ERR_COMM_FAILURE;29 }30 // Get the absolute path of input file31 char path[PATH_MAX + 1] = { 0x00 };32 if ((strlen(c) > PATH_MAX) || (realpath(c, path) == nullptr)) {33 LogError << "Failed to get image, the image path is (" << filePath << ").";34 return APP_ERR_COMM_NO_EXIST;35 }36 // Open file with reading mode37 FILE *fp = fopen(path, "rb");38 if (fp == nullptr) {39 LogError << "Failed to open file (" << path << ").";40 return APP_ERR_COMM_OPEN_FAIL;41 }42 // Get the length of input file43 fseek(fp, 0, SEEK_END);44 long fileSize = ftell(fp);45 fseek(fp, 0, SEEK_SET);46 // If file not empty, read it into FileInfo and return it47 if (fileSize > 0) {48 dataBuffer.dataSize = fileSize;49 dataBuffer.dataPtr = new (std::nothrow) uint32_t[fileSize];50 if (dataBuffer.dataPtr == nullptr) {51 LogError << "allocate memory with \"new uint32_t\" failed.";52 return APP_ERR_COMM_FAILURE;53 }5455 uint32_t readRet = fread(dataBuffer.dataPtr, 1, fileSize, fp);56 if (readRet <= 0) {57 fclose(fp);58 return APP_ERR_COMM_READ_FAIL;59 }60 fclose(fp);61 return APP_ERR_OK;62 }63 fclose(fp);64 return APP_ERR_COMM_FAILURE;65 }6667 std::string ReadPipelineConfig(const std::string& pipelineConfigPath)68 {69 std::ifstream file(pipelineConfigPath.c_str(), std::ifstream::binary);70 if (!file) {71 LogError << pipelineConfigPath <<" file dose not exist.";72 return "";73 }74 file.seekg(0, std::ifstream::end);75 uint32_t fileSize = file.tellg();76 file.seekg(0);77 std::unique_ptr<char[]> data(new char[fileSize]);78 file.read(data.get(), fileSize);79 file.close();80 std::string pipelineConfig(data.get(), fileSize);81 return pipelineConfig;82 }83 }8485 int main(int argc, char* argv[])86 {87 // read image file and build stream input88 MxStream::MxstDataInput dataBuffer;89 APP_ERROR ret = ReadFile("./test.jpg", dataBuffer);90 if (ret != APP_ERR_OK) {91 LogError << GetError(ret) << "Failed to read image file.";92 return ret;93 }94 // read pipeline config file95 std::string pipelineConfigPath = "../pipeline/yolov5s_example.pipeline";96 std::string pipelineConfig = ReadPipelineConfig(pipelineConfigPath);97 if (pipelineConfig == "") {98 LogError << "Read pipeline failed.";99 return APP_ERR_COMM_INIT_FAIL;
100 }
101 // init stream manager
102 MxStream::MxStreamManager mxStreamManager;
103 ret = mxStreamManager.InitManager();
104 if (ret != APP_ERR_OK) {
105 LogError << GetError(ret) << "Failed to init Stream manager.";
106 return ret;
107 }
108 // create stream by pipeline config file
109 ret = mxStreamManager.CreateMultipleStreams(pipelineConfig);
110 if (ret != APP_ERR_OK) {
111 LogError << GetError(ret) << "Failed to create Stream.";
112 return ret;
113 }
114 std::string streamName = "classification+detection";
115 int inPluginId = 0;
116 // send data into stream
117 auto startTime = std::chrono::high_resolution_clock::now();
118 ret = mxStreamManager.SendData(streamName, inPluginId, dataBuffer);
119 if (ret != APP_ERR_OK) {
120 LogError << GetError(ret) << "Failed to send data to stream.";
121 return ret;
122 }
123 // get stream output
124 MxStream::MxstDataOutput* output = mxStreamManager.GetResult(streamName, inPluginId);
125 auto endTime = std::chrono::high_resolution_clock::now();
126 double costMs = std::chrono::duration<double, std::milli>(endTime - startTime).count();
