search

yonseicasl / Nebula

Nebula: Lightweight Neural Network Benchmarks
https://casl.yonsei.ac.kr/nebula
Other
11 stars 2 forks source link

Questions about cudnnConvolutionForward in convolutional_layer.cu #1

Closed hangxu124 closed 3 years ago

hangxu124 commented 4 years ago

Hi, I just found that in your conv layer in: https://github.com/yonsei-icsl/nebula/blob/7a46adb8cfd6aa819a140ab2f912210aa01e2f1b/models/layers/convolutional_layer.cu#L32 you call the function "cudnnConvolutionForward" and use alpha = beta =1... I found some other code, they always set alpha=1 and beta=0, why you set beta =1 here? Thanks a lot !

bogilkim commented 3 years ago

According to the cuDNN documentation, beta is used to scale the output tensor according to the cuDNN documentation So we set beta = 1.