【R329开发板评测】实机测试MobileNetFaceNet人脸对比

经过几天周易AIPU仿真模拟，踩了不少坑（具体可以参考: 【周易AIPU 仿真】mobilefacenet人脸特征模型），终于可以拿到了R329来进行实机验证了。

系统镜像烧录

我烧录的系统镜像是r329-mainline-debian-20210802.img，具体可以参考https://aijishu.com/a/1060000000224083，友情提醒一个好的读卡器和SD卡是很有必要的，有些小伙伴反馈烧录镜像后老是启动异常，可以先换个读卡器或者SD卡试试。

实机测试

1. 模型生成

有了之前的模拟仿真经验，可以很快入手转换模型。

第一步，启动镜像，进入事先准备好的模型转换容器

docker run --gpus all -it --rm -e DISPLAY=$DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix  -v ${PWD}:${PWD}  -w ${PWD}  --name zhouyi  zepan/zhouyi

第二步，修改模型转换配置文件，与模拟仿真唯一不同的就是GBuilder字段下只需要配置outputs就行了

[Common]
mode = build

[Parser]
model_name = facenet
detection_postprocess = 
model_domain = image_classification
input_model = ./model/FaceMobileNet192_train_false.pb
input = img_inputs
input_shape = [1, 112, 112, 3]
output = MobileFaceNet/Logits/SpatialSqueeze
output_dir = ./

[AutoQuantizationTool]
quantize_method = SYMMETRIC
quant_precision = int8
ops_per_channel = DepthwiseConv
reverse_rgb = False
log = True
calibration_data = ./dataset/dataset.npy
calibration_label = ./dataset/label.npy
preprocess_mode = normalize

[GBuilder]
outputs=aipu.bin
target=Z1_0701

第三步，运行模型转换脚本

aipubuild facenet_run_pan.cfg

模型转换成功后，会有如下输出信息,重点要记录下输出层量化系数output_scale:[6.513236]，后面在做反量化时会用到

[I]     step1: get max/min statistic value DONE
[I]     step2: quantization each op DONE
[I]     step3: build quantization forward DONE
[I]     step4: show output scale of end node:
[I]             layer_id: 65, layer_top:MobileFaceNet/Logits/SpatialSqueeze_0, output_scale:[6.513236]
[I] ==== auto-quantization DONE =
[I] Quantize model complete
[I] Building ...
[I] [common_options.h: 276] BuildTool version: 4.0.175. Build for target Z1_0701 at frequency 800MHz
[I] [IRChecker] Start to check IR: /tmp/AIPUBuilder_1629203633.8358698/facenet_int8.txt
[I] [IRChecker] model_name: facenet
[I] [IRChecker] IRChecker: All IR pass
[I] [graph.cpp : 846] loading graph weight: /tmp/AIPUBuilder_1629203633.8358698/facenet_int8.bin size: 0x147200
[I] [builder.cpp:1059] Total memory for this graph: 0x7880b0 Bytes
[I] [builder.cpp:1060] Text   section:  0x00002370 Bytes
[I] [builder.cpp:1061] RO     section:  0x00003600 Bytes
[I] [builder.cpp:1062] Desc   section:  0x00016000 Bytes
[I] [builder.cpp:1063] Data   section:  0x00213900 Bytes
[I] [builder.cpp:1064] BSS    section:  0x00518a40 Bytes
[I] [builder.cpp:1065] Stack         :  0x00040400 Bytes
[I] [builder.cpp:1066] Workspace(BSS):  0x00000000 Bytes
Total errors: 0,  warnings: 0

至此你就得到了可实机运行的模型文件aipu.bin。

2. 更新应用程序zhouyi_cam

需要追加如下功能：

模型输出数值进行反量化操作，此时需要用到刚才模型转换时的量化系数output_scale

        std::vector<float> data;
        for(int i=0; i < dim; i++) {
            data.push_back((float)result[i] / output_scale);
        }

预先录入底库人脸，可以先针对底库照片打印出人脸特征数据后，硬编码到程序中
人脸特征归一化，由于采用余弦距离作为评价相似度的手段，需要对模型输出的人脸特征进行归一化操作

        float norm2 = 0;
        for (int i = 0; i < dim; i++) {
            norm2 += data[i] * data[i];
        }

        float norm = sqrt(norm2);
        for (int i = 0; i < dim; i++) {
            data[i] = data[i] / norm;
        }

编译输出