README.md

metadata

title: ResNet-18 CIFAR-100 Classifier
emoji: 🖼️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false

ResNet-18 CIFAR-100 Image Classifier 🎯

A high-performance image classifier trained on CIFAR-100 dataset, achieving 77.18% test accuracy.

Model Details

• Architecture: ResNet-18 (Custom CIFAR-100 variant)
• Parameters: ~11 million
• Test Accuracy: 77.18%
• Train Accuracy: 98.25%
• Training Time: ~70 minutes on RTX 4070 Laptop GPU
• Dataset: CIFAR-100 (60,000 32×32 color images in 100 classes)

Training Configuration

Advanced Techniques Used

1. OneCycle Learning Rate Policy - Gradual warmup + extended annealing
2. Cutout Augmentation - Randomly masks 8×8 patches
3. Label Smoothing (0.1) - Prevents overconfident predictions
4. Gradient Clipping - Stabilizes training during high-LR phase
5. Data Augmentation - Random crops, horizontal flips, normalization

Hyperparameters

• Epochs: 100
• Batch Size: 128
• Max Learning Rate: 0.1
• Weight Decay: 5e-4
• Optimizer: SGD with momentum (0.9)

100 Classes

The model can classify images into these categories:

Animals (42 classes):

• Mammals: bear, beaver, camel, cattle, chimpanzee, dolphin, elephant, fox, hamster, kangaroo, leopard, lion, mouse, otter, porcupine, possum, rabbit, raccoon, seal, shrew, skunk, squirrel, tiger, whale, wolf
• Aquatic: aquarium_fish, crab, crocodile, flatfish, lobster, ray, shark, trout
• Insects/Small creatures: bee, beetle, butterfly, caterpillar, cockroach, snail, snake, spider, turtle, worm
• Reptiles: dinosaur, lizard

Vehicles (5 classes): bicycle, bus, motorcycle, pickup_truck, streetcar, tank, tractor, train, rocket

Household Items (11 classes): bed, chair, clock, couch, cup, keyboard, lamp, plate, table, telephone, television, wardrobe

Food (5 classes): apple, mushroom, orange, pear, sweet_pepper

Nature (13 classes):

• Trees: maple_tree, oak_tree, palm_tree, pine_tree, willow_tree
• Flowers: orchid, poppy, rose, sunflower, tulip
• Landscapes: cloud, forest, mountain, plain, road, sea

People (3 classes): baby, boy, girl, man, woman

Structures (5 classes): bridge, castle, house, road, skyscraper

Other (16 classes): aquarium_fish, bottle, bowl, bridge, can, castle, house, lawn_mower, rocket, sea, tank

Usage

On Hugging Face Spaces

Simply upload an image and get instant predictions with confidence scores!

Local Usage

import torch
from PIL import Image
from torchvision import transforms

# Load model
checkpoint = torch.load('checkpoints/resnet18_best.pth')
model = ResNet(BasicBlock, [2, 2, 2, 2], num_classes=100)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Preprocess image
transform = transforms.Compose([
    transforms.Resize((32, 32)),
    transforms.ToTensor(),
    transforms.Normalize((0.5071, 0.4867, 0.4408), 
                        (0.2675, 0.2565, 0.2761))
])

image = Image.open('your_image.jpg')
img_tensor = transform(image).unsqueeze(0)

# Predict
with torch.no_grad():
    output = model(img_tensor)
    pred = output.argmax(dim=1)

Performance Notes

• Best for: Small objects, centered subjects, simple backgrounds
• Optimized for: 32×32 images (will be automatically resized)
• Categories: Works best with the 100 CIFAR-100 classes listed above

Training Curves

The model showed steady improvement throughout training:

• Epochs 1-30: Warmup phase (13.89% → 59.38%)
• Epochs 31-60: Peak learning (59.38% → 63.88%)
• Epochs 61-100: Fine-tuning (63.88% → 77.18%)

Key Achievements

✅ Exceeded 73% target accuracy by 4.18%
✅ Stable training with no divergence
✅ Effective use of OneCycleLR scheduler
✅ Combined regularization techniques
✅ Fast training (~70 minutes for 100 epochs)

Repository

Full training code, logs, and checkpoints available at: GitHub Repository

Citation

If you use this model, please cite:

@misc{resnet18-cifar100,
  author = {Your Name},
  title = {ResNet-18 CIFAR-100 Image Classifier},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/spaces/yourusername/resnet18-cifar100}}
}

License

MIT License - Feel free to use for research and educational purposes!