search

doonny / PipeCNN

An OpenCL-based FPGA Accelerator for Convolutional Neural Networks
Apache License 2.0
1.22k stars 369 forks source link

LRN piece-wise linear approximation function coefficients #94

Closed Mario1112 closed 5 years ago

Mario1112 commented 5 years ago

Hello,

I read your paper and also studied the code here [1] and here [2] but I am wondering how the values of h_inv, coef0 and coef1 are calculated. From the paper I can understand x_sample is derived from the 2^(-n) value with n = 2, but using only this subdivision I can at best get an maximum error of 8.432119111524738% as opposed to your value of a maximum error of 0.5%. Can you elaborate on how the values for h_inv, coef0 and coef1 are computed? Thank you.

[1] https://github.com/doonny/PipeCNN/blob/e63a9b7a7dc55e651193a85c474086f3e08f0835/project/device/conv_pipe.cl#L772 [2] https://github.com/doonny/PipeCNN/blob/e63a9b7a7dc55e651193a85c474086f3e08f0835/project/device/hw_param.cl#L80-L83

doonny commented 5 years ago

Hi, pls refer to the paper for more detailed informations.

Mario1112 commented 5 years ago

Thank your for your fast reply. Unfortunately I cannot find this information in the paper. In section 2.B. part 4) "LRN Kernel" you describe

we divide the function evaluation range by using power of 2^-n, where n is an integer that controls the accuracy.

and

In AlexNet implementation, a maximum approximation error of 0.5% is achieved by setting n = 2.

Can you help me understand in which section the information is given? I could not find it in the paper. Thank you for your time and effort.