Diego: Data in, IntElliGence Out.

简体中文

A fast framework that supports the rapid construction of automated learning tasks. Simply create an automated learning study (Study) and generate correlated trials (Trial). Then run the code and get a machine learning model. Implemented using Scikit-learn API glossary, using Bayesian optimization and genetic algorithms for automated machine learning.

Inspired by Fast.ai and MicroSoft nni.

• the classifier trained by a Study.
• AutoML classifier with support for scikit-learn api. Support for exporting models and use them directly.
• Hyperparametric optimization using Bayesian optimization and genetic algorithms
• Supports bucketing/binning algorithm and LUS sampling method for preprocessing
• Supports scikit-learn api classifier custom classifier for parameter search and super parameter optimization

You need to install swig first, and some rely on C/C interface compilation. Recommended to use conda installation

conda install --yes pip gcc swig libgcc=5.2.0
pip install diego

After installation, start with 6 lines of code to solve a machine learning classification problem.

Each task is considered to be a Study, and each Study consists of multiple Trial. It is recommended to create a Study first and then generate a Trial from the Study:

from diego.study import create_study
import sklearn.datasets
digits = sklearn.datasets.load_digits()
X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(digits.data, digits.target,train_size=0.75, test_size=0.25)

s = create_study(X_train, y_train)
# can use default trials in Study

# or generate one
# s.generate_trials(mode='fast')
s.optimize(X_test, y_test)
# all_trials = s.get_all_trials()
# for t in all_trials:
#     print(t.__dict__)
#     print(t.clf.score(X_test, y_test))

ideas for releases in the future

• 回归。
• add documents.
• 不同类型的Trial。TPE， BayesOpt， RandomSearch
• 自定义的Trial。Trials by custom Classifier (like sklearn, xgboost)
• 模型保存。model persistence
• 模型输出。model output
• basic Classifier
• fix mac os hanged in optimize pipeline
• add preprocessor
• add FeatureTools for automated feature engineering

Study:

Trial:

Since n_jobs>1 may get stuck during parallelization. Similar problems may occur in [scikit-learn] (https://scikit-learn.org/stable/faq.html#why-do-i-sometime-get-a-crash-freeze-with-n -jobs-1-under-osx-or-linux)

In Python 3.4 , one solution is to directly configure multiprocessing to use forkserver or spawn to start process pool management (instead of the default fork). For example, the forkserver mode is enabled globally directly in the code.

import multiprocessing
# other imports, custom code, load data, define model...
if __name__ == '__main__':
    multiprocessing.set_start_method('forkserver')

    # call scikit-learn utils with n_jobs > 1 here

more info :multiprocessing document

For each study, the data storage and parameters, and the model is additionally stored in the Storage object, which ensures that Study only controls trials, and each Trial updates the results in the storage after updating, and updates the best results.

When creating Study, you need to specify the direction of optimization maximize or minimize. Also specify the metrics for optimization when creating Trials. The default is maximize accuracy.

overview

1. fmfn/bayes
2. auto-sklearn

1. H2O.ai

1. hyperopt
2. mlbox

1. pybrain

1.tpot

1. ms nni