Interface Specifications

Before you start

Nipype is a young project maintained by an enthusiastic group of developers. Even though the documentation might be sparse or cryptic at times we strongly encourage you to contact us on the official nipype developers mailing list in case of any troubles: nipy-devel@neuroimaging.scipy.org (we are sharing a mailing list with the nipy community therefore please add [nipype] to the messsage title).

Overview

We’re using the Enthought Traits package for all of our inputs and outputs. Traits allows us to validate user inputs and provides a mechanism to handle all the special cases in a simple and concise way though metadata. With the metadata, each input/output can have an optional set of metadata attributes (described in more detail below). The machinery for handling the metadata is located in the base classes, so all subclasses use the same code to handle these cases. This is in contrast to our previous code where every class defined it’s own _parse_inputs, run and aggregate_outputs methods to handle these cases. Which of course leads to a dozen different ways to solve the same problem.

Traits is a big package with a lot to learn in order to take full advantage of. But don’t be intimidated! To write a Nipype Trait Specification, you only need to learn a few of the basics of Traits. Here are a few starting points in the documentation:

  • What are Traits? The Introduction in the User Manual gives a brief description of the functionality traits provides.
  • Traits and metadata. The second section of the User Manual gives more details on traits and how to use them. Plus there a section describing metadata, including the metadata all traits have.
  • If your interested in more of a big picture overview, Gael wrote a good tutorial that shows how to write a scientific application using traits for the benefit of the generated UI components. (For now, Nipype is not taking advantage of the generated UI feature of traits.)

Traits version

We’re using Traits version 3.x which can be install as part of EPD or from pypi

More documentation

Not everything is documented in the User Manual, in those cases the enthought-dev mailing list or the API docs is your next place to look.

Nipype Interface Specifications

Each interface class defines two specifications: 1) an InputSpec and 2) an OutputSpec. Each of these are prefixed with the class name of the interfaces. For example, Bet has these specs:

  • BETInputSpec
  • BETOutputSpec

Each of these Specs are classes, derived from a base TraitedSpec class (more on these below). The InputSpec consists of attributes which correspond to different parameters for the tool they wrap/interface. In the case of a command-line tool like Bet, the InputSpec attributes correspond to the different command-line parameters that can be passed to Bet. If you are familiar with the Nipype 0.2 code-base, these attributes are the same as the keys in the opt_map dictionaries. When an interfaces class is instantiated, the InputSpec is bound to the inputs attribute of that object. Below is an example of how the inputs appear to a user for Bet:

>>> from nipype.interfaces import fsl
>>> bet = fsl.BET()
>>> type(bet.inputs)
<class 'nipype.interfaces.fsl.preprocess.BETInputSpec'>
>>> bet.inputs.<TAB>
bet.inputs.__class__           bet.inputs.center
bet.inputs.__delattr__         bet.inputs.environ
bet.inputs.__doc__             bet.inputs.frac
bet.inputs.__getattribute__    bet.inputs.functional
bet.inputs.__hash__            bet.inputs.hashval
bet.inputs.__init__            bet.inputs.infile
bet.inputs.__new__             bet.inputs.items
bet.inputs.__reduce__          bet.inputs.mask
bet.inputs.__reduce_ex__       bet.inputs.mesh
bet.inputs.__repr__            bet.inputs.nooutput
bet.inputs.__setattr__         bet.inputs.outfile
bet.inputs.__str__             bet.inputs.outline
bet.inputs._generate_handlers  bet.inputs.outputtype
bet.inputs._get_hashval        bet.inputs.radius
bet.inputs._hash_infile        bet.inputs.reduce_bias
bet.inputs._xor_inputs         bet.inputs.skull
bet.inputs._xor_warn           bet.inputs.threshold
bet.inputs.args                bet.inputs.vertical_gradient

Each Spec inherits from a parent Spec. The parent Specs provide attribute(s) that are common to all child classes. For example, FSL InputSpecs inherit from interfaces.fsl.base.FSLTraitedSpec. FSLTraitedSpec defines an outputtype attribute, which stores the file type (NIFTI, NIFTI_PAIR, etc...) for all generated output files.

InputSpec class hierarchy

Below is the current class hierarchy for InputSpec classes (from base class down to subclasses).:

TraitedSpec: Nipype’s primary base class for all Specs. Provides initialization, some nipype-specific methods and any trait handlers we define. Inherits from traits.HasTraits.

BaseInterfaceInputSpec: Defines inputs common to all Interfaces (ignore_exception). If in doubt inherit from this.

CommandLineInputSpec: Defines inputs common to all command-line classes (args and environ)

FSLTraitedSpec: Defines inputs common to all FSL classes (outputtype)

SPMCommandInputSpec: Defines inputs common to all SPM classes (matlab_cmd, path, and mfile)

FSTraitedSpec: Defines inputs common to all FreeSurfer classes (sbjects_dir)

MatlabInputSpec: Defines inputs common to all Matlab classes (script, nodesktop, nosplash, logfile, single_comp_thread, mfile, script_file, and paths)

SlicerCommandLineInputSpec: Defines inputs common to all Slicer classes (module)

Most developers will only need to code at the the interface-level (i.e. implementing custom class inheriting from one of the above classes).

Output Specs

The OutputSpec defines the outputs that are generated, or possibly generated depending on inputs, by the tool. OutputSpecs inherit from interfaces.base.TraitedSpec directly.

Traited Attributes

Each specification attribute is an instance of a Trait class. These classes encapsulate many standard Python types like Float and Int, but with additional behavior like type checking. (See the documentation on traits for more information on these trait types.) To handle unique behaviors of our attributes we us traits metadata. These are keyword arguments supplied in the initialization of the attributes. The base classes BaseInterface and CommandLine (defined in nipype.interfaces.base) check for the existence/or value of these metadata and handle the inputs/outputs accordingly. For example, all mandatory parameters will have the mandatory = True metadata:

class BetInputSpec(FSLTraitedSpec):
  infile = File(exists=True,
                desc = 'input file to skull strip',
                argstr='%s', position=0, mandatory=True)

Common

exists

For files, use nipype.interfaces.base.File as the trait type. If the file must exist for the tool to execute, specify exists = True in the initialization of File (as shown in BetInputSpec above). This will trigger the underlying traits code to confirm the file assigned to that input actually exists. If it does not exist, the user will be presented with an error message:

>>> bet.inputs.infile = 'does_not_exist.nii'
------------------------------------------------------------
Traceback (most recent call last):
  File "<ipython console>", line 1, in <module>
  File "/Users/cburns/local/lib/python2.5/site-packages/nipype/interfaces/base.py", line 76, in validate
    self.error( object, name, value )
  File "/Users/cburns/local/lib/python2.5/site-packages/enthought/traits/trait_handlers.py", line 175, in error
    value )
TraitError: The 'infile' trait of a BetInputSpec instance must be a file
name, but a value of 'does_not_exist.nii' <type 'str'> was specified.
hash_files
To be used with inputs that are defining output filenames. When this flag is set to false any Nipype will not try to hash any files described by this input. This is useful to avoid rerunning when the specified output file already exists and has changed.
desc

All trait objects have a set of default metadata attributes. desc is one of those and is used as a simple, one-line docstring. The desc is printed when users use the help() methods.

Required: This metadata is required by all nipype interface
classes.
usedefault

Set this metadata to True when the default value for the trait type of this attribute is an acceptable value. All trait objects have a default value, traits.Int has a default of 0, traits.Float has a default of 0.0, etc... You can also define a default value when you define the class. For example, in the code below all objects of Foo will have a default value of 12 for x:

>>> import enthought.traits.api as traits
>>> class Foo(traits.HasTraits):
...     x = traits.Int(12)
...     y = traits.Int
...
>>> foo = Foo()
>>> foo.x
12
>>> foo.y
0

Nipype only passes inputs on to the underlying package if they have been defined (more on this later). So if you specify usedefault = True, you are telling the parser to pass the default value on to the underlying package. Let’s look at the InputSpec for SPM Realign:

class RealignInputSpec(BaseInterfaceInputSpec):
    jobtype = traits.Enum('estwrite', 'estimate', 'write',
                          desc='one of: estimate, write, estwrite',
                          usedefault=True)

Here we’ve defined jobtype to be an enumerated trait type, Enum, which can be set to one of the following: estwrite, estimate, or write. In a container, the default is always the first element. So in this case, the default will be estwrite:

>>> from nipype.interfaces import spm
>>> rlgn = spm.Realign()
>>> rlgn.inputs.infile
<undefined>
>>> rlgn.inputs.jobtype
'estwrite'
xor and requires
Both of these accept a list of trait names. The xor metadata reflects mutually exclusive traits, while the requires metadata reflects traits that have to be set together. When a xor-ed trait is set, all other traits belonging to the list are set to Undefined. The function check_mandatory_inputs ensures that all requirements (both mandatory and via the requires metadata are satisfied). These are also reflected in the help function.
copyfile
This is metadata for a File or Directory trait that is relevant only in the context of wrapping an interface in a Node and MapNode. copyfile can be set to either True or False. False indicates that contents should be symlinked, while True indicates that the contents should be copied over.
min_ver and max_ver

These metadata determine if a particular trait will be available when a given version of the underlying interface runs. Note that this check is performed at runtime.:

class RealignInputSpec(BaseInterfaceInputSpec):
    jobtype = traits.Enum('estwrite', 'estimate', 'write', min_ver='5',
                          usedefault=True)
deprecated and new_name

This is metadata for removing or renaming an input field from a spec.:

class RealignInputSpec(BaseInterfaceInputSpec):
    jobtype = traits.Enum('estwrite', 'estimate', 'write',
                          deprecated='0.8',
                          desc='one of: estimate, write, estwrite',
                          usedefault=True)

In the above example this means that the jobtype input is deprecated and will be removed in version 0.8. Deprecation should be set to two versions from current release. Raises TraitError after package version crosses the deprecation version.

For inputs that are being renamed, one can specify the new name of the field.:

class RealignInputSpec(BaseInterfaceInputSpec):
    jobtype = traits.Enum('estwrite', 'estimate', 'write',
                          deprecated='0.8', new_name='job_type',
                          desc='one of: estimate, write, estwrite',
                          usedefault=True)
    job_type = traits.Enum('estwrite', 'estimate', 'write',
                          desc='one of: estimate, write, estwrite',
                          usedefault=True)

In the above example, the jobtype field is being renamed to job_type. When new_name is provided it must exist as a trait, otherwise an exception will be raised.

Note

The version information for min_ver, max_ver and deprecated has to be provided as a string. For example, min_ver=‘0.1’.

CommandLine

argstr

The metadata keyword for specifying the format strings for the parameters. This was the value string in the opt_map dictionaries of Nipype 0.2 code. If we look at the FlirtInputSpec, the argstr for the reference file corresponds to the argument string I would need to provide with the command-line version of flirt:

class FlirtInputSpec(FSLTraitedSpec):
    reference = File(exists = True, argstr = '-ref %s', mandatory = True,
                     position = 1, desc = 'reference file')

Required: This metadata is required by all command-line interface classes.

position
This metadata is used to specify the position of arguments. Both positive and negative values are accepted. position = 0 will position this argument as the first parameter after the command name. position = -1 will position this argument as the last parameter, after all other parameters.
genfile
If True, the genfile metadata specifies that a filename should be generated for this parameter if-and-only-if the user did not provide one. The nipype convention is to automatically generate output filenames when not specified by the user both as a convenience for the user and so the pipeline can easily gather the outputs. Requires _gen_filename() method to be implemented. This way should be used if the desired file name is dependent on some runtime variables (such as file name of one of the inputs, or current working directory). In case when it should be fixed it’s recommended to just use usedefault.
sep
For List traits the string with witch elements of the list will be joined.
name_source
Indicates the list of input fields from which the value of the current File output variable will be drawn. This input field must be the name of a File. Chaining is allowed, meaning that an input field can point to another as name_source, which also points as name_source to a third field. In this situation, the templates for substitutions are also accumulated.
name_template
By default a %s_generated template is used to create the output filename. This metadata keyword allows overriding the generated name.
keep_extension
Use this and set it True if you want the extension from the input to be kept.

SPM

field

name of the structure refered by the SPM job manager

Required: This metadata is required by all SPM-mediated
interface classes.

Defining an interface class

Common

When you define an interface class, you will define these attributes and methods:

  • input_spec: the InputSpec
  • output_spec: the OutputSpec
  • _list_outputs(): Returns a dictionary containing names of generated files that are expected after package completes execution. This is used by BaseInterface.aggregate_outputs to gather all output files for the pipeline.

CommandLine

For command-line interfaces:

  • _cmd: the command-line command

If you used genfile:

  • _gen_filename(name): Generate filename, used for filenames that nipype generates as a convenience for users. This is for parameters that are required by the wrapped package, but we’re generating from some other parameter. For example, BET.inputs.outfile is required by BET but we can generate the name from BET.inputs.infile. Override this method in subclass to handle.

And optionally:

  • _redirect_x: If set to True it will make Nipype start Xvfb before running the interface and redirect X output to it. This is useful for

commandlines that spawn a graphical user interface.

  • _format_arg(name, spec, value): For extra formatting of the input values before passing them to generic _parse_inputs() method.

For example this is the class definition for Flirt, minus the docstring:

class FLIRTInputSpec(FSLCommandInputSpec):
    in_file = File(exists=True, argstr='-in %s', mandatory=True,
                   position=0, desc='input file')
    reference = File(exists=True, argstr='-ref %s', mandatory=True,
                     position=1, desc='reference file')
    out_file = File(argstr='-out %s', desc='registered output file',
                    name_source=['in_file'], name_template='%s_flirt',
                    position=2, hash_files=False)
    out_matrix_file = File(argstr='-omat %s',
                           name_source=['in_file'], keep_extension=True,
                           name_template='%s_flirt.mat',
                           desc='output affine matrix in 4x4 asciii format',
                           position=3, hash_files=False)
    out_log = File(name_source=['in_file'], keep_extension=True,
                   requires=['save_log'],
                   name_template='%s_flirt.log', desc='output log')
    ...

class FLIRTOutputSpec(TraitedSpec):
    out_file = File(exists=True,
                    desc='path/name of registered file (if generated)')
    out_matrix_file = File(exists=True,
                           desc='path/name of calculated affine transform '
                           '(if generated)')
    out_log = File(desc='path/name of output log (if generated)')

class Flirt(FSLCommand):
    _cmd = 'flirt'
    input_spec = FlirtInputSpec
    output_spec = FlirtOutputSpec

There are two possible output files outfile and outmatrix, both of which can be generated if not specified by the user.

Also notice the use of self._gen_fname() - a FSLCommand helper method for generating filenames (with extensions conforming with FSLOUTPUTTYPE).

See also How to wrap a command line tool.

SPM

For SPM-mediated interfaces:

  • _jobtype and _jobname: special names used used by the SPM job manager. You can find them by saving your batch job as an .m file and looking up the code.

And optionally:

  • _format_arg(name, spec, value): For extra formatting of the input values before passing them to generic _parse_inputs() method.

Undefined inputs

All the inputs and outputs that were not explicitly set (And do not have a usedefault flag - see above) will have Undefined value. To check if something is defined you have to explicitly call isdefiend function (comparing to None will not work).

Example of inputs

Below we have an example of using Bet. We can see from the help which inputs are mandatory and which are optional, along with the one-line description provided by the desc metadata:

>>> from nipype.interfaces import fsl
>>> fsl.BET.help()
Inputs
------

Mandatory:
 infile: input file to skull strip

Optional:
 args: Additional parameters to the command
 center: center of gravity in voxels
 environ: Environment variables (default={})
 frac: fractional intensity threshold
 functional: apply to 4D fMRI data
 mask: create binary mask image
 mesh: generate a vtk mesh brain surface
 nooutput: Don't generate segmented output
 outfile: name of output skull stripped image
 outline: create surface outline image
 outputtype: None
 radius: head radius
 reduce_bias: bias field and neck cleanup
 skull: create skull image
 threshold: apply thresholding to segmented brain image and mask
 vertical_gradient: vertical gradient in fractional intensity threshold (-1, 1)

Outputs
-------
maskfile: path/name of binary brain mask (if generated)
meshfile: path/name of vtk mesh file (if generated)
outfile: path/name of skullstripped file
outlinefile: path/name of outline file (if generated)

Here we create a bet object and specify the required input. We then check our inputs to see which are defined and which are not:

>>> bet = fsl.BET(infile = 'f3.nii')
>>> bet.inputs
args = <undefined>
center = <undefined>
environ = {'FSLOUTPUTTYPE': 'NIFTI_GZ'}
frac = <undefined>
functional = <undefined>
infile = f3.nii
mask = <undefined>
mesh = <undefined>
nooutput = <undefined>
outfile = <undefined>
outline = <undefined>
outputtype = NIFTI_GZ
radius = <undefined>
reduce_bias = <undefined>
skull = <undefined>
threshold = <undefined>
vertical_gradient = <undefined>
>>> bet.cmdline
'bet f3.nii /Users/cburns/data/nipype/s1/f3_brain.nii.gz'

We also checked the command-line that will be generated when we run the command and can see the generated output filename f3_brain.nii.gz.