Note
Machine Learning support is experimental in dask-sql.
We encourage you to try it out and report any issues on our
issue tracker.
Both the training as well as the prediction using Machine Learning methods play a crucial role in
many data analytics applications. dask-sql supports Machine Learning
applications in different ways, depending on how much you would like to do in Python or SQL.
Please also see :ref:`ml` for more information on the SQL statements used on this page.
If you are familiar with Python and the ML ecosystem in Python, this one is probably the simplest possibility. You can use the :func:`~dask_sql.Context.sql` call as described before to extract the data for your training or ML prediction. The result will be a Dask dataframe, which you can either directly feed into your model or convert to a pandas dataframe with .compute() before.
This gives you full control on the training process and the simplicity of using SQL for data manipulation. You can use this method in your python scripts or Jupyter notebooks, but not from the :ref:`server` or :ref:`cmd`.
In many companies/teams, it is typical that some team members are responsible for
creating/training a ML model, and others use it to predict unseen data.
It would be possible to create a custom function (see :ref:`custom`) to load and use the model,
which then can be used in SELECT queries.
However for convenience, dask-sql introduces a SQL keyword to do this work for you
automatically. The syntax is similar to the BigQuery Predict Syntax.
c.sql("""
SELECT * FROM PREDICT ( MODEL my_model,
SELECT x, y, z FROM data
)
""")This call will first collect the data from the inner SELECT call (which can be any valid
SELECT call, including JOIN, WHERE, GROUP BY, custom tables and views etc.)
and will then apply the model with the name "my_model" for prediction.
The model needs to be registered at the context before using :func:`~dask_sql.Context.register_model`.
c.register_model("my_model", model)The model registered here can be any valid python object, which follows the scikit-learn
interface, which is to have a predict() function.
Please note that the input will not be pandas dataframe, but a Dask dataframe.
See :ref:`ml` for more information.
This method, in contrast to the other two possibilities, works completely from SQL,
which allows you to also call it e.g. from your BI tool.
Additionally to the PREDICT keyword mentioned above, dask-sql also has a way to
create and train a model from SQL:
.. tabs::
.. group-tab:: CPU
.. code-block:: sql
CREATE MODEL my_model WITH (
model_class = 'sklearn.linear_model.LogisticRegression',
wrap_predict = True,
target_column = 'target'
) AS (
SELECT x, y, x*y > 0 as target
FROM timeseries
LIMIT 100
)
.. group-tab:: GPU
.. code-block:: sql
CREATE MODEL my_model WITH (
model_class = 'cuml.linear_model.LogisticRegression',
wrap_predict = True,
target_column = 'target'
) AS (
SELECT x, y, x*y > 0 as target
FROM timeseries
LIMIT 100
)
This call will create a new instance of linear_model.LogisticRegression
and train it with the data collected from the SELECT call (again, every valid SELECT
query can be given). The model can than be used in subsequent calls to PREDICT
using the given name.
We set wrap_predict = True here to parallelize post fit prediction task of non distributed models (sklearn/cuML etc) across workers.
Have a look into :ref:`ml` for more information.
After the model was trained, you can inspect and get model details by using the following sql statements
-- show the list of models which are trained and stored in the context.
SHOW MODELS
-- To get the hyperparameters of the trained MODEL, use
-- DESCRIBE MODEL <model_name>.
DESCRIBE MODEL my_modelWant to increase the performance of your model by tuning the parameters? Use the hyperparameter tuning directly in SQL using below SQL syntax, choose different tuners based on memory and compute constraints.
CREATE EXPERIMENT my_exp WITH (
model_class = 'sklearn.ensemble.GradientBoostingClassifier',
experiment_class = 'sklearn.model_selection.GridSearchCV',
tune_parameters = (n_estimators = ARRAY [16, 32, 2],
learning_rate = ARRAY [0.1,0.01,0.001],
max_depth = ARRAY [3,4,5,10]
),
target_column = 'target'
) AS (
SELECT x, y, x*y > 0 AS target
FROM timeseries
LIMIT 100
)Want to try different models with different parameters in SQL? Now you can
start automl experiments with the help of the tpot framework which trains
and evaluates a number of different sklearn compatible models and uses dask for
distributing the work across the dask clusters.
