cifar100-resnet34

ResNet34 for CIFAR-100 Classification

This model is a ResNet34 architecture trained from scratch on CIFAR-100 dataset, achieving 73.09% top-1 accuracy.

Model Details

• Architecture: ResNet34 with BasicBlock residual structure
• Parameters: 21,328,292
• Dataset: CIFAR-100 (100 classes)
• Training Epochs: 83
• Target Accuracy: 73.0% (✅ Achieved!)
• Final Accuracy: 73.09%

Training Configuration

• Optimizer: SGD with Nesterov momentum (0.9)
• Learning Rate: OneCycleLR (max_lr=0.1)
• Weight Decay: 5e-4
• Batch Size: 128
• Loss Function: CrossEntropyLoss with Label Smoothing (0.1)
• Data Augmentation: Heavy augmentation with Albumentations

Usage

import torch
from model import ResNet34  # You'll need the model definition

# Load the model
model = ResNet34(num_classes=100)
checkpoint = torch.load('best_model.pth', map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# For inference, use the transforms from the training

CIFAR-100 Classes

The model can classify images into 100 classes from CIFAR-100 dataset, including animals, vehicles, objects, and natural scenes.

Performance

• Top-1 Accuracy: 73.09%
• Training Time: ~3.75 hours on GPU
• Model Size: ~81MB

Repository

• GitHub: https://github.com/chethan180/ciphar-100
• Training Logs: See TRAINING_LOGS.md for detailed epoch-by-epoch progress