Plot Data Selection#

This page describes how the plot creators can make data for plotting available in a programmatic fashion.

Each plot creator is associated with a DataManager instance, which holds all the data that is currently available. This data is usually made available to you such that you can select data which you can then pass on to whatever you use for plotting.

While manual selection directly from the data manager suffices for specific cases, automation is often desired. The uniform interface of the DataManager paired with the TransformationDAG framework makes automated data selection and transformation for plotting possible while making available all the benefits of the Data Transformation Framework framework:

  • Generic application of transformations on data

  • Fully configuration-based interface

  • Caching of computationally expensive results

This functionality is embedded at the level of the BasePlotCreator, making it available for all plot creators and allowing subclasses to tailor it to their needs.

Additionally, result placeholders can be specified inside the plot configuration, thus allowing to use transformation results not only for data selection, but also for programmatically determining other configuration parameters.


General remarks#

This section holds information that is valid for all plot creators.

Enabling DAG usage#

If using the recommended plot function signature, the use_dag key can be specified right there and enables the data transformation framework. This declares that the plot function expects data selection to occur via the transformation framework.

After computation, the results are made available to the selected python plot function via the data keyword argument, which is a dictionary of the tags that were selected to be computed.

Setting use_dag in the plot configuration#

Alternatively, DAG usage can be controlled via the use_dag argument in the plot configuration:

# Some plot configuration file
---
my_plot:
  use_dag: true

  # ... more arguments

Hint

If setting use_dag in the plot configuration, take care not to create conflicts with the chosen plot function signature.

Arguments to control DAG behaviour#

You then have the following arguments available to control its behaviour:

  • select and transform: select data and perform transformations on it, see add_nodes().

  • compute_only: controls which tags are to be computed, see compute()

  • dag_options: passed to TransformationDAG initialization, e.g. to control file_cache_defaults, verbosity, or adding transformations via the define interface, see The define interface.

  • dag_visualization: controls visualization of the DAG which can be very helpful for debugging, see below. These arguments are passed to _generate_DAG_vis().

The creation of the DAG and its computation is controlled by the chosen plot creator and can be specialized to suit that plot creator’s needs.

Note

If DAG usage is enabled, these arguments will be used exclusively for the DAG, i.e.: they are not available downstream in the plot creator or the plot function.

Hint

To use meta-operations for plot data selection, define them under the dag_options.meta_operations key of a plot configuration.

Same for adding nodes via the define interface (see The define interface), which is also only available via dag_options.define.

Also check out the dantro base plot configs from which some pre-defined meta-operations can be included using based_on.

Hint

Specialized plot creators, like those based on paramspace operations may implement an expanded syntax.

Example#

Some example plot configuration to select some containers from the data manager, perform simple transformations on them and compute a result tag:

# Some plot configuration file
---
my_plot:
  creator: my_creator

  # ... some plot arguments here ...

  # Data selection via DAG framework
  use_dag: true
  select:
    foo: some/path/foo
    bar:
      path: some/path/bar
      transform:
        - mean: [!dag_prev ]
        - increment: [!dag_prev ]
  transform:
    - add: [!dag_tag foo, !dag_tag bar]
      tag: result
  compute_only: [result]
  dag_options:
    define:
      foo: bar
    verbosity: 3  # to show more profiling statistics (default: 1)
    file_cache_defaults:
      write: true
      read: true

    # ... other parameters here are passed on to TransformationDAG.__init__

DAG object caching#

For very complex data transformation sequences, DAGs can have many hundreds of thousands of nodes. In those cases, parsing the DAG configuration and creating the corresponding objects can be time-consuming and begin to noticeably prolong the plotting procedure.

To remedy this, the plotting framework implements memory-caching of TransformationDAG objects such that they can be re-used across multiple plots or repeated invocation of the same plot. The cache is used if the DAG-related configuration parameters (transform, select, …) are equal, i.e. have equal results when serialized using repr. In other words: if plots use the same data selection arguments, thus creating identical DAGs, the cache can be used.

