tensorflow - Convolution neural networks -- all feature maps are black(pixel value is 0) -


i doing project maps trained cnns on zynq soc. trained lenet in tensorflow , extracted weights , biases. far observed, value of weights close 0, none of them larger 1. input data of lenet gray scale image , pixel value 0 255.

when tried 2-d convolution between input image , kernel (trained weights), output feature maps black image since convolution result close 0. takes relu layer account. shown in picture below, value of weight in kernel , feature maps should value between 0 255 according brightness.

i wonder why got black(0 pixel value) feature maps?

enter image description here

the inputs normalized before doing convolution , feature maps created normalized create images. create similar images have find out how inputs normalized in network created weights using , normalize inputs in same way, min-max normalization on feature map , 0-255 range.


Comments

Popular posts from this blog

android - InAppBilling registering BroadcastReceiver in AndroidManifest -

python Tkinter Capturing keyboard events save as one single string -

sql server - Why does Linq-to-SQL add unnecessary COUNT()? -