README.md

Repository for reproduction of paper "Rethinking of Encoder-Based Warmstart Methods in Hyperparameter Optimization"

Running experiments

Full dataset

To reproduce experiments using full you need to have computer with linux opearing system and Python with version 3.10.12. First step is to install requirements using command

pip install -r requirements.txt

Second step is to run command which runs all steps and experiments:

./reproduction.sh

After that all results should be in directory results/. If you have permission denied error you need to run:

    chmod +x reproduction.sh

Toy dataset

To reproduce experiments with toy dataset you need:

• Rename directory data_toy to data i. e. mv data_toy data. NOTE: you need to rename current data folder to something else
• Rename file config/uci_splits_toy.json to config/uci_splits.json i. e. mv config/uci_splits_toy.json config/uci_splits.json. NOTE: you need to rename current config/uci_splits.json folder to something else.
• Run all steps from section Full dataset

Repository contents

├── bin - directory with all executable files used to reproduce results
├── config - files with models and data configurations
├── data - directory with data from openml and uci
├── data_toy - minimal example needed to run experiments
├── engine - directory with files containing some logic behind our experiments
│   ├── cd_plot.py - generating critical distance plots
│   ├── dataset2vec - dataset2vec implementation
│   └── random_hpo - generating random hyperparameters using conditional spaces
├── models - pretrained models
├── README.md
├── reproduction.sh - scripts to reproduce experiments
├── requirements.txt - requirements to python environment needed for experiments run