Installing Norse

1. Installing Norse#

Norse requires Python version 3.9 or higher. If this is a problem, it is recommended to install Norse via Docker, as described below.

Norse builds on top of the PyTorch deep learning library, which is also our primary dependency. This has the benefit that your models are hardware accelerated, providing the prerequisites are met.

1.1. Required dependencies#

To install Norse, you need the following two dependencies:

  1. pip - the Python package manager

    • This is preinstalled in most Linux and Unix distros

    • Note that Norse requires Python >= 3.9

  2. torch - the deep learning accelerator

1.2. Installing Norse#

Note that the following commands require access to a command line interface.

1.2.1. Installing with pip#

pip install norse

1.2.2. Installing from Conda#

conda install -c norse norse

1.2.3. Installing with Docker#

docker pull quay.io/norse/norse

Alternatively, build from source:

docker build -t norse -f publish/Dockerfile --build-arg VERSION=1.0.0 .

Replace 1.0.0 with the desired version number.

1.2.4. Installing from source#

git clone https://github.com/norse/norse
cd norse
pip install -e .

1.3. Optional dependencies#

Some of the tasks require additional dependencies like Pytorch Lightning, Torchtext and Torchvision. We also offer support for Tensorboard to make it easier to visualise the training and introspect models.

1.4. Running Norse with Docker#

Docker creates a closed environment for you, which also means that the network and filesystem is isolated. To run with a mounted volume:

docker run -it --rm -v $(pwd):/workspace quay.io/norse/norse

1.4.1. GPU acceleration in Docker#

The Docker image includes CUDA support by default. To use GPU hardware acceleration, you need:

  1. An NVIDIA GPU

  2. The NVIDIA Container Toolkit installed

  3. The --gpus all flag when running the container

docker run -it --rm --gpus all quay.io/norse/norse python -c "import torch; print(torch.cuda.is_available())"

The --ipc=host flag is recommended for multiprocessing and DataLoader with multiple workers:

docker run -it --rm --gpus all --ipc=host -v $(pwd):/workspace quay.io/norse/norse

For more information on hardware acceleration, please refer to our page on {ref}(page-hardware).

1.5. Installation troubleshooting#

Below follows a list of known problems that hopefully address your problem. If not, please do not hesitate to reach out either by

Because we are relying on optimised C++ for some of the hotspots in the library, you will need to download and install CMake and PyTorch before you can install Norse. For that reason, we recommend following the PyTorch “Get Started” guide as the first step.

You might also have to install Python headers if you have not already done that. In Debian-based distros (like Ubuntu), this can be done by running apt install python3-dev.

1.5.1. Common problems#

ImportError: … /norse_op.so: undefined symbol: ZN2at5addmmERKNS_6TensorES2_S2_RKN3c106ScalarES6

This is likely because an existing, incompatible version of PyTorch is interfering with the installation. Try to

  1. remove PyTorch (pip uninstall torch),

  2. install torch by following the official guide at pytorch.org, and

  3. reinstall Norse with your preferred method

See this issue for more information: norse/norse#280

UnsatisfiableError: The following specifications were found to be incompatible with each other

This can happen during installation from Conda of Norse<=0.07. A solution is to add the conda-forge channel, like so: conda install -c norse -c conda-forge norse

Or, when creating an environment: conda create -d -c norse -n temptest2 python==3.9 norse==0.0.7

Contents