AsangSingh/ResNet_CIFAR100

CIFAR-100 Image Classifier

This space hosts a ResNet18 model trained from scratch on the CIFAR-100 dataset. The model achieves 75.65% top-1 accuracy on the test set.

Model Details

• —Architecture: ResNet18 (modified for CIFAR-100)
• —Dataset: CIFAR-100 (100 classes)
• —Input Size: 32x32 RGB images
• —Performance: 75.65% Top-1 Accuracy

Usage

1. 1.Upload an image using the interface
2. 2.The model will predict the top 5 most likely classes
3. 3.Results show class names and confidence scores

Classes

The model can classify images into 100 different categories, including:

• —Animals (bear, butterfly, fish, etc.)
• —Vehicles (bicycle, bus, motorcycle, etc.)
• —Natural objects (cloud, forest, mountain, etc.)
• —Household items (bed, chair, table, etc.)
• —And many more!

Model Training

• —Trained for 100 epochs
• —Used SGD optimizer with momentum
• —Implemented OneCycleLR learning rate scheduling
• —Applied data augmentation (random crops, flips, rotations)

Limitations

• —Works best with clear, centered images
• —Designed for 32x32 images (larger images will be resized)
• —May have lower accuracy on images very different from CIFAR-100 style
• —Performance varies based on image quality and lighting