machinelearningnuremberg / WellTunedSimpleNets

[NeurIPS 2021] Well-tuned Simple Nets Excel on Tabular Datasets
Apache License 2.0
77 stars 14 forks source link
automl deep-learning feedforward-neural-network hpo hyperparameter-optimization neurips-2021 regularization state-of-the-art tabular-data

[NeurIPS 2021] Well-tuned Simple Nets Excel on Tabular Datasets

Introduction

This repo contains the source code accompanying the paper:

Well-tuned Simple Nets Excel on Tabular Datasets

Authors: Arlind Kadra, Marius Lindauer, Frank Hutter, Josif Grabocka

Tabular datasets are the last "unconquered castle" for deep learning, with traditional ML methods like Gradient-Boosted Decision Trees still performing strongly even against recent specialized neural architectures. In this paper, we hypothesize that the key to boosting the performance of neural networks lies in rethinking the joint and simultaneous application of a large set of modern regularization techniques. As a result, we propose regularizing plain Multilayer Perceptron (MLP) networks by searching for the optimal combination/cocktail of 13 regularization techniques for each dataset using a joint optimization over the decision on which regularizers to apply and their subsidiary hyperparameters.

We empirically assess the impact of these regularization cocktails for MLPs on a large-scale empirical study comprising 40 tabular datasets and demonstrate that: (i) well-regularized plain MLPs significantly outperform recent state-of-the-art specialized neural network architectures, and (ii) they even outperform strong traditional ML methods, such as XGBoost.

News: Our work is accepted in the Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS 2021).

Setting up the virtual environment

Our work is built on top of AutoPyTorch. To look at our implementation of the regularization cocktail ingredients, you can do the following:

git clone https://github.com/automl/Auto-PyTorch.git
cd Auto-PyTorch/
git checkout regularization_cocktails

To install the version of AutoPyTorch that features our work, you can use these additional commands:

# The following commands assume the user is in the cloned directory
conda create -n reg_cocktails python=3.8
conda activate reg_cocktails
conda install gxx_linux-64 gcc_linux-64 swig
cat requirements.txt | xargs -n 1 -L 1 pip install
python setup.py install

Running the Regularization Cocktail code

The main files to run the regularization cocktails are in the cocktails folder and are main_experiment.py and refit_experiment.py. The first module can be used to start a full HPO search, while, the other module can be used to refit on certain datasets when the time does not suffice to perform the full HPO search and to complete the refit of the incumbent hyperparameter configuration.

The main arguments for main_experiment.py:

A minimal example of running the regularization cocktails:

python main_experiment.py --task_id 233088 --wall_time 600 --func_eval_time 60 --epochs 10 --seed 42 --cash_cocktail True

The example above will run the regularization cocktails for 10 minutes, with a function evaluation limit of 50 seconds for task 233088. Every hyperparameter configuration will be evaluated for 10 epochs, the seed 42 will be used for the experiment and data splits.

A minimal example of running only one regularization method:

python main_experiment.py --task_id 233088 --wall_time 600 --func_eval_time 60 --epochs 10 --seed 42 --use_weight_decay

In case you would like to investigate individual regularization methods, you can look at the different arguments that control them in the main_experiment.py. Additionally, if you want to remove the limit on the number of hyperparameter configurations, you can remove the following lines:

smac_scenario_args={
    'runcount_limit': number_of_configurations_limit,
}

Plots

The plots that are included in our paper were generated from the functions in the module results.py. Although mentioned in most function documentations, most of the functions that plot the baseline diagrams and plots expect a folder structure as follows:

common_result_folder/baseline/results.csv

There are functions inside the module itself that generate the results.csv files.

Baselines

The code for running the baselines can be found in the baselines folder.

TabNet, XGBoost, CatBoost and AutoGluon have the same two main files as our regularization cocktails, main_experiment.py and refit_experiment.py.

Figures

alt text

Citation

@inproceedings{kadra2021well,
  title={Well-tuned Simple Nets Excel on Tabular Datasets},
  author={Kadra, Arlind and Lindauer, Marius and Hutter, Frank and Grabocka, Josif},
  booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
  year={2021}
}