Multiple aspects of caching can be controlled using the dag_object_cache parameter, passed via dag_options (see below):

  • read: whether to read from the cache (default: false)

  • write: whether to write from the cache (default: false)

  • use_copy: whether to read and write a deep copy of the TransformationDAG object to the cache (default: true).

  • clear: if set, will remove all objects from the cache (after reading from it) and trigger garbage collection (default: false)

  • collect_garbage: can be used to separately control garbage collection, e.g. to suppress it despite clear having been passed.

Warning

Only use use_copy: false if you can be certain that plot functions do not change the object; this would create side effects that may be very hard to track down.

Note

The clear option will also invoke general garbage collection (if not explicitly disabled). This will free up memory … but it may also take some time.

Example#

# Some plot configuration file
---
my_plot:
  # ... some plot arguments here ...

  # Data selection via DAG framework
  use_dag: true
  select:
    foo: some/path/foo
    bar:
      path: some/path/bar
      transform:
        - mean: [!dag_prev ]
        - increment: [!dag_prev ]
  transform:
    - add: [!dag_tag foo, !dag_tag bar]
      tag: result
  compute_only: [result]

  # Enable DAG object caching
  dag_options:
    dag_object_cache:
      read: true
      write: true

      # Other parameters (and their default values)
      # use_copy: true       # true:  cache a deep copy of the object
      # clear: false         # true:  clears the object cache and invokes
                             #        garbage collection
      # collect_garbage: ~   # true:  invokes garbage collection
                             # false: suppresses garbage collection even
                             #        if `clear` was set

my_other_plot_using_the_cache:
  based_on: my_plot          # --> identical DAG arguments (if not overwritten below)

  # ... some plot arguments ...

Defining a generic plot function#

Ideally, a plot function can focus on providing a bridge from data to a visual representation. Using the PyPlotCreator, this becomes feasible:

from dantro.plot import is_plot_func, PlotHelper

@is_plot_func(use_dag=True)
def my_plot_func(*, data: dict, hlpr: PlotHelper, **further_kwargs):
    """This is my custom plot function with preprocessed DAG data"""
    # ...
    pass

The only required arguments here are data and hlpr. The former contains all results from the DAG computation; the latter is the plot helper, which effectively is the interface to the visualization of the data.

Importantly, this makes the plot function averse to the specific choice of a creator: the plot function can be used with the PyPlotCreator and with its specializations, UniversePlotCreator and MultiversePlotCreator. In such cases, the creator should not be specified in the decorator, but it should be given in the plot configuration.

Specifying required tags#

If some specific tags are required, they can also be set in the decorator:

@is_plot_func(use_dag=True, required_dag_tags=('x', 'y'))
def simple_lineplot(*, data: dict, hlpr: "PlotHelper", **plt_kwargs):
    """Creates a simple line plot for selected x and y data"""
    hlpr.ax.plot(data['x'], data['y'], **plt_kwargs)

The DAG can be configured in the same way as in the general case.

Hint

If you want the computed tags to be directly available in the plot function signature, use the unpack_dag_results flag in the decorator:

@is_plot_func(use_dag=True, required_dag_tags=('x', 'y'),
              unpack_dag_results=True)
def simple_lineplot(*, x, y, hlpr: "PlotHelper", **plt_kwargs):
    """Creates a simple line plot for selected x and y data"""
    hlpr.ax.plot(x, y, **plt_kwargs)

Accessing the DataManager#

As visible from the plot function above, the PyPlotCreator does not pass along the current DataManager instance as first positional argument (dm) when DAG usage is enabled. This makes the plot function signature simpler and allows the creator-averse definition of plot functions while not restricting access to the data manager:

The data manager can still be accessed directly via the dm DAG tag. Make sure to specify that it should be included, e.g. via compute_only or the required_dag_tags argument to the decorator.

Special case: UniversePlotCreator#

For the UniversePlotCreator, data selection and transformation has to occur based on data from the currently selected universe. This is taken care of automatically by this creator: it dynamically sets the select_base() property to the current universe, not requiring any further user action. In effect, the select argument acts as if selections were to happen directly from the universe.

