andreasgrv / sigmoid-bottleneck

Sparse Multi Label Classification

Sparse Multi-Label Classification (spmlbl)

Herein we provide the code to reproduce our results for the paper:

Taming the Sigmoid Bottleneck: Provably Argmaxable Sparse Multi-Label Classification

Installation

Below installation is for main library.

Install Python Dependencies

python3.8 -m venv .env
source .env/bin/activate
# Adapt cuda in requirements to your use case
pip install -r requirements.txt
pip install -e .
# NOTE: For each specific dataset you may need to install more libraries
# See README.md file in each experiment section

Set Environment Variables

# Avoid pytorch using too many threads
export OMP_NUM_THREADS=1
export MKL_NUM_THREADS=1

# Random seed - needed for exps
export SEED=0
# What device to run models on
export MLBL_DEVICE="cuda:0"
# Number of threads
export MLBL_NUM_PROC=10

Install Gurobi

The linear programming algorithm that detects (un)argmaxable label assignments depends on Gurobi. It requires a license, see link above.

Run Tests

Tests require dependencies and Gurobi.

py.test tests

Experiments

For each dataset, see the README.md file in the corresponding folder under experiments.

• MIMIC-III
• BioASQ Task A
• OpenImages v6