abhilash88 / age-gender-prediction

🏆 ViT-Age-Gender-Prediction: Vision Transformer for Facial Analysis

✅ MODEL WEIGHTS NOW AVAILABLE - Trained model weights uploaded and ready for use!

🎯 Quick Usage

import torch
from transformers import ViTImageProcessor
from model import AgeGenderViTModel  # Use the model.py from this repo

# Load model
model = AgeGenderViTModel()
model.load_state_dict(torch.load("pytorch_model.bin"))
model.eval()

# Load processor
processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224")

# Predict on image
from PIL import Image
image = Image.open("your_image.jpg")
inputs = processor(images=image, return_tensors="pt")

with torch.no_grad():
    age_pred, gender_pred = model(inputs["pixel_values"])
    
age = int(age_pred.item())
gender = "Female" if gender_pred.item() > 0.5 else "Male"
confidence = gender_pred.item() if gender_pred.item() > 0.5 else 1 - gender_pred.item()

print(f"Age: {age} years, Gender: {gender}, Confidence: {confidence:.1%}")

🏆 Performance Achievements

• ✅ 94.3% Gender Accuracy - ELITE tier performance
• ✅ 4.5 Years Age MAE - Research-grade precision
• ✅ 86.8M Parameters - Optimally fine-tuned Vision Transformer
• ✅ Production Ready - Stable, consistent results

📊 Dataset & Training Details

Training Dataset: UTKFace

• Total Images : 23,687 facial images
• Age Range : 1-100 years
• Demographics : Balanced gender distribution (52.3% Male, 47.7% Female)
• Quality : High-resolution, diverse lighting and pose conditions

⚠️ Important Dataset Characteristics

The UTKFace dataset has a specific age distribution :

• Adults (21-50 years) : ~70% of data (majority)
• Young Adults (16-30 years) : ~20% of data
• Children (0-15 years) : ~5% of data (limited)
• Seniors (50+ years) : ~5% of data

🎯 Model Performance by Age Group

• Excellent : Adults and young adults (16-60 years) - 94.3% gender accuracy
• Good : Teenagers (13-20 years) - ~90% accuracy
• Limited : Children (0-12 years) - Reduced accuracy due to limited training data
• Good : Seniors (60+ years) - ~85% accuracy

🔄 Upcoming Improvements

Version 2.0 - Enhanced Children Support (In Development)

• 🎯 Training on FairFace Dataset - Better age distribution
• 👶 Children-Specific Fine-tuning - Focused 0-15 years training
• 📊 APPA-REAL Integration - Apparent age dataset inclusion
• 🎨 Data Augmentation - Synthetic children faces generation

Planned Enhancements

• Multi-Age Ensemble : Specialized models for different age ranges
• Cross-Cultural Training : Enhanced performance across ethnicities
• Age-Specific Confidence : Different confidence thresholds per age group
• Real-time Optimization : Mobile and edge device deployment

📈 Current Model Strengths

Best Use Cases

• ✅ Adult demographic analysis (primary strength)
• ✅ Social media content filtering (teen/adult classification)
• ✅ Marketing analytics (adult age segmentation)
• ✅ Security applications (adult age verification)

Architecture Advantages

• Vision Transformer : Superior to CNN-based approaches
• Multi-task Learning : Joint age and gender optimization
• Transfer Learning : Built on google/vit-base-patch16-224
• Robust Features : Handles various lighting and pose conditions

📊 Technical Specifications

Model Architecture

• Base : google/vit-base-patch16-224
• Parameters : 86.8M total
• Input : 224×224 RGB images
• Outputs : Age (regression) + Gender (binary classification)
• Attention Heads : 12
• Transformer Layers : 12

Training Configuration

• Epochs : 15 (fully converged)
• Optimizer : AdamW (lr=2e-5)
• Batch Size : 32
• Training Time : 2.95 hours on GPU
• Validation Split : 80/20 stratified

📊 Files Included

• pytorch_model.bin - Trained model weights (331MB)
• config.json - Model configuration and metadata
• training_logs.json - Complete training history and metrics
• model.py - Model architecture and usage code

⚠️ Usage Recommendations

Optimal Performance

• Primary Use : Adults and young adults (16-60 years)
• High Confidence : Gender classification across all ages
• Reasonable Accuracy : Age estimation for adults

Limitations to Consider

• Children (0-12 years) : Limited training data may affect accuracy
• Very elderly (70+ years) : Fewer training examples
• Extreme poses/lighting : May reduce performance

Best Practices

• Face Detection : Ensure clear, front-facing faces
• Image Quality : Use good lighting and resolution
• Age Context : Consider model strengths for your use case
• Confidence Thresholds : Adjust based on your application needs

🔬 Research & Citation

@misc{age-gender-prediction-2025,
  title={Age-Gender-Prediction: Vision Transformer for Facial Analysis},
  author={Abhilash Sahoo},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/abhilash88/age-gender-prediction}
}

🤝 Contributing & Feedback

We welcome contributions and feedback, especially:

• Children dataset suggestions for Version 2.0
• Performance evaluations on diverse datasets
• Use case feedback for model improvements
• Technical optimizations and enhancements

📈 Roadmap

• Q1 2025 : Children-focused fine-tuning (Version 2.0)
• Q2 2025 : Multi-cultural dataset integration
• Q3 2025 : Mobile optimization and edge deployment
• Q4 2025 : Real-time video analysis capabilities

---

Current Version : 1.0 (Adult-focused) | Next Version : 2.0 (Children-enhanced) | Status : Production Ready*

*Best performance on adults (16-60 years). Children support improved in upcoming Version 2.0.