Skip to content

GeorgeTheo99/AML_22

Repository files navigation

AML_22

Emotion Recongition System

  • Comprehensive emotino recognition that creates an ensemble method with CNN, MLP, and logistic regression. CLIP feature extraction is used for the MLP and LogReg training

Overview

  • Ensemble based emtion system that can classify based on
  • Angry, Disgust, Fear, Happy, Sad, Surprise, Neutral

Features Multi-Model Architecture:

  • Convolutional Neural Network (CNN) for direct image processing
  • Multi-Layer Perceptron (MLP) using CLIP features
  • Ensemble Logistic Regression using CLIP features

Advanced Ensemble Methods:

  • Weighted voting combination
  • Hierarchical decision making
  • Dynamic confidence-based weight adjustment

CLIP Integration: Utilizes OpenAI's CLIP model for robust feature extraction

Requirements: torch torchvision numpy pandas scikit-learn PIL clip ftfy regex tqdm

Model Architecture

CNN Model

  • Three convolutional blocks with batch normalization and dropout
  • Max pooling layers
  • Fully connected layers with dropout
  • BatchNorm and ReLU activation

Feature MLP

  • Multi-layer architecture with skip connections
  • Batch normalization and GELU activation
  • Dropout for regularization
  • Dynamic learning rate scheduling

Ensemble Logistic Regression

  • Multiple logistic regression models (3% performance improvement from standard LogReg)
  • Feature selection using variance threshold
  • Batch processing for memory efficiency
  • Class weight balancing

Performance The system achieves the following accuracies on the validation set: Through empircal testing, CNN performed better on stock images so was given preference in hierarchical model.

  • MLP with CLIP features: ~69%
  • Logistic Regression: ~65%
  • CNN: ~63%
  • Weighted Voting Ensemble: ~70%
  • Hierarchical Ensemble: ~69%

Model weights are saved with timestamps:

  • mlp_model_TIMESTAMP.pth
  • logreg_model_TIMESTAMP.pkl
  • cnn_model_TIMESTAMP.pth
  • ensemble_weights_TIMESTAMP.pkl

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages