This repository contains the code to reproduce the results in the paper.

In the datasets module:

• tumblr_extraction.py extracts Tumblr posts (url, date, image url, text, tags).
• download_images.py downloads the images using the image url.
• convert_images_tfrecords.py converts the posts to the TFRecord format for the TensorFlow Dataset pipeline.

• The Image model can be trained with train_image_model in the image_model module.
• The Text model can be trained with train_text_model in the text_model module.
• The Deep Sentiment model can be trained with train_deep_sentiment in the image_text_model module.

• The hierarchical clustering can be obtained using correlation_matrix in the image_text_model module.
• The top words can be obtained using word_most_relevant in the image_text_model module.
• The OASIS correlation can be obtained using oasis_evaluation in the image_text_model module.