How to calculate outlen

[Tez01]

In npu_ssd_det.py, how is outlen calculated?

My image input is (3,240,320)

[Pinnh]

run : python priorbox.py

you will find the outlen at final line

I0613 09:18:48.739490 4141 net.cpp:761] Ignoring source layer input
I0613 09:18:48.739521 4141 net.cpp:761] Ignoring source layer data_input_0_split
I0613 09:18:48.745203 4141 net.cpp:761] Ignoring source layer mbox_conf_reshape
I0613 09:18:48.745225 4141 net.cpp:761] Ignoring source layer mbox_conf_softmax
I0613 09:18:48.745229 4141 net.cpp:761] Ignoring source layer mbox_conf_flatten
I0613 09:18:48.745232 4141 net.cpp:761] Ignoring source layer detection_out
[[[-0.07368421 -0.07368421 0.12631579 ... -0.3227241 0.7742414
1.3227241 ]
[ 0.1 0.1 0.2 ... 0.1 0.2
0.2 ]]] (1, 2, 7668)

[Pinnh]

You should replace with your model and run priorbox.py first, then run npu_ssd_det.py to get rknn model and the detection

[Tez01]

Ok thanks

Why does the output of rknn model contains only 2 array : loc and conf
Where is priorbox
I ran the same prototxt on caffe its output contains loc, conf and priorbox

[Pinnh]

I guess NPU or RKNN framework not support priorbox layer , it may decrease NPU performance ,this layer is hard to parallel and not need to compute everytime .

[Tez01]

Okay

[Source82]

Hello,

Please does rknn allow one to get output of multiple layers at the same time?

[Pinnh]

@Source82 I have several months not worked on rknn , you can try to use this project to test your problem. for SSD model， rknn can get output of loc and conf layers concurrently.