18、darknet2ncnn将darknet 模型转换为ncnn模型，实现darknet网络...

东方耀

18、darknet2ncnn将darknet 模型转换为ncnn模型，实现darknet网络模型在移动端的快速部署




yolov3-tiny:Android端



转换模型(darknet ---> ncnn):
Use darknet2ncnn to convert darknet model to ncnn model
https://github.com/xiangweizeng/darknet2ncnn  这个不行了

这个才行: 重点看这个
https://blog.csdn.net/tugouxp/article/details/121143770








git clone https://gitee.com/damone/darknet2ncnn.git
cd darknet2ncnn


git submodule init


git submodule update




cd darknet


开启opencv 和 改架构CUDA的
ARCH= -gencode arch=compute_30,code=sm_30 \
      -gencode arch=compute_35,code=sm_35 \
      -gencode arch=compute_50,code=[sm_50,compute_50] \
      -gencode arch=compute_52,code=[sm_52,compute_52] \
      -gencode arch=compute_60,code=sm_60 \
      -gencode arch=compute_61,code=[sm_61,compute_61] \
      -gencode arch=compute_75,code=sm_75


make -j8


rm libdarknet.so




下面为我总结的：
data/yolov3-tiny.cfg  data/yolov3-tiny.weights 这两个文件没有变化 应该是输入的
example/zoo/yolov3-tiny.param  example/zoo/yolov3-tiny.bin 这两个文件已经变了 应该是输出的
example/data/dog.jpg 这个没变
命令的用法：./darknet2ncnn darknet.cfg darknet.weights ncnn.param ncnn.bin
命令的用法：./convert_verify [darknet.cfg] [darknet.weights] [ncnn.param] [ncnn.bin] [test.image]
ncnn模型需要的网络配置和权重文件的格式：[ncnn.param] [ncnn.bin]


make yolov3-tiny.net 执行之后：


./darknet2ncnn data/yolov3-tiny.cfg  data/yolov3-tiny.weights example/zoo/yolov3-tiny.param  example/zoo/yolov3-tiny.bin
Loading weights from data/yolov3-tiny.weights...Done!




./convert_verify data/yolov3-tiny.cfg  data/yolov3-tiny.weights example/zoo/yolov3-tiny.param  example/zoo/yolov3-tiny.bin example/data/dog.jpg
Loading weights from data/yolov3-tiny.weights...Done!


