Darknet2ncnn
Darknet2ncnn converts the darknet model to the ncnn model
Install / Use
/learn @xiangweizeng/Darknet2ncnnREADME
darknet2ncnn
Introduction
Darknet2ncnn converts the darknet model to the ncnn model, enabling rapid deployment of the darknet network model on the mobile device.
Gitee : https://gitee.com/damone/darknet2ncnn
- Support network layers except local/xor conv, rnn, lstm, gru, crnn and iseg
- Added all activation operations not directly supported by ncnn, implemented in the layer DarknetActivation
- Added the implementation of the shortcut layer, implemented in the layer DarknetShortCut
- Added yolo layer and detection layer implementation, support YOLOV1 and YOLOV3
- Provides a converted model verification tool, convert_verify, which supports checking the calculation output of each layer of the network, supports convolutional layer parameter checking, and facilitates rapid positioning of problems in model conversion.
NCNN, merged darknet layers https://github.com/xiangweizeng/ncnn
Technical communication QQ group
点击链接加入群聊【darknet2ncnn】：https://jq.qq.com/?_wv=1027&k=5Gou5zw
Install&Usage
-
Install opencv-dev, gcc, g++, make, cmake
-
Download source
git clone https://github.com/xiangweizeng/darknet2ncnn.git
- Init submodule
cd darknet2ncnn
git submodule init
git submodule update
- build darknet
cd darknet
make -j8
rm libdarknet.so
- build ncnn
# workspace darknet2ncnn
cd ncnn
mkdir build
cd build
cmake ..
make -j8
make install
cd ../../
- Build darknet2ncnn , convert_verify and libdarknet2ncnn.a
# workspace darknet2ncnn
make -j8
- Convert and verify
- Cifar
# workspace darknet2ncnn
make cifar
./darknet2ncnn data/cifar.cfg data/cifar.backup example/zoo/cifar.param example/zoo/cifar.bin
layer filters size input output
0 conv 128 3 x 3 / 1 28 x 28 x 3 -> 28 x 28 x 128 0.005 BFLOPs
1 conv 128 3 x 3 / 1 28 x 28 x 128 -> 28 x 28 x 128 0.231 BFLOPs
.
.
.
13 dropout p = 0.50 25088 -> 25088
14 conv 10 1 x 1 / 1 7 x 7 x 512 -> 7 x 7 x 10 0.001 BFLOPs
15 avg 7 x 7 x 10 -> 10
16 softmax 10
Loading weights from data/cifar.backup...Done!
./convert_verify data/cifar.cfg data/cifar.backup example/zoo/cifar.param example/zoo/cifar.bin example/data/21263_ship.png
layer filters size input output
0 conv 128 3 x 3 / 1 28 x 28 x 3 -> 28 x 28 x 128 0.005 BFLOPs
1 conv 128 3 x 3 / 1 28 x 28 x 128 -> 28 x 28 x 128 0.231 BFLOPs
.
.
.
13 dropout p = 0.50 25088 -> 25088
14 conv 10 1 x 1 / 1 7 x 7 x 512 -> 7 x 7 x 10 0.001 BFLOPs
15 avg 7 x 7 x 10 -> 10
16 softmax 10
Loading weights from data/cifar.backup...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 595.703918 Avg Diff: 0.005936
.
.
.
Layer: 14, Blob : conv_14_activation, Total Diff 35.058342 Avg Diff: 0.071548
Layer: 15, Blob : gloabl_avg_pool_15, Total Diff 0.235242 Avg Diff: 0.023524
Layer: 16, Blob : softmax_16, Total Diff 0.000001 Avg Diff: 0.000000
- Yolov3-tiny
make yolov3-tiny.net
./darknet2ncnn data/yolov3-tiny.cfg data/yolov3-tiny.weights example/zoo/yolov3-tiny.param example/zoo/yolov3-tiny.bin
layer filters size input output
0 conv 16 3 x 3 / 1 416 x 416 x 3 -> 416 x 416 x 16 0.150 BFLOPs
.
.
.
22 conv 255 1 x 1 / 1 26 x 26 x 256 -> 26 x 26 x 255 0.088 BFLOPs
23 yolo
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
layer filters size input output
0 conv 16 3 x 3 / 1 416 x 416 x 3 -> 416 x 416 x 16 0.150 BFLOPs
1 max 2 x 2 / 2 416 x 416 x 16 -> 208 x 208 x 16
.
.
.
20 route 19 8
21 conv 256 3 x 3 / 1 26 x 26 x 384 -> 26 x 26 x 256 1.196 BFLOPs
22 conv 255 1 x 1 / 1 26 x 26 x 256 -> 26 x 26 x 255 0.088 BFLOPs
23 yolo
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
.
.
.
conv_22 : weights diff : 0.000000
conv_22 : biases diff : 0.000000
Layer: 22, Blob : conv_22_activation, Total Diff 29411.240234 Avg Diff: 0.170619
- Build example
# workspace darknet2ncnn
cd example
make -j2
- Run classifier
# workspace example
make cifar.cifar
./classifier zoo/cifar.param zoo/cifar.bin data/32516_dog.png data/cifar_lable.txt
4 deer = 0.263103
6 frog = 0.224274
5 dog = 0.191360
3 cat = 0.180164
2 bird = 0.094251
- Run Yolo
- Run YoloV3-tiny
# workspace example
make yolov3-tiny.coco
./yolo zoo/yolov3-tiny.param zoo/yolov3-tiny.bin data/dog.jpg data/coco.names
3 [car ] = 0.64929 at 252.10 92.13 114.88 x 52.98
2 [bicycle ] = 0.60786 at 111.18 134.81 201.40 x 160.01
17 [dog ] = 0.56338 at 69.91 152.89 130.30 x 179.04
8 [truck ] = 0.54883 at 288.70 103.80 47.98 x 34.17
3 [car ] = 0.28332 at 274.47 100.36 48.90 x 35.03
- YoloV3-tiny figure
NCNN:
DARKNET:
- Build benchmark
# workspace darknet2ncnn
cd benchmark
make
- Run benchmark
- Firefly RK3399 thread2
firefly@firefly:~/project/darknet2ncnn/benchmark$ ./benchdarknet 10 2 &
[1] 4556
loop_count = 10
num_threads = 2
powersave = 0
firefly@firefly:~/project/darknet2ncnn/benchmark$ taskset -pc 4,5 4556
pid 4556's current affinity list: 0-5
pid 4556's new affinity list: 4,5
cifar min = 85.09 max = 89.15 avg = 85.81
alexnet min = 218.38 max = 220.96 avg = 218.88
darknet min = 88.38 max = 88.95 avg = 88.63
darknet19 min = 330.55 max = 337.12 avg = 333.64
darknet53 min = 874.69 max = 920.99 avg = 897.19
densenet201 min = 678.99 max = 684.97 avg = 681.38
extraction min = 332.78 max = 340.54 avg = 334.98
resnet18 min = 238.93 max = 245.66 avg = 240.32
resnet34 min = 398.92 max = 404.93 avg = 402.18
resnet50 min = 545.39 max = 558.67 avg = 551.90
resnet101 min = 948.88 max = 960.51 avg = 952.99
resnet152 min = 1350.78 max = 1373.51 avg = 1363.40
resnext50 min = 660.55 max = 698.07 avg = 669.49
resnext101-32x4d min = 1219.80 max = 1232.07 avg = 1227.58
resnext152-32x4d min = 1788.03 max = 1798.79 avg = 1795.48
vgg-16 min = 883.33 max = 903.98 avg = 895.03
yolov1-tiny min = 222.40 max = 227.51 avg = 224.67
yolov2-tiny min = 250.54 max = 259.84 avg = 252.38
yolov3-tiny min = 240.80 max = 249.98 avg = 245.08
- Firefly RK3399 thread4
firefly@firefly:~/project/darknet2ncnn/benchmark$ ./benchdarknet 10 4 &
[1] 4663
loop_count = 10
num_threads = 4
powersave = 0
firefly@firefly:~/project/darknet2ncnn/benchmark$ taskset -pc 0-3 4663
pid 4663's current affinity list: 0-5
pid 4663's new affinity list: 0-3
cifar min = 96.51 max = 108.22 avg = 100.60
alexnet min = 411.38 max = 432.00 avg = 420.11
darknet min = 101.89 max = 119.73 avg = 106.46
darknet19 min = 421.46 max = 453.59 avg = 433.74
darknet53 min = 1375.30 max = 1492.79 avg = 1406.82
densenet201 min = 1154.26 max = 1343.53 avg = 1218.28
extraction min = 399.31 max = 460.01 avg = 428.17
resnet18 min = 317.70 max = 376.89 avg = 338.93
resnet34 min = 567.30 max = 604.44 avg = 580.65
resnet50 min = 838.94 max = 978.21 avg = 925.14
resnet101 min = 1562.60 max = 1736.91 avg = 1642.27
resnet152 min = 2250.32 max = 2394.38 avg = 2311.42
resnext50 min = 993.34 max = 1210.04 avg = 1093.05
resnext101-32x4d min = 2207.74 max = 2366.66 avg = 2281.82
resnext152-32x4d min = 3139.89 max = 3372.58 avg = 3282.99
vgg-16 min = 1259.17 max = 1359.55 avg = 1300.04
yolov1-tiny min = 272.31 max = 330.71 avg = 295.98
yolov2-tiny min = 314.25 max = 352.12 avg = 329.02
yolov3-tiny min = 300.28 max = 349.13 avg = 322.54
Support network(Zoo)
Zoo(Baidu Cloud):https://pan.baidu.com/s/1BgqL8p1yB4gRPrxAK73omw
Cifar
- cifar
ImageNet
- alexnet
- darknet
- darknet19
- darknet53
- densenet201
- extraction
- resnet18
- resnet34
- resnet50
- resnet101
- resnet152
- resnext50
- resnext101-32x4d
- resnext152-32x4d
- vgg-16
YOLO
- yolov1-tiny
- yolov2-tiny
- yolov2
- yolov3-tiny
- yolov3
- yolov3-spp
Benchmark
Time： ms
Network | i7-7700K 4.20GHz 8thread | IMX6Q,Topeet 4thead | Firefly rk3399 2thread | Firefly rk3399 4thread ---------|----------|---------|---------|--------- cifar | 62 | 302 | 85 | 100 alexnet | 92 | 649 | 218 | 420 darknet | 28 | 297 | 88 | 106 darknet19 | 202 | 1218 | 333 | 433 darknet53 | 683 | 3235 | 897 | 1406 densenet201 |218 | 2647 | 681 | 12
