Private Github repository for the course Fairness, Accountability, Confidentiality and Transparency in AI at the University of Amsterdam.
- Albert Harkema (12854794) ([email protected])
- Anna Langedijk (12297402) ([email protected])
- Christiaan van der Vlist (12876658) ([email protected])
- Hinrik Snær (12675326) ([email protected])
Our implementation is based on the tensorflow code in https://github.com/OscarcarLi/PrototypeDL. It extends the original implementation by using hierarchical prototypes.
First, (create and then) activate the correct environment:
[conda env create -f environment_prototype.yml]
source activate prototype
Then, run the code either from the IPython notebook, or by running run.py:
python run.py [--hier true] [--seed <int>] [--dir <directory name>] ...
This will run the code with default parameters/seed for reproduction. Additional parameters can be set according to their descriptions, run
python run.py --help
for more information about all the different parameters.
All of our non-wrapper code is included in the src/ directory. The basic modules are in src/network. They are combined within the src/model.py file, together with all the files necessary for training.
I based this environment on the environment provided by the DL course and added jupyter, matplotlib for easy IPython notebooks.
This includes an older version of pillow, see python-pillow/Pillow#4130. This issue is encountered on older versions of packages (for instance on Lisa).