Start run all operation:
conv_0 : weights diff : 0.000000
conv_0_batch_norm : slope diff : 0.000000
conv_0_batch_norm : mean diff : 0.000000
conv_0_batch_norm : variance diff : 0.000000
conv_0_batch_norm : biases diff : 0.000000
Layer: 0, Blob : conv_0_activation, Total Diff 259359.718750 Avg Diff: 0.093669
Layer: 1, Blob : maxpool_1, Total Diff 85732.632812 Avg Diff: 0.123851
conv_2 : weights diff : 0.000000
conv_2_batch_norm : slope diff : 0.000000
conv_2_batch_norm : mean diff : 0.000000
conv_2_batch_norm : variance diff : 0.000000
conv_2_batch_norm : biases diff : 0.000000
Layer: 2, Blob : conv_2_activation, Total Diff 95723.023438 Avg Diff: 0.069142
Layer: 3, Blob : maxpool_3, Total Diff 33560.359375 Avg Diff: 0.096964
conv_4 : weights diff : 0.000000
conv_4_batch_norm : slope diff : 0.000000
conv_4_batch_norm : mean diff : 0.000000
conv_4_batch_norm : variance diff : 0.000000
conv_4_batch_norm : biases diff : 0.000000
Layer: 4, Blob : conv_4_activation, Total Diff 47419.449219 Avg Diff: 0.068503
Layer: 5, Blob : maxpool_5, Total Diff 17139.689453 Avg Diff: 0.099041
conv_6 : weights diff : 0.000000
conv_6_batch_norm : slope diff : 0.000000
conv_6_batch_norm : mean diff : 0.000000
conv_6_batch_norm : variance diff : 0.000000
conv_6_batch_norm : biases diff : 0.000000
Layer: 6, Blob : conv_6_activation, Total Diff 17113.324219 Avg Diff: 0.049444
Layer: 7, Blob : maxpool_7, Total Diff 6735.393066 Avg Diff: 0.077841
conv_8 : weights diff : 0.000000
conv_8_batch_norm : slope diff : 0.000000
conv_8_batch_norm : mean diff : 0.000000
conv_8_batch_norm : variance diff : 0.000000
conv_8_batch_norm : biases diff : 0.000000
Layer: 8, Blob : conv_8_activation, Total Diff 6129.817383 Avg Diff: 0.035421
Layer: 9, Blob : maxpool_9, Total Diff 2565.875244 Avg Diff: 0.059307
conv_10 : weights diff : 0.000000
conv_10_batch_norm : slope diff : 0.000000
conv_10_batch_norm : mean diff : 0.000000
conv_10_batch_norm : variance diff : 0.000000
conv_10_batch_norm : biases diff : 0.000000
Layer: 10, Blob : conv_10_activation, Total Diff 1663.026123 Avg Diff: 0.019220
Layer: 11, Blob : maxpool_11, Total Diff 2608.484619 Avg Diff: 0.030146
conv_12 : weights diff : 0.000000
conv_12_batch_norm : slope diff : 0.000000
conv_12_batch_norm : mean diff : 0.000000
conv_12_batch_norm : variance diff : 0.000000
conv_12_batch_norm : biases diff : 0.000000
Layer: 12, Blob : conv_12_activation, Total Diff 13798.509766 Avg Diff: 0.079734
conv_13 : weights diff : 0.000000
conv_13_batch_norm : slope diff : 0.000000
conv_13_batch_norm : mean diff : 0.000000
conv_13_batch_norm : variance diff : 0.000000
conv_13_batch_norm : biases diff : 0.000000
Layer: 13, Blob : conv_13_activation, Total Diff 1050.790405 Avg Diff: 0.024288
conv_14 : weights diff : 0.000000
conv_14_batch_norm : slope diff : 0.000000
conv_14_batch_norm : mean diff : 0.000000
conv_14_batch_norm : variance diff : 0.000000
conv_14_batch_norm : biases diff : 0.000000
Layer: 14, Blob : conv_14_activation, Total Diff 3824.680176 Avg Diff: 0.044202
conv_15 : weights diff : 0.000000
conv_15 : biases diff : 0.000000
Layer: 15, Blob : conv_15_activation, Total Diff 6452.077148 Avg Diff: 0.149718
Layer: 17, Blob : route_17, Total Diff 1050.790405 Avg Diff: 0.024288
conv_18 : weights diff : 0.000000
conv_18_batch_norm : slope diff : 0.000000
conv_18_batch_norm : mean diff : 0.000000
conv_18_batch_norm : variance diff : 0.000000
conv_18_batch_norm : biases diff : 0.000000
Layer: 18, Blob : conv_18_activation, Total Diff 1104.450073 Avg Diff: 0.051056
Layer: 19, Blob : upsample_19, Total Diff 4417.800293 Avg Diff: 0.051056
Layer: 20, Blob : route_20, Total Diff 10547.617188 Avg Diff: 0.040633
conv_21 : weights diff : 0.000000
conv_21_batch_norm : slope diff : 0.000000
conv_21_batch_norm : mean diff : 0.000000
conv_21_batch_norm : variance diff : 0.000000
conv_21_batch_norm : biases diff : 0.000000
Layer: 21, Blob : conv_21_activation, Total Diff 9796.311523 Avg Diff: 0.056608
conv_22 : weights diff : 0.000000
conv_22 : biases diff : 0.000000
Layer: 22, Blob : conv_22_activation, Total Diff 28997.115234 Avg Diff: 0.168216






make yolov3-tiny.coco


./yolo zoo/yolov3-tiny.param  zoo/yolov3-tiny.bin  data/dog.jpg  data/coco.names


命令的用法：./yolo [ncnn.param] [ncnn.bin] [imagepath] [lable.txt]


