tensorflow-wavenet icon indicating copy to clipboard operation
tensorflow-wavenet copied to clipboard

Receptive field calculation?

Open ronentk opened this issue 7 years ago • 5 comments

Hi, In the receptive field calculation, why is the extra addition made here? From what I understand, for an example of filter_width=2, a dilations block of [1, 2, 4, 8, 16, 32, 64, 128, 256, 512] should have a receptive field of 1024, not 1025.

Thanks! Ronen

ronentk avatar Apr 24 '17 14:04 ronentk

same question here.

gooooloo avatar Apr 26 '17 06:04 gooooloo

Take a look at this link

rafaelvalle avatar Dec 31 '17 11:12 rafaelvalle

Still did not get whiy there is an extra addition

KingStorm avatar Mar 01 '18 08:03 KingStorm

There is one more convolution with filter width = 2 before the dilation stack. That adds 1 to the receptive field.

aschoenebeck avatar Sep 27 '18 09:09 aschoenebeck

Hey, i was confused by the calculation for a long time as well. And i searched for a lot of related information. Now, I finally understand. image image The first picture, above is the general rule of receptive field calculation The second figure is how to convert the kernel size of ordinary convolution to dilated convolution. Therefore, the formula for calculating the receptive field of stacking dilated convolution is(stride=1): JTVDbGVmdCU which is supposed to be the receptive field calculation so for the dilation=[1,2,4...,512]. receptive field = 1+(1+2+4+...+512) * (2-1)=1+1023=1024 not 1025 So the plus one in the equation is not 1024 plus one, but 1023 plus one

WelkinYang avatar Jan 27 '20 15:01 WelkinYang