Exercise on Normalizing Flows, as conducted in the 2nd Terascale School of Machine Learning 2021 (Link) by S. Diefenbacher, G. Kasieczka, T. Quadfasel and M. Sommerhalder
There are two ways how these exercises can be performed:
- they can be run on locally using the jupyter notebooks provided in this repo directly
- they can be run in google colab
The colab notebook is available here. Note that the link only provides read-only access, so to be able to work on the exercise, one will need to create a copy of the notebook and save it (e.g. in the user's google drive).
The jupyter notebooks are also available to download and run locally in this repository. We recommend using a separate anaconda environment for this. The notebooks have been tested locally with the following packages:
python 3.8.8tensorflow 2.4.1tensorflow-probability 0.12.1numpy 1.19.5matplotlib 3.3.4seaborn 0.11.1pandas 1.2.3scikit-learn 0.24.1tables 3.6.1cudatoolkit 11.0.221cudnn 8.0.4
The cudatoolkit and cudnn packages are only needed if GPU ressources are available. It is highly recommended to run the notebooks using a GPU.
In order to run the notebooks inside an anaconda environment, it has to be installed first. Detailed instructions how to install anaconda can be found here.
After anaconda is set up, a new environment containing all the required packages can easily be created from the flow_environment.yml yaml file available in this repository. To do this, simply clone this repository, browse to the respective directory and issue the following command:
conda env create -n flow_env -f flow_environment.yml
This will create a new anaconda environment called flow_env that contains all the required packages already.
We also provided jupyter notebooks with the solutions to the tasks. In google colab, these can be accessed at this link.
They can also be accesed from this repository. The solutions are located in a separate solutions branch.