Xingu for automated ML model training

Xingu is a framework of a few classes that helps on full industrialization of your machine learning training and deployment pipelines. Just write your DataProvider class, mostly in a declarative way, that completely controls your training and deployment pipeline.

Notebooks are useful in EDA time, but when the modeling is ready to become a product, use Xingu proposed classes to organize interactions with DB (queries), data cleanup, feature engineering, hyper-parameters optimization, training algorithm, general and custom metrics computation, estimation post-processing.

Don’t save a pickle at the end of your EDA, let Xingu organize a versioned inventory of saved models (PKLs) linked and associated to commit hashes and branches of your code.

Don’t save metrics manually and in an informal way. Metrics are first class citizens, so use Xingu to write methods that compute metrics and let it store metrics in an organized database that can be queried and compared.

Don’t make ad-hoc plots to understand your data. Plots are important assets to measure the quality of your model, so use Xingu to write methods that formaly generate versioned plots.

Do not worry or even write code that loads pre-req models, use Xingu pre-req architecture to load pre-req models for you and package them together.

Don’t save ad-hoc hypermaters after optimizations. Let Xingu store and manage those for you in a way that can be reused in future trains.

Don’t change your code if you want different functionality. Use Xingu environment variables or command line parameters to strategize your trains.

Don’t manually copy PKLs to production environments on S3 or other object storage. Use Xingu’s deployment tools to automate the deployment step.

Don’t write database integration code. Just provide your queries and Xingu will give you the data. Xingu will also maintain a local cache of your data so you won’t hammer your database across multiple retrains. Do the same with static data files with parquet, CSV, on local filesystem or object storage.

Xingu can run anyware, from your laptop, with a plain SQLite database, to large scale cloud-powered training pipelines with GitOps, Jenkins, Docker etc. Xingu’s database is used only to cellect training information, it isn´t required later when model is used to predict.

Install

pip install https://github.com/avibrazil/xingu

or

pip install xingu

Use to Train a Model

Check your project has the necessary files and folders:

$ find
dataproviders/
dataproviders/my_dataprovider.py
estimators/
estimators/myrandomestimator.py
models/
data/
plots/

Train with DataProviders id_of_my_dataprovider1 and id_of_my_dataprovider2, both defined in dataproviders/my_dataprovider.py:

$ xingu \
--dps id_of_my_dataprovider1,id_of_my_dataprovider2 \
--databases athena "awsathena+rest://athena.us..." \
--query-cache-path data \
--trained-models-path models \
--debug

Use the API

See the proof of concept notebook with vairous usage scenarios:

• POC 1. Train some Models
• POC 2. Use Pre-Trained Models for Batch Predict
• POC 3. Assess Metrics and create Comparative Reports
• POC 4. Check and report how Metrics evolved
• POC 5. Play with Xingu barebones
• POC 6. Play with the ConfigManager
• POC 7. Xingu Estimators in the Command Line
• POC 8. Deploy Xingu Data and Estimators between environments (laptop, staging, production etc)

Procedures defined by Xingu

Xingu classes do all the heavy lifting while you focus on your machine learning code only.

Class Coach is responsible of coordinating the training process of one or multiple models. You control parallelism via command line or environment variables.

Class Model implements a standard pipelines for train, train with hyperparam optimization, load and save pickles, database access etc. These pipelines are is fully controlled by your DataProvider or the environment.

Class DataProvider is a base class that is constantly queried by the Model to determine how the Model should operate. Your should create a class derived from DataProvider and reimplement whatever you want to change. This will completely change behaviour of Model operation in a way that you´ll get a completelly different model.

• It is your DataProvider that defines the source of training data as SQL queries or URLs of parquets, CSVs, JSONs
• It is your DataProvider that defines how multi-source data should be integrated
• It is your DataProvider that defines how data should be split into train and test sets
• Your DataProvider defines which Estimator class to use
• Your DataProvider defines how the Estimator should be initialized and optimized
• Your DataProvider defines which metrics should be computed, how to compute them and against which dataset
• Your DataProvider defines which plots should be created and against which dataset
• See below when and how each method of your DataProvider will be called by xingu.Model

Class Estimator is another base class (that you can reimplement) to contain estimator-specific affairs. There will be an Estimator-derived class for an XGBoostRegressor, other for a CatBoostClassifier, other for a SciKit-Learn-specific algorithm, including hyperparam optimization logic and libraries. A concrete Estimator class can and should be reused across multiple different models.

The hierarchical diagrams below expose complete Xingu pipelines with all their steps. Steps marked with 💫 are were you put your code. All the rest is Xingu boilerplate code ready to use.

Coach.team_train():

Train various Models, all possible in parallel.

