cuDNN: Need training (backward propagation) sample code

Accelerated Computing GPU-Accelerated Libraries
1

I have download cudnn-6.5-win-v2-rc2 sample code and success implement(convert) on CUDA6.5 + VS2013(with windows form) + OpenCV-3.0-alpha, it’s can test MNIST sample without Caffe LIB correctly.

I hope no need Caffe then execute training, however, NVIDIA no provide training (backward propagation) sample code, is there someone can help me ?

I think maybe need like as cudnnActivationBackward() cudnnSoftmaxBackward() etc. function.

2

Who can help me?

3

MNIST provides great data for training/testing - the data is grayscale so scale the data by 0.00390625 to convert the byte values into a 1.0 to 0.0 range. I found it best to place the MNIST data into a SQL database for speed.

The data is in a form: channels = 1, height = 28, width = 28.

For code you might check out the RC3. Hope this helps.

-Gator

4

Actually I cant’ find sample code that demonstrates how to use cudnnActivationBackward and related backward calculation functions.
The sample code downloadable from https://developer.nvidia.com/rdp/assets/cudnn-code-samples-v2 only contains forward calculation in file mnistCUDNN.cpp.

Do you have sample code for backward calculation?

Thanks – Denis