TensorRT-5.1.5.0-YOLOv3

最新推荐文章于 2024-07-19 11:20:37 发布

知识在于分享

最新推荐文章于 2024-07-19 11:20:37 发布

阅读量1k

点赞数 1

分类专栏：深度学习

本文链接：https://blog.csdn.net/baidu_40840693/article/details/95889738

版权

深度学习专栏收录该内容

255 篇文章 18 订阅

订阅专栏

接上一个博客

TensorRT-5.1.5.0-SSD

https://blog.csdn.net/baidu_40840693/article/details/95642055

首先：

upsample层密码bwrd

https://pan.baidu.com/s/13GpoYoqKSCeFX0m0ves_fQ

然后：

https://github.com/lewes6369/TensorRT-Yolov3

https://github.com/lewes6369/tensorRTWrapper/

Models

Download the caffe model converted by official model:

Baidu Cloud here pwd: gbue
Google Drive here

里面有两种，input416*416和608*608

当然也可以在这里下载：

https://pan.baidu.com/s/1yiCrnmsOm0hbweJBiiUScQ

关于caffe版本-yolov3的运行，都可以在这里找到：

https://github.com/ChenYingpeng/caffe-yolov3

关于yolov3中的Upsample，记得添加对参数的注释：

layer {
    bottom: "layer85-conv"
    top: "layer86-upsample"
    name: "layer86-upsample"
    type: "Upsample"
    #upsample_param {
    #    scale: 2
    #}
}

TensorRT caffe parser can't check the param not in it's default proto file, although I added it as plugin.

TensorRT中的caffe解析器的没能识别，具体参数怎么用，我们可以通过TensorRT去使用

我们先测试yolov3在caffe中的前向速度：

配置好cpp和cu和hpp，还有caffe.proto后

  optional UpsampleParameter upsample_param = 149;
}

// added by chen 
message UpsampleParameter{
  optional int32 scale = 1 [default = 1];
}

我们，重新编译caffe，使用416×416的prototxt进行测试

测试机器-GTX TiTAN X

416×416分辨率：

./build/tools/caffe time -model=/home/boyun/code/caffe-ssd/yolov3/yolov3_416.prototxt --weights=/home/boyun/code/caffe-ssd/yolov3/yolov3_416.caffemodel --iterations=200 -gpu 0

I0714 22:10:42.351696 10668 caffe.cpp:412] Average Forward pass: 50.3751 ms.
I0714 22:10:42.351701 10668 caffe.cpp:414] Average Backward pass: 77.3334 ms.
I0714 22:10:42.351707 10668 caffe.cpp:416] Average Forward-Backward: 127.876 ms.
I0714 22:10:42.351712 10668 caffe.cpp:418] Total Time: 25575.2 ms.
I0714 22:10:42.351716 10668 caffe.cpp:419] *** Benchmark ends ***

顺便说一下在1070ti下进行416×416分辨率测试：

I0715 10:27:02.526588 20742 caffe.cpp:407] layer106-conv	backward: 0.205097 ms.
I0715 10:27:02.526610 20742 caffe.cpp:412] Average Forward pass: 35.5234 ms.
I0715 10:27:02.526621 20742 caffe.cpp:414] Average Backward pass: 43.3386 ms.
I0715 10:27:02.526648 20742 caffe.cpp:416] Average Forward-Backward: 78.9849 ms.
I0715 10:27:02.526661 20742 caffe.cpp:418] Total Time: 15797 ms.
I0715 10:27:02.526671 20742 caffe.cpp:419] *** Benchmark ends ***

yolov3一般采用32的倍数，因为多尺度的最大尺度是初始input的1/32，假如416×416,那么最后的feature map为13×13

我们把input改成320*320，重新测速：

320×320分辨率：

I0714 22:19:04.315593 10894 caffe.cpp:407] layer106-conv	backward: 0.18554 ms.
I0714 22:19:04.315603 10894 caffe.cpp:412] Average Forward pass: 42.3266 ms.
I0714 22:19:04.323930 10894 caffe.cpp:414] Average Backward pass: 62.2112 ms.
I0714 22:19:04.323948 10894 caffe.cpp:416] Average Forward-Backward: 104.755 ms.
I0714 22:19:04.323954 10894 caffe.cpp:418] Total Time: 20951 ms.
I0714 22:19:04.323959 10894 caffe.cpp:419] *** Benchmark ends **

顺便说一下在1070ti下进行320×320分辨率测试：

I0715 10:23:30.041016 20648 caffe.cpp:407] layer106-conv	backward: 0.14551 ms.
I0715 10:23:30.041043 20648 caffe.cpp:412] Average Forward pass: 25.4868 ms.
I0715 10:23:30.041054 20648 caffe.cpp:414] Average Backward pass: 49.6358 ms.
I0715 10:23:30.041083 20648 caffe.cpp:416] Average Forward-Backward: 75.2501 ms.
I0715 10:23:30.041097 20648 caffe.cpp:418] Total Time: 15050 ms.
I0715 10:23:30.041108 20648 caffe.cpp:419] *** Benchmark ends ***

608×608分辨率：

I0714 22:33:18.229004 11462 caffe.cpp:407] layer106-conv	backward: 0.293207 ms.
I0714 22:33:18.229012 11462 caffe.cpp:412] Average Forward pass: 75.0531 ms.
I0714 22:33:18.229017 11462 caffe.cpp:414] Average Backward pass: 109.845 ms.
I0714 22:33:18.229023 11462 caffe.cpp:416] Average Forward-Backward: 185.006 ms.
I0714 22:33:18.229028 11462 caffe.cpp:418] Total Time: 37001.2 ms.
I0714 22:33:18.229033 11462 caffe.cpp:419] *** Benchmark ends ***

