Multi-annotator Deep Learning: A Probabilistic Framework for Classification

Authors: Marek Herde, Denis Huseljic, and Bernhard Sick

Project Structure

• evaluation: collection of Python and Bash scripts required to perform experimental evaluation
• lfma: Python package consisting of several sub-packages
  • classifiers: implementation of multi-annotator supervised learning techniques according to scikit-learn interfaces
  • modules: implementation of multi-annotator supervised learning techniques as pytorch_lightning modules, pytorch data sets, and special layers
  • utils: helper functions
• notebooks:
  • annotator_simulation.ipynb: simulation of annotator sets for data sets without annotations from real-world annotators
  • classification.ipynb: visualization of MaDL's properties regarding the three research questions in the accompanied article
  • data_set_creation_download.ipynb: creation of artificial data sets and download of real-world data sets
  • evaluation.ipynb: loading and presentation of experimental results
• requirements.txt: list of Python packages required to reproduce experiments

How to execute experiments?

To accommodate a large number of experiments, we utilized a computer cluster equipped with multiple V100 and A100 GPUs, allowing us to execute many experiments simultaneously. Without access to such a cluster, reproducing all results presented in the accompanying article would take much more time.

In the following, we describe step-by-step how to execute all experiments presented in the accompanied article. As a prerequisite, we assume to have a Linux distribution as operating system and conda installed on your machine.

1. Setup Python environment: bash projectpath$ conda create --name madl python=3.9 projectpath$ conda activate madl First, we need to install torch with the build (1.13.1). For this purpose, we refer to pytorch. An exemplary command for a Linux operating system would be: bash projectpath$ pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116 Subsequently, we install the remaining requirements: bash projectpath$ pip install -r requirements.txt
2. Create and download data sets: Start jupyter-notebook and follow the instructions in the jupyter-notebook file notebooks/data_set_creation_download.ipynb. bash projectpath$ conda activate madl projectpath$ jupyter-notebook
3. Simulate annotators: Start jupyter-notebook and follow the instructions in the jupyter-notebook file notebooks/annotator_simulation.ipynb. bash projectpath$ conda activate madl projectpath$ jupyter-notebook
4. Create experiment scripts: We create a bash script for each triple of data set, annotator set, and multi-annotator supervised learning technique. Such a file contains the complete configuration of the respective experimental setup including hyperparameters. Depending on your machine or even compute cluster, you can optionally define your own SLURM setup. For this purpose, you need to define the corresponding variables in the evaluation/create_experiment_scripts.py accordingly. For example, you can adjust the amount of RAM allocated for a run of an experiment. Once, you have adjusted the variables in this script, you may execute it via: bash projectpath$ conda activate madl projectpath$ python evaluation/create_experiment_scripts.py
5. Execute experiment scripts: After the creation of the experiment scripts, there will be a folder for each multi-annotator supervised learning technique we want to evaluate. For example, the file evaluation/madl/madl_cifar10_none.py corresponds to evaluating MaDL on CIFAR10 with independent annotators. Such a file consists of multiple commands executing the file evaluation/run_experiment.py with different configurations. For a better understanding of these possible configurations, we refer to the explanations in the file evaluation/run_experiment.py. If you have disabled the use of slurm, you can now execute such a bash script via: bash projectpath$ conda activate madl projectpath$ ./evaluation/madl/madl_cifar10_none.sh Otherwise, you need to use the sbatch command: bash projectpath$ conda activate madl projectpath$ sbatch ./evaluation/madl/madl_cifar10_none.sh

How to investigate the experimental results?

Once, an experiment is completed, its associated results are saved as a .csv file at the directory specified by evaluation.run_experiment.RESULT_PATH. For getting a tabular and summarized presentation of these results, you need to start jupyter-notebook and follow the instructions in the jupyter-notebook file notebooks/evaluation.ipynb. bash projectpath$ conda activate madl projectpath$ jupyter-notebook

How to reproduce the visualizations?

Start jupyter-notebook and follow the instructions in the jupyter-notebook file notebooks/classification.ipynb. bash projectpath$ conda activate madl projectpath$ jupyter-notebook

Citing

If you use multi-annotator-deep-learning in one of your research projects or would like to reference the accompanied article, please use the following:

@article{ herde2023multiannotator, title={Multi-annotator Deep Learning: A Probabilistic Framework for Classification}, author={Marek Herde and Denis Huseljic and Bernhard Sick}, journal={Transactions on Machine Learning Research}, issn={2835-8856}, year={2023}, url={https://openreview.net/forum?id=MgdoxzImlK}, }