Use below SQL syntax for automl and for more details refer to the
tpot automl framework
CREATE EXPERIMENT my_exp WITH (
automl_class = 'tpot.TPOTClassifier',
automl_kwargs = (population_size = 2 ,
generations=2,
cv=2,
n_jobs=-1,
use_dask=True,
max_eval_time_mins=1),
target_column = 'target'
) AS (
SELECT x, y, x*y > 0 AS target
FROM timeseries
LIMIT 100
)After the experiment was completed, both hyperparameter tuner and automl experiments stores the best model of the experiment in the sql context with the name same as the experiment name, which can be used for prediction.
Once your model was trained and performs good in your validation dataset, you can export the model into a file with one of the supported model serialization formats like pickle, joblib, mlflow (framework-agnostic serialization format), etc.
Currently, dask-sql supports the pickle, joblib and mlflow format for exporting the trained model, which can then be deployed as microservices etc
Before training and exporting the models from different framework like lightgbm, catboost, please ensure the relevant packages are installed in the dask-sql environment, otherwise it will raise an exception on import and if you are using mlflow as format ensure mlflow was installed. Keep in mind that dask-sql supports only sklearn compatible model (i.e fit-predict style models) so far, so instead of using xgb.core.Booster consider using xgboost.XGBClassifier since later is sklearn compatible and used by dask-sql for training, prediction and exporting the model through standard sklearn interface
-- for pickle model serialization
EXPORT MODEL my_model WITH (
format ='pickle',
location = 'model.pkl'
)
-- for joblib model serialization
EXPORT MODEL my_model WITH (
format ='joblib',
location = 'model.pkl'
)
-- for mlflow model serialization
EXPORT MODEL my_model WITH (
format ='mlflow',
location = 'mlflow_dir'
)
-- Note you can pass more number of key value pairs
-- (parameters) which will be delegated to the respective
-- export functionsThe following SQL-only code gives an example on how the commands can play together.
We assume that you have created/registered a table "my_data" with the numerical columns x and y
and the boolean target label.
-- First, we create a new feature z out of x and y.
-- For convenience, we store it in another table
CREATE OR REPLACE TABLE transformed_data AS (
SELECT x, y, x + y AS z, label
FROM my_data
)
-- We split the data into a training set
-- by using the first 100 items.
-- Please note that this is just for a very quick-and-dirty
-- example - you would probably want to do something
-- more advanced here, maybe with TABLESAMPLE
CREATE OR REPLACE TABLE training_data AS (
SELECT * FROM transformed_data
LIMIT 15
)
-- Quickly check the data
SELECT * FROM training_data
-- We can now train a model from the sklearn package.
CREATE OR REPLACE MODEL my_model WITH (
model_class = 'sklearn.ensemble.GradientBoostingClassifier',
wrap_predict = True,
target_column = 'label'
) AS (
SELECT * FROM training_data
)
-- Now apply the trained model on all the data
-- and compare.
SELECT
*, (CASE WHEN target = label THEN True ELSE False END) AS correct
FROM PREDICT(MODEL my_model,
SELECT * FROM transformed_data
)
-- list models
SHOW MODELS
-- check parameters of the model
DESCRIBE MODEL my_model
-- experiment to tune different hyperparameters
CREATE EXPERIMENT my_exp WITH(
model_class = 'sklearn.ensemble.GradientBoostingClassifier',
experiment_class = 'sklearn.model_selection.GridSearchCV',
tune_parameters = (n_estimators = ARRAY [16, 32, 2],
learning_rate = ARRAY [0.1,0.01,0.001],
max_depth = ARRAY [3,4,5,10]
),
target_column = 'label'
) AS (
SELECT * FROM training_data
)
-- creates experiment with automl framework
CREATE EXPERIMENT my_exp WITH (
automl_class = 'tpot.TPOTRegressor',
automl_kwargs = (population_size = 2 ,
generations=2,
cv=2,
n_jobs=-1,
use_dask=True,
max_eval_time_mins=1),
target_column = 'z'
) AS (
SELECT * FROM training_data
)
-- checks the parameter of automl model
DESCRIBE MODEL automl_TPOTRegressor
-- export model
EXPORT MODEL my_model WITH (
format ='pickle',
location = 'model.pkl'
)