Counting 1,868 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
Oct. 17, 2017
Aug. 5, 2017

AutoML Service

Deploy automated machine learning (AutoML) as a service using Flask, for both pipeline training and pipeline serving.

The framework implements a fully automated time series classification pipeline, automating both feature engineering and model selection and optimization using Python libraries, TPOT and tsfresh.

Check out the blog post for more info.


  • TPOT– Automated feature preprocessing and model optimization tool
  • tsfresh– Automated time series feature engineering and selection
  • Flask– A web development microframework for Python


The application exposes both model training and model predictions with a RESTful API. For model training, input data and labels are sent via POST request, a pipeline is trained, and model predictions are accessible via a prediction route.

Pipelines are stored to a unique key, and thus, live predictions can be made on the same data using different feature construction and modeling pipelines.

An automated pipeline for time-series classification.

The model training logic is exposed as a REST endpoint. Raw, labeled training data is uploaded via a POST request and an optimal model is developed.

Raw training data is uploaded via a POST request and a model prediction is returned.

Using the app

View the Jupyter Notebook for an example.


# deploy locally
# deploy on cloud foundry
cf push


Train a pipeline:

train_url = ''
train_files = {'raw_data': open('data/data_train.json', 'rb'),
               'labels'  : open('data/label_train.json', 'rb'),
               'params'  : open('parameters/train_parameters_model2.yml', 'rb')}

# post request to train pipeline
r_train =, files=train_files)
result_df = json.loads(r_train.json())


{'featureEngParams': {'default_fc_parameters': "['median', 'minimum', 'standard_deviation', 
                                                 'sum_values', 'variance', 'maximum', 
                                                 'length', 'mean']",
                      'impute_function': 'impute',
 'mean_cv_accuracy': 0.865,
 'mean_cv_roc_auc': 0.932,
 'modelId': 1,
 'modelType': "Pipeline(steps=[('stackingestimator', StackingEstimator(estimator=LinearSVC(...))),
                               ('logisticregression', LogisticRegressionClassifier(solver='liblinear',...))])"
 'trainShape': [1647, 8],
 'trainTime': 1.953}

Serve pipeline predictions:

serve_url = ''
test_files = {'raw_data': open('data/data_test.json', 'rb'),
              'params' : open('parameters/test_parameters_model2.yml', 'rb')}

# post request to serve predictions from trained pipeline
r_test  =, files=test_files)
result = pd.read_json(r_test.json()).set_index('id')
example_id prediction
1 0.853
2 0.991
3 0.060
4 0.995
5 0.003
... ...

View all trained models:

r = requests.get('')
pipelines = json.loads(r.json())
    {'mean_cv_accuracy': 0.873,
     'modelType': "RandomForestClassifier(...),
    {'mean_cv_accuracy': 0.895,
     'modelType': "GradientBoostingClassifier(...),
    {'mean_cv_accuracy': 0.859,
     'modelType': "LogisticRegressionClassifier(...),

Running the tests

Supply a user argument for the host.

# use local app
py.test --host
# use cloud-deployed app
py.test --host http://ROUTE-HERE

Scaling the architecture

For production, I would suggest splitting training and serving into seperate applications, and incorporating a fascade API. Also it would be best to use a shared cache such as Redis or Pivotal Cloud Cache to allow other applications and multiple instances of the pipeline to access the trained model. Here is a potential architecture.

A scalable model training and model serving architecture.


Chris Rawles