🔬 Deep Neural Network from Scratch with PyTorch – Breast Cancer Classification

This project implements a deep feedforward neural network from scratch using PyTorch, without using high-level modules like torch.nn.Module or torch.optim. The model classifies breast cancer cases (malignant vs. benign) using the Breast Cancer Wisconsin Diagnostic Dataset from Kaggle.

---

📌 Project Highlights

• ✅ Pure PyTorch (no nn.Module, no optim)
• ✅ Manual forward pass, backward pass, and parameter updates
• ✅ Supports any architecture (you define layer sizes)
• ✅ Trained on real-world breast cancer dataset
• ✅ Visualized loss & accuracy over epochs

---

🧠 Model Architecture

The neural network used in this project is fully connected, with the following architecture:

Input Layer : 30 features (after preprocessing) Hidden Layer 1 : 4 neurons + ReLU Hidden Layer 2 : 3 neurons + ReLU Hidden Layer 3 : 2 neurons + ReLU Output Layer : 1 neuron + Sigmoid

yaml Copy Edit

Loss function: Binary Cross-Entropy (BCE)
Activation: ReLU (hidden), Sigmoid (output)
Optimizer: Manual gradient descent using loss.backward() and weight updates

---

📊 Dataset Overview

• Source: Kaggle – Breast Cancer Wisconsin (Diagnostic) Data Set
• Link: https://www.kaggle.com/datasets/uciml/breast-cancer-wisconsin-data

Preprocessing:

• ✅ Dropped non-informative columns (id)
• ✅ Encoded target: M → 1, B → 0
• ✅ Scaled features using StandardScaler (mean=0, std=1)
• ✅ Converted data to PyTorch tensors (dtype=torch.float64)

---

📈 Training Results

• Final Accuracy: ~61%
• Final Loss: ~0.72
• Plotted loss vs epochs and accuracy vs epochs
• Trained for 100 epochs with learning rate 0.01

Note: Since this is a pure-from-scratch implementation, there’s room for optimization by tuning learning rate, epochs, adding regularization, or using better architectures.

---

📦 File Structure

├── deep_nn_from_scratch.py # All model, training, and evaluation logic ├── breast_cancer.csv # Dataset (from Kaggle) ├── README.md # This file ├── requirements.txt # Minimal dependencies

yaml Copy Edit

---

📌 Key PyTorch Concepts Practiced

• requires_grad=True
• .backward() for autograd
• Manual parameter updates with torch.no_grad()
• Broadcasting and shape alignment
• ReLU and Sigmoid activation functions
• Manual loss function implementation (Binary Cross-Entropy)

---

🚀 Getting Started

1. Install dependencies

pip install torch matplotlib pandas scikit-learn
2. Run the training script
bash
Copy
Edit
python deep_nn_from_scratch.py
📊 Sample Output
yaml
Copy
Edit
Epoch 0 | Loss: 0.6932 | Accuracy: 0.56
Epoch 10 | Loss: 0.6821 | Accuracy: 0.61
...
Epoch 99 | Loss: 0.7234 | Accuracy: 0.61