1. Coach.team_train_parallel() (background, parallelism controled by PARALLEL_TRAIN_MAX_WORKERS):

2. Coach.team_load() (for pre-req models not trained in this session)

3. Per DataProvider requested to be trained:

4. Coach.team_train_member() (background):

5. Model.fit() calls:

6. 💫DataProvider.get_dataset_sources_for_train() return dict of queries and/or URLs

7. Model.data_sources_to_data(sources)

8. 💫DataProvider.clean_data_for_train(dict of DataFrames)

9. 💫DataProvider.feature_engineering_for_train(DataFrame)

10. 💫DataProvider.last_pre_process_for_train(DataFrame)

11. 💫DataProvider.data_split_for_train(DataFrame) return tuple of dataframes

12. Model.hyperparam_optimize() (decide origin of hyperparam)

13. 💫DataProvider.get_estimator_features_list()

14. 💫DataProvider.get_target()

15. 💫DataProvider.get_estimator_optimization_search_space()

16. 💫DataProvider.get_estimator_hyperparameters()

17. 💫Estimator.hyperparam_optimize() (SKOpt, GridSearch et all)

18. 💫Estimator.hyperparam_exchange()

19. 💫DataProvider.post_process_after_hyperparam_optimize()

20. 💫Estimator.fit()

21. 💫DataProvider.post_process_after_train()

22. Coach.post_train_parallel() (background, only if POST_PROCESS=true):

23. Per trained Model (parallelism controled by PARALLEL_POST_PROCESS_MAX_WORKERS):

24. Model.save() (PKL save in background)

25. Model.trainsets_save() (save the train datasets, background)

26. Model.trainsets_predict():

27. Model.predict_proba() or Model.predict() (see below)

28. 💫DataProvider.pre_process_for_trainsets_metrics()

29. Model.compute_and_save_metrics(channel=trainsets) (see below)

30. 💫DataProvider.post_process_after_trainsets_metrics()

31. Coach.single_batch_predict() (see below)

Coach.team_batch_predict():

Load from storage and use various pre-trained Models to estimate data from a pre-defined SQL query. The batch predict SQL query is defined into the DataProvider and this process will query the database to get it.

1. Coach.team_load() (for all requested DPs and their pre-reqs)

2. Per loaded model:

3. Coach.single_batch_predict() (background)

4. Model.batch_predict()

5. 💫DataProvider.get_dataset_sources_for_batch_predict()

6. Model.data_sources_to_data()

7. 💫DataProvider.clean_data_for_batch_predict()

8. 💫DataProvider.feature_engineering_for_batch_predict()

9. 💫DataProvider.last_pre_process_for_batch_predict()

10. Model.predict_proba() or Model.predict() (see below)

11. Model.compute_and_save_metrics(channel=batch_predict) (see below)

12. Model.save_batch_predict_estimations()

Model.predict() and Model.predict_proba():

1. Model.generic_predict()

2. 💫DataProvider.pre_process_for_predict() or DataProvider.pre_process_for_predict_proba()

3. 💫DataProvider.get_estimator_features_list()

4. 💫Estimator.predict() or Estimator.predict_proba()

5. 💫DataProvider.post_process_after_predict() or DataProvider.post_process_after_predict_proba()

Model.compute_and_save_metrics():

Sub-system to compute various metrics, graphics and transformations over a facet of the data.

This is executed right after a Model was trained and also during a batch predict.

Predicted data is computed before Model.compute_and_save_metrics() is called. By Model.trainsets_predict() and Model.batch_predict().

1. Model.save_model_metrics() calls:

2. Model.compute_model_metrics() calls:

3. Model.compute_trainsets_model_metrics() calls:

4. All Model.compute_trainsets_model_metrics_{NAME}()

5. All 💫DataProvider.compute_trainsets_model_metrics_{NAME}()

6. Model.compute_batch_model_metrics() calls:

7. All Model.compute_batch_model_metrics_{NAME}()

8. All 💫DataProvider.compute_batch_model_metrics_{NAME}()

9. Model.compute_global_model_metrics() calls:

10. All Model.compute_global_model_metrics_{NAME}()

11. All 💫DataProvider.compute_global_model_metrics_{NAME}()

12. Model.render_model_plots() calls:

13. Model.render_trainsets_model_plots() calls:

14. All Model.render_trainsets_model_plots_{NAME}()

15. All 💫DataProvider.render_trainsets_model_plots_{NAME}()

16. Model.render_batch_model_plots() calls:

17. All Model.render_batch_model_plots_{NAME}()

18. All 💫DataProvider.render_batch_model_plots_{NAME}()

19. Model.render_global_model_plots() calls:

20. All Model.render_global_model_plots_{NAME}()

21. All 💫DataProvider.render_global_model_plots_{NAME}()

22. Model.save_estimation_metrics() calls:

23. Model.compute_estimation_metrics() calls:

24. All Model.compute_estimation_metrics_{NAME}()

25. All 💫DataProvider.compute_estimation_metrics_{NAME}()