Except for the select_base and base_transform arguments, the full DAG interface is available via the UniversePlotCreator.

Hint

To restore parts of the functionality of the already-in-use select_base and base_transform arguments, the select_path_prefix argument of TransformationDAG can be used. It can be specified as part of dag_options and is prepended to all path arguments specified within select.

Example#

The following suffices to define a UniversePlotCreator-based plot function:

from dantro.plot import UniversePlotCreator

@is_plot_func(creator_type=UniversePlotCreator, use_dag=True)
def my_universe_plot(*, data: dict, hlpr: PlotHelper, **kwargs):
    """This is my custom universe plot function with DAG usage"""
    # ...
    pass

Hint

To not restrict the plot function to a specific creator, using the creator-averse plot function definition is recommended, which omits the creator_type in the decorator and instead specifies it in the plot configuration.

The DAG can be configured in the same way as in the general case.

Special case: MultiversePlotCreator#

The MultiversePlotCreator has a harder job: It has to select data from the whole multiverse subspace, apply transformations to it, and finally combine it, with optional further transformations following.

It does so fully within the DAG framework by building a separate DAG branch for each universe and bundling all of them into a transformation that combines the data. This happens via the select_and_combine argument.

Important: The select_and_combine argument behaves differently to the select argument of the DAG interface! This is because it has to accommodate various further configuration parameters that control the selection of universes and the multidimensional combination of the selected data.

The select_and_combine argument expects the following keys:

  • fields: all keys given here will appear as tags in the results dictionary. The values of these keys are dictionaries that contain the same parameters that can also be given to the select argument of the DAG interface. In other words: paths you would like to select from within each universe should be specified at select_and_combine.fields.<result_tag>.path rather than at select.<result_tag>.path.

  • base_path (optional): if given, this path is prepended to all paths given under fields

  • combination_method (optional, default: concat): how to combine the selected and transformed data from the various universes. Available parameters:

    • concat: attempts to preserve data types but is only possible if the universes fill a hypercube without holes

    • merge: which is always possible, but leads to the data type falling back to float. Missing data will be np.nan in the results.

    The combination method can also be specified for each tag under select_and_combine.<result_tag>.combination_method.

  • subspace (optional): which multiverse subspace to work on. This is evaluated fully by the paramspace.ParamSpace.activate_subspace method. The subspace can also be specified for each tag under select_and_combine.<result_tag>.subspace.

Remarks#

  • The select operations on each universe set the omit_tag flag in order not to create a flood of only-internally-used tags. Setting tags manually here does not make sense, as the tag names would collide with tags from other universe branches.

  • File caching is hard-coded to be disabled for the initial select operation and for the operation that attaches the parameter space coordinates to it. This behavior cannot be influenced.

  • The best place to cache is the result of the combination method.

  • The regular select argument is still available, but it is applied only after the select_and_combine-defined nodes were added and it does only act globally, i.e. not on each universe.

  • The select_path_prefix argument to TransformationDAG is not allowed for the MultiversePlotCreator. Use the select_and_combine.base_path argument instead.

Example#

A MultiversePlotCreator-based plot function can be implemented like this:

from dantro.plot import MultiversePlotCreator

@is_plot_func(creator_type=MultiversePlotCreator, use_dag=True)
def my_multiverse_plot(*, data: dict, hlpr: PlotHelper, **kwargs):
    """This is my custom multiverse plot function with DAG usage"""
    # ...
    pass

Hint

To not restrict the plot function to a specific creator, using the creator-averse plot function definition is recommended, which omits the creator_type in the decorator and instead specifies it in the plot configuration.

An associated plot configuration might look like this:

---
my_plot:
  # ... some plot arguments here ...

