cifar100-resnet18

CIFAR-100 ResNet18 Classifier

This space contains a ResNet18 model trained on CIFAR-100 dataset for image classification.

Model Details

• Architecture: ResNet18
• Dataset: CIFAR-100 (100 classes)
• Input Size: 32x32 RGB images
• Classes: 100 different object categories

Usage

Upload an image and the model will predict which of the 100 CIFAR-100 classes it belongs to.

Training Details

• Epochs: 120
• Optimizer: SGD with momentum (0.9)
• Learning Rate: 0.1 with MultiStepLR scheduler
• Weight Decay: 5e-4
• Data Augmentation: Random crops, flips, color jitter, rotation, and random erasing

Model Architecture

The model uses a ResNet18 architecture with:

• Initial convolution layer (3x3, 64 filters)
• 4 residual layers with [2, 2, 2, 2] blocks
• Global average pooling
• Fully connected layer (512 → 100 classes)

CIFAR-100 Classes

The model can classify images into 100 different categories including:

• Animals: bear, cat, dog, bird, fish, etc.
• Vehicles: car, truck, motorcycle, airplane, etc.
• Objects: chair, table, bottle, cup, etc.
• Nature: tree, flower, mountain, sea, etc.

Performance

The model achieves competitive accuracy on the CIFAR-100 test set with proper data augmentation and training techniques.

How to Use

1. Upload any image using the interface
2. The model will automatically resize it to 32x32 pixels
3. Get predictions with confidence scores
4. The model works best with images similar to CIFAR-100 dataset (small objects, clear backgrounds)

Technical Notes

• Images are automatically resized to 32x32 pixels
• Normalization is applied using CIFAR-100 mean and std values
• The model uses softmax activation for probability distribution
• Confidence scores represent the model's certainty in its prediction