Toybrick

标题: RKNN 转化Tensorflow模型python接口正常，C++接口错误 [打印本页]

作者: 润城 时间: 2020-9-15 18:02
标题: RKNN 转化Tensorflow模型python接口正常，C++接口错误
1、步骤：rknn将tensorflow的pb模型转换为*.rknn模型
2、步骤：加载到python接口运行模型可以正常运行：
-->loading model
loading model done
--> Init runtime environment
D RKNNAPI: ==============================================
D RKNNAPI: RKNN VERSION:
D RKNNAPI: API: 1.2.0 (1190a71 build: 2019-09-25 12:39:26)
D RKNNAPI: DRV: 1.3.0 (c4f8c23 build: 2019-11-25 10:39:29)
D RKNNAPI: ==============================================
done
(1, 19200) (1, 76800) (1, 307200)
rknn use time {} and fps: {} 13.186838022000302 7.583319051402974
(1, 19200) (1, 76800) (1, 307200)
rknn use time {} and fps: {} 13.250182258001587 7.547065999006278
(1, 19200) (1, 76800) (1, 307200)
rknn use time {} and fps: {} 13.264743579999049 7.538781235905894
3.同一个模型加载到C接口报错如下：
This system has 6 processor(s).
Bind NPU process(0) to CPU 4
Bind Display process to CPU 3
Bind VideoCapture process to CPU 2
D RKNNAPI: ==============================================
D RKNNAPI: RKNN VERSION:
D RKNNAPI: API: 1.3.3 (f20f0bd build: 2020-05-14 14:14:51)
D RKNNAPI: DRV: 1.3.0 (c4f8c23 build: 2019-11-25 10:39:29)
D RKNNAPI: ==============================================
The initialization of NPU Process 0 has been completed.
read camera / video start
frame cnt 40  idxinputImage 1
frame cnt 40  idxinputImage 2
frame cnt 40  idxinputImage 3
frame cnt 40  idxinputImage 4
frame cnt 40  idxinputImage 5
frame cnt 40  idxinputImage 6
frame cnt 40  idxinputImage 7
frame cnt 40  idxinputImage 8
frame cnt 40  idxinputImage 9
frame cnt 40  idxinputImage 10
frame cnt 40  idxinputImage 11
rknn use time: 0.157895
[32857.995978] Core dump to |/usr/lib/systemd/systemd-coredump 27067 1000 1000 11 1600164092 18446744073709551615 debian10.toybrick yolov3-512 pipe failed
Segmentation fault

作者: jefferyzhang 时间: 2020-9-15 20:41
你可以到置顶帖的QQ群里问下，那边能比较专业回答你的问题。
你这里可以看到的是API版本不一样，可以尝试c版本也用1.2.0的API试试，如果1.2.0能用1.3.0出问题，估计是rknn api的bug。

作者: 润城 时间: 2020-9-16 16:04
感谢，您的回复和提供的方式！

欢迎光临 Toybrick (https://t.rock-chips.com/)