11829570876

Committed 14 Nov 2024 02:25AM UTC coverage: 83.727% (-0.009%) from 83.736%

Build # 11829570876

Build Type

Pull #3114

github

Committed by

jdcpni

Commit Message

Merge branch 'fix/standard_output_ports_calculate' of https://github.com/PrincetonUniversity/PsyNeuLink into fix/standard_output_ports_calculate

# Conflicts:
#	psyneulink/library/components/mechanisms/processing/transfer/recurrenttransfermechanism.py

Pull Request Pull Request #3114: fix/standard_output_ports_calculate

Run Details

9409 of 12468 branches covered (75.47%)

Branch coverage included in aggregate %.

1 of 8 new or added lines in 2 files covered. (12.5%)

92 existing lines in 1 file now uncovered.

32561 of 37659 relevant lines covered (86.46%)

0.86 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

93.97

/psyneulink/core/components/functions/nonstateful/objectivefunctions.py

 #
# Princeton University licenses this file to You under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.  You may obtain a copy of the License at:
#     http://www.apache.org/licenses/LICENSE-2.0
# Unless required by applicable law or agreed to in writing, software distributed under the License is distributed
# on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and limitations under the License.
#
#
# *****************************************  OBJECTIVE FUNCTIONS *******************************************************
"""

* `Stability`
* `Energy`
* `Entropy`
* `Distance`

Functions that return a scalar evaluation of their input.

"""

import functools

import numpy as np
from beartype import beartype

from psyneulink._typing import Optional, Callable

from psyneulink.core import llvm as pnlvm
from psyneulink.core.components.component import DefaultsFlexibility
from psyneulink.core.components.functions.function import EPSILON, FunctionError, Function_Base, get_matrix
from psyneulink.core.globals.keywords import \
    CORRELATION, COSINE, COSINE_SIMILARITY, CROSS_ENTROPY, \
    DEFAULT_VARIABLE, DIFFERENCE, DISTANCE_FUNCTION, DISTANCE_METRICS, DOT_PRODUCT, \
    ENERGY, ENTROPY, EUCLIDEAN, HOLLOW_MATRIX, MATRIX, MAX_ABS_DIFF, NORMALIZE, \
    NORMED_L0_SIMILARITY, OBJECTIVE_FUNCTION_TYPE, INPUT_SHAPES, STABILITY_FUNCTION
from psyneulink.core.globals.parameters import FunctionParameter, Parameter, check_user_specified, copy_parameter_value
from psyneulink.core.globals.preferences.basepreferenceset import ValidPrefSet
from psyneulink.core.globals.utilities import DistanceMetricLiteral, safe_len, convert_to_np_array, convert_all_elements_to_np_array
from psyneulink.core.globals.utilities import is_iterable


__all__ = ['ObjectiveFunction', 'Stability', 'Distance', 'Energy', 'Entropy']


class ObjectiveFunction(Function_Base):
    """Abstract class of `Function` used for evaluating ports.
    """

    componentType = OBJECTIVE_FUNCTION_TYPE

    class Parameters(Function_Base.Parameters):
        """
            Attributes
            ----------

                metric
                    see `metric <ObjectiveFunction.metric>`

                    :default value: None
                    :type:

                normalize
                    see `normalize <ObjectiveFunction.normalize>`

                    :default value: False
                    :type: ``bool``
        """
        normalize = Parameter(False, stateful=False)
        metric = Parameter(None, stateful=False)


class Stability(ObjectiveFunction):
    """
    Stability(                                  \
        default_variable=None,                  \
        size=None,                              \
        matrix=HOLLOW_MATRIX,                   \
        metric=ENERGY                           \
        transfer_fct=None                       \
        normalize=False,                        \
        params=None,                            \
        owner=None,                             \
        prefs=None                              \
        )

    .. _StabilityFunction:

    Return the stability of `variable <Stability.variable>` based on a state transformation matrix.

    The value of `variable <Stability.variable>` is passed through the `matrix <Stability.matrix>`,
    transformed using the `transfer_fct <Stability.transfer_fct>` (if specified), and then compared with its initial
    value using the `distance metric <DistanceMetric>` specified by `metric <Stability.metric>`.  If `normalize
    <Stability.normalize>` is `True`, the result is normalized by the length of (number of elements in) `variable
    <Stability.variable>`.

    Arguments
    ---------

    variable : list or 1d array of numbers: Default class_defaults.variable
        specifies shape and default value of the array for which stability is calculated.

    input_shapes : int : None
        specifies length of the array over which stability is calculated;  can be used in place of default_value,
        in which case zeros are assigned as the value(s). An error is generated if both are specified but
        size != len(default_value).

    matrix : list, np.ndarray, or matrix keyword : default HOLLOW_MATRIX
        specifies the matrix of recurrent weights;  must be a square matrix with the same width as the
        length of `variable <Stability.variable>`.

    metric : keyword in DistanceMetrics : Default ENERGY
        specifies a `metric <DistanceMetrics>` from `DistanceMetrics` used to compute stability.

    transfer_fct : function or method : Default None
        specifies the function used to transform output of weight `matrix <Stability.matrix>`.

    normalize : bool : Default False
        specifies whether to normalize the stability value by the length of `variable <Stability.variable>`.

    params : Dict[param keyword: param value] : default None
        a `parameter dictionary <ParameterPort_Specification>` that specifies the parameters for the
        function.  Values specified for parameters in the dictionary override any assigned to those parameters in
        arguments of the constructor.

    owner : Component
        `component <Component>` to which to assign the Function.

    name : str : default see `name <Function.name>`
        specifies the name of the Function.

    prefs : PreferenceSet or specification dict : default Function.classPreferences
        specifies the `PreferenceSet` for the Function (see `prefs <Function_Base.prefs>` for details).


