ScratchANNet

scratchANNet is a minimalist deep learning framework written entirely from scratch in C++20 using module files (.ixx) — no external ML libraries or frameworks involved.

It includes a fully custom neural network engine with:

• Backpropagation
• Stratified dataset sampling
• ReLU and Softmax activations
• Gradient clipping
• Cross-entropy loss
• L2 regularization
• Performance metrics: Accuracy, Precision, Recall, F1-Score

---

🌱 Motivation

The goal of scratchANNet is to build a fully interpretable and performant neural network engine from first principles using modern C++20 features (modules, clean OOP, etc.).

---

🧠 Architecture Diagram

---

🚀 Features

• ✅ C++20 module-based architecture (export module ANNet;)
• ✅ Fully custom neuron, layer, and network classes
• ✅ Dynamic training with gradient clipping & softmax normalization
• ✅ Stratified sampling of datasets
• ✅ Real-time evaluation with confusion matrix

---

🧠 Model Architecture

Example architecture used for Iris dataset:

Input Layer:     4 features
Hidden Layer 1:  4 neurons (ReLU)
Hidden Layer 2:  4 neurons (ReLU)
Output Layer:    3 neurons (Softmax)

🏋️ Training

• Optimizer: Custom SGD with gradient clipping and L2 regularization
• Epochs: 200
• Learning rate: 0.005
• Accuracy: 96% on training data

---

📊 Evaluation Metrics

• Accuracy
• Precision (per class)
• Recall (per class)
• F1 Score (per class)

Also prints a full confusion matrix for test data.

---

🧪 Dataset: Iris Flower Classification

• Dataset: data.csv
• Format: 4 numerical features + 1 string class label
• Labels: setosa, versicolor, virginica
• Preprocessing: One-hot encoding, stratified train/test split

---

📁 Example Usage

// Load dataset
auto dataset = readCSVToDataset("data.csv", 4, 3);
auto [trainData, testData] = stratifiedSampling(dataset);

// Initialize and build network
Network model;
model.InputLayer(trainData);
model.HiddenLayer(4);
model.HiddenLayer(4);
model.OutputLayer(3);

// Train
model.train(0.005, 0.001, 0.5, 200);

// Evaluate
model.test(testData);
model.evaluation();
model.showMatrix();

// Predict
vector<double> input = {6.9, 3.2, 5.5, 2.2};
int prediction = model.predict(input);

---

🖥️ Sample Output

---

🛠 Build Instructions

• Requires C++20 compatible compiler (MSVC or Clang with module support)
• Compile with /std:c++20 and module-aware flags

---

📌 Highlights

• No dependencies: only STL
• Uses std::vector, std::random, and modern idioms
• Debug-friendly and readable
• Fully extensible for other datasets and architectures

---

💬 Author

Ansham Maurya Email: [email protected]

---

"Building intelligence from scratch helps us understand the soul of computation."

---

📜 License

This project is licensed under MIT - feel free to fork, use, extend, or contribute!