https://github.com/societe-generale/aikit
Automated machine learning package
https://github.com/societe-generale/aikit
automl data-science machine-learning python
Last synced: 5 months ago
JSON representation
Automated machine learning package
- Host: GitHub
- URL: https://github.com/societe-generale/aikit
- Owner: societe-generale
- License: bsd-2-clause
- Created: 2018-09-14T07:28:28.000Z (about 7 years ago)
- Default Branch: master
- Last Pushed: 2023-03-08T09:04:47.000Z (almost 3 years ago)
- Last Synced: 2024-11-20T09:22:31.401Z (about 1 year ago)
- Topics: automl, data-science, machine-learning, python
- Language: Python
- Homepage: http://aikit.readthedocs.io/
- Size: 1.5 MB
- Stars: 27
- Watchers: 10
- Forks: 11
- Open Issues: 31
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- awesome-automl - societe-generale/aikit
README
[](https://github.com/societe-generale/aikit)
[](https://badge.fury.io/py/aikit)
[](https://aikit.readthedocs.io/en/latest/?badge=latest)
[](https://mybinder.org/v2/gh/societe-generale/aikit/master?filepath=/notebooks)
# aikit
Automatic Tool Kit for Machine Learning and Datascience.
The objective is to provide tools to ease the repetitive part of the DataScientist job and so that he/she can focus on modelization. This package is still in alpha and more features will be added.
Its mains features are:
* improved and new "scikit-learn like" transformers ;
* GraphPipeline : an extension of sklearn Pipeline that handles more generic chains of tranformations ;
* an AutoML to automatically search throught several transformers and models.
Full documentation is available here: https://aikit.readthedocs.io/en/latest/
You can run examples [here](https://mybinder.org/v2/gh/societe-generale/aikit/master?filepath=/notebooks), thanks to [Binder](https://mybinder.org).
### GraphPipeline
The GraphPipeline object is an extension of `sklearn.pipeline.Pipeline` but the transformers/models can be chained with any directed graph.
The objects takes as input two arguments:
* models: dictionary of models (each key is the name of a given node, and each corresponding value is the transformer corresponding to that node)
* edges: list of tuples that links the nodes to each other
Example:
```python
gpipeline = GraphPipeline(
models = {
"vect": CountVectorizerWrapper(analyzer="char",
ngram_range=(1, 4),
columns_to_use=["text1", "text2"]),
"cat": NumericalEncoder(columns_to_use=["cat1", "cat2"]),
"rf": RandomForestClassifier(n_estimators=100)
},
edges = [("vect", "rf"), ("cat", "rf")]
)
```
### AutoML
Aikit contains an AutoML part which will test several models and transformers for a given dataset.
For example, you can create the following python script `run_automl_titanic.py`:
```python
from aikit.datasets import load_dataset, DatasetEnum
from aikit.ml_machine import MlMachineLauncher
def loader():
dfX, y, *_ = load_dataset(DatasetEnum.titanic)
return dfX, y
def set_configs(launcher):
""" modify that function to change launcher configuration """
launcher.job_config.score_base_line = 0.75
launcher.job_config.allow_approx_cv = True
return launcher
if __name__ == "__main__":
launcher = MlMachineLauncher(base_folder = "~/automl/titanic",
name = "titanic",
loader = loader,
set_configs = set_configs)
launcher.execute_processed_command_argument()
```
And then run the command:
```
python run_automl_titanic.py run -n 4
```
To run the automl using 4 workers, the results will be stored in the specified folder
You can aggregate those result using:
```
python run_automl_titanic.py result
```