    Attributes
    ----------

    variable : 1d array
        array for which stability is calculated.

    size : int
        length of array for which stability is calculated.

    matrix : list, np.ndarray, function keyword, or MappingProjection : default HOLLOW_MATRIX
        weight matrix from each element of `variable <Stability.variablity>` to each other;  if a matrix other
        than HOLLOW_MATRIX is assigned, it is convolved with HOLLOW_MATRIX to eliminate self-connections from the
        stability calculation.

    metric : keyword in DistanceMetrics
        metric used to compute stability; must be a `DistanceMetrics` keyword. The `Distance` Function is used to
        compute the stability of `variable <Stability.variable>` with respect to its value after its transformation
        by `matrix <Stability.matrix>` and `transfer_fct <Stability.transfer_fct>`.

    transfer_fct : function or method
        function used to transform output of weight `matrix <Stability.matrix>` prior to computing stability.

    normalize : bool
        if `True`, result of stability calculation is normalized by the length of `variable <Stability.variable>`.

    params : Dict[param keyword: param value] : default None
        a `parameter dictionary <ParameterPort_Specification>` that specifies the parameters for the
        function.  Values specified for parameters in the dictionary override any assigned to those parameters in
        arguments of the constructor.

    owner : Component
        `component <Component>` to which to assign the Function.

    prefs : PreferenceSet or specification dict : default Function.classPreferences
        specifies the `PreferenceSet` for the Function (see `prefs <Function_Base.prefs>` for details).
    """

    componentName = STABILITY_FUNCTION

    class Parameters(ObjectiveFunction.Parameters):
        """
            Attributes
            ----------

                matrix
                    see `matrix <Stability.matrix>`

                    :default value: `HOLLOW_MATRIX`
                    :type: ``str``

                metric
                    see `metric <Stability.metric>`

                    :default value: `ENERGY`
                    :type: ``str``

                metric_fct
                    see `metric_fct <Stability.metric_fct>`

                    :default value: None
                    :type:

                transfer_fct
                    see `transfer_fct <Stability.transfer_fct>`

                    :default value: None
                    :type:
        """
        matrix = HOLLOW_MATRIX
        metric = Parameter(ENERGY, stateful=False)
        metric_fct = Parameter(None, stateful=False, loggable=False)
        transfer_fct = Parameter(None, stateful=False, loggable=False)
        normalize = FunctionParameter(False, function_name='metric_fct')

    @check_user_specified
    @beartype
    def __init__(self,
                 default_variable=None,
                 input_shapes=None,
                 matrix=None,
                 # metric:is_distance_metric=None,
                 metric: Optional[DistanceMetricLiteral] = None,
                 transfer_fct: Optional[Callable] = None,
                 normalize: Optional[bool] = None,
                 params=None,
                 owner=None,
                 prefs:  Optional[ValidPrefSet] = None):

        if input_shapes:
            if default_variable is None:
                default_variable = np.zeros(input_shapes)
            elif input_shapes != len(default_variable):
                raise FunctionError(f"Both {repr(DEFAULT_VARIABLE)} ({default_variable}) and {repr(INPUT_SHAPES)} ({input_shapes}) "
                                    f"are specified for {self.name} but are {INPUT_SHAPES}!=len({DEFAULT_VARIABLE}).")

        super().__init__(
            default_variable=default_variable,
            matrix=matrix,
            metric=metric,
            transfer_fct=transfer_fct,
            normalize=normalize,
            params=params,
            owner=owner,
            prefs=prefs,
        )

        # MODIFIED 6/12/19 NEW: [JDC]
        self._variable_shape_flexibility = DefaultsFlexibility.FLEXIBLE
        # MODIFIED 6/12/19 END

    def _validate_variable(self, variable, context=None):
        """Validates that variable is 1d array
        """
        if len(np.atleast_2d(variable)) != 1:
            raise FunctionError("Variable for {} must contain a single array or list of numbers".format(self.name))
        return variable

    def _validate_params(self, variable, request_set, target_set=None, context=None):
        """Validate matrix param

        `matrix <Stability.matrix>` argument must be one of the following
            - 2d list, np.ndarray
            - ParameterPort for one of the above
            - MappingProjection with a parameterPorts[MATRIX] for one of the above

        Parse matrix specification to insure it resolves to a square matrix
        (but leave in the form in which it was specified so that, if it is a ParameterPort or MappingProjection,
         its current value can be accessed at runtime (i.e., it can be used as a "pointer")
        """

        # Validate matrix specification
        # (str can be automatically transformed to variable shape)
        if MATRIX in target_set and not isinstance(target_set[MATRIX], str):

            from psyneulink.core.components.projections.pathway.mappingprojection import MappingProjection
            from psyneulink.core.components.ports.parameterport import ParameterPort

            matrix = target_set[MATRIX]

            if isinstance(matrix, MappingProjection):
                try:
                    matrix = matrix._parameter_ports[MATRIX].value
                    param_type_string = "MappingProjection's ParameterPort"
                except KeyError:
                    raise FunctionError("The MappingProjection specified for the {} arg of {} ({}) must have a {} "
                                        "ParameterPort that has been assigned a 2d array or matrix".
                                        format(MATRIX, self.name, matrix.shape, MATRIX))

            elif isinstance(matrix, ParameterPort):
                try:
                    matrix = matrix.value
                    param_type_string = "ParameterPort"
                except KeyError:
                    raise FunctionError("The value of the {} parameterPort specified for the {} arg of {} ({}) "
                                        "must be a 2d array or matrix".
                                        format(MATRIX, MATRIX, self.name, matrix.shape))

            else:
                param_type_string = "array or matrix"

            matrix = np.array(matrix)
            if matrix.ndim != 2:
                raise FunctionError("The value of the {} specified for the {} arg of {} ({}) "
                                    "must be a 2d array or matrix".
                                    format(param_type_string, MATRIX, self.name, matrix))
            rows = matrix.shape[0]
            cols = matrix.shape[1]

