banalasaritha/afrnn

Advanced Fuzzy Relational Neural Network (AFRNN)

Implementation of paper presented at WILF2021

Important

The code in the repository is an update of the paper linked above. It contains some improvements and results that are better than those reported in the article. It is currently under development and research and any advances will be reported here

Differences

• The code is rewritten in PyTorch
• Captum library is used for exaplinability. Some layer can be slightly different and improved.
• Constraints over gradients are changed from MinMaxScaling to Clipping. This modification highly helps the network to learn and remove the heavy oscillation of the published results over the loss function.
• There are more layers such as Łukasiewicza/Yager ops.
• It is very experimental, it should be more pythonic

Differences on results

Table 1. Old results published in WILF2021

Table 2. New results, red cells represents tests not done on MNIST because it is a too easy dataset

Built-in Fuzzy Layers

• MaxMin2d
• MaxYager2d
• LeakyThreshold

Built-in Models

• Conv2d
• MaxMinAFRNN
• MaxMin2LAFRNN
• MaxLukAFRNN
• MaxLuk2LAFRNN
• MaxLuk3LAFRNN
• MaxLukMaxMinAFRNN
• MaxLearnableYagerAFRNN
• LeNet5AFRNN
• LeNet5SigAFRNN
• LeNet5_2AFRNN
• LeNet5_2SigAFRNN
• LeNet5
• LeNetLukAFRNN

Execution

For training, testing and explainability it is possible to use the following parameters:

Program_Parameters	Default	Usage
--name	''	Name the experiment
--weights	''	Path to weights to load the pretrained network
--resume_epoch	-1	Epoch from which resume training
--dataset	mnist	Dataset to use. It uses torchvision.datasets
--mode	train	Specifies the action to perform. Can be [train - test - explain - attack]
--path	''	Path where save the model and tests
--batch_size	8	Mini batch size
--model	fuzzy	Name of the model to run. All models are loaded from models.py. Insert new models there. Please use AFRNN as placeholder in new models
--epochs	1000	Number of epochs
--tnorm_p	1.	p of the t-norm (used in Yager)
--num_classes	10	Number of classes
--constraint	None	Constraint type. Can be [clip - minmax - sigmoid - gaussian - sinc2]
--log_every	10000	Number of iteration before log and save to file using tensorboard

Cite us

@inproceedings{di2021advanced,
  title={Advanced Fuzzy Relational Neural Network.},
  author={Di Nardo, Emanuel and Ciaramella, Angelo},
  booktitle={WILF},
  year={2021}
}