顺便说一下在1070ti下进行608×608分辨率测试：

I0715 10:20:45.098073 20529 caffe.cpp:407] layer106-conv	backward: 0.277381 ms.
I0715 10:20:45.098104 20529 caffe.cpp:412] Average Forward pass: 59.2867 ms.
I0715 10:20:45.098117 20529 caffe.cpp:414] Average Backward pass: 76.9649 ms.
I0715 10:20:45.098142 20529 caffe.cpp:416] Average Forward-Backward: 136.397 ms.
I0715 10:20:45.098160 20529 caffe.cpp:418] Total Time: 27279.3 ms.
I0715 10:20:45.098173 20529 caffe.cpp:419] *** Benchmark ends ***

接着我们测试yolov3-tiny

416×416分辨率：

./build/tools/caffe time -model=/home/boyun/code/caffe-ssd/yolov3-tiny/yolov3-tiny.prototxt --weights=/home/boyun/code/caffe-ssd/yolov3-tiny/yolov3-tiny.caffemodel --iterations=200 -gpu 0

I0714 22:27:39.310173 11273 caffe.cpp:407] layer23-conv	backward: 0.129828 ms.
I0714 22:27:39.310195 11273 caffe.cpp:412] Average Forward pass: 9.04005 ms.
I0714 22:27:39.310200 11273 caffe.cpp:414] Average Backward pass: 14.5964 ms.
I0714 22:27:39.310243 11273 caffe.cpp:416] Average Forward-Backward: 23.8689 ms.
I0714 22:27:39.310250 11273 caffe.cpp:418] Total Time: 4773.78 ms.
I0714 22:27:39.310253 11273 caffe.cpp:419] *** Benchmark ends ***

顺便说一下在1070ti下进行416×416分辨率测试：

I0715 10:29:10.839944 20802 caffe.cpp:407] layer23-conv	backward: 0.099241 ms.
I0715 10:29:10.839954 20802 caffe.cpp:412] Average Forward pass: 5.15907 ms.
I0715 10:29:10.839959 20802 caffe.cpp:414] Average Backward pass: 8.64283 ms.
I0715 10:29:10.840006 20802 caffe.cpp:416] Average Forward-Backward: 13.842 ms.
I0715 10:29:10.840013 20802 caffe.cpp:418] Total Time: 2768.4 ms.
I0715 10:29:10.840018 20802 caffe.cpp:419] *** Benchmark ends ***

320×320分辨率：

I0714 22:29:15.098641 11341 caffe.cpp:412] Average Forward pass: 8.37247 ms.
I0714 22:29:15.098646 11341 caffe.cpp:414] Average Backward pass: 11.5893 ms.
I0714 22:29:15.098657 11341 caffe.cpp:416] Average Forward-Backward: 20.1775 ms.
I0714 22:29:15.098662 11341 caffe.cpp:418] Total Time: 4035.49 ms.
I0714 22:29:15.098667 11341 caffe.cpp:419] *** Benchmark ends ***

顺便说一下在1070ti下进行320×320分辨率测试：

I0715 10:30:05.019290 20863 caffe.cpp:407] layer23-conv	backward: 0.0819661 ms.
I0715 10:30:05.019299 20863 caffe.cpp:412] Average Forward pass: 4.28037 ms.
I0715 10:30:05.019304 20863 caffe.cpp:414] Average Backward pass: 7.55383 ms.
I0715 10:30:05.019309 20863 caffe.cpp:416] Average Forward-Backward: 11.8756 ms.
I0715 10:30:05.019315 20863 caffe.cpp:418] Total Time: 2375.13 ms.
I0715 10:30:05.019322 20863 caffe.cpp:419] *** Benchmark ends ***

608×608分辨率：

I0714 22:30:52.382453 11410 caffe.cpp:407] layer23-conv	backward: 0.177674 ms.
I0714 22:30:52.382462 11410 caffe.cpp:412] Average Forward pass: 12.2339 ms.
I0714 22:30:52.382467 11410 caffe.cpp:414] Average Backward pass: 22.8627 ms.
I0714 22:30:52.382506 11410 caffe.cpp:416] Average Forward-Backward: 35.5057 ms.
I0714 22:30:52.382513 11410 caffe.cpp:418] Total Time: 7101.14 ms.
I0714 22:30:52.382515 11410 caffe.cpp:419] *** Benchmark ends ***

顺便说一下在1070ti下进行608×608分辨率测试：

I0715 10:30:49.398526 20891 caffe.cpp:404] layer23-conv	forward: 0.0902006 ms.
I0715 10:30:49.398531 20891 caffe.cpp:407] layer23-conv	backward: 0.144245 ms.
I0715 10:30:49.398541 20891 caffe.cpp:412] Average Forward pass: 8.15075 ms.
I0715 10:30:49.398546 20891 caffe.cpp:414] Average Backward pass: 15.6652 ms.
I0715 10:30:49.398555 20891 caffe.cpp:416] Average Forward-Backward: 23.8652 ms.
I0715 10:30:49.398561 20891 caffe.cpp:418] Total Time: 4773.03 ms.
I0715 10:30:49.398566 20891 caffe.cpp:419] *** Benchmark ends ***

接着我们测试yolov3-tiny所有channels减半的网络：

608×608分辨率：

./build/tools/caffe time -model=/home/boyun/code/caffe-ssd/yolov3-tiny/yolov3-tiny-half.prototxt --weights=/home/boyun/code/caffe-ssd/yolov3-tiny/yolov3-tiny-half.caffemodel --iterations=200 -gpu 0

I0714 23:15:48.347908 14637 caffe.cpp:407] layer23-conv	backward: 0.0562594 ms.
I0714 23:15:48.347916 14637 caffe.cpp:412] Average Forward pass: 8.46485 ms.
I0714 23:15:48.347920 14637 caffe.cpp:414] Average Backward pass: 15.7116 ms.
I0714 23:15:48.347964 14637 caffe.cpp:416] Average Forward-Backward: 24.4433 ms.
I0714 23:15:48.347970 14637 caffe.cpp:418] Total Time: 4888.66 ms.
I0714 23:15:48.347978 14637 caffe.cpp:419] *** Benchmark ends ***

