pandurangpatil/cifar100-wideresnet-session8

CIFAR-100 WideResNet-28-10

Model Description

WideResNet-28-10 trained on CIFAR-100 dataset with advanced augmentation techniques.

Model Architecture

• Architecture: WideResNet-28-10
• Parameters: 36.5M
• Depth: 28 layers
• Width Factor: 10
• Dropout: 0.3

Training Configuration

• Batch Size: 256
• Optimizer: SGD (momentum=0.9, weight_decay=1e-3)
• Learning Rate: Cosine annealing with warmup (0.01→0.1, min=1e-4)
• Scheduler: CosineAnnealingWarmRestarts (T_0=25)
• Augmentation: HorizontalFlip, ShiftScaleRotate, Cutout, ColorJitter
• MixUp: Alpha=0.2
• Label Smoothing: 0.1
• Mixed Precision: Enabled
• Gradient Clipping: 1.0

Performance

• Best Test Accuracy: 71.20%
• Total Epochs Trained: 100
• Final Train Accuracy: 69.26%
• Final Test Accuracy: 71.20%

Available Checkpoints

• best_model.pth - Best performing model
• checkpoint_epoch10.pth - Epoch 10 checkpoint
• checkpoint_epoch25.pth - Epoch 25 checkpoint (end of cycle 1)
• checkpoint_epoch50.pth - Epoch 50 checkpoint (mid-training)
• checkpoint_epoch75.pth - Epoch 75 checkpoint (late training)
• final_model.pth - Final epoch model

Usage

import torch
from huggingface_hub import hf_hub_download

# Download model
checkpoint_path = hf_hub_download(
    repo_id="pandurangpatil/cifar100-wideresnet-session8",
    filename="best_model.pth"
)

# Load checkpoint
checkpoint = torch.load(checkpoint_path, map_location='cpu')

# Load model (define WideResNet class first)
model = WideResNet(depth=28, widen_factor=10, num_classes=100)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

Training Details

• Dataset: CIFAR-100 (50,000 train, 10,000 test)
• Classes: 100
• Image Size: 32×32
• Normalization: mean=(0.5071, 0.4865, 0.4409), std=(0.2673, 0.2564, 0.2761)

Files

• training_curves.png - Training/test accuracy and loss curves
• metrics.json - Complete training history
• config.json - Hyperparameter configuration

License

MIT

Citation

@misc{wideresnet-cifar100,
  author = {Pandurang Patil},
  title = {CIFAR-100 WideResNet-28-10},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/pandurangpatil/cifar100-wideresnet-session8}
}