Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


dswizard is an efficient solver for machine learning (ML) pipeline synthesis inspired by human behaviour. It automatically derives a pipeline structure, selects algorithms and performs hyperparameter optimization. This repository contains the source code and data used in our publication Iterative Search Space Construction for Pipeline Structure Search.

How to install

The code has only be tested with Python 3.8, but any version supporting type hints should work. We recommend using a virtual environment.

python3 -m virtualenv venv
source venv/bin/activate

dswizard is available on PyPI, you can simply install it via

pip install dswizard

Alternatively, you can checkout the source code and install it directly via

pip install -e dswizard

Now you are ready to go.


dswizard contains an optional pipeline search space visualization functionality intended for debugging and explainability. If you don't need this feature, you can skip this step. To use the visualization you have to install Graphviz manually and add the additional visualization libraries using

pip install dswizard[visualization]


In the folder scripts, we have provided scripts to showcase usage of dswizard. The most important script is scripts/ This script solves the pipeline synthesis for a given task. To get usage information use

python dswizard/scripts/ --help

yielding a output similar to

usage: [-h] [--wallclock_limit WALLCLOCK_LIMIT] [--cutoff CUTOFF] [--log_dir LOG_DIR] task

Example 1 - dswizard optimization.

positional arguments:
  task                  OpenML task id

optional arguments:
  -h, --help            show this help message and exit
  --wallclock_limit WALLCLOCK_LIMIT
                        Maximum optimization time for in seconds
  --cutoff CUTOFF       Maximum cutoff time for a single evaluation in seconds
  --log_dir LOG_DIR     Directory used for logging
  --fold FOLD           Fold of OpenML task to optimize

You have to pass an OpenML task id. For example, to create pipelines for the kc2 data set use python dswizard/scripts/ 3913. Via the optional parameter you can change the total optimization time (default 300 seconds), maximum evaluation time for a single configuration (default 60 seconds), the directory to store optimization artifacts (default run/{TASK}) and the fold to evaluate (default 0).

The optimization procedure prints the best found pipeline structure with the according configuration and test performance to the console., similar to

2020-11-13 16:45:55,312 INFO     root            MainThread Best found configuration: [('22', KBinsDiscretizer), ('26', PCAComponent), ('28', AdaBoostingClassifier)]
  22:encode, Value: 'ordinal'
  22:n_bins, Value: 32
  22:strategy, Value: 'kmeans'
  26:keep_variance, Value: 0.9145797030897109
  26:whiten, Value: True
  28:algorithm, Value: 'SAMME'
  28:learning_rate, Value: 0.039407336108331845
  28:n_estimators, Value: 138
 with loss -0.8401893431635389
2020-11-13 16:45:55,312 INFO     root            MainThread A total of 20 unique structures where sampled.
2020-11-13 16:45:55,312 INFO     root            MainThread A total of 58 runs where executed.
2020-11-13 16:45:55,316 INFO     root            MainThread Final pipeline:
FlexiblePipeline(configuration={'22:encode': 'ordinal', '22:n_bins': 32,
                                '22:strategy': 'kmeans',
                                '26:keep_variance': 0.9145797030897109,
                                '26:whiten': True, '28:algorithm': 'SAMME',
                                '28:learning_rate': 0.039407336108331845,
                                '28:n_estimators': 138},
                 steps=[('22', KBinsDiscretizer(encode='ordinal', n_bins=32, strategy='kmeans')),
                        ('26', PCAComponent(keep_variance=0.9145797030897109, whiten=True)),
                        ('28', AdaBoostingClassifier(algorithm='SAMME', learning_rate=0.039407336108331845, n_estimators=138))])
2020-11-13 16:45:55,828 INFO     root            MainThread Final test performance -0.8430735930735931

Additionally, an ensemble of the evaluated pipeline candidates is constructed.

2020-11-13 16:46:06,371 DEBUG    Ensemble        MainThread Building bagged ensemble
2020-11-13 16:46:09,606 DEBUG    Ensemble        MainThread Ensemble constructed
2020-11-13 16:46:10,472 INFO     root            MainThread Final ensemble performance -0.8528138528138528 based on 11 pipelines

In the log directory (default run/{task}) four files are stored:

  1. log.txt contains the complete logging output
  2. results.json contains detailed information about all evaluated hyperparameter configurations.
  3. search_graph.pdf is a visual representation of the internal pipeline structure graph.
  4. structures.json contains all tested pipeline structures including the list of algorithms and the complete configuration space.


To assess the performance of dswizard we have implemented an adapter for the OpenML automlbenchmark available here. Please refer to that repository for benchmarking dswizard. The file scripts/, scripts/ and scripts/ are used to compare dswizard with autosklearn and tpot, both also evaluated via automlbenchmark.


The meta-learning base used in this repository is created using meta-learning-base. Please see this repository on how to create the required meta-learning models.

For simplicity, we directly provide a random and sgd forest regression model trained on all available data ready to use. It is available in in dswizard/assets/. scripts/ is already configured to use this model.

The data used to train the regression model is also available online. Please refer to meta-learning-base to see how to train the model from the raw data.


This repository only contains the optimization logic. The actual basic ML components to be optimized are available in dswizard-components. Currently, only sklearn components are supported.


No description, website, or topics provided.




No releases published


No packages published