  # Data selection via DAG framework
  select_and_combine:
    fields:
      foo: some/path/foo
      bar:
        path: some/path/bar
        transform:
          - mean: [!dag_prev ]
          - increment: [!dag_prev ]

    combination_method: concat  # can be ``concat`` (default) or ``merge``
    subspace: ~                 # some subspace selection

  transform:
    - add: [!dag_tag foo, !dag_tag bar]
      tag: result

Handling missing data#

In some cases, the ParamSpaceGroup associated with the MultiversePlotCreator might miss some states. This can happen, for instance, if the to-be-plotted data is the result of a simulation for each point in parameter space and the simulation was stopped before visiting all these points. In such a case, select_and_combine will typically fail.

Another reason for errors during this operation may be that the data structures between the different points in parameter space are different, such that a valid path within one ParamSpaceStateGroup (or: “universe”) is not a valid path in another.

To be able to plot the partial data in both of these cases, this plot creator makes use of the error handling feature in the data transformation framework. It’s as simple as adding the allow_missing_or_failing key to select_and_combine:

# Select the creator and use the generic facet grid plotting function
based_on:
  - .creator.multiverse
  - .plot.facet_grid

# Select data, allowing for missing universes or failing .mean operation
select_and_combine:
  allow_missing_or_failing: true
  combination_method: merge        # needed with allow_missing_or_failing
  fields:
    data:
      path: labelled/randints
      transform:
        - .mean: [!dag_prev , [x]]

This option kicks in when any of the following scenarios occur:

  • A universe from the selected subspace is missing altogether

  • The getitem operation for the given path within a universe fails

  • Any operation within transform fails

In any of these cases, the data for the whole universe is discarded. Instead, an empty xr.Dataset with the coordinates of that universe is used as fallback, with the following effect: The corresponding coordinates will be present in the final xr.Dataset, but they contain no data (or NaNs). The latter is also the reason why the merge combination method is required here.

Note

The rationale behind this behavior is that coordinate information is valuable, as it shows which data would have been available. If desired, null-like data can be dropped afterwards using the .dropna operation.

In case of missing data, the error message will come from the dantro.expand_dims operation and contain information on the failure.

..warning:

If *all* data is missing, ``select_and_combine`` will not be able to succeed, because there will be nothing to combine and insufficient information to create a null-like output instead.
This feature is explicitly meant for data *partially* missing.

The expected error message for such a case will be coming from ``dantro.merge``:

::

    The Dataset resulting from the xr.merge operation can only be reduced
    to a DataArray, if one and only one data variable is present in the
    Dataset! However, the merged Dataset contains 0 data variables.

Hint

The allow_missing_or_failing argument accepts the same values as the allow_failure argument of the error handling framework; in fact, it sets exactly that argument internally.

Thus, the messaging behavior can be influenced as follows:

select_and_combine:
  allow_missing_or_failing: silent        # other options: warn, log

Hint

Same as combination_method and subspace, the allow_missing_or_failing argument can also be specified separately for each field, overwriting the default value from the select_and_combine root level:

select_and_combine:
  allow_missing_or_failing: silent
  fields:
    some_data:
      allow_missing_or_failing: warn   # overwrites default from above
      path: path/to/some/data

Applying transformations after combination of data#

In some cases, it can be useful to define postprocessing transformations on the combined data. For that purpose, there is the transform_after_combine option which can be added for each individual field or as a default on the select_and_combine level. While this postprocessing can of course also be done alongside transform, it is often easier to define this alongside the field.

Some example use cases:

  • Perform some postprocessing on all fields, without having to repeat the definitions.

  • Use print to see the result of the combination directly, without having to touch the transform definition.

  • Call .squeeze to reduce the one-sized dimensions of a combination, which can simplify some plotting calls.

Custom combination method#

Apart from the merge and concat combination methods, a custom combination method can also be used by specifying the name of an operation that is capable of combining the data in a desired way:

select_and_combine:
  # further kwargs are passed on to the chosen custom operation

  fields:
    some_data:
      path: path/to/some_data
      combination_method:
        operation: my_combination_operation
        pass_pspace: false  # default: false. If true, will pass additional
                            # keyword argument ``pspace``.
        # further kwargs passed to combination operation
      combination_kwargs:

Such a combination operation needs to have the following signature:

def my_combination_function(objs: list, **kwargs) -> xr.DataArray:
    # ...

