Lprnet custom dataset Output lprnet/avgpool1 UINT8, FCR(1x19x37)

General
1

Hello,

I trained lprnet with a custom dataset, compiled it to hef and it does not work out of the box with tappas licence_plate_recognition pipeline. The output differs, see below.
I followed the trainig guide available here

and compiled it with docker hailo_ai_sw_sui

Best_LPRNet_model.onnx
Best_LPRNet_model.onnx882×8438 243 KB

te_2025-01

Have I done something wrong or it is a compile problem ?
I will not use that pipeline but it would be nice to just work.

LPRNet compiled by Hailo
hailortcli parse-hef lprnet.hef
Architecture HEF was compiled for: HAILO8
Network group name: lprnet, Single Context
Network name: lprnet/lprnet
VStream infos:
Input lprnet/input_layer1 UINT8, NHWC(75x300x3)
Output lprnet/conv31 UINT8, FCR(5x19x11)

LPRNet compiled by me
hailortcli parse-hef lprnet.hef
Architecture HEF was compiled for: HAILO8
Network group name: lprnet, Single Context
Network name: lprnet/lprnet
VStream infos:
Input lprnet/input_layer1 UINT8, NHWC(75x300x3)
Output lprnet/avgpool1 UINT8, FCR(1x19x37)

2

Hi @razvans
Welcome to the Hailo community. I am guessing you also added letters of English alphabet to the set and hence you have 37(=11+26) in your output. If you can provide the hef file, we can give you a reference code using our PySDK.

3

Hello,

You are right, I used 36 chars, 10 numeric + 26 letters.

CHARS = [
‘0’, ‘1’, ‘2’, ‘3’, ‘4’, ‘5’, ‘6’, ‘7’, ‘8’, ‘9’,
‘A’, ‘B’, ‘C’, ‘D’, ‘E’, ‘F’, ‘G’, ‘H’, ‘I’, ‘J’,
‘K’, ‘L’, ‘M’, ‘N’, ‘O’, ‘P’, ‘Q’, ‘R’, ‘S’, ‘T’,
‘U’, ‘V’, ‘W’, ‘X’, ‘Y’, ‘Z’
]

You can find here the hef file and some syntetic data.

I am trying a new approach with yolov11 that traines much much faster.

Thank you for your help.

PS: the model is not fully trained yet. it is trained on about 10% of the data.