顺便说一下在1070ti下进行608×608分辨率测试：

I0715 10:32:16.474354 20914 caffe.cpp:404] layer23-conv	forward: 0.0535805 ms.
I0715 10:32:16.474359 20914 caffe.cpp:407] layer23-conv	backward: 0.0452762 ms.
I0715 10:32:16.474370 20914 caffe.cpp:412] Average Forward pass: 4.50442 ms.
I0715 10:32:16.474375 20914 caffe.cpp:414] Average Backward pass: 11.1228 ms.
I0715 10:32:16.474383 20914 caffe.cpp:416] Average Forward-Backward: 15.6752 ms.
I0715 10:32:16.474391 20914 caffe.cpp:418] Total Time: 3135.05 ms.
I0715 10:32:16.474402 20914 caffe.cpp:419] *** Benchmark ends ***

还在这种减半网络下测试了backbone的前向，就是conv1-conv16的速度：

I0714 23:24:19.573200 14933 caffe.cpp:407] layer16-conv	backward: 0.054884 ms.
I0714 23:24:19.573209 14933 caffe.cpp:412] Average Forward pass: 6.96861 ms.
I0714 23:24:19.573212 14933 caffe.cpp:414] Average Backward pass: 14.954 ms.
I0714 23:24:19.573729 14933 caffe.cpp:416] Average Forward-Backward: 22.1895 ms.
I0714 23:24:19.573735 14933 caffe.cpp:418] Total Time: 4437.9 ms.
I0714 23:24:19.573740 14933 caffe.cpp:419] *** Benchmark ends ***

顺便说一下使用darknet进行测试，网络前向一次运行时间为：5-7ms

因为caffe和darknet的卷积和池化后feature map大小的机制有一些不同，所以做了一些修改

name: "Darkent2Caffe"
input: "data"
input_dim: 1
input_dim: 3
input_dim: 608
input_dim: 608

layer {
    bottom: "data"
    top: "layer1-conv"
    name: "layer1-conv"
    type: "Convolution"
    convolution_param {
        num_output: 8
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer1-conv"
    top: "layer1-conv"
    name: "layer1-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer1-conv"
    top: "layer1-conv"
    name: "layer1-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer1-conv"
    top: "layer1-conv"
    name: "layer1-act"
    type: "ReLU"
}
layer {
    bottom: "layer1-conv"
    top: "layer2-maxpool"
    name: "layer2-maxpool"
    type: "Pooling"
    pooling_param {
        kernel_size: 2
        stride: 2
        pool: MAX
    }
}
layer {
    bottom: "layer2-maxpool"
    top: "layer3-conv"
    name: "layer3-conv"
    type: "Convolution"
    convolution_param {
        num_output: 16
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer3-conv"
    top: "layer3-conv"
    name: "layer3-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer3-conv"
    top: "layer3-conv"
    name: "layer3-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer3-conv"
    top: "layer3-conv"
    name: "layer3-act"
    type: "ReLU"
}
layer {
    bottom: "layer3-conv"
    top: "layer4-maxpool"
    name: "layer4-maxpool"
    type: "Pooling"
    pooling_param {
        kernel_size: 2
        stride: 2
        pool: MAX
    }
}
layer {
    bottom: "layer4-maxpool"
    top: "layer5-conv"
    name: "layer5-conv"
    type: "Convolution"
    convolution_param {
        num_output: 32
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer5-conv"
    top: "layer5-conv"
    name: "layer5-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer5-conv"
    top: "layer5-conv"
    name: "layer5-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer5-conv"
    top: "layer5-conv"
    name: "layer5-act"
    type: "ReLU"
}
layer {
    bottom: "layer5-conv"
    top: "layer6-maxpool"
    name: "layer6-maxpool"
    type: "Pooling"
    pooling_param {
        kernel_size: 2
        stride: 2
        pool: MAX
    }
}
layer {
    bottom: "layer6-maxpool"
    top: "layer7-conv"
    name: "layer7-conv"
    type: "Convolution"
    convolution_param {
        num_output: 64
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer7-conv"
    top: "layer7-conv"
    name: "layer7-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer7-conv"
    top: "layer7-conv"
    name: "layer7-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer7-conv"
    top: "layer7-conv"
    name: "layer7-act"
    type: "ReLU"
}
layer {
    bottom: "layer7-conv"
    top: "layer8-maxpool"
    name: "layer8-maxpool"
    type: "Pooling"
    pooling_param {
        kernel_size: 2
        stride: 2
        pool: MAX
    }
}
layer {
    bottom: "layer8-maxpool"
    top: "layer9-conv"
    name: "layer9-conv"
    type: "Convolution"
    convolution_param {
        num_output: 128
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer9-conv"
    top: "layer9-conv"
    name: "layer9-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer9-conv"
    top: "layer9-conv"
    name: "layer9-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer9-conv"
    top: "layer9-conv"
    name: "layer9-act"
    type: "ReLU"
}
layer {
    bottom: "layer9-conv"
    top: "layer10-maxpool"
    name: "layer10-maxpool"
    type: "Pooling"
    pooling_param {
        kernel_size: 2
        stride: 2
        pool: MAX
    }
}
layer {
    bottom: "layer10-maxpool"
    top: "layer11-conv"
    name: "layer11-conv"
    type: "Convolution"
    convolution_param {
        num_output: 256
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer11-conv"
    top: "layer11-conv"
    name: "layer11-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer11-conv"
    top: "layer11-conv"
    name: "layer11-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer11-conv"
    top: "layer11-conv"
    name: "layer11-act"
    type: "ReLU"
}
layer {
    bottom: "layer11-conv"
    top: "layer12-maxpool"
    name: "layer12-maxpool"
    type: "Pooling"
    pooling_param {
        kernel_size: 3
        stride: 1
        pad: 1
        pool: MAX
    }
}
layer {
    bottom: "layer12-maxpool"
    top: "layer13-conv"
    name: "layer13-conv"
    type: "Convolution"
    convolution_param {
        num_output: 512
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer13-conv"
    top: "layer13-conv"
    name: "layer13-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer13-conv"
    top: "layer13-conv"
    name: "layer13-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer13-conv"
    top: "layer13-conv"
    name: "layer13-act"
    type: "ReLU"
}
layer {
    bottom: "layer13-conv"
    top: "layer14-conv"
    name: "layer14-conv"
    type: "Convolution"
    convolution_param {
        num_output: 128
        kernel_size: 1
        pad: 0
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer14-conv"
    top: "layer14-conv"
    name: "layer14-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer14-conv"
    top: "layer14-conv"
    name: "layer14-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer14-conv"
    top: "layer14-conv"
    name: "layer14-act"
    type: "ReLU"
}
layer {
    bottom: "layer14-conv"
    top: "layer15-conv"
    name: "layer15-conv"
    type: "Convolution"
    convolution_param {
        num_output: 256
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer15-conv"
    top: "layer15-conv"
    name: "layer15-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer15-conv"
    top: "layer15-conv"
    name: "layer15-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer15-conv"
    top: "layer15-conv"
    name: "layer15-act"
    type: "ReLU"
}
layer {
    bottom: "layer15-conv"
    top: "layer16-conv"
    name: "layer16-conv"
    type: "Convolution"
    convolution_param {
        num_output: 72
        kernel_size: 1
        pad: 0
        stride: 1
        bias_term: true
    }
}
layer {
    bottom: "layer14-conv"
    top: "layer18-route"
    name: "layer18-route"
    type: "Concat"
}
layer {
    bottom: "layer18-route"
    top: "layer19-conv"
    name: "layer19-conv"
    type: "Convolution"
    convolution_param {
        num_output: 64
        kernel_size: 1
        pad: 0
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer19-conv"
    top: "layer19-conv"
    name: "layer19-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer19-conv"
    top: "layer19-conv"
    name: "layer19-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer19-conv"
    top: "layer19-conv"
    name: "layer19-act"
    type: "ReLU"
}
layer {
    bottom: "layer19-conv"
    top: "layer20-upsample"
    name: "layer20-upsample"
    type: "Upsample"
    upsample_param {
        scale: 2
    }
}
layer {
    bottom: "layer20-upsample"
    bottom: "layer9-conv"
    top: "layer21-route"
    name: "layer21-route"
    type: "Concat"
}
layer {
    bottom: "layer21-route"
    top: "layer22-conv"
    name: "layer22-conv"
    type: "Convolution"
    convolution_param {
        num_output: 128
        kernel_size: 3
        pad: 1
        stride: 1
        bias_term: false
    }
}
layer {
    bottom: "layer22-conv"
    top: "layer22-conv"
    name: "layer22-bn"
    type: "BatchNorm"
    batch_norm_param {
        use_global_stats: true
    }
}
layer {
    bottom: "layer22-conv"
    top: "layer22-conv"
    name: "layer22-scale"
    type: "Scale"
    scale_param {
        bias_term: true
    }
}
layer {
    bottom: "layer22-conv"
    top: "layer22-conv"
    name: "layer22-act"
    type: "ReLU"
}
layer {
    bottom: "layer22-conv"
    top: "layer23-conv"
    name: "layer23-conv"
    type: "Convolution"
    convolution_param {
        num_output: 48
        kernel_size: 1
        pad: 0
        stride: 1
        bias_term: true
    }
}