            # this mirrors the transformation in _function
            # it is a hack, and a general solution should be found
            squeezed = np.array(self.defaults.variable)
            if squeezed.ndim > 1:
                squeezed = np.squeeze(squeezed)

            size = safe_len(squeezed)

            if rows != size:
                raise FunctionError("The value of the {} specified for the {} arg of {} is the wrong size;"
                                    "it is {}x{}, but must be square matrix of size {}".
                                    format(param_type_string, MATRIX, self.name, rows, cols, size))

            if rows != cols:
                raise FunctionError("The value of the {} specified for the {} arg of {} ({}) "
                                    "must be a square matrix".
                                    format(param_type_string, MATRIX, self.name, matrix))

        super()._validate_params(request_set=request_set,
                                 target_set=target_set,
                                 context=context)

    def _parse_metric_fct_variable(self, variable):
        return convert_to_np_array([variable, variable])

    def _instantiate_attributes_before_function(self, function=None, context=None):
        """Instantiate matrix

        Specified matrix is convolved with HOLLOW_MATRIX
            to eliminate the diagonal (self-connections) from the calculation.
        The `Distance` Function is used for all calculations except ENERGY (which is not really a distance metric).
        If ENTROPY is specified as the metric, convert to CROSS_ENTROPY for use with the Distance Function.
        :param function:

        """
        from psyneulink.core.components.projections.pathway.mappingprojection import MappingProjection
        from psyneulink.core.components.ports.parameterport import ParameterPort

        # this mirrors the transformation in _function
        # it is a hack, and a general solution should be found
        squeezed = np.array(self.defaults.variable)
        if squeezed.ndim > 1:
            squeezed = np.squeeze(squeezed)

        size = safe_len(squeezed)

        matrix = self.parameters.matrix._get(context)

        if isinstance(matrix, MappingProjection):
            matrix = matrix._parameter_ports[MATRIX]
        elif isinstance(matrix, ParameterPort):
            pass
        else:
            matrix = get_matrix(matrix, size, size)

        self.parameters.matrix._set(matrix, context)

        self._hollow_matrix = get_matrix(HOLLOW_MATRIX, size, size)

        default_variable = [self.defaults.variable,
                            self.defaults.variable]

        if self.metric == ENTROPY:
            self.metric_fct = Distance(default_variable=default_variable, metric=CROSS_ENTROPY, normalize=self.parameters.normalize.default_value)
        elif self.metric in DISTANCE_METRICS._set():
            self.metric_fct = Distance(default_variable=default_variable, metric=self.metric, normalize=self.parameters.normalize.default_value)
        else:
            assert False, "Unknown metric"

        #FIXME: This is a hack to make sure metric-fct param is set
        self.parameters.metric_fct.set(self.metric_fct)

    def _update_default_variable(self, new_default_variable, context):
        from psyneulink.core.components.projections.pathway.mappingprojection import MappingProjection
        from psyneulink.core.components.ports.parameterport import ParameterPort

        # this mirrors the transformation in _function
        # it is a hack, and a general solution should be found
        new_default_variable = convert_all_elements_to_np_array(new_default_variable)
        size = safe_len(np.squeeze(new_default_variable))
        matrix = self.parameters.matrix._get(context)

        if isinstance(matrix, MappingProjection):
            matrix = matrix._parameter_ports[MATRIX]
        elif isinstance(matrix, ParameterPort):
            pass
        else:
            matrix = get_matrix(copy_parameter_value(self.defaults.matrix), size, size)

        self.parameters.matrix._set(matrix, context)

        self._hollow_matrix = get_matrix(HOLLOW_MATRIX, size, size)

        super()._update_default_variable(new_default_variable, context)

    def _gen_llvm_function_body(self, ctx, builder, params, state, arg_in, arg_out, *, tags:frozenset):
        # Dot product
        dot_out = builder.alloca(arg_in.type.pointee)
        matrix = ctx.get_param_or_state_ptr(builder, self, MATRIX, param_struct_ptr=params, state_struct_ptr=state)

        # Convert array pointer to pointer to the fist element
        matrix = builder.gep(matrix, [ctx.int32_ty(0), ctx.int32_ty(0)])
        vec_in = builder.gep(arg_in, [ctx.int32_ty(0), ctx.int32_ty(0)])
        vec_out = builder.gep(dot_out, [ctx.int32_ty(0), ctx.int32_ty(0)])

        input_length = ctx.int32_ty(arg_in.type.pointee.count)
        output_length = ctx.int32_ty(arg_in.type.pointee.count)
        builtin = ctx.import_llvm_function("__pnl_builtin_vxm")
        builder.call(builtin, [vec_in, matrix, input_length, output_length, vec_out])

        # Prepare metric function
        metric_fun = ctx.import_llvm_function(self.metric_fct)
        metric_in = builder.alloca(metric_fun.args[2].type.pointee)

        # Transfer Function if configured
        if self.transfer_fct is not None:
            #FIXME: implement this
            assert False, "Support for transfer functions is not implemented"
        else:
            # Check that transfer_fct is absent from the compiled parameter
            # structure or represented by an empty structure
            assert "transfer_fct" not in self.llvm_param_ids or ctx.get_param_or_state_ptr(builder, self, "transfer_fct", param_struct_ptr=params).type.pointee.elements == ()

            trans_out = builder.gep(metric_in, [ctx.int32_ty(0), ctx.int32_ty(1)])
            builder.store(builder.load(dot_out), trans_out)

        # Copy original variable
        builder.store(builder.load(arg_in), builder.gep(metric_in, [ctx.int32_ty(0), ctx.int32_ty(0)]))

        # Distance Function
        metric_params, metric_state = ctx.get_param_or_state_ptr(builder, self, "metric_fct", param_struct_ptr=params, state_struct_ptr=state)
        metric_out = arg_out
        builder.call(metric_fun, [metric_params, metric_state, metric_in, metric_out])
        return builder

    def _function(self,
                 variable=None,
                 context=None,
                 params=None,
                 ):
        """Calculate the stability of `variable <Stability.variable>`.

