neural-network

---

neural-network is a Python package on TestPyPi that provides a Multi-Layer Perceptron (MLP) framework built using only NumPy. The framework supports Gradient Descent, Momentum, RMSProp, Adam optimizers.

Table of Contents

1. Installation
   • Dependencies
   • User installation
2. Simple Usage
   • Designing the Model Architecture
   • Training the Model
   • Making predictions
3. Beyond the Framework
   • Activation functions
   • Loss functions
   • 2D Decision Boundary
4. License

 

Installation

Dependencies

python>= 3.8
numpy>= 1.22.1
matplotlib>= 3.5.1 

User installation

You can install neural-network using pip:

pip install -i https://test.pypi.org/simple/neural-network

Simple Usage

Designing the Model Architecture

To define your MLP model, you need to specify the number of layers, and the number of neurons in each one.
Unless you want to manually set up the parameters, the size of the input layer is not needed, as it will be automatically determined in the initial training process.

from neural_network import NeuralNetwork
model = NeuralNetwork(neurons=[64, 120, 1])

In this example, we have a four-layer neural network containing auto-defined input neurons, first hidden layer with 64 neurons, second hidden layer with 120 neurons, and one output neuron.

Training the Model

To train the model, you need to provide the input data and the corresponding target (or label) data.

X = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([[0], [1], [1], [0]])

model.fit(X, y, epochs=1000, learning_rate=0.1, optimizer='adam')

When training the model without setting the activation functions or/and the loss functions, the framework will automatically do the job for you. It will initialize the parameters and the functions according to the type of model (regression or classification) and its architecture.

Making predictions

Once the model has been trained, you can use it to make predictions by simple call predict method.

predictions = model.predict(X)

Beyond the Framework

Apart from the neural network framework, the package also provides:

Activation functions

Sigmoid function	sigmoid()
Hyperbolic tangent function	tanh()
Rectified linear unit	relu()
Leaky Rectified linear unit	leaky_relu()
Softmax function	softmax()
Gaussian error linear unit	gelu()

All above functions have 2 parameters:

• x: The input values. Even though some functions can accept numeric primitive data type, it is advised to use NumPy array.
• derivative: A boolean value indicating whether the function computes the derivative on input x. Default is False.

Loss functions

Logistic loss function	log_loss()
Cross-entropy loss function	cross_entropy_loss()
Quadratic loss function	quadratic_loss()

All above functions have 3 parameters:

• y_pred: Predicted labels. It must be a 2D NumPy array and have the same size as y_true.
• y_true: True labels. It must be a 2D NumPy array and have the same size as y_pred.
• derivative: A boolean value indicating whether the function computes the derivative. Default is False.

2D Decision Boundary

This utility function is used for illustrative purpose. It takes a trained binary classification model, a 2D NumPy input data with 2 attributes, and the corresponding binary label data as input. The function then will plot a 2D decision boundary based on the prediction of the model.
The input model is not necessarily an instance of NeuralNetwork, but it must have predict method that accepts a 2D NumPy array as input.

plot_decision_boundary(model, train_x, train_y)

License

This project is licensed under the MIT License, as found in the LICENSE file.