Torchwnn: Weightless Neural Network

Torchwnn is a Python library for Weightless Neural Network (also known as RAM-based and N-tuple based Neural Network ).

Usage

Installation

First, install PyTorch using their installation instructions. Then, use the following command to install Torchwnn:

pip install torchwnn

Requirements: PyTorch and ucimlrepo to load datasets from UCI repository.

Quick Start

Iris Example

To quickly get started with Torchwnn, here's an example using the Iris dataset. Full training code is available in the examples/iris.py file.

import torch
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

from torchwnn.datasets.iris import Iris
from torchwnn.classifiers import Wisard
from torchwnn.encoding import Thermometer

# Use the GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Using {} device".format(device))

iris = Iris()
X = iris.features
X = torch.tensor(X.values).to(device)
y = list(iris.labels)
y = torch.tensor(y).squeeze().to(device)

bits_encoding = 20
encoding = Thermometer(bits_encoding).fit(X)    
X_bin = encoding.binarize(X).flatten(start_dim=1)

X_train, X_test, y_train, y_test = train_test_split(X_bin, y, test_size=0.3, random_state = 0)  

entry_size = X_train.shape[1]
tuple_size = 8
model = Wisard(entry_size, iris.num_classes, tuple_size)

with torch.no_grad():
    model.fit(X_train,y_train)
    predictions = model.predict(X_test)  
    acc = accuracy_score(predictions, y_test)
    print("Wisard: Accuracy = ", acc)

Examples

There are several examples in the repository.

Bleaching

import torch
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score

from torchwnn.datasets.iris import Iris
from torchwnn.classifiers import Wisard
from torchwnn.encoding import Thermometer

# Use the GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Using {} device".format(device))

iris = Iris()
X = iris.features
X = torch.tensor(X.values).to(device)
y = list(iris.labels)
y = torch.tensor(y).squeeze().to(device)

bits_encoding = 20
encoding = Thermometer(bits_encoding).fit(X)    
X_bin = encoding.binarize(X).flatten(start_dim=1)

X_train, X_test, y_train, y_test = train_test_split(X_bin, y, test_size=0.3, random_state = 0)  

entry_size = X_train.shape[1]
tuple_size = 8
model = Wisard(entry_size, iris.num_classes, tuple_size, bleaching=True)

with torch.no_grad():
    model.fit(X_train,y_train)
    predictions = model.predict(X_test)  
    acc = accuracy_score(predictions, y_test)
    print("Wisard: Accuracy = ", acc)
    
    # Applying bleaching
    model.fit_bleach(X_train,y_train)
    print("Selected bleach: ", model.bleach)
    predictions = model.predict(X_test)  
    acc = accuracy_score(predictions, y_test)
    print("Wisard with bleaching = ", model.bleach,": Accuracy = ", acc)

BloomWisard

Example using BloomWisard is available in the examples/iris_filters.py file.

Supported WNN models

Currently, the library supports the following WNN models:

• WiSARD - Neurons based on dictionary.
• BloomWiSARD - Neurons based on Bloom Filters. Reference: Weightless Neural Networks as Memory Segmented Bloom Filters

Supported techniques:

• B-bleaching - Bleaching based on binary search. Reference: B-bleaching : Agile Overtraining Avoidance in the WiSARD Weightless Neural Classifier.
  • WiSARD
  • BloomWiSARD