Shared templating library¶

Pyconstruct provides a shared Jinja2 templating API with a number of utilities for building domains.

To use the available macros, simply import them from the available libraries, e.g.:

{% from 'globals.pmzn' import domain, solve %}
{% from 'linear.pmzn' import linear_model %}
{% from 'metrics.pmzn' import hamming %}

globals.pmzn¶

Collection of template macros to be used to define domains for structured classification. Implements standard procedures for handling different inference problems.

PROBLEMS¶

phi: Gets x and y as data, returns a feature vector phi(x, y).
n_features: Returns the number of features defined in the domain.
map: Gets x as data, computes the argmax of the score w.r.t. the given model.
loss_augmented_map: Gets x as data, computes the argmax of the score + loss.

CONSTANTS¶

BOOL_SET
    An integer set of boolean values {0, 1}.
INT_SET
    A standard set of integers.
FLOAT_SET
    A standard set of floats.

MACROS¶

domain(
    problem, allowed=('phi', 'map', 'loss_augmented_map')
)
    Boilerplate for domain definition.

    PARAMETERS
    ----------
    problem : str in ['phi', 'n_features', 'map', 'loss_augmented_map']
        The problem to solve. This parameter is usually passed by the Python
        domain class.
    allowed : tuple
        Allowed problems for this domain.

    USAGE
    -----
        {% call domain(problem) %}
            % Your domain definition, e.g.
            % int: x;
            % var 0 .. 10: y;
            % constraint x + y <= 10;
        {% endcall %}


solve(
    problem, score='score', loss='loss'
)
    Boilerplate for problem dependent solve statement.

    PARAMETERS
    ----------
    problem : str in ['phi', 'n_features', 'map', 'loss_augmented_map']
        The problem to solve. This parameter is usually passed by the Python
        domain class.
    score : str
        The score function. Default to variable name 'score'.
    loss : str
        Formula for the loss. Used in loss_augmented_map problems.

    USAGE
    -----
        {% set loss = 'some loss' %}
        {{ solve(problem, loss=loss) }}

linear.pmzn¶

MACROS¶

feature_set(
    params=none, n_features_var='N_FEATURES', feature_set_var='FEATURES'
)
    Boilerplate for feature number definition.

    PARAMETERS
    ----------
    params : dict
        Parameters of the model.
    n_feature_var : str
        The variable name of the feature array length. Default is 'N_FEATURES'.
    feature_set_var : str
        The variable name of the feature index set: Default is 'FEATURES'.

    USAGE
    -----
    Only fixed domain features:
        {% call feature_set() %}
            N_FEATURE_SET_1 + N_FEATURE_SET_2 + 5
        {% endcall %}

    Only model-dependent features:
        {{ call feature_set(params) }}

    Combination of both:
        {% call feature_set(params) %}
            N_FEATURE_SET_1 + N_FEATURE_SET_2 + 5
        {% endcall %}

features(
    params=none, feature_var='phi', feature_set_var='FEATURES',
    feature_type=none
)
    Boilerplate for features definition.

    PARAMETERS
    ----------
    params : dict
        Parameters of the model.
    feature_var : str
        The variable name of the feature array. Default is 'phi'.
    feature_set_var : str
        The variable name of the feature index set: Default is 'FEATURES'.
    feature_type : str
        The type of the features. Default is FLOAT_SET.

    USAGE
    -----
    Only fixed domain features:
        {% call features() %}
            % Your features definition, e.g.
            % [ feature1, feature2 ] ++ [ feature3 ]
        {% endcall %}

    Only model-dependent features:
        {{ call features(params) }}

    Combination of both:
        {% call features(params) %}
            % Your features definition, e.g.
            % [ feature1, feature2 ] ++ [ feature3 ]
        {% endcall %}


linear_model(
    problem, params=none, n_features=none, n_features_var='N_FEATURES',
    feature_var='phi', feature_set_var='FEATURES', feature_type=none,
    discretize=False, factor=100, score_var='score',
    allowed=('phi', 'map', 'loss_augmented_map')
)
    Boilerplate for defining a linear model.

    PARAMETERS
    ----------
    problem : str in ['phi', 'n_features', 'map', 'loss_augmented_map']
        The problem to solve. This parameter is usually passed by the Python
        domain class.
    params : dict
        Parameters of the model.
    n_features : str
        The number of feautures (or a fomula to compute it).
    n_feature_var : str
        The variable name of the feature array length. Default is 'N_FEATURES'.
    feature_var : str
        The variable name of the feature array. Default is 'phi'.
    feature_set_var : str
        The variable name of the feature index set: Default is 'FEATURES'.
    feature_type : str
        The type of the features. Default is FLOAT_SET.
    discretize : bool
    Wheter to discretize the weight vector and the scoring function. Use only
    if feature_type and loss_type are int. Default 'False'.
    factor : int
        The pre-multiplicative factor before discretization.
    score_var : str
        The variable name of the score to be used for inference (dot product
        between weights and features). Default 'score'.
    allowed : tuple
        Allowed problems for this model.

    USAGE
    -----
        {% call linear_model(problem, params, n_features='20') %}
            % Your features definition, e.g.
            % [ feature1, feature2 ] ++ [ feature3 ]
        {% endcall %}

metrics.pmzn¶

MACROS¶

hamming(
    sequence_set='SEQUENCE', sequence='sequence', true_sequence='true_sequence'
)
    Formula for the hamming loss between two sequences.

    PARAMETERS
    ----------
    sequence_set : str
        The set to iterate over.
    sequence : str
        The name of predicted sequence.
    true_sequence : str
        The name of the true sequence.

chain.pmzn¶

MACROS¶

n_emission_features(
    n_attributes='N_ATTRIBUTES', n_labels='N_LABELS'
)
    Prints the number of emission feautures.

    PARAMETERS
    ----------
    n_attributes : str
        The value or variable name containing the number of input attributes.
    n_labels : str
        The value or variable name containing the number of output labels.

emission_features(
    attributes_var='attributes', sequence_var='sequence',
    attributes_set='ATTRIBUTES', sequence_set='SEQUENCE', labels_set='LABELS'
)
    Prints the emission feautures.

    PARAMETERS
    ----------
    attributes_var : str
        The variable name of the attribute matrix. Default 'attributes'.
    sequence_var : str
        The variable name of the sequence array. Default 'sequence'.
    attributes_set : str
        The value or variable name containing the indexset of input attributes.
    sequence_set : str
        The value or variable name containing the indexset of output labels.
    labels_set : str
        The value or variable name containing the set of label symbols.

n_transition_features(n_labels='N_LABELS')
    Prints the number of transition feautures.

    PARAMETERS
    ----------
    n_labels : str
        The value or variable name containing the number of output labels.

transition_features(
    sequence_var='sequence', sequence_set='SEQUENCE', labels_set='LABELS'
)
    Prints the transition feautures.

    PARAMETERS
    ----------
    sequence_var : str
        The variable name of the sequence array. Default 'sequence'.
    sequence_set : str
        The value or variable name containing the indexset of output labels.
    labels_set : str
        The value or variable name containing the set of label symbols.