Deep Learning Project
This project proposes a deep learning-based framework for Emotional Valence Prediction in images using popular CNN architectures including VGG16, ResNet50, and InceptionV3. This paper utilized the ArtEmis dataset. The framework utilizes a pre-processing procedure to convert the dataset(image data and labels) stored as compressed NumPy arrays on NYU HPC. In this implementation, a CNN model is constructed by combining a selected base model architecture with custom classification layers. The base model is initialized with pre-trained weights from the ImageNet dataset and its layers are frozen to preserve the learned features during training. Niche techniques have been implemented like, early stopping to prevent overfitting and compute the validation curacy and a classification report for performance evaluation. The framework further includes a model selection mechanism that iterates through the specified model names and tracks the best-performing model based on validation accuracy. The framework allows researchers to easily experiment with different CNN architectures, leverage pre-trained models, and assess model performance using appropriate evaluation metrics. Overall, this framework provides a valuable tool for deep learning-based emotional valence prediction research and its applications.