autotraino 0.1.0

AutoML for Tabular datasets.

Autotraino :truck:

:warning: Warning, alpha version, everything brakes. :warning:

autotraino is a small wrapper library for AutoML on tabular datasets.

from autotraino.gluon import AutogluonTrainer 
from datasets import load_dataset

train = load_dataset("mstz/adult", "income")["train"].to_pandas() 

# train the model
trainer = AutogluonTrainer()
trainer = trainer.fit(train, target_feature="over_threshold", time_limit=100)

When fitting we can control basic parameters such as where to store the resulting models (parameter save_path of the trainer constructor) or the time budget assigned to the trainer (parameter time_limit, expressed in seconds).

Once trained, we can access the single models

# trained models
print(trainer.names)

print(trainer["LightGBM"])

and predict directly from the Trainer itself:

train_x = train.copy().drop("over_threshold", axis="columns")
predictions = trainer.predict(train_x, with_models=["LightGBM", "RandomForest"])

Quickstart

You can install autotraino via test-pypi:

# optional
mkvirtualenv -p python3.10 autotraino

pip install --extra-index-url https://test.pypi.org/simple/autotraino

Datasets

autotraino is based off pandas.DataFrames. You can find a large collection on a Huggingface repository I'm curating at huggingface.co/mstz. Datasets are sourced from UCI, Kaggle, and OpenML. Most are still to be updated (especially dataset cards).

What model families to train?

Currently based on Autogluon, autotraino currently trains the following models:

• Boosting models
  • LightGBM
  • CatBoost
• Bagging models
  • Random Forest
  • ExtraTree Classifier
• Neural Network
  • FastAI
  • NNTorch
• Classical AI models
  • k-NN
  • Logistic Regression

Preprocessing

autotraino automatically detects feature types and performs the necessary feature preprocessing per model. To ease the process, consider setting the appropriate dtypes in the input pandas.DataFrame.

In the works

Future developments include:

• Fitting arbitrary functions (ray tune)
• Fitting multi-output models.