Here, objs is a list of the data from each individual parameter space state (“universe”), ready with attached coordinates.

Note

While the given objs already have coordinates assigned, you might be interested in some macroscopic information about the shape of the target data. To that end, an additional argument can be passed to the combination function by setting combination_method.pass_pspace: true.

The pspace argument is then a ParamSpace object (from the paramspace package) which contains information about the dimensionality of the data and the names and coordinates of the dimensions. The data in objs is ordered in the same way as the iteration over pspace.

Full DAG configuration interface for multiverse selection#

An example of all options available in the MultiversePlotCreator.

# Full DAG specification for multiverse selection
---
my_plot:
  # ... some plot arguments here ...

  # DAG parameters
  # Selection from multiple universes with subsequent combination
  select_and_combine:
    fields:
      # Define a tag 'foo' that will use the defaults defined directly on
      # the ``select_and_combine`` level, see below
      foo: foo                       # ``base_path`` will be prepended here
                                     # resulting in: some/path/foo

      # Define a tag 'bar' that overwrites some of the defaults
      bar:
        path: bar
        subspace:                    # only use universes from a subspace
          seed: [0, 10]
          my_param: [-42., 42.]
        combination_method: merge    # overwriting default specified below
        combination_kwargs:          # passed to Transformation.__init__
                                     # of the *tagged* output node
          file_cache:
            read: true
            write:
              enabled: true
              # Configure the file cache to only be written if this
              # operation took a large amount of time.
              min_cumulative_compute_time: 20.
        allow_missing_or_failing: silent  # transformations or path lookup
                                          # is allowed to fail
        transform:
          - mean: !dag_prev
          - increment: [!dag_prev ]
          - some_op_with_kwargs:
              data: !dag_prev
              foo: bar
              spam: 42
          - operation: my_operation
            args: [!dag_prev ]
            file_cache: {}           # can configure file cache here

        transform_after_combine:     # applied after combination
          - increment
          - print

    base_path: some_path        # if given, prepended to ``path`` in ``fields``

    # Default arguments, can be overwritten in each ``fields`` entry
    combination_method: concat  # can be ``concat`` (default), ``merge``.
                                # If a dict, may contain the key
                                # ``operation`` which will then be used as
                                # the operation to use for combination; any
                                # further arguments are passed on to that
                                # operation call.
    subspace: ~                 # some subspace selection
    allow_missing_or_failing: ~ # whether to allow missing universes or
                                # failing transformations; can be: boolean,
                                # ``log``, ``warn``, ``silent``
    transform_after_combine: ~

  # Additional selections, now based on ``dm`` tag
  select: {}

  # Additional transformations; all tags from above available here
  transform: []

  # Other DAG-related parameters: ``compute_only``, ``dag_options``
  # ...

Note

This does not include all possible options for DAG configuration but focusses on those options added by MultiversePlotCreator to work with multiverse data, e.g. subspace, combination_kwargs.

For other arguments, see Full syntax specification of a single Transformation.


Using data transformation results in the plot configuration#

The data transformation framework can not only be used for the selection of plot data: using so-called “result placeholders”, data transformation results can be used as part of the plot configuration.

One use case is to include a computation result, e.g. some mean value, into the title of the plot via the plot helper. In general, this feature allows to automate further parts of the plot configuration by giving access to the capabilities of the transformation framework.