接下来进行TensorRT

下载好我们前面说的TensorRT-Yolov3，其中子目录tensorRTWrapper也要下好文件

然后在tensorRTWrapper子目录下的

./TensorRT-Yolov3/tensorRTWrapper/code/CMakeLists.txt

加一句话，让它能找到你的TnesorRT

set(TENSORRT_ROOT "/home/boyun/NVIDIA/TensorRT-5.1.5.0")

cmake_minimum_required(VERSION 2.8)
project(trtNet)

set(CMAKE_BUILD_TYPE Release)

#include
include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include)

#src
set(PLUGIN_SOURCES
  src/EntroyCalibrator.cpp
  src/UpsampleLayer.cpp
  src/UpsampleLayer.cu
  src/YoloLayer.cu
  src/TrtNet.cpp
)

#
# CUDA Configuration
#
find_package(CUDA REQUIRED)

set(CUDA_VERBOSE_BUILD ON)

# Specify the cuda host compiler to use the same compiler as cmake.
set(CUDA_HOST_COMPILER ${CMAKE_CXX_COMPILER})
set(TENSORRT_ROOT "/home/boyun/NVIDIA/TensorRT-5.1.5.0")

然后在后面需要对coco数据集处理，安装库：

sudo pip install pycocotools  -i https://pypi.tuna.tsinghua.edu.cn/simple

接着就是和平常一样：

mkdir build
cd build && cmake .. && make && make install
cd ..

然后自己找个图片，还有yolov3_608.caffemodel和yolov3_608.prototxt

#for yolov3-608
./install/runYolov3 --caffemodel=./yolov3_608.caffemodel --prototxt=./yolov3_608.prototxt --input=./test.jpg --W=608 --H=608 --class=80

#for fp16
./install/runYolov3 --caffemodel=./yolov3_608.caffemodel --prototxt=./yolov3_608.prototxt --input=./test.jpg --W=608 --H=608 --class=80 --mode=fp16

#for int8 with calibration datasets
./install/runYolov3 --caffemodel=./yolov3_608.caffemodel --prototxt=./yolov3_608.prototxt --input=./test.jpg --W=608 --H=608 --class=80 --mode=int8 --calib=./calib_sample.txt

#for yolov3-416 (need to modify include/YoloConfigs for YoloKernel)
./install/runYolov3 --caffemodel=./yolov3_416.caffemodel --prototxt=./yolov3_416.prototxt --input=./test.jpg --W=416 --H=416 --class=80

