FreeNN

Content

The repository is structured as follows. We only describe the most important files for a new user.

./
|-- experiments
|  |-- run_generate_architectures.py: Samples random architectures and compute FPT metrics.
|  |-- run_randomized_experiment.py: Trains several neural networks based on architectures described in json files.
|-- freenn: Contains the numerical routines aimed at computing FPT densities. 
|-- ipynb: Contains Python notebooks which demonstrate how the code works
|  |-- DemoFPT.ipynb: Illustrates the measure concentration in FPT.
|  |-- mc_vs_lilypads.ipynb: Benchmarks our method compared to Monte-Carlo sampling. Used to generate the benchmark figure (Fig. 1).
|  |-- LossStatistics.ipynb: Once all the experiments have been run, this notebook computes correlation statistics and gives the scatter plot (Fig. 2).
|-- tests: Unit tests
|-- README.md: This file

Note that the dependencies have been left to a bare minimum in order to run the package freenn. Running the experiments however requires the installation of torch, torchvision and click via:

$ pip install torch torchvision click

Installation

1. Create new virtual environment

$ python3 -m venv .venv

(Do sudo apt install python3-venv if needed)

3. Activate virtual environment

$ source .venv/bin/activate

4. Upgrade pip, wheel and setuptools

$ pip install --upgrade pip
$ pip install --upgrade setuptools
$ pip install wheel

5. Install the freenn package.

python setup.py develop

6. (Optional) In order to use Jupyter with this virtual environment .venv

pip install --user ipykernel
python -m ipykernel install --user --name=myenv

(see https://janakiev.com/blog/jupyter-virtual-envs/ for details)

Configuration

Nothing to do

Credits

Later