MODEL_CARD.md

Model Card for Kolam AI Generator

Model Details

Model Description

The Kolam AI Generator is a Generative Adversarial Network (GAN) specifically designed to generate traditional Indian Kolam designs. It consists of a generator network that creates Kolam patterns from random noise and a discriminator network that ensures realistic pattern generation.

• Model Type: Generative Adversarial Network (GAN)
• Architecture: Custom CNN-based generator and discriminator
• Framework: PyTorch
• Input: Random noise vector (100 dimensions)
• Output: 64x64 grayscale Kolam image
• Parameters: ~4.3M total (Generator: ~2.5M, Discriminator: ~1.8M)

Model Architecture

Generator Network

Input: Random Noise (100D)
  ↓
Linear: 100 → 512*4*4
  ↓
Reshape: 512 × 4 × 4
  ↓
ConvTranspose2d: 512 → 256 (8×8)
  ↓
BatchNorm + ReLU
  ↓
ConvTranspose2d: 256 → 128 (16×16)
  ↓
BatchNorm + ReLU
  ↓
ConvTranspose2d: 128 → 64 (32×32)
  ↓
BatchNorm + ReLU
  ↓
ConvTranspose2d: 64 → 1 (64×64)
  ↓
Tanh Activation
  ↓
Output: Kolam Image (64×64)

Discriminator Network

Input: Kolam Image (64×64)
  ↓
Conv2d: 1 → 64 (32×32)
  ↓
LeakyReLU
  ↓
Conv2d: 64 → 128 (16×16)
  ↓
BatchNorm + LeakyReLU
  ↓
Conv2d: 128 → 256 (8×8)
  ↓
BatchNorm + LeakyReLU
  ↓
Conv2d: 256 → 512 (4×4)
  ↓
BatchNorm + LeakyReLU
  ↓
Flatten + Linear
  ↓
Output: Real/Fake Score

Intended Use

Primary Use Cases

• Art Education: Teaching traditional Indian Kolam patterns
• Cultural Preservation: Digitizing and preserving traditional art forms
• Design Inspiration: Source for modern design elements
• Therapeutic Art: Relaxing and meditative art creation
• Cultural Events: Festival decorations and celebrations

Out-of-Scope Use Cases

• Commercial Art Production: Not intended for mass commercial use
• High-Resolution Generation: Limited to 64x64 resolution
• Color Generation: Currently generates grayscale images only
• 3D Applications: 2D pattern generation only

Training Data

Dataset Composition

• Traditional Kolam Images: Photographs of real Kolam designs
• Synthetic Patterns: Algorithmically generated geometric patterns
• Data Augmentation: Rotation, scaling, and noise addition
• Size: 64x64 grayscale images
• Format: PNG files with transparency support

Data Preprocessing

• Resizing: All images resized to 64x64 pixels
• Normalization: Pixel values normalized to [-1, 1] range
• Grayscale Conversion: RGB images converted to grayscale
• Augmentation: Random rotations, flips, and noise addition

Training Procedure

Training Configuration

• Optimizer: Adam (Generator: lr=0.0002, Discriminator: lr=0.0002)
• Batch Size: 64
• Epochs: 200
• Training Time: 2-4 hours on GPU
• Hardware: NVIDIA GPU with CUDA support

Training Process

1. Adversarial Training: Generator and discriminator trained alternately
2. Feature Matching: Generator trained to match real image features
3. Learning Rate Scheduling: Exponential decay after 100 epochs
4. Gradient Clipping: Applied to prevent training instability

Evaluation Metrics

• Symmetry Analysis: Horizontal and vertical correlation coefficients
• Complexity Measurement: Edge detection and pattern density
• Balance Assessment: Visual weight distribution analysis
• Quality Score: Overall design quality (0-1 scale)

Performance

Quantitative Results

• Overall Quality Score: 0.789
• Horizontal Symmetry: 0.999
• Vertical Symmetry: 1.000
• Complexity: 0.297
• Balance: 1.000
• Rhythm: 1.000

Generation Performance

• Inference Speed: <1 second per image
• Memory Usage: <2GB during inference
• Model Size: ~50MB total
• Batch Generation: Supports batch processing

Limitations and Bias

Known Limitations

• Resolution: Limited to 64x64 pixel output
• Color: Generates grayscale images only
• Style: Focused on traditional geometric patterns
• Cultural Scope: Specific to Indian Kolam art

Potential Bias

• Training Data: May reflect bias in source images
• Cultural Representation: Limited to specific Kolam styles
• Geometric Patterns: May favor certain pattern types
• Symmetry: Strong bias toward symmetrical designs

Environmental Impact

Carbon Footprint

• Training: Estimated 2-4 hours on GPU
• Inference: Minimal computational requirements
• Storage: ~50MB model size
• Energy: Low energy consumption during inference

Technical Specifications

Hardware Requirements

• Training: NVIDIA GPU with CUDA support
• Inference: CPU or GPU (GPU recommended)
• Memory: 2GB RAM minimum, 4GB recommended
• Storage: 100MB for model and dependencies

Software Requirements

• Python: 3.9+
• PyTorch: 1.9.0+
• CUDA: 11.0+ (for GPU training)
• Dependencies: See requirements.txt

Model Card Contact

Maintainer

• Name: Rishi
• Email: [email protected]
• GitHub: [Your GitHub Profile]
• Hugging Face: [Your HF Profile]

Citation

@misc{kolam-ai-generator,
  title={Kolam AI Generator: Traditional Indian Art Generation using GANs},
  author={Rishi},
  year={2024},
  url={https://huggingface.co/spaces/yourusername/kolam-ai-generator}
}

License

This model is released under the MIT License. See the LICENSE file for details.

Acknowledgments

• Traditional Kolam artists for cultural guidance
• PyTorch team for the deep learning framework
• Hugging Face for the deployment platform
• Open source community for inspiration and support

---

Model Card Version: 1.0
Last Updated: December 2024
Model Version: 1.0.0