Single-Dataset Applications of Typhon and Parallel Transfer to Mitigate Overfitting and Improve Sample Efficiency Across Deep Learning

This repository contains the code of the work done during my Master thesis at the University of Fribourg. This is a direct following-up and extension of the original Typhon framework. The details of the algorithms and the implementation can be read in the full thesis, 2024_MSc_Christophe_Broillet.pdf. In its current state, Typhon can be applied to the following tasks:

1. Classification (in standard Typhon, Two Levels Typhon, or Ultra Typhon mode, see 2024_MSc_Christophe_Broillet.pdf for more details)
2. Segmentation
3. Autoencoding

Description of the files and folders

The folder experiments contains all the files for the different desired experiments. An example file basefile_0.py is given. These are the files that need to be called in order to launch the experiments (see the Running experiments section below).

The folder architectures contains all blocks composing the different architectures for the different tasks. The user can of course add customized ones. See more details on the architectures in the file architectures/README.md

The file experiment.py is the generic experiment script that launches the main experiment.

The file architecture_loader.py handles the loading of the neural network architecture given as parameters.

The file loop_loader.py takes care of loading the data in various ways depending on the tasks, and implements the loop loader which continuously gives batch for all datasets.

The file metrics.py implements the different metrics and custom loss functions for the experiments.

The file results.ipynb is used to check the results of the experiments.

The file typhon.py is the core file containing the algorithmic parts of Typhon.

The file typhon_model.py handles the Typhon model, and the relations between the feature extractor and the decision makers.

The file utils.py contains some utilitary and miscellaneous functions.

Running experiments

To run the code and the experiments, a Conda enviroment must first be setup using

conda env create -f environment.yml

after cloning this repository.

One can then activate the environment using

conda activate typhon

The example experiment can finally be run using the following command:

python experiments/basefile_0.py

Hyperparameters and many other options can be modified in the aformentioned file.

N.B. that you can do a shorter run by simply adding your OS name by going into the Experiment class implementation in the experiment.py file (line 32).

Finally one can check the results of the experiments by running the results.ipynb Jupyter Notebook using

jupyter notebook