ImageNet-100 PyTorch Dataloader

Streamlined PyTorch implementation for ImageNet-100 from parquet files. Efficient dataloaders for training and validation.

🚀 Quick Start

from scripts.pytorch_dataloader import ImageNet100Parquet
from torch.utils.data import DataLoader
from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
])

train_dataset = ImageNet100Parquet("data", "train", transform)
val_dataset = ImageNet100Parquet("data", "validation", transform)

train_loader = DataLoader(train_dataset, batch_size=8, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=8, shuffle=False)

for x, y in train_loader:  # x: [batch, 3, 224, 224], y: [batch]
    pass

📊 Dataset Details

• Classes: 100 ImageNet classes (balanced)
• Training: 126,689 images
• Validation: 5,000 images
• Image sizes: Variable (standard output: 224×224)

🛠️ Utilities

Data Inspection

python scripts/utils.py                    # Run all utilities
python scripts/utils.py debug              # Debug structure only
python scripts/utils.py sizes              # Check image sizes only
python scripts/utils.py memory             # Analyze memory usage

Programmatic Usage

from scripts.utils import debug_structure, check_image_sizes, analyze_memory_usage

debug_structure()                          # Inspect parquet structure
check_image_sizes(num_samples=20)          # Analyze image dimensions
analyze_memory_usage(batch_size=32)        # Memory usage analysis

🎯 Key Features

• Efficient Loading: Direct parquet file reading with proper image decoding
• Memory Optimized: Lazy loading with efficient tensor memory usage
• Robust Error Handling: Comprehensive validation and error messages
• Type Safe: Full type hints for better IDE support and debugging
• Flexible Transforms: Easy to customize preprocessing pipeline
• Data Inspection: Built-in utilities for dataset structure analysis
• PyTorch Native: Seamless integration with PyTorch training workflows

🔄 Future PT File Support

Planned conversion to eliminate pandas dependency:

# Future usage (backlog item)
train_dataset = ImageNet100PT("data", "train")  # Direct torch.load()
val_dataset = ImageNet100PT("data", "validation")

📝 Data Format

Parquet structure:

{
    'image': {'bytes': b'\x89PNG...', 'path': None},
    'label': 0  # Integer class label (0-99)
}

🔧 Configuration

# Different image sizes
transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
])

# Custom preprocessing  
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.RandomHorizontalFlip(),
    transforms.ColorJitter(brightness=0.2, contrast=0.2),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                         std=[0.229, 0.224, 0.225])
])

📚 Dataset Information

• Classes: 100 ImageNet classes (balanced)
• Training: 126,689 images
• Validation: 5,000 images
• Homepage: https://github.com/HobbitLong/CMC
• Paper: Contrastive multiview coding

📄 License

Original ImageNet license terms. Non-commercial research and educational use only.

🙏 Acknowledgments

• Original ImageNet team
• 🤗 Transformers (parquet format reference)
• CMC paper (ImageNet-100 subset)