127 LogInfo << "[SendData-GetResult] cost: " << costMs << "ms. ";
128 LogInfo << "[SendData-GetResult] fps: " << 1000/costMs << "fps";
129 if (output == nullptr) {
130 LogError << "Failed to get pipeline output.";
131 return ret;
132 }
133
134 std::string result = std::string((char *)output->dataPtr, output->dataSize);
135 LogInfo << "Results:" << result;
136
137 web::json::value jsonText = web::json::value::parse(result);
138 if (jsonText.is_object()) {
139 web::json::object textObject = jsonText.as_object();
140 auto itInferObject = textObject.find("MxpiObject");
141 if (itInferObject == textObject.end() || (!itInferObject->second.is_array())) {
142 return 0;
143 }
144 auto iter = itInferObject->second.as_array().begin();
145 cv::Mat src = cv::imread("./test.jpg");
146 for (; iter != itInferObject->second.as_array().end(); iter++) {
147 if (iter->is_object()) {
148 auto modelInferObject = iter->as_object();
149 float x0 = 0;
150 float x1 = 0;
151 float y0 = 0;
152 float y1 = 0;
153 auto it = modelInferObject.find("x0");
154 if (it != modelInferObject.end()) {
155 x0 = float(it->second.as_double());
156 }
157 it = modelInferObject.find("x1");
158 if (it != modelInferObject.end()) {
159 x1 = float(it->second.as_double());
160 }
161 it = modelInferObject.find("y0");
162 if (it != modelInferObject.end()) {
163 y0 = float(it->second.as_double());
164 }
165 it = modelInferObject.find("y1");
166 if (it != modelInferObject.end()) {
167 y1 = float(it->second.as_double());
168 }
169 cv::Rect rect(x0, y0, x1 - x0, y1 - y0);
170 cv::rectangle(src, rect, cv::Scalar(0, 255, 0),5, cv::LINE_8,0);
171 }
172 }
173 cv::imwrite("./result.jpg", src);
174 }
175 // destroy streams
176 mxStreamManager.DestroyAllStreams();
177 delete dataBuffer.dataPtr;
178 dataBuffer.dataPtr = nullptr;
179
180 delete output;
181 return 0;
182 }再回到yolov5s_C++目录，运行bash run.sh，日志显示, 处理时间为：49.491ms. 平均帧率为：20.2057fps， test.jpg的分辨率为：1024x683
$ bash run.sh
省略部分输出...
I0518 05:48:41.255576 8093 MxGstBase.cpp:632] element(mxpi_imageresize0) dynamicImageSize=640,640;
I0518 05:48:41.255784 8093 MxGstBase.cpp:764] element(mxpi_imagedecoder0) gst_change_state NULL_TO_READY.
I0518 05:48:41.256076 8093 MxStreamManager.cpp:384] Creates stream(classification+detection) successfully.
I0518 05:48:41.256207 8093 MxStreamManager.cpp:434] Creates streams successfully.
I0518 05:48:41.305809 8093 main.cpp:131] [SendData-GetResult] cost: 49.491ms.
I0518 05:48:41.305980 8093 main.cpp:132] [SendData-GetResult] fps: 20.2057fps
I0518 05:48:41.306020 8093 main.cpp:135] Results:{"MxpiObject":[{"classVec":[{"classId":16,"className":"dog","confidence":0.67103165399999998,"headerVec":[]}],"x0":71.558685299999993,"x1":943.03247099999999,"y0":132.90640300000001,"y1":604.41442900000004}]}
I0518 05:48:41.377825 8093 MxsmStream.cpp:688] Begin to destroy stream(classification+detection).
I0518 05:48:41.378067 8093 MxsmStream.cpp:743] Send custom eos to the Stream successfully.
I0518 05:48:41.378247 8093 MxsmStream.cpp:749] Send eos to the Stream successfully.
I0518 05:48:41.378386 8093 MxsmStream.cpp:755] Flushes the Stream data successfully.
I0518 05:48:41.378731 8093 MxGstBase.cpp:770] element(mxpi_dataserialize0) gst_change_state PLAYING_TO_PAUSED.
