Training DNN using cuDNN

We can’t train the deep neural network by cuDNN? but only CAFFE?
cuDNN doesn’t provide the function to train the net?
if it doesn’t, it can classify using training data by only CAFFE? not Theano, or torch?
And also, i will be appreciating if you tell me how to train the net by CAFFE.
Thanks in advance.

This is an involved topic, not likely something that can be completely covered in a paragraph. cuDNN is a library (something like CUBLAS) that is designed to provide low-level operations that are useful for DNNs, including training. But does it provide a complete training framework? No it does not. Does it provide functions that can be used to accelerate training? Yes, it does.

AFAIK cuDNN v2 has been integrated into CAFFE, Theano, and Torch. From here:

“Several widely used machine learning frameworks support cuDNN acceleration, including UC Berkeley’s Caffe, NYU’s Torch, and University of Montreal’s Theano, as well as many other software projects. Please see the documentation for each framework/application on enabling cuDNN acceleration.”

Teaching you how to use CAFFE is not something I can cover here.

You can get started at the BVLC caffe site:

Thanks txbob.
Your answer contains all what i wanted to know.