Phase diagram of different possible configurations of a single-layer attention model learned with SGD.
Requirements are listed in requirements.txt
. You can install them using pip install -r requirements.txt
.
The file structure of this repository is as follows:
analytic/
contains the files for computing the populations loss.experiments/
: contains the hydra configuration files for running the experiments.simulate.py
: is the main script to run the experiments, and should be run with hydra, i.e.python simulate.py --config-name=experiments/[config_name] --multirun
example-training.ipynb
is a Jupyter notebook that shows how to run the experiments and plot the results.example-analytic.ipynb
is a Jupyter notebook that shows how to compute the populations loss and plot the results.