下载COCO数据集到TensorRT-Yolov3/scripts/

https://blog.csdn.net/weixin_36474809/article/details/90262591

还可以测试eval也就是map，两个阈值下的测试 0.5 和 0.75

首先是label.txt的生成：

将coco数据集解压到scripts，我们只用val2014

其中scripts/annotations放置的是coco的json

scripts/images放置的是coco的val2014图片库，images中有

然后scripts下新建一个label

python ./createCOCOlabels.py ./annotations/instances_val2014.json ./label val2014

运行完会在label下生成txt

16 [214.15,41.29,348.26,243.78]

其中images是40504个文件

其中label是40137个文件

因为

接着

makelabels.py修改一句话，否则python2会报错：

    with open('labels.txt', 'w') as f:
        f.write(str)
        # print(str, file=f)

cd TensorRT-Yolov3/scripts/

../install/runYolov3 --caffemodel=../yolov3_608.caffemodel --prototxt=../yolov3_608.prototxt --W=608 --H=608 --class=80 --evallist=./labels.txt

至于为什么要这样运行代码，可以打开TensorRT-Yolov3/scripts/labels.txt

我使用的使相对路径./images 大家可以改为绝对路径，就可以像readme那样运行了，比如/home/boyun/data/coco2014/val2014

Time taken for nms is 0.002925 ms.
layer1-conv                              0.329ms
layer1-act                               0.476ms
layer2-conv                              0.539ms
layer2-act                               0.239ms
layer3-conv                              0.195ms
layer3-act                               0.121ms
layer4-conv                              0.578ms
layer4-act                               0.238ms
layer5-shortcut                          0.346ms
layer6-conv                              0.506ms
layer6-act                               0.121ms
layer7-conv                              0.118ms
layer7-act                               0.061ms
layer8-conv                              0.328ms
layer8-act                               0.120ms
layer9-shortcut                          0.277ms
layer10-conv                             0.134ms
layer10-act                              0.061ms
layer11-conv                             0.337ms
layer11-act                              0.120ms
layer12-shortcut                         0.175ms
layer13-conv                             0.548ms
layer13-act                              0.061ms
layer14-conv                             0.094ms
layer14-act                              0.031ms
layer15-conv                             0.313ms
layer15-act                              0.059ms
layer16-shortcut                         0.096ms
layer17-conv                             0.115ms
layer17-act                              0.049ms
layer18-conv                             0.365ms
layer18-act                              0.059ms
layer19-shortcut                         0.092ms
layer20-conv                             0.095ms
layer20-act                              0.031ms
layer21-conv                             0.308ms
layer21-act                              0.059ms
layer22-shortcut                         0.092ms
layer23-conv                             0.092ms
layer23-act                              0.031ms
layer24-conv                             0.307ms
layer24-act                              0.058ms
layer25-shortcut                         0.092ms
layer26-conv                             0.093ms
layer26-act                              0.031ms
layer27-conv                             0.311ms
layer27-act                              0.081ms
layer28-shortcut                         0.145ms
layer29-conv                             0.102ms
layer29-act                              0.033ms
layer30-conv                             0.312ms
layer30-act                              0.058ms
layer31-shortcut                         0.092ms
layer32-conv                             0.091ms
layer32-act                              0.030ms
layer33-conv                             0.307ms
layer33-act                              0.059ms
layer34-shortcut                         0.092ms
layer35-conv                             0.091ms
layer35-act                              0.030ms
layer36-conv                             0.308ms
layer36-act                              0.059ms
layer37-shortcut                         0.093ms
layer38-conv                             0.716ms
layer38-act                              0.032ms
layer39-conv                             0.081ms
layer39-act                              0.015ms
layer40-conv                             0.387ms
layer40-act                              0.028ms
layer41-shortcut                         0.053ms
layer42-conv                             0.086ms
layer42-act                              0.015ms
layer43-conv                             0.389ms
layer43-act                              0.028ms
layer44-shortcut                         0.053ms
layer45-conv                             0.081ms
layer45-act                              0.015ms
layer46-conv                             0.384ms
layer46-act                              0.028ms
layer47-shortcut                         0.053ms
layer48-conv                             0.086ms
layer48-act                              0.016ms
layer49-conv                             0.491ms
layer49-act                              0.028ms
layer50-shortcut                         0.053ms
layer51-conv                             0.079ms
layer51-act                              0.015ms
