CE-ImageProducer

Python scripts and code to produce and transform images in a format for "Hierarchical Image Classification using Entailment Cone Embeddings"

Reference

The paper that describes this method can be found on arxiv here.

The repository that implements this learning algorithm can be found here.

Reproduction

This repository contains specific scripts that were required to reproduce results of the paper on a different dataset. We chose to attempt to reproduce the reuslts and attempt to learn a hierarchical structure of a set of wines. The hierarchical representation could be country -> region -> winery -> wine. We have access to a large dataset of wines and corresponding images of high quality.

Production of the dataset in steps

1. Fetches a feed containing wines, information about wines and their unique identifiers.
2. Filter out underrepresented wines to make sure we have enough data
3. Download images from remote storage for each wine and use these images to produce enough samples by adding noise and performing image transformations
4. Upload the images to a remote location and keep them there so work can be done simultaneously
5. Produce a JSON file in the format the is expected for the learning_embeddings algorithm using the chosen hierarchical information.