Deepy

Deepy is a minimalistic deep learning framework built from scratch to demonstrate the core concepts of neural networks and backpropagation. It was developed as part of the EPFL course EE-559 Deep Learning in Spring 2020.

Authors

• Jerome van der Elst
• Simon Janssen
• Christoph Tietz

Features

• Custom implementation of neural network modules including:
  • Linear layers
  • Activation functions (ReLU, LeakyReLU, Tanh)
• Manual backpropagation using the chain rule.
• Data handling and splitting for training, validation, and testing.
• Gradient descent optimization.
• Example task: Circle classification.

Requirements

• Python 3.x
• PyTorch for tensor operations.
• Matplotlib (optional) for visualization.

Usage

1. Clone the repository:

   git clone https://github.com/jeromevde/Deepy.git
   cd Deepy
   pip install torch matplotlib

2. run the notebook

Example

inputs, targets = create_disc(20000)
model = Sequential(
    Linear(input_dim=2, output_dim=25, epsilon=1),
    ReLu(),
    Linear(input_dim=25, output_dim=25, epsilon=1),
    ReLu(),
    Linear(input_dim=25, output_dim=25, epsilon=1),
    Tanh(),
    Linear(input_dim=25, output_dim=2, epsilon=1),
    Tanh()
)
tr_error, v_error = train(model, tr_i, binaryToOneHot(tr_t), LEARNING_RATE, EPOCHS, v_i, binaryToOneHot(v_t))
draw_circle(t_i, t_t_pred, "Predicted data")

Results