layer52-conv                             0.382ms
layer52-act                              0.028ms
layer53-shortcut                         0.053ms
layer54-conv                             0.079ms
layer54-act                              0.015ms
layer55-conv                             0.385ms
layer55-act                              0.028ms
layer56-shortcut                         0.053ms
layer57-conv                             0.079ms
layer57-act                              0.015ms
layer58-conv                             0.384ms
layer58-act                              0.030ms
layer59-shortcut                         0.058ms
layer60-conv                             0.103ms
layer60-act                              0.018ms
layer61-conv                             0.469ms
layer61-act                              0.028ms
layer62-shortcut                         0.053ms
layer63-conv                             0.619ms
layer63-act                              0.014ms
layer64-conv                             0.110ms
layer64-act                              0.005ms
layer65-conv                             0.612ms
layer65-act                              0.014ms
layer66-shortcut                         0.026ms
layer67-conv                             0.110ms
layer67-act                              0.007ms
layer68-conv                             0.760ms
layer68-act                              0.014ms
layer69-shortcut                         0.025ms
layer70-conv                             0.107ms
layer70-act                              0.005ms
layer71-conv                             0.615ms
layer71-act                              0.015ms
layer72-shortcut                         0.025ms
layer73-conv                             0.107ms
layer73-act                              0.005ms
layer74-conv                             0.620ms
layer74-act                              0.015ms
layer75-shortcut                         0.029ms
layer76-conv                             0.121ms
layer76-act                              0.005ms
layer77-conv                             0.827ms
layer77-act                              0.014ms
layer78-conv                             0.107ms
layer78-act                              0.005ms
layer79-conv                             0.619ms
layer79-act                              0.014ms
layer80-conv                             0.105ms
layer80-act                              0.005ms
layer81-conv                             0.630ms
layer81-act                              0.017ms
layer82-conv                             0.089ms
layer80-conv copy                        0.008ms
layer85-conv                             0.057ms
layer85-act                              0.004ms
layer86-upsample                         0.021ms
layer86-upsample copy                    0.013ms
layer88-conv                             0.112ms
layer88-act                              0.015ms
layer89-conv                             0.385ms
layer89-act                              0.028ms
layer90-conv                             0.077ms
layer90-act                              0.015ms
layer91-conv                             0.383ms
layer91-act                              0.028ms
layer92-conv                             0.077ms
layer92-act                              0.015ms
layer93-conv                             0.381ms
layer93-act                              0.028ms
layer94-conv                             0.077ms
layer92-conv copy                        0.019ms
layer97-conv                             0.031ms
layer97-act                              0.005ms
layer98-upsample                         0.037ms
layer98-upsample copy                    0.035ms
layer100-conv                            0.123ms
layer100-act                             0.032ms
layer101-conv                            0.330ms
layer101-act                             0.071ms
layer102-conv                            0.086ms
layer102-act                             0.031ms
layer103-conv                            0.306ms
layer103-act                             0.059ms
layer104-conv                            0.085ms
layer104-act                             0.031ms
layer105-conv                            0.305ms
layer105-act                             0.059ms
layer106-conv                            0.153ms
yolo-det                                 0.184ms
Time over all layers: 26.481
evalMAPResult:
class:  0 iou thresh-0.5 AP: 0.7614 recall: 0.7236 precision: 0.8038
class:  1 iou thresh-0.5 AP:   0.75 recall:   0.25 precision:   0.75
class:  2 iou thresh-0.5 AP: 0.8815 recall:  0.629 precision: 0.8966
class:  3 iou thresh-0.5 AP: 0.8598 recall: 0.7407 precision: 0.8696
class:  4 iou thresh-0.5 AP:      1 recall: 0.7692 precision:      1
class:  5 iou thresh-0.5 AP:   0.75 recall: 0.8571 precision:   0.75
class:  6 iou thresh-0.5 AP:      1 recall: 0.8667 precision:      1
class:  7 iou thresh-0.5 AP:  0.905 recall:    0.6 precision:  0.913
class:  8 iou thresh-0.5 AP: 0.5815 recall:   0.36 precision: 0.6923
class:  9 iou thresh-0.5 AP: 0.8105 recall:  0.641 precision: 0.8333
class: 10 iou thresh-0.5 AP:      1 recall:    0.8 precision:      1
