the differences between rknn v1.3 vs lower rknn version

Hi,
I am trying to run a rknn model in lower version of rknn.
In rknn version 1.3.0, it works perfectly. Output is a array with shape = (1,320,320,3)
In rknn version 1.2.0 or 0.9.9: output shape = (1, 307200).

I tried to reshape the output with shape = (1, 307200) to has same output of rknnv1.3.0. But the result is totally different.

1. out_boxes = rknn.inference(inputs=[image])
   
2. out_boxes = out_boxes[0].reshape(320, 320, 3)

复制代码
Please give me advices, how should I reshape my rknn 1.2 outpur? Thanks

Version 1.3.0 fixes many issues and add pytorch & mxnet support .
In the older version, you can reshape output by yourself, and be careful about NHWC or NCHW.
Maybe you can try reshape with (3, 320, 320) and check again.

jefferyzhang 发表于 2020-2-28 10:25
Version 1.3.0 fixes many issues and add pytorch & mxnet support .
In the older version, you can resh ...

Thanks. I tried again. reshape works in rknn v1.20, but not 0.9.9. How should I reshape in version 0.9.9? Thanks again!

nvnn 发表于 2020-2-28 11:47
Thanks. I tried again. reshape works in rknn v1.20, but not 0.9.9. How should I reshape in version ...

We don't suggest use old version.
I found there was something changed in output shape in version 1.1.0, I'm no sure about if need reshape in 0.9.9 which is too old.

jefferyzhang 发表于 2020-2-28 12:20
We don't suggest use old version.
I found there was something changed in output shape in version 1 ...

Thanks for your suggestion. By the way, What if I want to use rknn v1.3 for c++ api? Should I rebuild the rootfs.img? Or just get rknn-appi.so (for rknn v1.3.0) and link it into my program?

Update system by using apt upgrade/ dnf update, which will update librknn-api.so and npu-driver, both need 1.3.x.

jefferyzhang 发表于 2020-2-28 14:24
Update system by using apt upgrade/ dnf update, which will update librknn-api.so and npu-driver, bot ...

thank you very much