🧠 CNN Image Classification on CIFAR-10 and CIFAR-100

This project explores deep learning techniques for image classification using PyTorch. A custom Convolutional Neural Network (CNN) is first built and trained from scratch on CIFAR-10. The learned features from this model are then transferred to CIFAR-100 to classify a larger set of categories using transfer learning.

---

📁 Datasets

• CIFAR-10: 60,000 32x32 color images across 10 balanced classes
• CIFAR-100: 60,000 32x32 color images across 100 fine-grained classes
  (Both split into 50,000 train and 10,000 test images)

---

💠 Project Highlights

• ✅ Built CNN from scratch for CIFAR-10

  • 6-level convolutional structure with progressive filters: 16 → 32 → 64 → 128 → 256
  • Experiments with deeper models (starting from 32 filters up to 512)
  • Explore different activation functions: ReLU, LEakyRelu, ELU
  • Use of Batch Normalization, and Dropout for training stability and regularization
  • MaxPooling layers to downsample and reduce spatial dimensions

• 🔄 Transfer Learning on CIFAR-100

  • Used trained CIFAR-10 model weights
  • Retained convolutional layers, fine-tuned classification head for 100 classes
  • Enabled faster convergence and better generalization on the more complex dataset

• 📊 Training Strategies

  • Data normalization and augmentation
  • Early stopping based on validation accuracy
  • Accuracy used as the main performance metric (due to balanced datasets)

---

📊 Results

Dataset	Model Strategy	Test Accuracy
CIFAR-10	Custom CNN (trained from scratch)	85.6%
CIFAR-100	Transfer Learning from CIFAR-10	67.1%

---

📦 Libraries

• PyTorch
• torchvision
• torch
• torcham
• numpy
• matplotlib
• scikit-learn

---