🧠 CNN COMSYS Hackathon

A robust deep learning pipeline for gender classification and face verification under challenging, real-world conditions, inspired by Vedic principles of fairness and accuracy.

---

All the submission files are under PROUDUCTION and the two Releases with tags publish101 and publish102.

📂 Repository Structure

CNN_vedic/
├── PRODUCTION/
│   ├── Task_A/               # Gender classification (VGG19-based)
│   │   ├── training/         # Training scripts, logs, plots
│   │   ├── inference/        # Inference & Grad-CAM scripts
│   │   └── test/             # Test images for inference
│   ├── Task_B/               # Face verification (Triplet Network)
│   │   ├── training/         # Triplet model training, embedding extraction
│   │   ├── inference/        # Face matching demo
│   │   └── test/             # Test/reference images for verification
├── [DEPRECATED MODELS]/      # Old/experimental models & scripts
├── Data/                     # (Git-ignored) Training/validation/test data
├── TASK.md                   # Challenge description & dataset structure
└── README.md                 # (You are here)
├── app/                      # Contains the server.py for the flask server inference
├── index.html                # frontend entry to check inference using a GUI

---

🚀 Project Overview

This project addresses two core computer vision tasks:

Task A: Gender Classification

• Goal: Predict gender (male/female) from face images, even in poor lighting or weather.
• Approach: Fine-tuned VGG19 with focal loss, heavy augmentation, and class balancing.
• Key Features:
  • Handles class imbalance (more male images than female).
  • Auto-threshold tuning for optimal F1 score.
  • Grad-CAM for model explainability.
  • Visual Explainability: Uses Grad-CAM to highlight which regions of the face the model focuses on to determine gender, displaying a heatmap overlay. See PRODUCTION/Task_A/inference/inference_vgg19_updated.py for implementation details.
• Relevant Files:
  • PRODUCTION/Task_A/training/train_vgg19_updated.py — Full training pipeline.
  • PRODUCTION/Task_A/inference/inference_vgg19_updated.py — Inference & Grad-CAM.

  Put the model in this folder from the link:
  vgg19_final_epoch

  • PRODUCTION/models/vgg19_final_epoch.h5 — Final model weights.
  • PRODUCTION/Task_A/training/vgg19_training_logs.txt — Training/evaluation logs.

Task B: Face Matching (Verification)

• Goal: Match distorted face images to correct identity using embeddings (not classification).

• Approach: Triplet Network with ResNet50 backbone, trained to minimize intra-class and maximize inter-class distances.

• Key Features:

  • Embedding extraction for similarity-based matching.
  • Automatic threshold tuning for best accuracy.

• Relevant Files:

  • PRODUCTION/Task_B/training/tripletNetwork_updated.py — Triplet training pipeline.
  • PRODUCTION/Task_B/training/embedding_model_extract.py — Extracts embedding submodel.
  • PRODUCTION/Task_B/inference/verify_face.py — Face verification demo.

  Put the model in this folder from the link

  Note: You only need the embedding_sequel.h5 model from here: embedding_sequel.h5

  • PRODUCTION/models/tripletNetwork.h5 — Full triplet model.
  • PRODUCTION/models/embedding_sequel.h5 — Embedding-only model.
  • PRODUCTION/Task_B/training/tripletTrainingLogs.txt — Training/evaluation logs.

---

📁 Folder Details

PRODUCTION/

• Task_A/: All scripts, logs, and models for gender classification.
  • training/: Training code, logs, and plots.
  • inference/: Inference script with Grad-CAM visualization.
• Task_B/: All scripts, logs, and models for face verification.
  • training/: Triplet network training, embedding extraction.
  • inference/: Face verification demo using embeddings.
• models/: Final production models (.h5), ready for inference.

[DEPRECATED MODELS]/

• Contains experimental, ablation, and legacy models/scripts (Keras, PyTorch, EfficientNet, etc.).
• Not for production use; see README.md inside for details.

Data/

• Not included in repo (see .gitignore).
• Structure for Task A and Task B as described in Task.md.

Task.md

• Full challenge description, dataset structure, and objectives for both tasks.

---

🏗️ How to Use

1. Gender Classification (Task A)

• Train:
  Run PRODUCTION/Task_A/training/train_vgg19_updated.py (requires data in Data/Task_A/).
• Inference:
  Place test image in PRODUCTION/Task_A/test/, run PRODUCTION/Task_B/inference/inference_vgg19_updated.py.
• Model:
  Download vgg19_final_epoch.h5 from releases.

2. Face Verification (Task B)

• Train:
  Run PRODUCTION/Task_B/training/tripletNetwork_updated.py (requires data in Data/Task_B/).
• Extract Embedding Model:
  Run embedding_model_extract.py after training.
• Inference:
  Place reference/test images, run PRODUCTION/Task_B/inference/verify_face.py.
• Model:
  Download embedding_sequel.h5 from releases.

---

📝 Key Features

• Class Imbalance Handling:
  Automatic class weights and focal loss for fair gender classification.
• Data Augmentation:
  Robust to real-world distortions (rotation, brightness, etc.).
• Auto Threshold Tuning:
  Finds best probability/distance threshold for optimal F1/accuracy.
• Explainability:
  Grad-CAM visualizations for gender model, showing which parts of the face are being scanned to determine gender using a heatmap.
• Production-Ready:
  All production models and scripts are in PRODUCTION/.

---

📈 GRAD-CAM Result Showcase (Male and Female)

• MALE
• FEMALE

📊 Results

• Gender Classification:
  • Accuracy: ~97% (see vgg19_training_logs.txt
  • Balanced precision/recall for both genders.
• Face Verification:
  • Accuracy: ~96% (see tripletTrainingLogs.txt)
  • High precision/recall for identity matching.

---

🔧 Development Environment

Component	Version
Python	3.10.0
TensorFlow	2.9.0
CUDA Toolkit	11.8
cuDNN	8.6.0
GPU	NVIDIA RTX (6GB or higher) recommended

---

📦 Requirements

CPU Environment

Install with:

pip install -r requirements.txt

---

GPU Environment

Install with:

pip install -r requirements_GPU.txt

OR

conda env create -f facecom_env.yml

Check Runtime File to get detailed steps on 100% reproducable inference smoothly

Ensure CUDA 11.8 and cuDNN 8.6.0 are installed and properly configured for GPU support (only if you are accessing the training scripts)

---

📚 References

• See Task.md for dataset structure and challenge details.
• Model weights and releases:
  GitHub Releases

---

🤝 Acknowledgements

Developed by Circuit-Overtime and contributors.
For academic/educational use only.

---