README.md

Machine Learning & Deep Learning Projects

This repository contains three projects developed in Python using Jupyter Notebooks.
They explore computer vision, classification, and regression tasks using PyTorch & FastAI.

---

📂 Projects Overview

1. Pet Image Classification (Single-Label Classification)

• Dataset: Oxford-IIIT Pet Dataset
• Techniques:
  • Transfer Learning (ResNet18, ResNet34).
  • Data augmentation (flips, lighting, zoom).
  • Mixed precision training (to_fp16()).
  • One-cycle learning rate policy.
• Concepts Explored:
  • FastAI DataBlock API.
  • Error analysis (confusion matrix, most confused classes).
  • Architecture & hyperparameter comparison.

---

2. PASCAL VOC 2007 Multi-Label Classification

• Dataset: PASCAL VOC 2007 (objects with multiple labels per image).
• Techniques:
  • Multi-category classification (MultiCategoryBlock).
  • Custom label parsing from CSV.
  • Custom splitter for train/validation sets.
  • Data augmentation with RandomResizedCrop.
• Concepts Explored:
  • Multi-label learning with one-hot encodings.
  • Interpreting predictions for multiple classes.
  • Training CNNs for complex image datasets.

---

3. Head Pose Estimation (Regression)

• Dataset: BIWI Head Pose Dataset.
• Techniques:
  • CNN regression with PointBlock.
  • Custom label extraction from calibration files.
  • Subject-based validation splitting.
  • Data normalization & augmentations.
• Concepts Explored:
  • Using CNNs for regression tasks (predicting coordinates).
  • Handling custom datasets.
  • Visual evaluation of predictions.

---

🛠️ Tools & Libraries

• Python
• PyTorch
• FastAI
• Scikit-learn, Pandas, Numpy
• Matplotlib, Seaborn

---

📊 Skills Demonstrated

• Single-label & multi-label classification.
• CNN regression for pose estimation.
• Data preprocessing & augmentation.
• Transfer learning & fine-tuning.
• Model evaluation & interpretation.
• Experimentation with architectures & hyperparameters.