FrankCCCCC/colored_mnist_28

Colored MNIST Dataset

A comprehensive dataset of MNIST digits with RGB colored backgrounds, designed for multi-objective classification tasks including digit recognition and background color classification.

Overview

This dataset extends the classic MNIST digit dataset by adding colored backgrounds (blue, green, red) to create a multi-label classification challenge. Each image contains:

• Digit label: 0-9 (digit recognition)
• Color label: blue, green, red (background color classification)
• Image format: 28×28 RGB PNG files

Dataset Statistics

Training Set (8,835 samples)

Digit	Count	Digit	Count
0	868	5	764
1	1,000	6	893
2	858	7	948
3	909	8	835
4	883	9	877

Test Set (10,000 samples)

Digit	Count	Digit	Count
0	980	5	892
1	1,135	6	958
2	1,032	7	1,028
3	1,010	8	974
4	982	9	1,009

Color Distribution

• Blue: ~2,946 training samples
• Green: ~2,968 training samples
• Red: ~2,921 training samples

Directory Structure

colored_mnist/
├── README.md
├── colored_mnist.py          # HuggingFace dataset loader
├── images/                   # Example images
│   ├── blue_312.png         # Blue background, digit 3
│   ├── green_8218.png       # Green background, digit 8  
│   └── red_8428.png         # Red background, digit 4
├── training/                 # Training data
│   ├── 0/                   # Digit 0 samples
│   │   ├── blue_1015.png
│   │   ├── green_1000.png
│   │   └── red_1028.png
│   └── ...                  # Digits 1-9
├── testing/                  # Test data
│   ├── 0/                   # Digit 0 samples
│   │   ├── blue_10.png
│   │   ├── green_1084.png
│   │   └── red_1001.png
│   └── ...                  # Digits 1-9
├── training.h5              # HDF5 format (optional)
└── testing.h5               # HDF5 format (optional)

Usage

Loading with HuggingFace Datasets

from datasets import load_dataset

# Load the dataset using the custom script
dataset = load_dataset("FrankCCCCC/colored_mnist", trust_remote_code=True)

print(f"Train samples: {len(dataset['train'])}")
print(f"Test samples: {len(dataset['test'])}")

# Access a sample
sample = dataset['train'][0]
print(f"Image shape: {sample['image'].shape}")  # PIL.Image: Raw Image with shape (28, 28, 3)
print(f"Digit Label: {sample['digit_label']}")  # Integer: Digit class 0-9
print(f"Color Label: {sample['color_label']}")  # String: Color class: 'blue', 'green', 'red'
print(f"filename: {sample['filename']}")        # String: file name
print(f"image_path: {sample['image_path']}")    # String: file path

Applications

This dataset is ideal for:

1. Multi-task Learning: Train models to simultaneously classify digits and colors
2. Domain Adaptation: Study how background colors affect digit recognition
3. Bias Analysis: Investigate color-digit correlations and model robustness
4. Transfer Learning: Pre-train on one task (digit/color) and fine-tune on another
5. Curriculum Learning: Start with single-task then progress to multi-task
6. Attention Mechanisms: Visualize what models focus on (digit vs. background)

Features

• Multi-label Classification: Both digit and color labels for each image
• Balanced Colors: Approximately equal distribution across blue, green, red
• Standard Format: Compatible with PyTorch, TensorFlow, and HuggingFace
• PIL Integration: Images loaded as PIL objects for easy preprocessing
• Metadata: Includes original filename and image path information

Dataset Format

Each sample contains:

{
    'image': PIL.Image,           # 28×28 RGB image
    'digit_label': int,           # Digit class (0-9)
    'color_label': str,           # Color class ('blue'|'green'|'red')
    'filename': str,              # Original filename
    'image_path': str             # Full path to image file
}

Requirements

pip install datasets pillow torch torchvision matplotlib numpy

Citation

@misc{colorized-mnist,
  title={Colorized MNIST},
  author={Original MNIST Dataset},
  year={2024},
  url={https://github.com/jayaneetha/colorized-MNIST}
}

License

This dataset follows the same license as the original MNIST dataset. Please cite both the original MNIST work and this colored extension when using this dataset.

Acknowledgments

• Based on the original MNIST dataset by Yann LeCun et al.
• Color augmentation creates additional research opportunities for multi-task learning
• Compatible with modern deep learning frameworks and HuggingFace ecosystem