INSTALL.md

Install

It is recommended to use a virtual environment for Python, such as Anaconda.

Step-by-step

First, we create a virtual environment.

conda create --name diffusion python=3
conda activate diffusion

The second command changes the prompt from $ to (diffusion)$. All further commands are assumed to be executed in this new environment.

Next, we install the dependencies. We use the conda versions of the major packages as these usually work better. In addition, the conda version of numpy comes with support for the Intel MKL, offering extra speed up on supported systems.

conda install numpy scipy matplotlib

Optionally, one can install IPython and/or Jupyter for interactive use.

conda install ipython jupyter

Now, we can install the code in this repository.

pip install git+https://github.com/janniklasrose/diffusion-models.git

Alternatively, from the local cloned repository, we can execute:

pip install --editable .

The -e/--editable flag tells pip to install the package path. This leaves the source files editable.

Reproducible environment

We provide the file diffusion.yml to quickly create a tested conda environment. Simply execute:

conda env create --file diffusion.yml --name diffusion

The file was created using

conda env export --name diffusion > diffusion.yml

and then edited to remove the prefix key and move pip requirements to the requirements.txt file.