Speed-up with cuDNN ---------------------------------------------------------------------------------------------------------------------- cuDNN is a NVIDIA library for GPU-accelerated deep learning. One of its highlights is the optimized convolution operations tuned for speed up on NVIDIA GPUs. Here are 3 simple steps by which you can install cuDNN and make it work with PDNN: 1. Upgrade your CUDA driver if the version is <6.5 2. Download the cuDNN package and decompress it to /path/to/cuDNN. Add the path to the following environment variabiles:     > export CPATH=/path/to/cuDNN:$CPATH > export LIBRARY_PATH=/path/to/cuDNN:$LIBRARY_PATH > export LD_LIBRARY_PATH=/path/to/cuDNN:$LD_LIBRARY_PATH 3. Now cuDNN will be used automatically when you run PDNN. To verify if cuDNN is truly being used, add " optimizer_including=cudnn" to THEANO_FLAGS Depending on your GPU cards and experiment setups, you can expect 2~3x speed up on CNNs training.