I0518 05:48:41.378816 8093 MxGstBase.cpp:770] element(mxpi_modelinfer0) gst_change_state PLAYING_TO_PAUSED.
I0518 05:48:41.378883 8093 MxGstBase.cpp:770] element(mxpi_imageresize0) gst_change_state PLAYING_TO_PAUSED.
I0518 05:48:41.378957 8093 MxGstBase.cpp:770] element(mxpi_imagedecoder0) gst_change_state PLAYING_TO_PAUSED.
I0518 05:48:41.379127 8093 MxGstBase.cpp:773] element(mxpi_dataserialize0) gst_change_state PAUSED_TO_READY.
省略部分输出...
和当前目录生成的result.jpg图片如下。
5. 纯推理测试
4.2中的推理性能是端到端的性能，其中包括了图像预处理和推理结果的后处理时间，现在我们测试模型的纯推理时间。
5.1 下载benchmark工具
从网盘下载benchmark工具，提取码：5uzu，并解压缩到/home/HwHiAiUser目录
$ cd /home/HwHiAiUser
$ unzip Ascend-cann-benchmark_5.0.1-Linux-aarch64.zip
5.2 性能测试
$ cd /home/HwHiAiUser/mxManufacture/samples/mxManufacture/models/yolov5s
$ sudo /home/HwHiAiUser/benchmark_tools/benchmark.aarch64 -batch_size=1 -om_path=./yolov5s_sim_t.om -round=50 -device_id=0
[INFO][Inference] PureInfer Init SUCCESS
[INFO] Dataset number: 0 finished cost 24.818ms
[INFO] Dataset number: 1 finished cost 9.682ms
[INFO] Dataset number: 2 finished cost 9.714ms
[INFO] Dataset number: 3 finished cost 9.707ms
[INFO] Dataset number: 4 finished cost 9.658ms
[INFO] Dataset number: 5 finished cost 9.704ms
[INFO] Dataset number: 6 finished cost 9.686ms
[INFO] Dataset number: 7 finished cost 9.612ms
[INFO] Dataset number: 8 finished cost 9.588ms
[INFO] Dataset number: 9 finished cost 9.631ms
[INFO] Dataset number: 10 finished cost 9.614ms
[INFO] Dataset number: 11 finished cost 9.721ms
[INFO] Dataset number: 12 finished cost 9.677ms
[INFO] Dataset number: 13 finished cost 9.717ms
[INFO] Dataset number: 14 finished cost 9.671ms
[INFO] Dataset number: 15 finished cost 9.652ms
[INFO] Dataset number: 16 finished cost 9.672ms
[INFO] Dataset number: 17 finished cost 9.594ms
[INFO] Dataset number: 18 finished cost 9.622ms
[INFO] Dataset number: 19 finished cost 9.624ms
[INFO] Dataset number: 20 finished cost 9.671ms
[INFO] Dataset number: 21 finished cost 9.621ms
[INFO] Dataset number: 22 finished cost 9.682ms
[INFO] Dataset number: 23 finished cost 9.75ms
[INFO] Dataset number: 24 finished cost 9.736ms
[INFO] Dataset number: 25 finished cost 9.645ms
[INFO] Dataset number: 26 finished cost 9.657ms
[INFO] Dataset number: 27 finished cost 9.59ms
[INFO] Dataset number: 28 finished cost 9.631ms
[INFO] Dataset number: 29 finished cost 9.61ms
[INFO] Dataset number: 30 finished cost 9.616ms
[INFO] Dataset number: 31 finished cost 9.651ms
[INFO] Dataset number: 32 finished cost 9.671ms
[INFO] Dataset number: 33 finished cost 9.661ms
[INFO] Dataset number: 34 finished cost 9.71399ms
[INFO] Dataset number: 35 finished cost 9.691ms
[INFO] Dataset number: 36 finished cost 9.69ms
[INFO] Dataset number: 37 finished cost 9.666ms
[INFO] Dataset number: 38 finished cost 9.62ms
[INFO] Dataset number: 39 finished cost 9.607ms
[INFO] Dataset number: 40 finished cost 9.591ms
[INFO] Dataset number: 41 finished cost 9.63ms
