Deep Learning Project: Face Attributes Prediction

This repository contains the implementation of a deep learning project for face attribute prediction using models like DenseNet121 and ResNet50. It includes notebooks, trained models, evaluation metrics, and outputs for comprehensive analysis.

Table of Contents

1. Introduction
2. Project Structure
3. Model Details
4. Evaluation
5. Datasets
6. Visualizations
7. Dependencies

Introduction

The project focuses on predicting facial attributes using deep learning models. It utilizes popular architectures such as DenseNet121 and ResNet50, and includes pre-trained models fine-tuned on the CelebA dataset.

Project Structure

• model_evaluation/: Contains evaluation notebooks and results.
• model_images/: Includes visual outputs generated during modeling and evaluation.
• models_trained/: Stores pre-trained and fine-tuned models.
• output_metrics/: Detailed metrics and performance outputs.
• testing_notebooks/: Notebooks for testing and debugging the models.
• Notebooks:
  • MIS_548_DL_PROJECT_DenseNet121.ipynb: Implementation of DenseNet121.
  • MIS_548_DL_PROJECT_ResNet50.ipynb: Implementation of ResNet50.
  • MIS_548_Metrics.ipynb: Evaluation metrics for the models.
• Datasets:
  • list_attr_celeba.csv: Attributes of CelebA dataset.
  • list_bbox_celeba.csv: Bounding boxes for face detection.
  • list_eval_partition.csv: Partitioning of the CelebA dataset.
  • list_landmarks_align_celeba.csv: Facial landmarks for alignment.

Model Details

DenseNet121

• A convolutional neural network that connects each layer to every other layer.
• Fine-tuned for facial attribute prediction.

ResNet50

• A deep residual network that facilitates training deeper networks.
• Used for extracting features and predicting attributes.

Evaluation

Model performance is evaluated using:

• Accuracy
• Precision
• Recall
• F1 Score

Results and metrics are stored in the output_metrics/ directory and summarized in MIS_548_Metrics.ipynb.

Datasets

The CelebA dataset is used for training and evaluation. It includes:

• Over 200,000 labeled images with 40 binary attributes.
• Metadata files for attributes, bounding boxes, and landmarks.

Visualizations

ResNet50 Predictions on Unseen Data

DenseNet121 Predictions on Unseen Data

Model Testing Result Comparison

Dependencies

The project uses:

• TensorFlow
• Keras
• PyTorch
• scikit-learn
• matplotlib
• pandas
• numpy

Ensure all dependencies are installed using requirements.txt.