There are plenty of tutorial and worked examples on the Caffe home page: http://caffe.berkeleyvision.... See for instance the MNIST LeNet tutorial for training a classic convolutional neural net to recognize handwritten digits http://caffe.berkeleyvision... or the Filter Visualization notebook to see how to extract features and classify images with a modern ImageNet scaled CNN http://nbviewer.ipython.org.... For cuDNN acceleration install the library then install Caffe while turning on USE_CUDNN := 1 in the Caffe Makefile.config as detailed here: http://caffe.berkeleyvision...
I am not looking for Caffe examples. I am looking for cuDNN examples without reverse engineering Caffe. Installation of Caffe is a tricky task in itself to say the least.
Hi Alex! It's not clear to us what you are trying to accomplish. If you want to run DNNs on a cluster, then I suggest you stick with an existing framework, like Caffe. If you are a DNN framework developer, and you are having trouble with cuDNN, then please elaborate on the specific problem you are having (and I recommend you use https://devtalk.nvidia.com/... for that). If you just post sarcastic comments, then there's not much we can do to help.
Mark,
I am trying to see if we could use this library within our machine learning cluster. We are not using caffe, we have our own CUDA implementation of a neural network.
For this task I clearly need some documentation which goes slightly beyond the function names. Some simple working examples would be extremely helpful.
Alex
Thanks for the great article and useful information. For some of your readers who are new to this fascinating science:
Hi Alex, we hope to include code samples with future releases of cuDNN. For now, feel free to ask specific questions here or on devtalk.nvidia.com.
A small error: "State-of-the-art DNNs and CNNs can have can have"
I am trying to build mnistCUDNN for windows. I added some include directories, so it can compile everything, but it won't link. Lots of LNK2019 errors:
mnistCUDNN.obj : error LNK2019: unresolved external symbol _cudnnSetLRNDescriptor@32 referenced in function "public: void __thiscall network_t<struct half1="">::lrnForward(int,int,int,int,struct half1 *,struct half1 * *)" (?lrnForward@?$network_t@Uhalf1@@@@QAEXHHHHPAUhalf1@@PAPAU2@@Z)
Caffe with cudnn is 1.6x faster than that without cudnn, when batch-size was 64. However, when I set batch-size to 1, caffe with cudnn is 2x slower. Dose cudnn have heavy overhead?
Any help would be much appreciated!
Hi Feng, what GPU(s) are you testing on?
Hi, Mark, it's Tesla K40m and CUDA 6.5.
If you are interested in NVIDIA GPU systems with cuDNN, Digits, Caffe etc preinstalled there are some good turn key solutions here: http://exxactcorp.com/index...
Thank you!