for using readNET, should i have to train model with channel first input type?

yagorgy
1

Hello guys

I use two models Unet and googleNET V4 on python 3.8(openCV 4.2.2)

i want to make a dll file operating independently in device so i choose openCV

My question is for using openCV readNET function, should i have to train model on python channe first input type?

or using channel last input trained model on python is fine with openCV readNET?

thank you

1 answer

Sort by » oldest newest most voted

berak
32993 ●7 ●81 ●312

to my knowledge, it is impossible to train convolutional layers using images in NHWC order

both GoogleNet and UNet use NCHW channel order, so rather leave it like that

I trained images(512x512x3) on model unet and googleNETV4.(input shape : channel last) so i got a pb file(model structure and weights) from .h5 file.

can i use this pb file with openCV readNET function to predict? of course i load image as channel first type in openCV C++ env

please give me a solution!

yagorgy ( 2020-06-08 04:16:55 -0600 )edit

of course i load image as channel first type in openCV C++ env

no idea how you do that, opencv's samples use imread() / blobFromImage() for this purpose

berak ( 2020-06-08 04:30:11 -0600 )edit

1 follower

Asked: 2020-06-07 21:17:34 -0600

Seen: 188 times

Last updated: Jun 08 '20