class: 11 iou thresh-0.5 AP: 0.8207 recall: 0.7778 precision:  0.875
class: 12 iou thresh-0.5 AP:      1 recall: 0.2857 precision:      1
class: 13 iou thresh-0.5 AP: 0.7858 recall: 0.2564 precision: 0.8333
class: 14 iou thresh-0.5 AP: 0.8182 recall:    0.3 precision: 0.8182
class: 15 iou thresh-0.5 AP:    0.9 recall: 0.8182 precision:    0.9
class: 16 iou thresh-0.5 AP: 0.7692 recall: 0.7692 precision: 0.7692
class: 17 iou thresh-0.5 AP:      1 recall: 0.9474 precision:      1
class: 18 iou thresh-0.5 AP: 0.8074 recall: 0.6557 precision: 0.8511
class: 19 iou thresh-0.5 AP: 0.9068 recall: 0.6364 precision:  0.913
class: 20 iou thresh-0.5 AP:   0.96 recall: 0.5854 precision:   0.96
class: 21 iou thresh-0.5 AP: 0.6667 recall:      1 precision: 0.6667
class: 22 iou thresh-0.5 AP:      1 recall:  0.871 precision:      1
class: 23 iou thresh-0.5 AP:      1 recall:   0.85 precision:      1
class: 24 iou thresh-0.5 AP: 0.5283 recall: 0.2143 precision:    0.6
class: 25 iou thresh-0.5 AP:   0.88 recall: 0.6286 precision:   0.88
class: 26 iou thresh-0.5 AP: 0.7957 recall: 0.1556 precision:  0.875
class: 27 iou thresh-0.5 AP: 0.8276 recall: 0.5926 precision: 0.8421
class: 28 iou thresh-0.5 AP:  0.839 recall: 0.6071 precision: 0.8947
class: 29 iou thresh-0.5 AP: 0.8571 recall:   0.75 precision: 0.8571
class: 30 iou thresh-0.5 AP: 0.7783 recall:    0.4 precision: 0.8889
class: 31 iou thresh-0.5 AP:      1 recall: 0.6667 precision:      1
class: 32 iou thresh-0.5 AP: 0.7597 recall:   0.75 precision:    0.8
class: 33 iou thresh-0.5 AP:      1 recall:   0.75 precision:      1
class: 34 iou thresh-0.5 AP:      1 recall:   0.75 precision:      1
class: 35 iou thresh-0.5 AP: 0.6667 recall:    0.4 precision: 0.6667
class: 36 iou thresh-0.5 AP: 0.9294 recall: 0.6522 precision: 0.9375
class: 37 iou thresh-0.5 AP: 0.8815 recall: 0.5909 precision: 0.9286
class: 38 iou thresh-0.5 AP: 0.9107 recall: 0.6667 precision: 0.9231
class: 39 iou thresh-0.5 AP:  0.764 recall:  0.451 precision:  0.807
class: 40 iou thresh-0.5 AP: 0.8311 recall: 0.5152 precision:   0.85
class: 41 iou thresh-0.5 AP: 0.7388 recall: 0.5244 precision: 0.7544
class: 42 iou thresh-0.5 AP: 0.6759 recall:   0.32 precision: 0.7273
class: 43 iou thresh-0.5 AP: 0.7675 recall: 0.2353 precision:    0.8
class: 44 iou thresh-0.5 AP: 0.7292 recall: 0.3214 precision:   0.75
class: 45 iou thresh-0.5 AP: 0.6084 recall: 0.3864 precision: 0.6538
class: 46 iou thresh-0.5 AP:      1 recall: 0.1818 precision:      1
class: 47 iou thresh-0.5 AP: 0.9231 recall: 0.6667 precision: 0.9231
class: 48 iou thresh-0.5 AP: 0.8664 recall: 0.5625 precision:    0.9
class: 49 iou thresh-0.5 AP: 0.8333 recall:    0.5 precision: 0.8333
class: 50 iou thresh-0.5 AP:      1 recall: 0.3333 precision:      1
class: 51 iou thresh-0.5 AP: 0.8182 recall: 0.3214 precision: 0.8182
class: 52 iou thresh-0.5 AP: 0.6006 recall: 0.5714 precision:   0.75
class: 53 iou thresh-0.5 AP: 0.7398 recall: 0.5417 precision: 0.7647
class: 54 iou thresh-0.5 AP:      1 recall: 0.8182 precision:      1
class: 55 iou thresh-0.5 AP:    0.6 recall:  0.375 precision:    0.6
class: 56 iou thresh-0.5 AP: 0.7776 recall: 0.4729 precision: 0.8133
class: 57 iou thresh-0.5 AP: 0.8447 recall: 0.5833 precision:  0.875
class: 58 iou thresh-0.5 AP: 0.8789 recall: 0.2647 precision:    0.9
class: 59 iou thresh-0.5 AP: 0.9091 recall: 0.4348 precision: 0.9091
class: 60 iou thresh-0.5 AP:  0.738 recall: 0.4219 precision: 0.7941
class: 61 iou thresh-0.5 AP:      1 recall: 0.8696 precision:      1
class: 62 iou thresh-0.5 AP:      1 recall: 0.9333 precision:      1
class: 63 iou thresh-0.5 AP:      1 recall:    0.9 precision:      1
class: 64 iou thresh-0.5 AP:      1 recall:  0.625 precision:      1
class: 65 iou thresh-0.5 AP: 0.7904 recall: 0.5294 precision:    0.9
class: 66 iou thresh-0.5 AP:  0.875 recall: 0.7778 precision:  0.875
class: 67 iou thresh-0.5 AP:      1 recall: 0.4242 precision:      1
class: 68 iou thresh-0.5 AP:      1 recall:      1 precision:      1
class: 69 iou thresh-0.5 AP: 0.8571 recall: 0.4286 precision: 0.8571
class: 71 iou thresh-0.5 AP: 0.7921 recall:    0.5 precision: 0.8421
class: 72 iou thresh-0.5 AP: 0.8889 recall: 0.7273 precision: 0.8889
class: 73 iou thresh-0.5 AP: 0.5855 recall: 0.2436 precision: 0.7037
class: 74 iou thresh-0.5 AP:      1 recall: 0.5172 precision:      1
class: 75 iou thresh-0.5 AP: 0.6417 recall: 0.3333 precision:   0.75
class: 76 iou thresh-0.5 AP: 0.7692 recall: 0.6667 precision: 0.7692
class: 77 iou thresh-0.5 AP:      1 recall: 0.6154 precision:      1
class: 79 iou thresh-0.5 AP:   0.25 recall:    0.2 precision:    0.5
MAP:0.831
evalMAPResult:
class:  0 iou thresh-0.75 AP: 0.5343 recall: 0.5459 precision: 0.6064
class:  1 iou thresh-0.75 AP: 0.5063 recall: 0.2083 precision:  0.625
class:  2 iou thresh-0.75 AP: 0.4526 recall: 0.3871 precision: 0.5517
class:  3 iou thresh-0.75 AP: 0.5801 recall: 0.5185 precision: 0.6087
class:  4 iou thresh-0.75 AP:      1 recall: 0.7692 precision:      1
class:  5 iou thresh-0.75 AP:   0.75 recall: 0.8571 precision:   0.75
class:  6 iou thresh-0.75 AP: 0.8669 recall:    0.8 precision: 0.9231
class:  7 iou thresh-0.75 AP: 0.6159 recall: 0.4571 precision: 0.6957