./yolo-param-bin [ncnn.param.bin] [ncnn.bin] [imagepath] [lable.txt]
./classifier [ncnn.param] [ncnn.bin] [imagepath] [lable.txt]


./classifier zoo/yolov3-tiny.param  zoo/yolov3-tiny.bin  data/dog.jpg  data/coco.names




mkdir -p build-android-armv7


cd build-android-armv7


cmake -DCMAKE_TOOLCHAIN_FILE=/home/dfy888/Android_Sdk/ndk/android-ndk-r19c/build/cmake/android.toolchain.cmake -DANDROID_ABI="armeabi-v7a" -DANDROID_ARM_NEON=ON -DANDROID_PLATFORM=android-20 ..
cmake -DCMAKE_TOOLCHAIN_FILE= -DANDROID_ABI= -DANDROID_ARM_NEON=ON -DANDROID_PLATFORM=android-14 ..
ANDROID_ABI 是移动端硬件架构名字，"armeabi-v7a" 支持绝大部分手机硬件，项目中运用RK3399是v8硬件架构,这边用手机测试所以用v7a
ANDROID_ARM_NEON 是否使用 NEON 指令集，设为 ON 支持绝大部分手机硬件
ANDROID_PLATFORM 指定最低系统版本，"android-14" 就是 android-4.0

我们推荐大部分的开发者删简 x86，x86_64，和arm32 的ABIs。您可以通过如下的Gradle配置实现，
这个配置只包括了 armeabi-v7a和arm64-v8a，该配置能涵盖住大部分的现代Android设备
ndk {
            abiFilters 'armeabi-v7a', 'arm64-v8a'
        }


lib/libncnn.a文件　　以我理解这个文件就是相当于把ncnn打包成android可导入的形式
include文件夹　里面包含各种常用的.h文件


app闪退 报错：java.lang.UnsatisfiedLinkError: dlopen failed: library "libomp.so" not found


还是小屁孩回复 hzmsfdl1年前
闪退根据我们实验室的使用，应该是你的ndk版本问题   的确我降级后可以了 19都可以




参考：
https://gitee.com/damone/darknet2ncnn
https://github.com/Tencent/ncnn
http://github-mirror.bugkiller.org/chehongshu/ncnnforandroid_objectiondetection_Mobilenetssd




目前ncnn对Caffe的支持是最好的 模型还可以加密等



对yolov4-tiny模型转换 日志:

1. jiang@jiang-Ubuntu:~/jjj_darknet_works/darknet2ncnn$ ./darknet2ncnn data/yolov4-tiny.cfg  data/yolov4-tiny.weights example/zoo/yolov4-tiny-radar.param  example/zoo/yolov4-tiny-radar.bin
   
2. layer     filters    size              input                output
   
3.     0 conv     32  3 x 3 / 2   608 x 608 x   3   ->   304 x 304 x  32  0.160 BFLOPs
   
4.     1 conv     64  3 x 3 / 2   304 x 304 x  32   ->   152 x 152 x  64  0.852 BFLOPs
   
5.     2 conv     64  3 x 3 / 1   152 x 152 x  64   ->   152 x 152 x  64  1.703 BFLOPs
   
6.     3 route  2
   
7. Unused field: 'groups = 2'
   
8. Unused field: 'group_id = 1'
   
9.     4 conv     32  3 x 3 / 1   152 x 152 x  64   ->   152 x 152 x  32  0.852 BFLOPs
   
10.     5 conv     32  3 x 3 / 1   152 x 152 x  32   ->   152 x 152 x  32  0.426 BFLOPs
    
11.     6 route  5 4
    
12.     7 conv     64  1 x 1 / 1   152 x 152 x  64   ->   152 x 152 x  64  0.189 BFLOPs
    