Let’s look at an example plot configuration:

# Select the creator and use the generic errorbar plotting function
based_on:
  - .creator.universe
  - .plot.facet_grid.errorbars

select:
  # 3D data with random integers
  some_data: randints

transform:
  # Compute the mean and standard deviation
  - .mean: [!dag_tag some_data, [x, z]]
    tag: mean
  - .std: [!dag_tag some_data, [x, z]]
    tag: stddev

  # Assemble them into a Dataset for the errorbars plot
  - xr.Dataset:
      - mean: !dag_tag mean
        stddev: !dag_tag stddev
    tag: data

  # Additional transformations for ResultPlaceholders
  - .mean: [!dag_tag mean]
  - .item  # ... otherwise it's still an xr.DataArray
  - .format: ["Some Data (total mean: {:.3g})", !dag_prev ]
    tag: title_str

# Specify which data variable to plot as line and which as errorbands
y: mean
yerr: stddev
use_bands: true

# Now, use the place holder in the helper configuration
helpers:
  set_title:
    title: !dag_result title_str

As can be seen here, there are additional operations defined within transform, which lead to the title_str tag. In the helper configuration, that tag is referred to via the !dag_result YAML tag, thus creating a placeholder at the helpers.set_title.title key.

This illustrates the basic idea. Of course, multiple placeholders can be used and they can be used almost everywhere inside the plot configuration; however, make sure to have a look at the caveats to learn about current limitations.

Hint

When adding placeholders, you will notice additional log messages which inform about the placeholder names and their computation profile.

Caveats#

Where in the plot configuration can placeholders be used?#

Placeholders can be used in wide parts of the plot configuration, but not everywhere. If you encounter errors that refer to an unexpected ResultPlaceholder object, this is probably because they were defined in a part of the plot configuration where they cannot be resolved.

Where can (✅) placeholders always be used? Where can they never (❌) be used?

  • ✅ They can be used in all configuration entries that are passed through to the selected plot function of the The PyPlotCreator and derived plot creators.

  • ✅ They can be used within the helpers argument that controls the The PlotHelper.

  • ❌ They can not be used for entries related to data transformation (select, transform, dag_options, …) because these need to be evaluated in order to set up the TransformationDAG.

  • ❌ They can not be used for entries evaluated by the The PlotManager (out_path, etc) or the plot creator prior to data selection (animation, style, module, etc).

Why is my placeholder not resolved?#

The identification and replacement of placeholders happens by recursively iterating through list-like and dict-like objects in the plot configuration dict. Typically, this reaches all places where these placeholders could be defined. The only exception being if the placeholder is in some part of an object that does not behave like a list or a dict.

Implementation details#

Under the hood, the !dag_result YAML tag is read as a ResultPlaceholder object, which simply stores the name of the tag that should come in its place. After the plot data was computed, the BasePlotCreator inspects the plot configuration and recursively collects all these placeholder objects. The compute() method is then invoked to retrieve the specified results. Subsequently, the placeholder entries in the plot configuration are replaced with the result from the computation.

For the above operations, functions from the paramspace package are used, specifically: paramspace.tools.recursive_collect and paramspace.tools.recursive_replace.

DAG Visualization#

The DAG used for plot data selection and transformation can also be visualized. This can be helpful to understand what kind of operations are carried out on which kind of data; this can be a big assistance during debugging.

By default, DAG visualization is enabled and will generate output if there was an error during the computation of data transformation results. This can be controlled; see below.

However, there are many ways to further control when a visualization is created and how it looks like. All parameters for controlling DAG visualization can be passed via the dag_visualization in a plot configuration.

Such a plot may look like these:

DAG visualization DAG visualization DAG visualization

DAG generation#

The way the DAG is generated is controlled by the generation arguments, which are evaluated by generate_nx_graph(). Also see Graph representation and visualization for more information.

Controlling when to generate a DAG plot#

For instance, if we’d like to always generate a DAG plot upon a computation, we can pass the following parameters:

my_dag_plot:
  # ...
  dag_visualization:
    when:
      only_once: true            # only generate a single DAG plot
      on_compute_error: true     # ... either upon failing computation
      on_compute_success: true   # ... or upon a successful one.