class:  8 iou thresh-0.75 AP: 0.3966 recall:   0.24 precision: 0.4615
class:  9 iou thresh-0.75 AP: 0.3891 recall: 0.4359 precision: 0.5667
class: 10 iou thresh-0.75 AP: 0.4792 recall:    0.6 precision:   0.75
class: 11 iou thresh-0.75 AP: 0.4375 recall: 0.4444 precision:    0.5
class: 12 iou thresh-0.75 AP:    0.5 recall: 0.1429 precision:    0.5
class: 13 iou thresh-0.75 AP: 0.4232 recall: 0.1795 precision: 0.5833
class: 14 iou thresh-0.75 AP: 0.1515 recall:    0.1 precision: 0.2727
class: 15 iou thresh-0.75 AP:    0.8 recall: 0.7273 precision:    0.8
class: 16 iou thresh-0.75 AP: 0.6923 recall: 0.6923 precision: 0.6923
class: 17 iou thresh-0.75 AP: 0.7934 recall: 0.7895 precision: 0.8333
class: 18 iou thresh-0.75 AP: 0.3753 recall: 0.3934 precision: 0.5106
class: 19 iou thresh-0.75 AP: 0.4418 recall: 0.4242 precision: 0.6087
class: 20 iou thresh-0.75 AP: 0.8643 recall: 0.5366 precision:   0.88
class: 21 iou thresh-0.75 AP: 0.6667 recall:      1 precision: 0.6667
class: 22 iou thresh-0.75 AP: 0.9199 recall: 0.8065 precision: 0.9259
class: 23 iou thresh-0.75 AP: 0.9377 recall:    0.8 precision: 0.9412
class: 24 iou thresh-0.75 AP:  0.253 recall: 0.1429 precision:    0.4
class: 25 iou thresh-0.75 AP: 0.4121 recall: 0.3714 precision:   0.52
class: 26 iou thresh-0.75 AP: 0.4583 recall:0.08889 precision:    0.5
class: 27 iou thresh-0.75 AP: 0.5228 recall: 0.4074 precision: 0.5789
class: 28 iou thresh-0.75 AP: 0.6633 recall:    0.5 precision: 0.7368
class: 29 iou thresh-0.75 AP: 0.3714 recall:  0.375 precision: 0.4286
class: 30 iou thresh-0.75 AP:      0 recall:      0 precision:      0
class: 31 iou thresh-0.75 AP:    0.5 recall: 0.3333 precision:    0.5
class: 32 iou thresh-0.75 AP: 0.5012 recall: 0.5625 precision:    0.6
class: 33 iou thresh-0.75 AP:      1 recall:   0.75 precision:      1
class: 34 iou thresh-0.75 AP: 0.5608 recall:    0.5 precision: 0.6667
class: 35 iou thresh-0.75 AP: 0.4583 recall:    0.3 precision:    0.5
class: 36 iou thresh-0.75 AP: 0.5326 recall: 0.4783 precision: 0.6875
class: 37 iou thresh-0.75 AP: 0.3845 recall: 0.3182 precision:    0.5
class: 38 iou thresh-0.75 AP: 0.4077 recall: 0.3333 precision: 0.4615
class: 39 iou thresh-0.75 AP: 0.3815 recall: 0.2843 precision: 0.5088
class: 40 iou thresh-0.75 AP: 0.6853 recall: 0.4242 precision:    0.7
class: 41 iou thresh-0.75 AP: 0.4974 recall: 0.3902 precision: 0.5614
class: 42 iou thresh-0.75 AP:  0.471 recall:   0.24 precision: 0.5455
class: 43 iou thresh-0.75 AP: 0.5225 recall: 0.1765 precision:    0.6
class: 44 iou thresh-0.75 AP: 0.3673 recall: 0.2143 precision:    0.5
class: 45 iou thresh-0.75 AP: 0.2551 recall:   0.25 precision: 0.4231
class: 46 iou thresh-0.75 AP:    0.5 recall:0.09091 precision:    0.5
class: 47 iou thresh-0.75 AP:  0.422 recall: 0.4444 precision: 0.6154
class: 48 iou thresh-0.75 AP: 0.6711 recall:    0.5 precision:    0.8
class: 49 iou thresh-0.75 AP: 0.6333 recall:    0.4 precision: 0.6667
class: 50 iou thresh-0.75 AP:   0.71 recall: 0.2667 precision:    0.8
class: 51 iou thresh-0.75 AP: 0.3838 recall: 0.1786 precision: 0.4545
class: 52 iou thresh-0.75 AP: 0.1869 recall: 0.3333 precision: 0.4375
class: 53 iou thresh-0.75 AP: 0.3933 recall: 0.3333 precision: 0.4706
class: 54 iou thresh-0.75 AP: 0.8627 recall: 0.7273 precision: 0.8889
class: 55 iou thresh-0.75 AP:    0.6 recall:  0.375 precision:    0.6
class: 56 iou thresh-0.75 AP: 0.3677 recall: 0.3256 precision:   0.56
class: 57 iou thresh-0.75 AP: 0.3568 recall:  0.375 precision: 0.5625
class: 58 iou thresh-0.75 AP: 0.3494 recall: 0.1765 precision:    0.6
class: 59 iou thresh-0.75 AP: 0.5091 recall: 0.2609 precision: 0.5455
class: 60 iou thresh-0.75 AP: 0.5487 recall: 0.3281 precision: 0.6176
class: 61 iou thresh-0.75 AP: 0.9211 recall: 0.8261 precision:   0.95
class: 62 iou thresh-0.75 AP: 0.8516 recall:    0.8 precision: 0.8571
class: 63 iou thresh-0.75 AP: 0.8889 recall:    0.8 precision: 0.8889
class: 64 iou thresh-0.75 AP:    0.6 recall:  0.375 precision:    0.6
class: 65 iou thresh-0.75 AP: 0.3296 recall: 0.2941 precision:    0.5
class: 66 iou thresh-0.75 AP: 0.2923 recall: 0.4444 precision:    0.5
class: 67 iou thresh-0.75 AP: 0.7691 recall: 0.3636 precision: 0.8571
class: 68 iou thresh-0.75 AP:      1 recall:      1 precision:      1
class: 69 iou thresh-0.75 AP: 0.7143 recall: 0.3571 precision: 0.7143
class: 71 iou thresh-0.75 AP:  0.342 recall: 0.2812 precision: 0.4737
class: 72 iou thresh-0.75 AP: 0.8889 recall: 0.7273 precision: 0.8889
class: 73 iou thresh-0.75 AP:0.06203 recall:0.05128 precision: 0.1481
class: 74 iou thresh-0.75 AP:  0.484 recall: 0.3103 precision:    0.6
class: 75 iou thresh-0.75 AP: 0.4083 recall: 0.2222 precision:    0.5
class: 76 iou thresh-0.75 AP: 0.2231 recall: 0.2667 precision: 0.3077
class: 77 iou thresh-0.75 AP:    0.6 recall: 0.4615 precision:   0.75
class: 79 iou thresh-0.75 AP:   0.25 recall:    0.2 precision:    0.5
MAP:0.5363

TensorRT已经支持的激活函数：

知识在于分享

关注

1
点赞
踩
8

收藏

觉得还不错? 一键收藏
1
评论
TensorRT-5.1.5.0-YOLOv3

接上一个博客TensorRT-5.1.5.0-SSDhttps://blog.csdn.net/baidu_40840693/article/details/95642055首先：upsample层密码bwrdhttps://pan.baidu.com/s/13GpoYoqKSCeFX0m0ves_fQ然后：https://github.com/lewes636...
复制链接

扫一扫

专栏目录