13.     8 route  2 7
    
14.     9 max          2 x 2 / 2   152 x 152 x 128   ->    76 x  76 x 128
    
15.    10 conv    128  3 x 3 / 1    76 x  76 x 128   ->    76 x  76 x 128  1.703 BFLOPs
    
16.    11 route  10
    
17. Unused field: 'groups = 2'
    
18. Unused field: 'group_id = 1'
    
19.    12 conv     64  3 x 3 / 1    76 x  76 x 128   ->    76 x  76 x  64  0.852 BFLOPs
    
20.    13 conv     64  3 x 3 / 1    76 x  76 x  64   ->    76 x  76 x  64  0.426 BFLOPs
    
21.    14 route  13 12
    
22.    15 conv    128  1 x 1 / 1    76 x  76 x 128   ->    76 x  76 x 128  0.189 BFLOPs
    
23.    16 route  10 15
    
24.    17 max          2 x 2 / 2    76 x  76 x 256   ->    38 x  38 x 256
    
25.    18 conv    256  3 x 3 / 1    38 x  38 x 256   ->    38 x  38 x 256  1.703 BFLOPs
    
26.    19 route  18
    
27. Unused field: 'groups = 2'
    
28. Unused field: 'group_id = 1'
    
29.    20 conv    128  3 x 3 / 1    38 x  38 x 256   ->    38 x  38 x 128  0.852 BFLOPs
    
30.    21 conv    128  3 x 3 / 1    38 x  38 x 128   ->    38 x  38 x 128  0.426 BFLOPs
    
31.    22 route  21 20
    
32.    23 conv    256  1 x 1 / 1    38 x  38 x 256   ->    38 x  38 x 256  0.189 BFLOPs
    
33.    24 route  18 23
    
34.    25 max          2 x 2 / 2    38 x  38 x 512   ->    19 x  19 x 512
    
35.    26 conv    512  3 x 3 / 1    19 x  19 x 512   ->    19 x  19 x 512  1.703 BFLOPs
    
36.    27 conv    256  1 x 1 / 1    19 x  19 x 512   ->    19 x  19 x 256  0.095 BFLOPs
    
37.    28 conv    512  3 x 3 / 1    19 x  19 x 256   ->    19 x  19 x 512  0.852 BFLOPs
    
38.    29 conv     18  1 x 1 / 1    19 x  19 x 512   ->    19 x  19 x  18  0.007 BFLOPs
    
39.    30 yolo
    
40. Unused field: 'scale_x_y = 1.05'
    
41. Unused field: 'cls_normalizer = 1.0'
    
42. Unused field: 'iou_normalizer = 0.07'
    
43. Unused field: 'iou_loss = ciou'
    
44. Unused field: 'resize = 1.5'
    
45. Unused field: 'nms_kind = greedynms'
    
46. Unused field: 'beta_nms = 0.6'
    
47.    31 route  27
    
48.    32 conv    128  1 x 1 / 1    19 x  19 x 256   ->    19 x  19 x 128  0.024 BFLOPs
    
49.    33 upsample            2x    19 x  19 x 128   ->    38 x  38 x 128
    
50.    34 route  33 23
    
51.    35 conv    256  3 x 3 / 1    38 x  38 x 384   ->    38 x  38 x 256  2.555 BFLOPs
    
52.    36 conv     18  1 x 1 / 1    38 x  38 x 256   ->    38 x  38 x  18  0.013 BFLOPs
    
53.    37 yolo
    
54. Unused field: 'scale_x_y = 1.05'
    
55. Unused field: 'cls_normalizer = 1.0'
    
56. Unused field: 'iou_normalizer = 0.07'
    
57. Unused field: 'iou_loss = ciou'
    
58. Unused field: 'resize = 1.5'
    
59. Unused field: 'nms_kind = greedynms'
    
60. Unused field: 'beta_nms = 0.6'
    
61. Loading weights from data/yolov4-tiny.weights...Done!
    

复制代码