CIFAR-10 Image Classification — PyTorch CNN

A custom Convolutional Neural Network (CNN) built from scratch in PyTorch to classify images from the CIFAR-10 dataset across 10 categories (planes, cars, birds, cats, deer, dogs, frogs, horses, ships, trucks).

Result

Test Accuracy: 86.22%

Baseline for random guessing on CIFAR-10 is 10%. A typical introductory CNN achieves ~70-75%. This model achieves 86.22% without any pretrained weights.

---

Architecture

4 convolutional blocks with progressively increasing filter depth (32 → 64 → 128 → 256), followed by a fully connected classifier.

Input (3×32×32)
  → Conv Block 1: Conv2d(3, 32)  → ReLU → MaxPool → BatchNorm
  → Conv Block 2: Conv2d(32, 64) → ReLU → MaxPool → BatchNorm
  → Conv Block 3: Conv2d(64,128) → ReLU → MaxPool → BatchNorm
  → Conv Block 4: Conv2d(128,256)→ ReLU → MaxPool → BatchNorm
  → Flatten
  → Dropout(0.3)
  → Linear(1024, 256) → ReLU
  → Linear(256, 10)
Output: 10-class softmax

---

Key Design Choices

Choice	Reason
Batch normalisation after each pooling layer	Stabilises training, allows higher learning rates
Dropout (0.3) before classifier	Reduces overfitting on training set
Data augmentation (crops, flips, colour jitter)	Applied to training set only to improve generalisation
Train/val split with fixed seed (42)	Ensures reproducible evaluation
Adam optimiser, lr=0.001	Adaptive learning rate, fast convergence
CrossEntropyLoss	Standard for multi-class classification

---

Training Details

Parameter	Value
Epochs	50
Batch size	32
Optimiser	Adam (lr=0.001)
Train/Val split	80/20 (40,000/10,000)
Device	CPU

---

How to Run

Requirements

pip install torch torchvision

Train the model

python train.py

This will download the CIFAR-10 dataset automatically, train for 50 epochs, save the model weights to IMLO_Coursework.pth, and print test accuracy.

Test a saved model

python test.py

---

Technologies

• Python
• PyTorch
• torchvision