Beta loss regularisation for CNN

Unimodal regularisation based on beta distribution for ordinal regression with Deep Learning

Algorithms included

This repo contains the code to run experiments using a deep neural network with an ordinal structure in the output layer and the regularised cross-entropy loss based on the beta distribution.

• Stick-breaking output structure.
• Poisson regularisation for cross-entropy loss function.
• Binomial regularisation for cross-entropy loss function.
• Exponential regularisation for cross-entropy loss function.
• Beta regularisation for cross-entropy loss function.

Installation

Dependencies

This repo basically requires:

• Python (>= 3.6.8)
• Keras (==2.2.4)
• numpy (>=1.17.2)
• opencv-python (>=4.1.2)
• pandas (>=0.23.4)
• scikit-image (>=0.15.0)
• scikit-learn (>=0.21.3)
• tensorflow (==1.13.1)

Compilation

To install the requirements, use: pip install -r requirements.txt

Development

Contributions are welcome. Pull requests are encouraged to be formatted according to PEP8, e.g., using yapf.

Usage

You can run all the experiments using the file run.sh.

Citation

The paper titled "Unimodal regularisation based on beta distribution for ordinal regression with Deep Learning" has been submitted to Pattern Recognition.

Contributors

Unimodal regularisation based on beta distribution for ordinal regression with Deep Learning

• Víctor Manuel Vargas (@victormvy)
• Pedro Antonio Gutiérrez (@pagutierrez)
• César Hervás-Martínez ([email protected])