GlasXC

GlasXC is an implementation of paper Breaking the Glass Ceiling for Embedding-Based Classifiers for Large Output Spaces - NIPS'19 in Pytorch:

• GlasXC: is for solving the extreme classification using simple Deep Learning architecture + GLAS Regularizer

Setup

Prerequisites

• Python 2.7 or 3.5
• Requirements for the project are listed in requirements.txt. In addition to these, PyTorch 0.4.1 or higher is necessary. The requirements can be installed using pip:

  $ pip install -r requirements.txt 

  or using conda:

  $ conda install --file requirements.txt

Installing extreme_classification

• Clone the repository.

  $ git clone https://github.com/pyschedelicsid/GlasXC
  $ cd GlasXC

• Install

  $ python setup.py install

Using the scripts

GlasXC

Use train_GlasXC.py. A description of the options available can be found using:

$ python train_GlasXC.py --help

This script trains (and optionally evaluates) evaluates a model on a given dataset using the GlasXC algorithm.

Reported Results

To run GlasXC in the configuration used in the report, use:

$ ./train_GlasXC_with_args.sh

To run the baseline model, use:

$ python baseline.py

Links to downloading each dataset used can be found here, and the project report can be found here. The configuration files used (described below) for each dataset can be found here.

Data Format

The input data must be in the LIBSVM format. An example of such a dataset is the Bibtex dataset found here.

The first row in the LIBSVM format specifies dataset size and input and output dimensions. This row must be removed, and this information must be provided through configuration files, as explained below.

Configuration files

Deep Neural Network Architecture Configurations

For using GlasXC through GlasXC.py, you need to have valid neural network configurations for the encoding of the inputs, labels in the latent space and the regressor in the YAML format. An example configuration file is:

- name: Linear
  kwargs:
    in_features: 500
    out_features: 1152

- name: LeakyReLU
  kwargs:
    negative_slope: 0.2
    inplace: True

- name: Linear
  kwargs:
    in_features: 1152
    out_features: 1836

- name: Sigmoid

Please note that the name and kwargs attributes have to resemble the same names as those in PyTorch.

Optimizer Configurations

Optimizer configurations are very similar to the neural network configurations. Here you have to retain the same naming as PyTorch for optimizer names and their parameters - for example: lr for learning rate. Below is a sample:

name: SGD
args:
  lr: 0.01
  momentum: 0.9

Dataset Configurations

In both the scripts, you are required to specify a data root (data_root), dataset information file (dataset_info). data_root corresponds to the folder containing the datasets. dataset_info requires a YAML file in the following format:

train_filename:
train_opts:
  num_data_points:
  input_dims:
  output_dims:

test_filename:
test_opts:
  num_data_points:
  input_dims:
  output_dims:

If the test dataset doesn't exist, then please remove the fields test_filename and test_opts. An example for the Bibtex dataset would be:

train_filename: bibtex_train.txt
train_opts:
  num_data_points: 4880
  input_dims: 1836
  output_dims: 159

test_filename: bibtex_test.txt
test_opts:
  num_data_points: 2515
  input_dims: 1836
  output_dims: 159

---

Written By: Siddhant Katyan

GLaS