[INFO] Dataset number: 42 finished cost 9.633ms
[INFO] Dataset number: 43 finished cost 9.66ms
[INFO] Dataset number: 44 finished cost 9.71ms
[INFO] Dataset number: 45 finished cost 9.706ms
[INFO] Dataset number: 46 finished cost 9.688ms
[INFO] Dataset number: 47 finished cost 9.665ms
[INFO] Dataset number: 48 finished cost 9.616ms
[INFO] Dataset number: 49 finished cost 9.617ms
[INFO] PureInfer result saved in ./result/PureInfer_perf_of_yolov5s_sim_t_in_device_0.txt
-----------------PureInfer Performance Summary------------------
[INFO] ave_throughputRate: 103.547samples/s, ave_latency: 9.9612ms
----------------------------------------------------------------6. 基于SDK的推理代码开发
6.1 下载样例代码
从https://gitee.com/shiner-chen/APISamples获取最新的样例代码。
$ git clone git@gitee.com:shiner-chen/APISamples.git
6.2 编译运行
$ cd APISamples/src/Samples/InferObjectDetection
$ bash build.sh YOLOV5
$ cd dist
$ ./main -i test.jpg -t 2
[Info ][2021-05-20 11:39:19:411328][ResourceManager.cpp InitResource:75] Initialized acl successfully.
[Info ][2021-05-20 11:39:19:505191][ResourceManager.cpp InitResource:84] Open device 0 successfully.
[Info ][2021-05-20 11:39:19:506968][ResourceManager.cpp InitResource:91] Created context for device 0 successfully
[Info ][2021-05-20 11:39:19:507094][ResourceManager.cpp InitResource:102] Init resource successfully.
[Info ][2021-05-20 11:39:19:507272][AclProcess.cpp InitResource:161] Create context successfully
[Info ][2021-05-20 11:39:19:507957][AclProcess.cpp InitResource:167] Create stream successfully
[Info ][2021-05-20 11:39:19:523990][AclProcess.cpp InitModule:121] Initialize dvppCommon_ successfully
[Info ][2021-05-20 11:39:19:524167][ModelProcess.cpp Init:240] ModelProcess:Begin to init instance.
[Info ][2021-05-20 11:39:19:839716][AclProcess.cpp InitModule:139] Initialize ModelProcess_ successfully
[Info ][2021-05-20 11:39:19:856578][AclProcess.cpp InitModule:145] Loaded label successfully.
[Info ][2021-05-20 11:39:19:884394][AclProcess.cpp YoloPostProcess:503] The number of output buffers of yolov3 model is 3
[Info ][2021-05-20 11:39:19:903552][AclProcess.cpp YoloPostProcess:528] #Obj0, box(33.537, 122.282, 955.744, 612.067) confidence: 0.777343 label: dog
[Info ][2021-05-20 11:39:19:903891][AclProcess.cpp YoloPostProcess:535] [Process Delay] cost: 47.1801ms fps: 21.1954
[Info ][2021-05-20 11:39:19:942970][AclProcess.cpp Release:71] Destroy stream successfully
[Info ][2021-05-20 11:39:19:943135][ModelProcess.cpp DeInit:150] Model[yolov3][0] deinit begin
[Info ][2021-05-20 11:39:19:965709][ModelProcess.cpp DeInit:189] Model[yolov3][0] deinit success
[Info ][2021-05-20 11:39:19:973285][ResourceManager.cpp Release:44] Finalized acl successfully.$ ls ./result/result_test.jpg
result_test.jpg生成的结果图片：
DLAP221上Yolov5s性能测试相关推荐
- 使用OpenVINO优化和部署DenseNet模型并在DevCloud上完成性能测试
概述 Intel® DevCloud for the Edge支持在英特尔的硬件平台上主动构建原型并试验面向计算机视觉的AI 工作负载.其硬件平台托管在云环境中,专门用于深度学习,用户可以全面访问这些 ...
