Extreme Classification via Adversarial Softmax Approximation

This repository contains the code for our paper

• R. Bamler and S. Mandt, Extreme Classification via Adversarial Softmax Approximation, ICLR 2020.

Dependencies

This code was tested with TensorFlow version 1.15 on python 3.6. The code to fit the auxiliary model was tested with Julia version 1.1.

Directory Overview

• Directory dat:
  • Contains a textual version of pre-extracted XML-CNN features, originally downloaded from https://github.com/siddsax/XML-CNN (the download link there is currently broken, which is why we provide a mirror of the data set). Due to their large size, the data sets are included via Git Large File Storage (git lfs).
• Directory preprocess-extreme-predicton:
  • Contains code to reproduce the exact binary representation of the data sets used in the paper, using the textual representation in the directory dat as input.
  • Contains a jupyter notebook pca.ipynb that was used to generate the low-dimensional feature vectors for the auxiliary model as described in the paper.
• Directory aux_model:
  • Contains both Julia code to fit the auxiliary model and python code to use the fitted model during training of the main model, as described in the paper.
• File train.py:
  • The main file to train the proposed model. See paper for hyperparameters.
• Directory main_model:
  • Contains internal utilities used by train.py. You shouldn't usually need to run any of the python scripts in this directory manually.

License

The source code in this repository is released under the MIT License. If you use this software for a scientific publication, please consider citing the following paper: R. Bamler and S. Mandt, Extreme Classification via Adversarial Softmax Approximation, ICLR 2020.

Authors

• Robert Bamler
• Stephan Mandt