.. DO NOT EDIT. .. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY. .. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE: .. "auto_examples/data_ops/11_choices.py" .. LINE NUMBERS ARE GIVEN BELOW. .. only:: html .. note:: :class: sphx-glr-download-link-note :ref:`Go to the end ` to download the full example code. or to run this example in your browser via JupyterLite or Binder .. rst-class:: sphx-glr-example-title .. _sphx_glr_auto_examples_data_ops_11_choices.py: .. currentmodule:: skrub .. _example_tuning_pipelines: Tuning DataOps ================ A machine-learning pipeline typically contains some values or choices which may influence its prediction performance, such as hyperparameters (e.g. the regularization parameter ``alpha`` of a ``RidgeClassifier``, the ``learning_rate`` of a ``HistGradientBoostingClassifier``), which estimator to use (e.g. ``RidgeClassifier`` or ``HistGradientBoostingClassifier``), or which steps to include (e.g. should we join a table to bring additional information or not). We want to tune those choices by trying several options and keeping those that give the best performance on a validation set. Skrub :ref:`DataOps ` provide a convenient way to specify the range of possible values, by inserting it directly in place of the actual value. For example we can write: ``RidgeClassifier(alpha=skrub.choose_from([0.1, 1.0, 10.0], name='α'))`` instead of: ``RidgeClassifier(alpha=1.0)``. Skrub then inspects our DataOps plan to discover all the places where we used objects like ``skrub.choose_from()`` and builds a grid of hyperparameters for us. .. GENERATED FROM PYTHON SOURCE LINES 35-40 We will illustrate hyperparameter tuning on the "toxicity" dataset. This dataset contains 1,000 texts and the task is to predict if they are flagged as being toxic or not. We start from a very simple pipeline without any hyperparameters. .. GENERATED FROM PYTHON SOURCE LINES 42-55 .. code-block:: Python from sklearn.ensemble import HistGradientBoostingClassifier import skrub import skrub.datasets data = skrub.datasets.fetch_toxicity().toxicity # This dataset is sorted -- all toxic tweets appear first, so we shuffle it data = data.sample(frac=1.0, random_state=1) texts = data[["text"]] labels = data["is_toxic"] .. GENERATED FROM PYTHON SOURCE LINES 56-62 We mark the ``texts`` column as the input variable and the ``labels`` column as the target variable. See `the previous example <10_expressions.html>`_ for a more detailed explanation of ``skrub.X`` and ``skrub.y``. We then encode the text with a ``MinHashEncoder`` and fit a ``HistGradientBoostingClassifier`` on the resulting features. .. GENERATED FROM PYTHON SOURCE LINES 64-67 .. code-block:: Python X = skrub.X(texts) X .. raw:: html
<Var 'X'>
Show graph X: Var 'X'

Result:

Please enable javascript

The skrub table reports need javascript to display correctly. If you are displaying a report in a Jupyter notebook and you see this message, you may need to re-execute the cell or to trust the notebook (button on the top right or "File > Trust notebook").



.. GENERATED FROM PYTHON SOURCE LINES 68-71 .. code-block:: Python y = skrub.y(labels) y .. raw:: html
<Var 'y'>
Show graph y: Var 'y'

Result:

Please enable javascript

The skrub table reports need javascript to display correctly. If you are displaying a report in a Jupyter notebook and you see this message, you may need to re-execute the cell or to trust the notebook (button on the top right or "File > Trust notebook").



.. GENERATED FROM PYTHON SOURCE LINES 72-78 .. code-block:: Python pred = X.skb.apply(skrub.MinHashEncoder()).skb.apply( HistGradientBoostingClassifier(), y=y ) pred.skb.cross_validate(n_jobs=4)["test_score"] .. rst-class:: sphx-glr-script-out .. code-block:: none 0 0.635 1 0.590 2 0.645 3 0.595 4 0.585 Name: test_score, dtype: float64 .. GENERATED FROM PYTHON SOURCE LINES 79-106 For the sake of the example, we will focus on the number of ``MinHashEncoder`` components and the ``learning_rate`` of the ``HistGradientBoostingClassifier`` to illustrate the ``skrub.choose_from(...)`` objects. When we use a scikit-learn hyperparameter-tuner like ``GridSearchCV`` or ``RandomizedSearchCV``, we need to specify a grid of hyperparameters separately from the estimator, with something similar to ``GridSearchCV(my_pipeline, param_grid={"encoder__n_components: [5, 10, 20]"})``. Instead, within a skrub DataOps plan we can use ``skrub.choose_from(...)`` directly where the actual value would normally go. Skrub then takes care of constructing the ``GridSearchCV``'s parameter grid for us. Several utilities are available: - :func:`choose_from` to choose from a discrete set of values - :func:`choose_float` and :func:`choose_int` to sample numbers in a given range - :func:`choose_bool` to choose between ``True`` and ``False`` - :func:`optional` to choose between something and ``None``; typically to make a transformation step optional such as ``X.skb.apply(skrub.optional(StandardScaler()))`` Choices can be given an optional name which is used to display hyperparameter search results and plots, or to override their outcome (ADD A REFERENCE HERE). Note that :func:`skrub.choose_float()` and :func:`skrub.choose_int()` can be given a ``log`` argument to sample in log scale, and that it is possible to specify the number of steps with the ``n_steps`` argument. .. GENERATED FROM PYTHON SOURCE LINES 108-118 .. code-block:: Python X, y = skrub.X(texts), skrub.y(labels) encoder = skrub.MinHashEncoder( n_components=skrub.choose_int(5, 15, n_steps=5, name="N components") ) classifier = HistGradientBoostingClassifier( learning_rate=skrub.choose_float(0.01, 0.9, log=True, name="lr") ) pred = X.skb.apply(encoder).skb.apply(classifier, y=y) .. GENERATED FROM PYTHON SOURCE LINES 119-125 From here, the ``pred`` DataOp can be used to perform hyperparameter search with ``.skb.make_grid_search()`` or ``.skb.make_randomized_search()``. They accept the same arguments as their scikit-learn counterparts (e.g. ``scoring``, ``cv``, ``n_jobs``). Also, like ``.skb.make_learner()``, they accept a ``fitted`` argument: if``fitted=True``, the search is fitted on the data we provided when initializing our pipeline's variables. .. GENERATED FROM PYTHON SOURCE LINES 125-131 .. code-block:: Python search = pred.skb.make_randomized_search( n_iter=8, n_jobs=4, random_state=1, fitted=True ) search.results_ .. raw:: html
N components lr mean_test_score
0 15 0.059618 0.595
1 12 0.255675 0.593
2 15 0.520061 0.585
3 15 0.057289 0.576
4 10 0.105948 0.570
5 10 0.450710 0.551
6 8 0.038979 0.547
7 5 0.015151 0.533


.. GENERATED FROM PYTHON SOURCE LINES 132-145 If the plotly library is installed, we can visualize the results of the hyperparameter search with ``.plot_results()``. In the plot below, each line represents a combination of hyperparameters (in this case, only ``N components`` and ``learning rate``), and each column of points represents either a hyperparameter, or the score of a given combination of hyperparameters. The color of the line represents the score of the combination of hyperparameters. The plot is interactive, and it is possible to select only a subset of the hyperparameters to visualize by dragging the mouse over each column to select the desired range. This is particularly useful when there are many combinations of hyperparameters, and we are interested in understanding which hyperparameters have the largest impact on the score. .. GENERATED FROM PYTHON SOURCE LINES 147-149 .. code-block:: Python search.plot_results() .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 150-172 Default choice values --------------------- The goal of using the different ``choose_*`` functions is to tune choices on validation metrics with randomized or grid search. However, even when our expression contains such choices we can still use it without tuning, for example in previews or to get a quick first result before spending the computation time to run the search. When we use :meth:`.skb.get_pipeline() `, we get a pipeline that does not perform any tuning and uses those default values. This default pipeline is used for :meth:`.skb.eval() `. We can control what should be the default value for each choice. For :func:`choose_int`, :func:`choose_float` and :func:`choose_bool`, we can use the ``default`` parameter. For :func:`choose_from`, the default is the first item from the list or dict of outcomes we provide. For :func:`optional`, we can pass ``default=None`` to force the default to be the alternative outcome, ``None``. When we do not set an explicit default, skrub picks one for depending on the kind of choice, as detailed in :ref:`this table` in the User Guide. .. GENERATED FROM PYTHON SOURCE LINES 174-175 As mentioned we can control the default value: .. GENERATED FROM PYTHON SOURCE LINES 177-179 .. code-block:: Python skrub.choose_float(1.0, 100.0, default=12.0).default() .. rst-class:: sphx-glr-script-out .. code-block:: none 12.0 .. GENERATED FROM PYTHON SOURCE LINES 180-193 Choices can appear in many places --------------------------------- Choices are not limited to selecting estimator hyperparameters. They can also be used to choose between different estimators, or in place of any value used in our pipeline. For example, here we pass a choice to pandas DataFrame's ``assign`` method. We want to add a feature that captures the length of the text, but we are not sure if it is better to count length in characters or in words. We do not want to add both because it would be redundant. We can add a column to the dataframe, which will be chosen among the length in characters or the length in words: .. GENERATED FROM PYTHON SOURCE LINES 195-204 .. code-block:: Python X, y = skrub.X(texts), skrub.y(labels) X.assign( length=skrub.choose_from( {"words": X["text"].str.count(r"\b\w+\b"), "chars": X["text"].str.len()}, name="length", ) ) .. raw:: html
<CallMethod 'assign'>
Show graph X: Var 'X' GetItem 'text' GetItem 'text' CallMethod 'assign' GetAttr 'str' CallMethod 'count' GetAttr 'str' CallMethod 'len'

Result:

Please enable javascript

The skrub table reports need javascript to display correctly. If you are displaying a report in a Jupyter notebook and you see this message, you may need to re-execute the cell or to trust the notebook (button on the top right or "File > Trust notebook").