- 在DLAP221上部署容器进行推理操作
1. Install toolbox 根据设备驱动版本,安装对应版本的Ascend-cann-toolbox_x.x.x_linux-aarch64.run,我的驱动版本是21.0.1,在此我安装的t ...
- 3测试图片显示置信度_云上的移动性能测试平台
1. 功能决定现在,性能决定未来 性能测试在移动测试领域一直是一个大难题,它最直观的表现是用户在前台使用 App 时的主观体验,然而决定体验优劣的背后,涉及到了许许多多的技术变迁. 当我们习惯于诺基亚 ...
- ldap客户端工具_什么是性能测试？性能测试主流工具有哪些？
性能测试的概念 一般来说,性能是一种指标,表明软件系统或构件对其及时性要求的符合程度;其次,性能是软件产品的一种特性,可以用时间来进行度量.性能的及时性用响应时间或吞吐量来衡量.响应时间是指服务器对请 ...
- 七月新增开源项目：你学习的速度跟得上前端造轮子的速度吗？
2019独角兽企业重金招聘Python工程师标准>>> 每月新增开源项目.顾名思义,每月更新一期.我们会从社区上个月新收录的开源项目中,挑选出有价值的.有用的.优秀的.或者好玩的开源 ...
- python递归合并排序_python 归并排序的递归法与迭代法(利用队列)实现，以及性能测试...
递归排序核心 递归排序的核心是 分与合 分的最终结果 就是将原数组中每一个数字分作一个数组, 合就是 所有小数组不断排序,合并的过程. 合并的过程是先将两个含有一个数字的数组排序,合并(每次比较两个数 ...
- “纹身贴皮电路“：未来在皮肤上画个电路就能监测身体健康状况
导读 从人体皮肤中准确提取生理和物理信号对于健康监测.疾病预防和治疗至关重要.可穿戴生物电子学的最新进展直接嵌入到表皮表面是未来表皮传感的一个很有前途的解决方案.然而,现有的可穿戴生物电子学容易受到运 ...
- 聊聊性能测试、操作系统优化对性能测试的影响，以及如何优化操作系统
序 男生追女生方法名堂层出不穷,而女生追男生好像招数不多.写情书寻常女子肯定低不下这个头,于是找男生"修电脑"成了跨域祖国大江南北,甚至我们台湾人民的不二选择.很可惜,我读大学的时 ...
- 在做性能测试之前需要知道什么
以下是我自己录制的关于这篇文章的一小段视频,有兴趣的可以下载看看 https://yunpan.cn/cPQc4mm2DjbMu 访问密码 a76f //此篇摘抄于虫师博客,个人觉得通俗易懂 关于理 ...
最新文章
- c语言指针实现字符串倒序输出,菜鸟求助-如何用指针法将一串字符按单词的倒序输出？如：i love yo...
- 图像配准----NCC
- 内核-syn-ack RTO修改
- 多分类下的ROC曲线和AUC
- 什么是倾斜45度的火山图？
- 数据科学 IPython 笔记本 8.11 多个子图
- Helm 3 完整教程（二十）：在 Helm 模板中定义和使用变量
- 十分钟用Django创建一个简单的职位管理系统
- java liveness_并发 - 活跃度（Liveness） - 《Java 编程要点（Essential Java）》 - 书栈网 · BookStack...
- 卷积神经网络(CNN)卷积核与滤波器理解
- 超级经典的Word技巧
- ArcGIS Server 发布服务失败
- 降噪蓝牙耳机对比测评：南卡和OPPO哪款降噪耳机好？
- 解决thinkpad或者其他笔记本电脑无线网络不可用问题
- 我的汉语名字叫东坤，可以帮我起一个读音相近的英语名字吗
- HashMap和TreeMap
- RNC/SGSN/PDP/IUPS概念
- Java阿凯_彩虹城堡之七彩宝石篇最终版攻略
- 【BUG解决】 RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor)
- learnopengl 中 pbr的球体算法