Hint

To only plot if the creator runs in debug mode (i.e., with raise_exc set), set the scenario to debug instead of a boolean.

my_dag_plot:
  # ...
  dag_visualization:
    when:
      on_compute_error: debug

In the on_compute_error scenario, it is advisable to activate the show_node_status option for visualization, which will indicate at which node an error occurred:

DAG visualization

The colors indicate the following node status, as detailed in the legend:

  • green: computation succeeded

  • yellow: computation failed but a fallback value was used

  • red: computation failed in this node

  • dark red: computation failed in a node that this node depends on

Hint

To adjust the status colors, set the node_status_colors argument; see visualize() docstring for more info.

Changing plot content#

What is shown in the plot depends mostly on the label attribute of the nodes. By default, that content is generated via the get_description() operation function, which takes into account the name of the tag, the operation, and potential results.

What is shown in the plot is the label attribute, so in order to show something else there, we need to tell the visualize() method to use something else for the label. By default, the description attribute is shown. In the following example, we will instead show simply the operation attribute by setting the drawing.labels.from_attr entry of the configuration:

dag_visualization:
  drawing:
    labels:
      from_attr: operation
      # available attributes: tag, description, operation

Using the manipulate_attributes() function, we can also generate custom attributes. In the following example, the name of that attribute is my_custom_attr, which is then also set as the label.

dag_visualization:
  generation:
    include_results: true
    manipulate_attrs:
      map_node_attrs:
        my_custom_attr:
          # Invoke *some* function; as an example, use a lambda to copy
          # over some node attribute data into `my_attr`
          call_lambda: "lambda *, attrs: attrs.get('result', '(no result)')"

  drawing:
    labels:
      # Use the custom attribute as a label
      from_attr: my_custom_attr

Note

If not setting drawing.labels.from_attr explicitly, it will always use the description attribute as the label.

Setting plot aesthetics#

The looks of the DAG plot are set via the drawing keyword, which end up in the visualize() method:

my_dag_plot:
  # ...

  dag_visualization:
    drawing:
      # Whether to include default values for nodes, edges, and labels.
      # If true, will recursively update these defaults with the values
      # given below.
      # Set to false to use the networkx defaults instead.
      use_defaults: true

      # Arguments to networkx.draw_networkx_*
      nodes:
        node_color: blue
        # ...
      edges:
        width: 2.5
        # ...
      labels:
        font_size: 10
        # ...

Note

With networkx using matplotlib as drawing backend, there are a number of limitations: For instance, it is not possible to let edges terminate exactly at the edge of the label’s box.

If this is desired, you may want to have a look at Exporting a DAG representation.

Exporting a DAG representation#

For more control over the looks of the DAG, you can use the export keyword and use whatever other program you like to look at the plot output. This will invoke export_graph().

In that case you may want to set plot_enabled: False as well:

my_dag_plot:
  # ...

  dag_visualization:
    plot_enabled: false
    export_enabled: true

    # ...

    export:
      manipulate_attrs:
        # Use the description as label and keep only that attribute
        map_node_attrs:
          label:
            attr_mapper.copy_from_attr: description
        keep_node_attrs:
          - label

      # Export formats
      graphml: true
      dot: true
      # ...

Remarks#

For more information on possible arguments, see _generate_DAG_vis(). For a background on DAG representation as a networkx.DiGraph, see Graph representation and visualization.

Note

The layouting algorithm cannot be changed yet.

If GraphViz and pygraphviz are installed, graphviz_layout() is used with the dot algorithm. If those are not installed, a multipartite_layout() is carried out.

Full Interface#

The following documents the full interface and the corresponding default values:

# DAG Visualization interface; values given here are default values
dag_visualization:
  enabled: true           # Main toggle
  plot_enabled: true      # Whether to generate a plot
  export_enabled: true    # Whether to export the graph

  # Whether to raise an exception if graph generation, plotting or
  # exporting failed. If None, will use the creator's setting.
  raise_exc: ~

  # Whether to *additionally* export the graph
  export:
    # Manipulate node or edge attributes (for export only)
    manipulate_attrs:
      map_node_attrs: {}
      map_edge_attrs: {}
      keep_node_attrs: True
      keep_edge_attrs: True

    # Export formats
    # ... need to be specified here.
    # Examples:
    gml: True
    graphml:  # arguments passed on to writer
      infer_numeric_types: True
    # dot: True  # Needs pygraphviz
    # ... more formats here ...

