functionalfilet 0.5.2

Evolutionnary Neural Network Model with PyTorch

Functional Filet is a Python module for machine learning build on top of Pytorch.

This project explores a new method of constructing linear neural networks using evolutionary algorithms and graphs with gradient descent for adjustment of network weights.

More details in the Arxiv preprint An Artificial Neural Network Functionalized by Evolution (model based).

Installation

Dependencies

Functional-Filet requires:

• Python (>= 3.5)

• NumPy

• Pandas

• PyTorch (>= 1.0.0)

• Torchvision

• Matplotlib

• Networkx

Optionally, you need:

• Scikit-learn

• Seaborn

• Gym

---

Functional-Filet is stable only from version 0.5.2, any previous version corresponds to the development phase.

However, there are several possible optimizations, in particular on the restructuring of Torch tensors in Python which could be done in C++. For this, it is possible that there will be several code modifications in the future.

The documentation includes more detailed installation and examples instructions.

User installation

If you already have a working installation of numpy and pytorch, the easiest way to install scikit-learn is using pip:

python3 -m pip install functionalfilet

Development

We welcome new contributors of all experience levels.

Important links

• Official source code repo: https://github.com/fabienfrfr/functionalfilet

• Download releases: https://pypi.org/project/functionalfilet

Source code

You can check the latest sources with the command:

git clone https://github.com/fabienfrfr/functionalfilet.git

Utilization

Once installed, if we consider two variables feature X and label y already executed upstream of the code, here is a simple example of use in the case of a classification problem:

# package
from functionalfilet import model as ff
# init model
model = ff.FunctionalFilet()
# train
model.fit(X,y)
# test
y_pred = model.predict(X, index=seeder_idx)

Existing code

There is in the example directory of the git, several code to play with the learning parameters in simple cases. A brief summary is described at the top of each file:

python3 -m IPython
# universal approximation theorem
run example/uat_regression.py
# classification with overlapping and unbalance
run example/blob_classification.py
# reinforcment leaning with time dependancy
run example/gym_RL-CartPole-v0.py

Citation

If you take inspiration from my machine learning algorithm for a scientific publication, we would appreciate citations:

@article{furfaro2022artificial,
title={An Artificial Neural Network Functionalized by Evolution},
author={Furfaro, Fabien and Bar-Hen, Avner and Berthelot, Geoffroy},
journal={arXiv preprint arXiv:2205.10118},
year={2022}
}

Attribution required : Fabien Furfaro (CC 4.0 BY NC ND SA)