This project involves building a dynamic neural network from scratch using the numpy library, applied on the MNIST dataset. The neural network is designed as a dynamic class, featuring various activation functions and capable of handling different network architectures. It's an ideal resource for learning the fundamentals of neural networks.
- Dynamic Architecture: Create neural networks with customizable layers and neurons.
- Activation Functions: ReLU and Sigmoid functions for activation, including their derivatives for backpropagation.
- Softmax Output Layer: For multi-class classification.
- Batch Learning: Supports mini-batch and stochastic gradient descent (SGD) updates.
- Learning Rate Adjustment: Update learning rate using decay parameters.
- Epoch-Based Training: Set a specific number of epochs for training.
- Performance Metrics: Loss and accuracy calculations included.
- Uses the MNIST dataset, which comprises handwritten digit images.
- Python 3.x
- Numpy
pip install numpy
To use this neural network implementation:
- Clone the repository and navigate to the project directory.
- Import the neural network class from the provided script.
- Load the MNIST dataset.
- Create an instance of the neural network with desired parameters.
- Train the model using the
trainmethod with specified learning parameters. - Evaluate the model's performance using the
testmethod.
Here's an example of how to use the neural network:
from Task1 import MyNeuralNetwork
# Initialize the neural network
nn1 = MyNeuralNetwork(input_size=784, hidden_sizes=[4], output_size=10, activation_function='sigmoid', dropout=0.0)
# Train the neural network
losses, accuracies = nn1.train(X_train, y_train, learning_rate=1, epochs=25, batch_size=64, decay_rate=0.1, decay_step=10)
# Append last entry of train loss to dictionary for plotting
last_epoch_loss = losses[-1]
training_loss.append(last_epoch_loss)
# Test the neural network
loss, acc = nn1.test(X_test, y_test)
print(f'Test Loss: {loss}, Test Accuracy: {acc}')
