Creating a task

This document describes how to write a data processing task. For an introduction to data processing tasks, please read Overview of the task framework. It also helps to have a basic understanding of data repositories and how to use the butler to read and write data (to be written; for now read existing tasks to see how it is done).

After reading this document you may wish to read Creating a PipelineTask.

Introduction

Tasks are subclasses of lsst.pipe.base.Task in general or lsst.pipe.base.PipelineTask for tasks that should be run within pipelines.

Tasks are constructed by calling __init__ with the task configuration. Occasionally additional arguments are required (see the task’s documentation for details). lsst.pipe.base.Task has a few other arguments that are usually only specified when a task is created as a subtask of another task; you will probably never have to specify them yourself.

Tasks typically have a run method that executes the task’s main function. See Methods for more information.

Configuration

Every task requires a configuration: a task-specific set of configuration parameters. The configuration is read-only; once you construct a task, the same configuration will be used to process all data with that task. This makes the data processing more predictable: it does not depend on the order in which items of data are processed.

The task’s configuration is specified using a task-specific subclass of lsst.pex.config.Config. The task class specifies its configuration class using class variable ConfigClass. If the task has no configuration parameters then it may use lsst.pex.config.Config as its configuration class.

Some important details of configurations:

  • Supply useful defaults for all config parameters if at all possible. Your task will be much easier to use if the default configuration usually suffices.

  • Document each field of the configuration carefully. Pretend you don’t know anything about your task and ask yourself what you would need to be told to change the parameter. What does the parameter do and why might you change it? What units does it have?

  • Subtasks are specified in your configuration as fields of type lsst.pex.config.ConfigurableField or (less commonly) lsst.pex.config.RegistryField. This allows subtasks to be retargeted (replaced with a variant subtask). For more information see Subtasks.

    ExampleSigmaClippedStatsTask uses configuration class ExampleSigmaClippedStatsConfig:

    class ExampleSigmaClippedStatsConfig(pexConfig.Config):
        """Configuration for ExampleSigmaClippedStatsTask.
        """
        badMaskPlanes = pexConfig.ListField(
            dtype=str,
            doc="Mask planes that, if set, indicate the associated pixel should "
            "not be included when the calculating statistics.",
            default=("EDGE",),
        )
        numSigmaClip = pexConfig.Field(
            doc="number of sigmas at which to clip data",
            dtype=float,
            default=3.0,
        )
        numIter = pexConfig.Field(
            doc="number of iterations of sigma clipping",
            dtype=int,
            default=2,
        )
    

The configuration class is specified as ExampleSigmaClippedStatsTask class variable ConfigClass, as described in Class variables.

Here is the config for ExampleTask, a task that calls one subtask named stats; notice the lsst.pex.config.ConfigurableField:

class ExampleConfig(pexConfig.Config):
    """Configuration for ExampleTask.
    """
    stats = pexConfig.ConfigurableField(
        doc="Subtask to compute statistics of an image",
        target=ExampleSigmaClippedStatsTask,
    )
    doFail = pexConfig.Field(
        doc="Raise an lsst.base.TaskError exception when processing each image? "
        + "This allows one to see the effects of the --doraise command line flag",
        dtype=bool,
        default=False,
    )

Class variables

Tasks require several class variables to function:

  • ConfigClass: the configuration class used by the task.

  • _DefaultName: a string used as the default name for the task. This is required for a pipeline task (PipelineTask), and strongly recommended for other tasks because it makes them easier to construct for unit tests. Note that when a task creates a subtask, it ignores the subtask’s _DefaultName and assigns the name of the config parameter as the subtask’s name. For example exampleTask.ExampleConfig creates the statistics subtask with name stats because the config field for that subtask is stats = lsst.pex.config.ConfigurableField(...).

    Task names are used for the hierarchy of task and subtask metadata. Also, for pipeline tasks the name may be used as a component of the dataset type for saving the task’s configuration and metadata.

Here are the class variables for ExampleTask:

ConfigClass = ExampleConfig
_DefaultName = "exampleTask"

Methods

Tasks have the following important methods:

  • __init__: construct and initialize a task.

  • run: process one item of data.

These methods are described in more depth below.

The __init__ method

Use the __init__ method (task constructor) to do the following:

Here is exampleTask.ExampleTask.__init__:

def __init__(self, *args, **kwargs):
    """Construct an ExampleTask

    Call the parent class constructor and make the "stats" subtask from the config field of the same name.
    """
    super().__init__(self, *args, **kwargs)
    self.makeSubtask("stats")

That task creates a subtask named stats to compute image statistics. Here is the __init__ method for the default version of the stats subtask exampleTask.ExampleSigmaClippedStatsTask, which is slightly more interesting:

def __init__(self, *args, **kwargs):
    """Construct an ExampleSigmaClippedStatsTask

    The init method may compute anything that that does not require data.
    In this case we create a statistics control object using the config
    (which cannot change once the task is created).
    """
    super().__init__(self, *args, **kwargs)
    self._badPixelMask = MaskU.getPlaneBitMask(self.config.badMaskPlanes)
    self._statsControl = afwMath.StatisticsControl()
    self._statsControl.setNumSigmaClip(self.config.numSigmaClip)
    self._statsControl.setNumIter(self.config.numIter)
    self._statsControl.setAndMask(self._badPixelMask)

This creates a binary mask identifying bad pixels in the mask plane and an lsst.afw.math.StatisticsControl, specifying how statistics are computed. Both of these are constants, and thus are the same for each invocation of the run method; this is strongly recommended, as explained in the next section.