        Compare the value of `variable <Stability.variable>` with its value after transformation by
        `matrix <Stability.matrix>` and `transfer_fct <Stability.transfer_fct>` (if specified), using the specified
        `metric <Stability.metric>`.  If `normalize <Stability.normalize>` is `True`, the result is divided
        by the length of `variable <Stability.variable>`.

        Returns
        -------

        stability : scalar

        """

        # MODIFIED 6/12/19 NEW: [JDC]
        variable = np.array(variable)
        if variable.ndim > 1:
            variable = np.squeeze(variable)
        # MODIFIED 6/12/19 END

        matrix = self._get_current_parameter_value(MATRIX, context)
        if matrix is None:
            matrix = self.matrix

        current = variable

        transformed = np.dot(matrix * self._hollow_matrix, variable)
        if self.transfer_fct is not None:
            transformed = self.transfer_fct(transformed)

        result = self.metric_fct(variable=[current, transformed], context=context)

        return self.convert_output_type(result)


class Energy(Stability):
    """
    Energy(                           \
        default_variable=None,        \
        size=None,                    \
        matrix=INVERSE_HOLLOW_MATRIX, \
        transfer_fct=None             \
        normalize=False,              \
        params=None,                  \
        owner=None,                   \
        prefs=None                    \
        )

    .. _EnergyFunction:

    Subclass of `Stability` Function that returns the energy of an array.

    Arguments
    ---------

    variable : list or 1d array of numbers: Default class_defaults.variable
        specifies shape and default value of the array for which energy is calculated.

    input_shapes : int : None
        specifies length of the array over which energy is calculated;  can be used in place of default_value,
        in which case zeros are assigned as the value(s). An error is generated if both are specified but
        size != len(default_value).

    matrix : list, np.ndarray, or matrix keyword : default INVERSE_HOLLOW_MATRIX
        specifies the matrix of recurrent weights;  must be a square matrix with the same width as the
        length of `variable <Stability.variable>`.

    transfer_fct : function or method : Default None
        specifies the function used to transform output of `matrix <Stability.matrix>` prior to the energy calculation
        (see `Stablility <Stability>` for explanation).

    normalize : bool : Default False
        specifies whether to normalize the energy value by the length of `variable <Stability.variable>`.

    params : Dict[param keyword: param value] : default None
        a `parameter dictionary <ParameterPort_Specification>` that specifies the parameters for the
        function.  Values specified for parameters in the dictionary override any assigned to those parameters in
        arguments of the constructor.

    owner : Component
        `component <Component>` to which to assign the Function.

    name : str : default see `name <Function.name>`
        specifies the name of the Function.

    prefs : PreferenceSet or specification dict : default Function.classPreferences
        specifies the `PreferenceSet` for the Function (see `prefs <Function_Base.prefs>` for details).


    Attributes
    ----------

    variable : 1d array
        array for which energy is calculated.

    size : int
        length of array for which energy is calculated.

    matrix : list, np.ndarray, or matrix keyword
        weight matrix from each element of `variable <Energy.variablity>` to each other;  if a matrix other
        than INVERSE_HOLLOW_MATRIX is assigned, it is convolved with HOLLOW_MATRIX to eliminate self-connections from
        the energy calculation.

    transfer_fct : function or method
        function used to transform output of `matrix <Stability.matrix>` prior to computing energy
        (see `Stability` for explanation).

    normalize : bool
        if `True`, result of energy calculation is normalized by the length of `variable <Energy.variable>`.

    params : Dict[param keyword: param value] : default None
        a `parameter dictionary <ParameterPort_Specification>` that specifies the parameters for the
        function.  Values specified for parameters in the dictionary override any assigned to those parameters in
        arguments of the constructor.

    owner : Component
        `component <Component>` to which to assign the Function.

    prefs : PreferenceSet or specification dict : default Function.classPreferences
        specifies the `PreferenceSet` for the Function (see `prefs <Function_Base.prefs>` for details).
    """

    @check_user_specified
    def __init__(self,
                 default_variable=None,
                 input_shapes=None,
                 normalize:bool=None,
                 # transfer_fct=None,
                 matrix=None,
                 params=None,
                 owner=None,
                 prefs=None):

        super().__init__(
            default_variable=default_variable,
            input_shapes=input_shapes,
                         metric=ENERGY,
                         matrix=matrix,
                         # transfer_fct=transfer_fct,
                         normalize=normalize,
                         params=params,
                         owner=owner,
                         prefs=prefs)


class Entropy(Stability):
    """
    Entropy(                          \
        default_variable=None,        \
        input_shapes=None,                    \
        matrix=INVERSE_HOLLOW_MATRIX, \
        transfer_fct=None             \
        normalize=False,              \
        params=None,                  \
        owner=None,                   \
        prefs=None                    \
        )

    .. _EntropyFunction:

    Subclass of `Stability` Function that returns the entropy of an array.

    Arguments
    ---------

    variable : list or 1d array of numbers: Default class_defaults.variable
        specifies shape and default value of the array for which entropy is calculated.

    input_shapes : int : None
        specifies length of the array over which entropy is calculated;  can be used in place of default_value,
        in which case zeros are assigned as the value(s). An error is generated if both are specified but
        input_shapes != len(default_value).

    matrix : list, np.ndarray, or matrix keyword : default INVERSE_HOLLOW_MATRIX
        specifies the matrix of recurrent weights;  must be a square matrix with the same width as the
        length of `variable <Stability.variable>`.

    transfer_fct : function or method : Default None
        specifies the function used to transform output of `matrix <Stability.matrix>` prior to the entropy calculation
        (see `Stability` for explanation).

    normalize : bool : Default False
        specifies whether to normalize the entropy value by the length of `variable <Stability.variable>`.

    params : Dict[param keyword: param value] : default None
        a `parameter dictionary <ParameterPort_Specification>` that specifies the parameters for the
        function.  Values specified for parameters in the dictionary override any assigned to those parameters in
        arguments of the constructor.

    owner : Component
        `component <Component>` to which to assign the Function.

    name : str : default see `name <Function.name>`
        specifies the name of the Function.

    prefs : PreferenceSet or specification dict : default Function.classPreferences
        specifies the `PreferenceSet` for the Function (see `prefs <Function_Base.prefs>` for details).


