DeepCIFAR: Image Classification with ResNet18

Overview

DeepCIFAR is a deep learning project for image classification using the CIFAR-10 dataset.
It leverages PyTorch and transfer learning with a pretrained ResNet18 to classify images into 10 categories:

airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck

This project demonstrates a complete end-to-end pipeline:

• Data preprocessing and augmentation
• Training, validation, and test evaluation
• Fine-tuning a deep CNN (ResNet18)
• Visualization of results (curves, confusion matrix, predictions)
• Inference on custom images

---

Dataset

• CIFAR-10: 60,000 color images (32x32 pixels) across 10 classes
• Split: 50,000 training + 10,000 test images
• During training, images were resized (128–160 px) and augmented with:
  • Random crop & flip
  • Color jitter
  • Normalization

---

Model

• ResNet18 pretrained on ImageNet

• Last fully-connected layer replaced with 10 output classes

• Two training modes:

  1. Frozen backbone (only last layer trained → faster, good baseline)
  2. Fine-tuning entire network (higher accuracy)

• Optimizer: AdamW

• Scheduler: CosineAnnealingLR

• Loss: CrossEntropyLoss with optional label smoothing

---

Training Setup

• Device: Apple Silicon GPU (MPS)
• Epochs: 6
• Batch size: 96
• Image size: 160
• Mixed precision: off (for MPS stability)

---

Results

• Validation Accuracy: 96.62%
• Test Accuracy: 96.02%

Curves

Confusion Matrix

Sample Predictions

---

Inference (Single Image)

Run predictions on a custom image:

python3 inference.py --image samples/cat1.jpg --ckpt results/best.pt --device mps

Example output:

('cat', 0.7955046892166138) ('airplane', 0.029839668422937393) ('automobile', 0.02575363218784332)]

---

Inference Demo (Batch + Visualization)

You can also test multiple images at once (placed in samples/) and generate a grid visualization + text report:

python3 inference_demo.py

• The script will run inference on all images in samples/ plus 3 random CIFAR-10 test images.
• Outputs are saved in results/sample_preds_custom.png and results/sample_preds_custom.txt.
• Note: All custom photos used in the demo (airplane1.jpg, cat1.jpg, dog1.jpg, etc.) were downloaded from Unsplash.

---

Requirements

Install dependencies (Mac/CPU/MPS version):

pip install torch torchvision torchaudio
pip install matplotlib scikit-learn tqdm

Or simply:

pip install -r requirements.txt

---

Future Work

• Extend to CIFAR-100 (100 classes)
• Apply Test-Time Augmentation (TTA) for more robust predictions
• Add a GUI application for real-time inference
• Explore other architectures: ResNet34, EfficientNet

---

👤 Author

Developed by Nikan as part of Deep Learning projects for Co-op applications.