FAQs

How to install a CUDA version of kaldifeat from source

You need to first install a CUDA version of PyTorch and then install kaldifeat.

Note

You can use a CUDA version of kaldifeat on machines with no GPUs.

How to install a CPU version of kaldifeat from source

You need to first install a CPU version of PyTorch and then install kaldifeat.

How to fix Caffe2: Cannot find cuDNN library

Your installed Caffe2 version uses cuDNN but I cannot find the cuDNN
libraries.  Please set the proper cuDNN prefixes and / or install cuDNN.

You will have such an error when you want to install a CUDA version of kaldifeat by pip install kaldifeat or from source.

You need to first install cuDNN. Assume you have installed cuDNN to the path /path/to/cudnn. You can fix the error by using one of the following commands.

  1. Fix for installation using pip install

export KALDIFEAT_CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DCUDNN_LIBRARY_PATH=/path/to/cudnn/lib/libcudnn.so -DCUDNN_INCLUDE_PATH=/path/to/cudnn/include"
pip install --verbose kaldifeat

  1. Fix for installation from source

mkdir /some/path
git clone https://github.com/csukuangfj/kaldifeat.git
cd kaldifeat
export KALDIFEAT_CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DCUDNN_LIBRARY_PATH=/path/to/cudnn/lib/libcudnn.so -DCUDNN_INCLUDE_PATH=/path/to/cudnn/include"
python setup.py install