    Attributes
    ----------

    variable : 1d array
        array for which entropy is calculated.

    input_shapes : int
        length of array for which energy is calculated.

    matrix : list, np.ndarray, or matrix keyword
        weight matrix from each element of `variable <Entropy.variablity>` to each other;  if a matrix other
        than INVERSE_HOLLOW_MATRIX is assigned, it is convolved with HOLLOW_MATRIX to eliminate self-connections from
        the entropy calculation.

    transfer_fct : function or method
        function used to transform output of `matrix <Stability.matrix>` prior to computing entropy
        (see `Stability` for explanation).

    normalize : bool
        if `True`, result of entropy calculation is normalized by the length of `variable <Entropy.variable>`.

    params : Dict[param keyword: param value] : default None
        a `parameter dictionary <ParameterPort_Specification>` that specifies the parameters for the
        function.  Values specified for parameters in the dictionary override any assigned to those parameters in
        arguments of the constructor.

    owner : Component
        `component <Component>` to which to assign the Function.

    prefs : PreferenceSet or specification dict : default Function.classPreferences
        specifies the `PreferenceSet` for the Function (see `prefs <Function_Base.prefs>` for details).
    """

    @check_user_specified
    def __init__(self,
                 default_variable=None,
                 normalize:bool=None,
                 transfer_fct=None,
                 params=None,
                 owner=None,
                 prefs=None):

        super().__init__(
            default_variable=default_variable,
            # matrix=matrix,
                         metric=ENTROPY,
                         transfer_fct=transfer_fct,
                         normalize=normalize,
                         params=params,
                         owner=owner,
                         prefs=prefs)


class Distance(ObjectiveFunction):
    """
    Distance(                                    \
       default_variable=None,                    \
       metric=EUCLIDEAN                          \
       normalize=False,                          \
       params=None,                              \
       owner=None,                               \
       prefs=None                                \
       )

    .. _Distance:

    Return the distance between the vectors in the two items of `variable <Distance.variable>` using the `distance
    metric <DistanceMetrics>` specified in the `metric <Stability.metric>` attribute.  If `normalize
    <Distance.normalize>` is `True`, the result is normalized by the length of (number of elements in) `variable
    <Stability.variable>`.

    Arguments
    ---------

    variable : 2d array with two items : Default class_defaults.variable
        the arrays between which the distance is calculated.

    metric : keyword in DistancesMetrics : Default EUCLIDEAN
        specifies a `distance metric <DistanceMetrics>` used to compute the distance between the two items in `variable
        <Distance.variable>`.

    normalize : bool : Default False
        specifies whether to normalize the distance by the length of `variable <Distance.variable>`.

    params : Dict[param keyword: param value] : default None
        a `parameter dictionary <ParameterPort_Specification>` that specifies the parameters for the
        function.  Values specified for parameters in the dictionary override any assigned to those parameters in
        arguments of the constructor.

    owner : Component
        `component <Component>` to which to assign the Function.

    name : str : default see `name <Function.name>`
        specifies the name of the Function.

    prefs : PreferenceSet or specification dict : default Function.classPreferences
        specifies the `PreferenceSet` for the Function (see `prefs <Function_Base.prefs>` for details).

    Attributes
    ----------

    variable : 2d array with two items
        contains the arrays between which the distance is calculated.

    metric : keyword in DistanceMetrics
        determines the `metric <DistanceMetrics>` used to compute the distance between the two items in `variable
        <Distance.variable>`.

    normalize : bool
        determines whether the distance is normalized by the length of `variable <Distance.variable>`.

    params : Dict[param keyword: param value] : default None
        a `parameter dictionary <ParameterPort_Specification>` that specifies the parameters for the
        function.  Values specified for parameters in the dictionary override any assigned to those parameters in
        arguments of the constructor.

    owner : Component
        `component <Component>` to which to assign the Function.

    prefs : PreferenceSet or specification dict : default Function.classPreferences
        specifies the `PreferenceSet` for the Function (see `prefs <Function_Base.prefs>` for details).
    """

    componentName = DISTANCE_FUNCTION

    class Parameters(ObjectiveFunction.Parameters):
        """
            Attributes
            ----------

                variable
                    see `variable <Distance.variable>`

                    :default value: numpy.array([[0], [0]])
                    :type: ``numpy.ndarray``
                    :read only: True

                metric
                    see `metric <Distance.metric>`

                    :default value: `DIFFERENCE`
                    :type: ``str``
        """
        variable = Parameter(np.array([[0], [0]]), read_only=True, pnl_internal=True, constructor_argument='default_variable')
        metric = Parameter(DIFFERENCE, stateful=False)

    @check_user_specified
    @beartype
    def __init__(self,
                 default_variable=None,
                 metric: Optional[DistanceMetricLiteral] = None,
                 normalize: Optional[bool] = None,
                 params=None,
                 owner=None,
                 prefs:  Optional[ValidPrefSet] = None):
        super().__init__(
            default_variable=default_variable,
            metric=metric,
            normalize=normalize,
            params=params,
            owner=owner,
            prefs=prefs,
        )

    def _validate_params(self, request_set, target_set=None, variable=None, context=None):
        """Validate that variable has two items of equal length

