mehrdadmoradii/neural-network

A minimalistic deep learning framework developed from scratch for educational purposes.

Neural Network

Minimalistic deep learning framework developed from scratch, inspired by Keras and built upon the foundations of the CS231n assignments. Created for educational purposes, this lightweight framework provides a simple and intuitive API for understanding the intricacies of neural network construction and training.

Getting Started

from src.models import Sequential
from src.layers import Dense, ReLU
from src.optimizers import SGD
from src.losses import CrossEntropyLoss


model = Sequential([
    Dense(20, activation=ReLU()),
    Dense(20, activation=ReLU()),
    Dense(20, activation=ReLU()),
    Dense(3)
])

model.compile(
    loss=CrossEntropyLoss(),
    optimizer=SGD(learning_rate=5e-2),
)

history = model.fit(X_train, y_train, epochs=100, batch_size=32, x_val=X_test, y_val=y_test)

Models

• Sequential: Linear stack of layers for building the neural network.

Layers

• Dense: Fully connected layer.
• ReLU: Positive part function.

Optimizers

• SGD: Stochastic Gradient Descent optimizer.

Losses

• CrossEntropyLoss: Cross-entropy loss for classification tasks.

Contributing

Contributions are welcome! If you have ideas, suggestions, or find issues, please open a pull request. Your input is highly valued.

License

This project is licensed under the MIT License - see the LICENSE file for details.