TPOT stands for Tree-based Pipeline Optimization Tool. TPOT2 is a Python Automated Machine Learning tool that optimizes machine learning pipelines using genetic programming. Consider TPOT2 your Data Science Assistant.
TPOT2 is a rewrite of TPOT with some additional functionality. Notably, we added support for graph-based pipelines and additional parameters to better specify the desired search space. TPOT2 is currently in Alpha. This means that there will likely be some backwards incompatible changes to the API as we develop. Some implemented features may be buggy. There is a list of known issues written at the bottom of this README. Some features have placeholder names or are listed as "Experimental" in the doc string. These are features that may not be fully implemented and may or may work with all other features.
If you are interested in using the current stable release of TPOT, you can do that here: https://github.com/EpistasisLab/tpot/.
Please see the repository license for the licensing and usage information for TPOT2. Generally, we have licensed TPOT2 to make it as widely usable as possible.
TPOT2 requires a working installation of Python.
We recommend using conda environments for installing TPOT2, though it would work equally well if manually installed without it.
More information on making anaconda environments found here.
conda create --name tpot2env python=3.10
conda activate tpot2env
You need to install the lightgbm package directly from conda using the following command before installing TPOT2.
This is to ensure that you get the version that is compatible with your system.
conda install --yes -c conda-forge 'lightgbm>=3.3.3'
pip install -e /path/to/tpot2repo
If you downloaded with git pull, then the repository folder will be named TPOT2. (Note: this folder is the one that includes setup.py inside of it and not the folder of the same name inside it). If you downloaded as a zip, the folder may be called tpot2-main.
See the Tutorials Folder for more instructions and examples.
TPOT2 uses dask for parallel processing. When Python is parallelized, each module is imported within each processes. Therefore it is important to protect all code within a if __name__ == "__main__"
when running TPOT2 from a script. This is not required when running TPOT2 from a notebook.
For example:
#my_analysis.py
import tpot2
if __name__ == "__main__":
X, y = load_my_data()
est = tpot2.TPOTClassifier()
est.fit(X,y)
#rest of analysis
When designing custom objective functions, avoid the use of global variables.
Don't Do:
global_X = [[1,2],[4,5]]
global_y = [0,1]
def foo(est):
return my_scorer(est, X=global_X, y=global_y)
Instead use a partial
from functools import partial
def foo_scorer(est, X, y):
return my_scorer(est, X, y)
if __name__=='__main__':
X = [[1,2],[4,5]]
y = [0,1]
final_scorer = partial(foo_scorer, X=X, y=y)
Similarly when using lambda functions.
Dont Do:
def new_objective(est, a, b)
#definition
a = 100
b = 20
bad_function = lambda est : new_objective(est=est, a=a, b=b)
Do:
def new_objective(est, a, b)
#definition
a = 100
b = 20
good_function = lambda est, a=a, b=b : new_objective(est=est, a=a, b=b)
TPOT2 will not check if your data is correctly formatted. It will assume that you have passed in operators that can handle the type of data that was passed in. For instance, if you pass in a pandas dataframe with categorical features and missing data, then you should also include in your configuration operators that can handle those feautures of the data. Alternatively, if you pass in preprocessing = True
, TPOT2 will impute missing values, one hot encode categorical features, then standardize the data. (Note that this is currently fitted and transformed on the entire training set before splitting for CV. Later there will be an option to apply per fold, and have the parameters be learnable.)
Setting verbose
to 5 can be helpful during debugging as it will print out the error generated by failing pipelines.
We welcome you to check the existing issues for bugs or enhancements to work on. If you have an idea for an extension to TPOT, please file a new issue so we can discuss it.
- TPOT2 uses the func_timeout package to terminate long running pipelines. The early termination signal may fail on particular estimators and cause TPOT2 to run for longer than intended. If you are using your own custom configuration dictionaries, and are noticing that TPOT2 is running for longer than intended, this may be the issue. We are currently looking into it. Sometimes restarting TPOT2 resolves the issue.
- Periodic checkpoint folder may not correctly resume if using budget and/or initial_population size.
- Population class is slow to add new individuals. The Population class needs to be updated to use a dictionary for storage rather than a pandas dataframe.
- Crossover may sometimes go over the size restrictions.
- Memory caching with GraphPipeline may miss some nodes where the ordering on inputs happens to be different between two nodes.
TPOT2 was developed in the Artificial Intelligence Innovation (A2I) Lab at Cedars-Sinai with funding from the NIH under grants U01 AG066833 and R01 LM010098. We are incredibly grateful for the support of the NIH and the Cedars-Sinai during the development of this project.
The TPOT logo was designed by Todd Newmuis, who generously donated his time to the project.