        """
        super()._validate_params(request_set=request_set, target_set=target_set, context=context)

        err_two_items = FunctionError("variable for {} ({}) must have two items".format(self.name, variable))

        try:
            if len(variable) != 2:
                raise err_two_items
        except TypeError:
            raise err_two_items

        try:
            if len(variable[0]) != len(variable[1]):
                raise FunctionError(
                    "The lengths of the items in the variable for {0} ({1},{2}) must be equal".format(
                        self.name,
                        variable[0],
                        variable[1]
                    )
                )
        except TypeError:
            if is_iterable(variable[0]) ^ is_iterable(variable[1]):
                raise FunctionError(
                    "The lengths of the items in the variable for {0} ({1},{2}) must be equal".format(
                        self.name,
                        variable[0],
                        variable[1]
                    )
                )

    def cosine(v1, v2):
        numer = np.sum(v1 * v2)
        denom = np.sqrt(np.sum(v1 ** 2)) * np.sqrt(np.sum(v2 ** 2)) or EPSILON
        return numer / denom

    def correlation(v1, v2):
        v1_norm = v1 - np.mean(v1)
        v2_norm = v2 - np.mean(v2)
        denom = np.sqrt(np.sum(v1_norm ** 2) * np.sum(v2_norm ** 2)) or EPSILON
        return np.sum(v1_norm * v2_norm) / denom

    def __gen_llvm_sum_difference(self, builder, index, ctx, v1, v2, acc):
        ptr1 = builder.gep(v1, [index])
        ptr2 = builder.gep(v2, [index])
        val1 = builder.load(ptr1)
        val2 = builder.load(ptr2)

        sub = builder.fsub(val1, val2)
        abs_f = ctx.get_builtin("fabs", [sub.type])
        abs_val = builder.call(abs_f, [sub])
        acc_val = builder.load(acc)
        new_acc = builder.fadd(acc_val, abs_val)
        builder.store(new_acc, acc)

    def __gen_llvm_sum_diff_squares(self, builder, index, ctx, v1, v2, acc):
        ptr1 = builder.gep(v1, [index])
        ptr2 = builder.gep(v2, [index])
        val1 = builder.load(ptr1)
        val2 = builder.load(ptr2)

        sub = builder.fsub(val1, val2)
        sqr = builder.fmul(sub, sub)
        acc_val = builder.load(acc)
        new_acc = builder.fadd(acc_val, sqr)
        builder.store(new_acc, acc)

    def __gen_llvm_cross_entropy(self, builder, index, ctx, v1, v2, acc):
        ptr1 = builder.gep(v1, [index])
        ptr2 = builder.gep(v2, [index])
        val1 = builder.load(ptr1)
        val2 = builder.load(ptr2)

        log_f = ctx.get_builtin("log", [ctx.float_ty])
        log = builder.call(log_f, [val2])
        prod = builder.fmul(val1, log)

        acc_val = builder.load(acc)
        new_acc = builder.fsub(acc_val, prod)
        builder.store(new_acc, acc)

    def __gen_llvm_sum_product(self, builder, index, ctx, v1, v2, acc):
        ptr1 = builder.gep(v1, [index])
        ptr2 = builder.gep(v2, [index])
        val1 = builder.load(ptr1)
        val2 = builder.load(ptr2)

        prod = builder.fmul(val1, val2)

        acc_val = builder.load(acc)
        new_acc = builder.fadd(acc_val, prod)
        builder.store(new_acc, acc)

    def __gen_llvm_cosine(self, builder, index, ctx, v1, v2, numer_acc, denom1_acc, denom2_acc):
        ptr1 = builder.gep(v1, [index])
        ptr2 = builder.gep(v2, [index])
        val1 = builder.load(ptr1)
        val2 = builder.load(ptr2)

        # Numerator
        numer = builder.load(numer_acc)
        val = builder.fmul(val1, val2)
        numer = builder.fadd(numer, val)
        builder.store(numer, numer_acc)

        # Denominator1
        denom = builder.load(denom1_acc)
        val = builder.fmul(val1, val1)
        denom = builder.fadd(denom, val)
        builder.store(denom, denom1_acc)

        # Denominator2
        denom = builder.load(denom2_acc)
        val = builder.fmul(val2, val2)
        denom = builder.fadd(denom, val)
        builder.store(denom, denom2_acc)

    def __gen_llvm_max_diff(self, builder, index, ctx, v1, v2, max_diff_ptr):
        ptr1 = builder.gep(v1, [index])
        ptr2 = builder.gep(v2, [index])
        val1 = builder.load(ptr1)
        val2 = builder.load(ptr2)

        # Get the difference
        diff = builder.fsub(val1, val2)

        # Get absolute value
        fabs = ctx.get_builtin("fabs", [ctx.float_ty])
        diff = builder.call(fabs, [diff])

        # Get maximum
        old_max = builder.load(max_diff_ptr)
        fmax = ctx.get_builtin("maxnum", [ctx.float_ty])
        max_diff = builder.call(fmax, [diff, old_max])

        builder.store(max_diff, max_diff_ptr)

    def __gen_llvm_pearson(self, builder, index, ctx, v1, v2, acc_x, acc_y, acc_xy, acc_x2, acc_y2):
        ptr1 = builder.gep(v1, [index])
        ptr2 = builder.gep(v2, [index])
        val1 = builder.load(ptr1)
        val2 = builder.load(ptr2)

        # Sum X
        acc_x_val = builder.load(acc_x)
        acc_x_val = builder.fadd(acc_x_val, val1)
        builder.store(acc_x_val, acc_x)

        # Sum Y
        acc_y_val = builder.load(acc_y)
        acc_y_val = builder.fadd(acc_y_val, val2)
        builder.store(acc_y_val, acc_y)

        # Sum XY
        acc_xy_val = builder.load(acc_xy)
        xy = builder.fmul(val1, val2)
        acc_xy_val = builder.fadd(acc_xy_val, xy)
        builder.store(acc_xy_val, acc_xy)

        # Sum X2
        acc_x2_val = builder.load(acc_x2)
        x2 = builder.fmul(val1, val1)
        acc_x2_val = builder.fadd(acc_x2_val, x2)
        builder.store(acc_x2_val, acc_x2)