Task execution methods

For a detailed overview of creating a PipelineTask see Creating a PipelineTask.

The run method

All tasks are required to have a run method which acts as their primary point of entry. This method takes, as explicit arguments, everything that the task needs to perform one unit of execution (for example, processing a single image), and returns the result to the caller. The run method should not perform I/O, and, in particular, should not be expected to have access to the Data Butler for storing and retrieving data. Instead, results are returned as an lsst.pipe.base.struct.Struct object, with a named field for each item of data. This is safer than returning a tuple of items, and allows adding fields without affecting existing code.

Note

In some, unusual, circumstances, it may be necessary for run to have access to the Data Butler, or for a task not to provide a run method. In code released by DM, these cases should be approved by the DM Change Control Board through the RFC process.

If your task’s processing can be divided into logical units, then we recommend that you provide methods for each unit. run can then call each method to do its work. This allows your task to be more easily adapted: a subclass can override just a few methods. Any method that is likely to take significant time or memory should be preceded by this python decorator: lsst.utils.timer.timeMethod. This automatically records the execution time and memory of the method in the task’s metadata attribute.

We strongly recommend that you make your task stateless, by not using instance variables as part of your data processing. Pass data between methods by calling and returning it. This makes the task much easier to reason about, since processing one item of data cannot affect future items of data.

Debug variables

Debug variables are variables the user may set while running your task, to enable additional debug output. To have your task support debug variables, have it import lsstDebug and call lsstDebug.Info(__name__).varname to get the debug variable varname specific to your task. If you look for a variable the user has not specified, it will have a value of False. For example, to look for a debug variable named “display”:

import lsstDebug
display = lsstDebug.Info(__name__).display
if display:
   # ...
   pass

Documentation

To be written.

Subtasks

Each subtask is specified in the configuration as a field of type lsst.pex.config.ConfigurableField or (less commonly) lsst.pex.config.RegistryField. There are advantages to each:

  • lsst.pex.config.ConfigurableField advantages:

    • It is easier for the user to override settings of the subtask; simply use dotted name notation:

      config.configurableSubtask.subtaskParam1 = ...
      

      In contrast, to override configuration for a subtask specified as an lsst.pex.config.RegistryField you must either specify the name of the subtask to configure:

      config.registrySubtask[nameOfSelectedSubtask].subtaskParam1 = ...
      

      Or use the active attribute to modify the configuration of the currently selected (active) subtask:

      config.registrySubtask.active.subtaskParam1 = ...
      
  • lsst.pex.config.RegistryField advantages:

    • You can specify overrides for any registered subtask and they are remembered if you retarget subtasks. In comparison if the subtask is specified as an lsst.pex.config.ConfigurableField then you can only override parameters for the currently retargeted subtask, and all overrides are lost each time you retarget. Thus using an lsst.pex.config.RegistryField offers the opportunity to specify suitable overrides for more than one variant subtask, making it safer for the user to use those variants. Of course this can get out of hand if there are many variants, so users should not assume that all variants have suitable overrides.

    • Retargeting a subtask can be done using --config on the pipetask command line, as long as the module containing the desired subtask has been imported:

      config.registrySubtask.name = "foo"
      

      By comparison, a subtask specified as an lsst.pex.config.ConfigurableField can only be retargeted from a config override file (e.g. using --config-file with pipetask, never --config):

      from ... import FooTask
      config.configurableSubtask.retarget(FooTask)
      
    • Variants subtasks are kept together in one registry, making it easier to find them.

Our recommendation: if you anticipate that users will often wish to override the subtask with a variant, then use an lsst.pex.config.RegistryField. Otherwise use an lsst.pex.config.ConfigurableField to keep config overrides simpler and easier to read.

For example PSF determiners and star selectors are perhaps best specified using lsst.pex.config.RegistryField because there are several variants users may wish to select from. However, calibration and instrument signature removal are best specified using lsst.pex.config.ConfigurableField because (for a given camera) there is likely to be only one logical variant, and that variant is specified in a camera-specific configuration override file, so the user need not specify it.

Variant tasks

When there are (or are expected to be) different versions of a given task, those tasks should inherit from an abstract base class that defines the interface and is itself a subclass of lsst.pipe.base.Task. Star selectors (lsst.meas.algorithms.BaseStarSelectorTask) and PSF determiners (lsst.meas.algorithms.BasePsfDeterminerTask) are two examples of tasks with multiple variants. The abstract base class should be written using abc.ABC or abc.ABCMeta. The same module that defines the abstract base class should also define a registry, using lsst.pex.config.RegistryField, and all implementations should register themselves with that registry. Examples include lsst.meas.algorithms.starSelectorRegistry and lsst.meas.algorithms.psfDeterminerRegistry.