Project Overview
This project details the construction of a custom Convolutional Neural Network (CNN) from scratch using PyTorch 2.0+, designed to perform binary classification on hair textures (Curly vs. Straight). It emphasizes understanding foundational deep learning mechanics over using pre-trained models (Transfer Learning).
Key Technical Implementations
<ul>
<li>Architecture: A lightweight model featuring a single Convolutional block (32 filters, 3 x 3 kernel), Max Pooling (2 x 2), and a dense classification head. It outputs a single neuron using Sigmoid activation.</li>
<li>Loss Function: Utilizes <code>nn.BCELoss()</code> (Binary Cross Entropy), specifically chosen to pair with the Sigmoid output for binary probability.</li>
<li>Data Pipeline: Implements a strict separation between Training (heavy augmentation via RandomRotation and RandomResizedCrop) and Testing (deterministic resizing) to ensure valid benchmarking and prevent overfitting.</li>
<li>Engineering Standards: Focuses on reproducibility by setting fixed random seeds for PyTorch, NumPy, and CUDA, and tracks Median Training Accuracy to filter out training noise.</li>
</ul>
Core Takeaway
The project demonstrates that simple, well-architected CNNs can be highly effective for texture recognition without the computational overhead of Transformers, provided that data augmentation and layer mathematics are handled correctly.

Austin Deyan

A hands-on guide to building a binary image classifier using PyTorch 2.0+ without transfer learning. The implementation covers creating a custom CNN architecture with convolutional layers, max pooling, and sigmoid activation for hair texture classification. Key focus areas include proper data augmentation strategies, reproducibility through seed management, and understanding the mathematical foundations of binary cross-entropy loss paired with sigmoid outputs.

Back to Basics: Building a Hair Texture Classifier from Scratch with PyTorch