Tina SDK 使用V833/V831 NPU跑YOLO网络

1.Tina SDK打开内核NNA驱动支持make kernel_menuconfig->打开 CONFIG_SUNXI_NNA这样，内核才会出现/dev/nna设备节点2.编译YOLO算法用例库算法用例库的布局如下图所示，编译前，需要调整Makefile中默认的编译器路径未正确路径:编译过程：编译生成了可执行测试文件yolo3将整个目录拷贝到TF卡上，卡挂载到V833平台上，输出警告不用管

papaofdoudou

2461人浏览 · 2021-08-31 12:25:13

papaofdoudou · 2021-08-31 12:25:13 发布

1.Tina SDK打开内核NNA驱动支持

make kernel_menuconfig->打开 CONFIG_SUNXI_NNA

这样，内核才会出现/dev/nna设备节点

2.编译YOLO算法用例库

算法用例库的布局如下图所示，

编译前，需要调整Makefile中默认的编译器路径未正确路径:

编译过程：

编译生成了可执行测试文件yolo3

将整个目录拷贝到TF卡上，卡挂载到V833平台上，输出警告不用管.

2.运行测试用例

首次执行yolo3用例，运行失败，会出现找不到库的问题。

问题的原因是用例找不到NNA的runtime库，我们需要将runtime库添加到LD_LIBRARY_PATH环境变量下面，根据上面的目录结构，我们找到runtime库的location在路径：

添加runtime库路径到环境变量，重新运行用例，执行成功：

3.用例解读

首先看用例目录中唯一的一张图片，图片内容展示的是一名勇敢的冲浪运动员在冲浪，比较明显的物体有人和冲浪板。

我们看用例识别出了什么？

root@(none):/mnt/extsd/YOLOV3_Demo# ./yolov3
[  282.552689] NNA: nna_open irqnum = 312
nna_fd = 4[  282.557041] NNA: enable nna  irq

AwG2dCreate!
Load model param and bin...
data scale: 127.404358
conv 81_599 scale: 4.844433
conv 93_685 scale: 3.955845
conv 105_772 scale: 3.481591

Read and transform data...
image width: 612, image height: 612
resized image width: 416, resized image height: 416
padded image width: 416, padded image height: 416

Infer by IPU...
Infer sucessfully!

Dequantize...
0, output width 13, output height 13, output channel 255
1, output width 26, output height 26, output channel 255
2, output width 52, output height 52, output channel 255

Infer by CPU...
detect object num: 2
0, label 1, class person, prob 0.999197, x1 213, y1 138, x2 399, y2 375
1, label 38, class surfboard, prob 0.770885, x1 258, y1 322, x2 516, y2 410
[  285.378372] NNA: disable nna  irq
[  285.382302] NNA: nna_release irqnum = 312
root@(none):/mnt/extsd/YOLOV3_Demo#

可以看到，识别除了两个物体，第一个%99.9的概率是一个人，并且给出了对角线的坐标值，识别出的第二个物体是滑板，它的概率是%77.088，并且也给出了对角线的坐标值.

我们根据提供的坐标信息，绘制出矩形框，直观的看一下人形检测算法的结果是什么：

import cv2
 
fname = './test.bmp'
img = cv2.imread(fname)

pt1 = (213, 138)
pt2 = (399, 375)
cv2.rectangle(img, pt1, pt2, (0, 255, 0), 2)

pt3 = (258, 322)
pt4 = (516, 410)
cv2.rectangle(img, pt3, pt4, (255, 0, 0), 2)

cv2.imwrite('22.jpg', img)

绘制结果：

可以看到结论还是蛮精确的。

用例调试：

打开Tina strace通过strace探测用例需要用到哪些设备：

strace跟踪的结果是：

root@(none):/mnt/extsd/YOLOV3_Demo# strace -f -e trace=open,openat ./yolov3
open("/usr/lib/eyesee-mpp/libawipubsp.so", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = -1 ENOENT (No such file or directory)
open("/mnt/extsd/YOLOV3_Demo/3rdparty/lib/libawipubsp.so", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = 3
open("/usr/lib/eyesee-mpp/libawnn.so", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = -1 ENOENT (No such file or directory)
open("/mnt/extsd/YOLOV3_Demo/3rdparty/lib/libawnn.so", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = 3
open("/usr/lib/eyesee-mpp/libyolov3.so", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = -1 ENOENT (No such file or directory)
open("/mnt/extsd/YOLOV3_Demo/3rdparty/lib/libyolov3.so", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = 3
open("/usr/lib/eyesee-mpp/libstdc++.so.6", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = -1 ENOENT (No such file or directory)
open("/mnt/extsd/YOLOV3_Demo/3rdparty/lib/libstdc++.so.6", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = -1 ENOEN[  771.885964] NNA: nna_open irqnum = 312
T (No such file or directory)
open("/etc/ld-musl-armhf.path", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = -1 ENOENT (No such file or dire[  771.894841] NNA: enable nna  irq
ctory)
open("/lib/libstdc++.so.6", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = 3
open("/usr/lib/eyesee-mpp/libgcc_s.so.1", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = -1 ENOENT (No such file or directory)
open("/mnt/extsd/YOLOV3_Demo/3rdparty/lib/libgcc_s.so.1", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = -1 ENOENT (No such file or directory)
open("/lib/libgcc_s.so.1", O_RDONLY|O_LARGEFILE|O_CLOEXEC) = 3
open("/dev/mem", O_RDWR|O_NONBLOCK|O_LARGEFILE|O_CLOEXEC) = 3
open("/dev/nna", O_RDWR|O_NONBLOCK|O_LARGEFILE|O_CLOEXEC) = 4
nna_fd = 4
open("/dev/ion", O_RDONLY|O_LARGEFILE)  = 5
open("/dev/cedar_dev", O_RDONLY|O_LARGEFILE) = 6
AwG2dCreate!
open("/dev/g2d", O_RDWR|O_LARGEFILE)    = 7
Load model param and bin...
open("./model/yolov3_int8.param", O_RDONLY|O_LARGEFILE) = 8
open("./model/yolov3_int8.bin", O_RDONLY|O_LARGEFILE) = 8
data scale: 127.404358
conv 81_599 scale: 4.844433
conv 93_685 scale: 3.955845
conv 105_772 scale: 3.481591

Read and transform data...
open("./data/000000004765.bmp", O_RDONLY|O_LARGEFILE) = 8
image width: 612, image height: 612
resized image width: 416, resized image height: 416
padded image width: 416, padded image height: 416

Infer by IPU...
Infer sucessfully!

Dequantize...
0, output width 13, output height 13, output channel 255
1, output width 26, output height 26, output channel 255
2, output width 52, output height 52, output channel 255

Infer by CPU...
detect object num: 2
0, label 1, class person, prob 0.999197, x1 213, y1 138, x2 399, y2 375
1, label 38, class surfboard, prob 0.770885, x1 258, y1 322, x2 516, y2 410
[  774.230116] NNA: disable nna  irq
[  774.234289] NNA: nna_release irqnum = 312
+++ exited with 0 +++
root@(none):/mnt/extsd/YOLOV3_Demo#