  # Output arguments
  output:
    plot_dir: ~  # None: Output will be aside the just-generated plot

    # A format string that is used to create the actual output path.
    # The `plot_dir` key is the one evaluated from the above argument.
    path_fstr: "{plot_dir:}/{name:}_dag_{scenario:}.pdf"

  # When to generate the visualization
  when:
    # General toggles
    always: false       # If true: always generate a DAG plot
    only_once: false    # If true: only generate one DAG plot

    # Scenarios: After which events to generate a DAG plot
    # Values can be: false, true, debug.
    # In case of 'debug', output is only generated if the creator was in
    # debug mode itself.
    on_compute_error: true
    on_compute_success: false
    on_plot_error: false
    on_plot_success: false

  # Generation kwargs
  generation:
    tags_to_include: all
    include_results: false
    lookup_tags: true
    manipulate_attrs:
      map_node_attrs:
        # Default operations: these are set by default
        operation: attr_mapper.dag.get_operation
        layer: attr_mapper.dag.get_layer
        description: attr_mapper.dag.get_description

        # Other available operations:
        # meta_operation: attr_mapper.dag.get_meta_operation
        # arguments: attr_mapper.dag.format_arguments
        # some_attr: attr_mapper.copy_from_attrs
        # another_attr: attr_mapper.set_value

        # ... or any other registered data operation:
        # my_attr:
        #   call_lambda: "lambda *, attrs: attrs.get('foo')"

  # Whether to base layouting and visualization on optimized default
  # values or not. For illustration, the actual default values are used
  # below; they do NOT have to be set explicitly as done here!
  use_defaults: true

  # Whether to show the node status and which colors to use for it
  show_node_status: true
  node_status_color:
    initialized: lightskyblue
    queued: cornflowerblue
    computed: limegreen
    looked_up: forestgreen
    failed_here: red
    failed_in_dependency: firebrick
    used_fallback: gold
    no_status: silver

  # Layouting algorithm (and fallback)
  layout:
    model: graphviz_dot      # requires graphviz and pygraphviz

    # In case the above model fails, silently switch to another one
    fallback: multipartite
    silent_fallback: true

    # Arguments for the respective layouting models
    model_kwargs:
      graphviz_dot: {}
      multipartite:
        align: horizontal
        subset_key: layer
        scale: -1

        # Whether to wiggle layouted positions to reduce edge overlap.
        # This is recommended for the multipartite layout, because it
        # does not handle edges going over multiple layers very well,
        # producing confusing edge overlaps ...
        wiggle:
          x: 0.005
          y: ~
          seed: 123  # set to None to always get new wiggles

  # Drawing, using networkx.draw_networkx_<...>
  drawing:
    nodes:
      alpha: 0.
      node_size: &node_size 600
    edges:
      arrows: true
      arrowsize: 12
      min_target_margin: 20
      min_source_margin: 20
      node_size: *node_size

    labels:
      # Which attribute to use as node label
      from_attr: description

      # Aesthetics; see matplotlib.patches.FancyBboxPatch
      font_size: 7
      bbox:
        fc: "#fffa"
        ec: "#666"
        linewidth: 0.5
        boxstyle: round

  # Figure creation via matplotlib.pyplot.figure
  figure_kwargs:
    figsize: [9, 7]

  # Scale figure size with "width" and "height" of the resulting graph
  # to avoid node overlapping; using these scaling factors.
  # Set to False to disable.
  scale_figsize: [0.25, 0.22]

  # Figure-level plot annotations: suptitle, figure legend for node color
  annotate_kwargs:
    # Title
    title: my custom DAG visualization
    title_kwargs: {}

    # Legend
    add_legend: true
    legend_kwargs: {}
    handle_kwargs: {}

  # Saving via matplotlib.pyplot.savefig
  save_kwargs:
    bbox_inches: tight