.. GENERATED FROM PYTHON SOURCE LINES 205-214 ``choose_from`` can be given a dictionary if we want to provide names for the individual outcomes, or a list, when names are not needed: ``choose_from([1, 100], name='N')``, ``choose_from({'small': 1, 'big': 100}, name='N')``. Choices can be nested arbitrarily. For example, here we want to choose between 2 possible encoder types: the ``MinHashEncoder`` or the ``StringEncoder``. Each of the possible outcomes contains a choice itself: the number of components. .. GENERATED FROM PYTHON SOURCE LINES 216-229 .. code-block:: Python X, y = skrub.X(texts), skrub.y(labels) n_components = skrub.choose_int(5, 15, name="N components") encoder = skrub.choose_from( { "minhash": skrub.MinHashEncoder(n_components=n_components), "lse": skrub.StringEncoder(n_components=n_components), }, name="encoder", ) X.skb.apply(encoder, cols="text") .. raw:: html
<Apply MinHashEncoder>
Show graph X: Var 'X' Apply MinHashEncoder

Result:

Please enable javascript

The skrub table reports need javascript to display correctly. If you are displaying a report in a Jupyter notebook and you see this message, you may need to re-execute the cell or to trust the notebook (button on the top right or "File > Trust notebook").



.. GENERATED FROM PYTHON SOURCE LINES 230-234 In a similar vein, we might want to choose between a HGB classifier and a Ridge classifier, each with its own set of hyperparameters. We can then define a choice for the classifier and a choice for the hyperparameters of each classifier. .. GENERATED FROM PYTHON SOURCE LINES 236-246 .. code-block:: Python from sklearn.linear_model import RidgeClassifier hgb = HistGradientBoostingClassifier( learning_rate=skrub.choose_float(0.01, 0.9, log=True, name="lr") ) ridge = RidgeClassifier(alpha=skrub.choose_float(0.01, 100, log=True, name="α")) classifier = skrub.choose_from({"hgb": hgb, "ridge": ridge}, name="classifier") pred = X.skb.apply(encoder).skb.apply(classifier, y=y) print(pred.skb.describe_param_grid()) .. rst-class:: sphx-glr-script-out .. code-block:: none - encoder: 'minhash' N components: choose_int(5, 15, name='N components') classifier: 'hgb' lr: choose_float(0.01, 0.9, log=True, name='lr') - encoder: 'minhash' N components: choose_int(5, 15, name='N components') classifier: 'ridge' α: choose_float(0.01, 100, log=True, name='α') - encoder: 'lse' N components: choose_int(5, 15, name='N components') classifier: 'hgb' lr: choose_float(0.01, 0.9, log=True, name='lr') - encoder: 'lse' N components: choose_int(5, 15, name='N components') classifier: 'ridge' α: choose_float(0.01, 100, log=True, name='α') .. GENERATED FROM PYTHON SOURCE LINES 247-252 .. code-block:: Python search = pred.skb.make_randomized_search( n_iter=16, n_jobs=4, random_state=1, fitted=True ) search.plot_results() .. raw:: html


.. GENERATED FROM PYTHON SOURCE LINES 253-259 Now that we have a more complex plan, we can draw more conclusions from the parallel coordinate plot. For example, we can see that the ``HistGradientBoostingClassifier`` performs better than the ``RidgeClassifier`` in most cases, that the ``StringEncoder`` outperforms the ``MinHashEncoder``, and that the choice of the additional ``length`` feature does not have a significant impact on the score. .. GENERATED FROM PYTHON SOURCE LINES 261-273 Concluding, we have seen how to use skrub's ``choose_from`` objects to tune hyperparameters, choose optional configurations, and nest choices. We then looked at how the different choices affect the plan and the prediction scores. There is more to say about skrub choices than what is covered in this example. In particular, choices are not limited to choosing estimators and their hyperparameters: they can be used anywhere DataOps are used, such as the argument of a :func:`deferred` function, or the argument of other DataOps' method or operator. Finally, choices can be inter-dependent. Please find more information in the :ref:`user guide `. .. rst-class:: sphx-glr-timing **Total running time of the script:** (0 minutes 41.374 seconds) .. _sphx_glr_download_auto_examples_data_ops_11_choices.py: .. only:: html .. container:: sphx-glr-footer sphx-glr-footer-example .. container:: binder-badge .. image:: images/binder_badge_logo.svg :target: https://mybinder.org/v2/gh/skrub-data/skrub/main?urlpath=lab/tree/notebooks/auto_examples/data_ops/11_choices.ipynb :alt: Launch binder :width: 150 px .. container:: lite-badge .. image:: images/jupyterlite_badge_logo.svg :target: ../../lite/lab/index.html?path=auto_examples/data_ops/11_choices.ipynb :alt: Launch JupyterLite :width: 150 px .. container:: sphx-glr-download sphx-glr-download-jupyter :download:`Download Jupyter notebook: 11_choices.ipynb <11_choices.ipynb>` .. container:: sphx-glr-download sphx-glr-download-python :download:`Download Python source code: 11_choices.py <11_choices.py>` .. container:: sphx-glr-download sphx-glr-download-zip :download:`Download zipped: 11_choices.zip <11_choices.zip>` .. only:: html .. rst-class:: sphx-glr-signature `Gallery generated by Sphinx-Gallery `_