Feedforward Neural Network

A minimal feedforward neural network for multiclass classification, implemented in PyTorch and demonstrated on the Iris dataset.

---

Overview

This project implements a custom feedforward NN from scratch (using PyTorch tensors and autograd) to classify Iris flowers into three species: setosa, versicolor, and virginica. All details—architecture, data loading, training loop, and evaluation—are in the notebook.

---

Features

• Custom architecture: Configurable layers and widths (no nn.Linear stack; uses nn.Parameter and manual forward pass).
• Iris dataset: Loaded via scikit-learn; 90% train / 10% test split.
• Training: SGD optimizer, cross-entropy loss, ReLU in hidden layers, softmax at output (via CrossEntropyLoss).
• Metrics: Training and test loss plus accuracy logged every 10 epochs.

---

Requirements

• Python 3.x
• PyTorch
• scikit-learn
• NumPy
• Matplotlib

Install with:

pip install torch scikit-learn numpy matplotlib

---

Quick Start

1. Clone the repository:

   git clone https://github.com/AugusGuarna/Feedforward_NN.git
   cd Feedforward_NN

2. Open and run the notebook:

   • Locally: Open NN_multiclass_pred.ipynb in Jupyter or VS Code.
   • Online: Use the “Open in Colab” badge at the top of the notebook.

No extra setup or data download is required; the notebook loads the Iris dataset automatically.

---

Project Structure

Feedforward_NN/
├── README.md
└── NN_multiclass_pred.ipynb   # Full implementation and walkthrough

---

Notebook Contents

Section	Description
Architecture	Input (4) → hidden (6, 8, 10) → output (3); ReLU and softmax.
Data	Iris loaded, encoded, split into train/test, wrapped in DataLoaders.
Model	FeedForwardNeuralNetwork class and parameter check.
Training	Loss, optimizer, 200-epoch loop with train/test loss and accuracy.
Results	Short discussion of performance and data size.

---

License

This project is open source. Use and adapt as needed.