        # Sum Y2
        acc_y2_val = builder.load(acc_y2)
        y2 = builder.fmul(val2, val2)
        acc_y2_val = builder.fadd(acc_y2_val, y2)
        builder.store(acc_y2_val, acc_y2)

    def _gen_llvm_function_body(self, ctx, builder, params, _, arg_in, arg_out, *, tags:frozenset):
        assert isinstance(arg_in.type.pointee, pnlvm.ir.ArrayType)
        assert isinstance(arg_in.type.pointee.element, pnlvm.ir.ArrayType)
        assert arg_in.type.pointee.count == 2

        v1 = builder.gep(arg_in, [ctx.int32_ty(0), ctx.int32_ty(0), ctx.int32_ty(0)])
        v2 = builder.gep(arg_in, [ctx.int32_ty(0), ctx.int32_ty(1), ctx.int32_ty(0)])

        acc_ptr = builder.alloca(ctx.float_ty)
        builder.store(ctx.float_ty(-0.0), acc_ptr)

        kwargs = {"ctx": ctx, "v1": v1, "v2": v2, "acc": acc_ptr}
        if self.metric == DIFFERENCE or self.metric == NORMED_L0_SIMILARITY:
            inner = functools.partial(self.__gen_llvm_sum_difference, **kwargs)
        elif self.metric == EUCLIDEAN:
            inner = functools.partial(self.__gen_llvm_sum_diff_squares, **kwargs)
        elif self.metric == ENERGY or self.metric == DOT_PRODUCT:
            inner = functools.partial(self.__gen_llvm_sum_product, **kwargs)
        elif self.metric == CROSS_ENTROPY:
            inner = functools.partial(self.__gen_llvm_cross_entropy, **kwargs)
        elif self.metric in {COSINE, COSINE_SIMILARITY}:
            del kwargs['acc']
            numer_acc = builder.alloca(ctx.float_ty)
            denom1_acc = builder.alloca(ctx.float_ty)
            denom2_acc = builder.alloca(ctx.float_ty)
            for loc in numer_acc, denom1_acc, denom2_acc:
                builder.store(ctx.float_ty(-0.0), loc)
            kwargs['numer_acc'] = numer_acc
            kwargs['denom1_acc'] = denom1_acc
            kwargs['denom2_acc'] = denom2_acc
            inner = functools.partial(self.__gen_llvm_cosine, **kwargs)
        elif self.metric == MAX_ABS_DIFF:
            del kwargs['acc']
            max_diff_ptr = builder.alloca(ctx.float_ty)
            builder.store(ctx.float_ty(float("NaN")), max_diff_ptr)
            kwargs['max_diff_ptr'] = max_diff_ptr
            inner = functools.partial(self.__gen_llvm_max_diff, **kwargs)
        elif self.metric == CORRELATION:
            acc_x_ptr = builder.alloca(ctx.float_ty)
            acc_y_ptr = builder.alloca(ctx.float_ty)
            acc_xy_ptr = builder.alloca(ctx.float_ty)
            acc_x2_ptr = builder.alloca(ctx.float_ty)
            acc_y2_ptr = builder.alloca(ctx.float_ty)
            for loc in [acc_x_ptr, acc_y_ptr, acc_xy_ptr, acc_x2_ptr, acc_y2_ptr]:
                builder.store(ctx.float_ty(-0.0), loc)
            del kwargs['acc']
            kwargs['acc_x'] = acc_x_ptr
            kwargs['acc_y'] = acc_y_ptr
            kwargs['acc_xy'] = acc_xy_ptr
            kwargs['acc_x2'] = acc_x2_ptr
            kwargs['acc_y2'] = acc_y2_ptr
            inner = functools.partial(self.__gen_llvm_pearson, **kwargs)
        else:
            raise RuntimeError('Unsupported metric')

        input_length = arg_in.type.pointee.element.count
        vector_length = ctx.int32_ty(input_length)
        with pnlvm.helpers.for_loop_zero_inc(builder, vector_length, self.metric) as args:
            inner(*args)

        sqrt = ctx.get_builtin("sqrt", [ctx.float_ty])
        fabs = ctx.get_builtin("fabs", [ctx.float_ty])
        ret = builder.load(acc_ptr)
        if self.metric == NORMED_L0_SIMILARITY:
            ret = builder.fdiv(ret, ret.type(4))
            ret = builder.fsub(ret.type(1), ret)
        elif self.metric == DOT_PRODUCT:
            pass # the dot product has already been computed above by __gen_llvm_sum_product
        elif self.metric == ENERGY:
            ret = builder.fmul(ret, ret.type(-0.5))
        elif self.metric == EUCLIDEAN:
            ret = builder.call(sqrt, [ret])
        elif self.metric == MAX_ABS_DIFF:
            ret = builder.load(max_diff_ptr)
        elif self.metric in {COSINE, COSINE_SIMILARITY}:
            numer = builder.load(numer_acc)
            denom1 = builder.load(denom1_acc)
            denom1 = builder.call(sqrt, [denom1])
            denom2 = builder.load(denom2_acc)
            denom2 = builder.call(sqrt, [denom2])
            denom = builder.fmul(denom1, denom2)

            ret = builder.fdiv(numer, denom)
            ret = builder.call(fabs, [ret])
            ret = builder.fsub(ret.type(1), ret)

        elif self.metric == CORRELATION:
            n = ctx.float_ty(input_length)
            acc_xy = builder.load(acc_xy_ptr)
            acc_x = builder.load(acc_x_ptr)
            acc_y = builder.load(acc_y_ptr)
            acc_x2 = builder.load(acc_x2_ptr)
            acc_y2 = builder.load(acc_y2_ptr)

            # We'll need meanx,y below
            mean_x = builder.fdiv(acc_x, n)
            mean_y = builder.fdiv(acc_y, n)

            # Numerator: sum((x - mean(x))*(y - mean(y)) =
            # sum(x*y - x*mean(y) - y*mean(x) + mean(x)*mean(y)) =
            # sum(x*y) - sum(x)*mean(y) - sum(y)*mean(x) + mean(x)*mean(y)*n
            b = builder.fmul(acc_x, mean_y)
            c = builder.fmul(acc_y, mean_x)
            d = builder.fmul(mean_x, mean_y)
            d = builder.fmul(d, n)

            numerator = builder.fsub(acc_xy, b)
            numerator = builder.fsub(numerator, c)
            numerator = builder.fadd(numerator, d)

            # Denominator: sqrt(D_X * D_Y)
            # D_X = sum((x - mean(x))^2) = sum(x^2 - 2*x*mean(x) + mean(x)^2) =
            # sum(x^2) - 2 * sum(x) * mean(x) + n * mean(x)^2
            dxb = builder.fmul(acc_x, mean_x)
            dxb = builder.fadd(dxb, dxb)        # *2
            dxc = builder.fmul(mean_x, mean_x)  # ^2
            dxc = builder.fmul(dxc, n)

            dx = builder.fsub(acc_x2, dxb)
            dx = builder.fadd(dx, dxc)

            # Similarly for y
            dyb = builder.fmul(acc_y, mean_y)
            dyb = builder.fadd(dyb, dyb)        # *2
            dyc = builder.fmul(mean_y, mean_y)  # ^2
            dyc = builder.fmul(dyc, n)

            dy = builder.fsub(acc_y2, dyb)
            dy = builder.fadd(dy, dyc)

            # Denominator: sqrt(D_X * D_Y)
            denominator = builder.fmul(dx, dy)
            denominator = builder.call(sqrt, [denominator])

            corr = builder.fdiv(numerator, denominator)

            # ret =  1 - abs(corr)
            ret = builder.call(fabs, [corr])
            ret = builder.fsub(ctx.float_ty(1), ret)

        if arg_out.type.pointee != ret.type:
            # Some instances use 2d output values
            arg_out = builder.gep(arg_out, [ctx.int32_ty(0), ctx.int32_ty(0),
                                            ctx.int32_ty(0)])

        normalize_ptr = ctx.get_param_or_state_ptr(builder, self, NORMALIZE, param_struct_ptr=params)
        normalize = builder.load(normalize_ptr)
        normalize_b = builder.fcmp_ordered("!=", normalize, normalize.type(0))

        # MAX_ABS_DIFF, CORRELATION, and COSINE/COSINE_SIMILARITY ignore normalization
        allow_normalize_b = normalize_b.type(self.metric not in {MAX_ABS_DIFF, CORRELATION, COSINE, COSINE_SIMILARITY})
        normalize_b = builder.and_(normalize_b, allow_normalize_b)
        with builder.if_else(normalize_b) as (then, otherwise):
            with then:
                norm_factor = input_length ** 2 if self.metric == ENERGY else input_length
                normalized = builder.fdiv(ret, ctx.float_ty(norm_factor), name="normalized")
                builder.store(normalized, arg_out)
            with otherwise:
                builder.store(ret, arg_out)

        return builder

    def _function(self,
                 variable=None,
                 context=None,
                 params=None,
                 ):
        """Calculate the distance between the two vectors in `variable <Stability.variable>`.

        Use the `distance metric <DistanceMetrics>` specified in `metric <Distance.metric>` to calculate the distance.
        If `normalize <Distance.normalize>` is `True`, the result is divided by the length of `variable
        <Distance.variable>`.

        Returns
        -------

        distance : scalar

        """

        try:
            v1 = np.hstack(variable[0])
        except TypeError:
            v1 = variable[0]

        try:
            v2 = np.hstack(variable[1])
        except TypeError:
            v2 = variable[1]

        # Maximum of Hadamard (elementwise) difference of v1 and v2
        if self.metric == MAX_ABS_DIFF:
            result = np.max(np.fabs(v1 - v2))

        # Simple Hadamard (elementwise) difference of v1 and v2
        elif self.metric == DIFFERENCE:
            result = np.sum(np.fabs(v1 - v2))

        # Similarity (used specifically for testing Compilation of Predator-Prey Model)
        elif self.metric == NORMED_L0_SIMILARITY:
            result = 1.0 - np.sum(np.abs(v1 - v2)) / 4.0

        # Simple dot product of v1 and v2
        elif self.metric == DOT_PRODUCT:
            result = np.dot(v1, v2)

        # Euclidean distance between v1 and v2
        elif self.metric == EUCLIDEAN:
            result = np.linalg.norm(v2 - v1)

        # Cosine similarity of v1 and v2
        elif self.metric in {COSINE, COSINE_SIMILARITY}:
            result = 1.0 - np.fabs(Distance.cosine(v1, v2))

        # Correlation of v1 and v2
        elif self.metric == CORRELATION:
            # result = np.correlate(v1, v2)
            result = 1.0 - np.fabs(Distance.correlation(v1, v2))

        # FIX: IMPLEMENT VERSION THAT DIRECTLY COMPUTES THE LUCE RATIO
        # Cross-entropy of v1 and v2
        elif self.metric == CROSS_ENTROPY:
            # FIX: VALIDATE THAT ALL ELEMENTS OF V1 AND V2 ARE 0 TO 1
            if not self.is_initializing:
                v1 = np.where(v1 == 0, EPSILON, v1)
                v2 = np.where(v2 == 0, EPSILON, v2)
            # MODIFIED CW 3/20/18: avoid divide by zero error by plugging in two zeros
            # FIX: unsure about desired behavior when v2 = 0 and v1 != 0
            # JDC: returns [inf]; leave, and let it generate a warning or error message for user
            both_zero = np.logical_and(v1 == 0, v2 == 0)
            result = -np.sum(v1 * np.where(both_zero, 0.0, np.log(v2, where=np.logical_not(both_zero))))

        # Energy
        elif self.metric == ENERGY:
            result = -np.sum(v1 * v2) / 2.0

        else:
            assert False, '{} not a recognized metric in {}'.format(self.metric, self.__class__.__name__)

        if self.normalize and self.metric not in {MAX_ABS_DIFF, CORRELATION, COSINE, COSINE_SIMILARITY}:
            if self.metric == ENERGY:
                result /= len(v1) ** 2.0
            else:
                result /= len(v1)

        return self.convert_output_type(result)

PrincetonUniversity / PsyNeuLink / 11829570876

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous