Merge a6215dc into 29c9fc8

setup.py

@@ -26,6 +26,7 @@
         'numpy>=1.8.0',
         'pandas>=0.13.1',
         'patsy>=0.2.1',
+        'prettytable>=0.7.2',
         'pyyaml>=3.10',
         'scipy>=0.13.3',
         'shapely>=1.3.0',

urbansim/exceptions.py

@@ -0,0 +1,2 @@
+class ModelEvaluationError(Exception):
+    pass

urbansim/models/__init__.py

@@ -0,0 +1,2 @@
+from .regression import RegressionModel, RegressionModelGroup
+from .lcm import MNLLocationChoiceModel

urbansim/models/lcm.py

@@ -0,0 +1,266 @@
+from __future__ import print_function, division
+
+import numpy as np
+import pandas as pd
+from patsy import dmatrix
+from prettytable import PrettyTable
+
+from . import util
+from ..urbanchoice import interaction, mnl
+
+
+def unit_choice(chooser_ids, alternative_ids, probabilities):
+    """
+    Have a set of choosers choose from among alternatives according
+    to a probability distribution. Choice is binary: each
+    alternative can only be chosen once.
+
+    Parameters
+    ----------
+    chooser_ids : 1d array_like
+        Array of IDs of the agents that are making choices.
+    alternative_ids : 1d array_like
+        Array of IDs of alternatives among which agents are making choices.
+    probabilities : 1d array_like
+        The probability that an agent will choose an alternative.
+        Must be the same shape as `alternative_ids`. Unavailable
+        alternatives should have a probability of 0.
+
+    Returns
+    -------
+    choices : pandas.Series
+        Mapping of chooser ID to alternative ID. Some choosers
+        will map to a nan value when there are not enough alternatives
+        for all the choosers.
+
+    """
+    chooser_ids = np.asanyarray(chooser_ids)
+    alternative_ids = np.asanyarray(alternative_ids)
+    probabilities = np.asanyarray(probabilities)
+
+    choices = pd.Series([np.nan] * len(chooser_ids), index=chooser_ids)
+
+    if probabilities.sum() == 0:
+        # return all nan if there are no available units
+        return choices
+
+    # probabilities need to sum to 1 for np.random.choice
+    probabilities = probabilities / probabilities.sum()
+
+    # need to see if there are as many available alternatives as choosers
+    n_available = np.count_nonzero(probabilities)
+    n_choosers = len(chooser_ids)
+    n_to_choose = n_choosers if n_choosers < n_available else n_available
+
+    chosen = np.random.choice(
+        alternative_ids, size=n_to_choose, replace=False, p=probabilities)
+
+    # if there are fewer available units than choosers we need to pick
+    # which choosers get a unit
+    if n_to_choose == n_available:
+        chooser_ids = np.random.choice(
+            chooser_ids, size=n_to_choose, replace=False)
+
+    choices[chooser_ids] = chosen
+
+    return choices
+
+
+class MNLLocationChoiceModel(object):
+    """
+    A location choice model with the ability to store an estimated
+    model and predict new data based on the model.
+    Based on multinomial logit.
+
+    Parameters
+    ----------
+    model_expression : str
+        A patsy model expression. Should contain only a right-hand side.
+    sample_size : int
+        Number of choices to sample for estimating the model.
+    location_id_col : str, optional
+        Name of a column in the choosers table that corresponds to the
+        index of the location being chosen. If given, this is used to
+        make sure that during prediction only choosers that have nan
+        in this column choose new alternatives.
+    choosers_fit_filters : list of str, optional
+        Filters applied to choosers table before fitting the model.
+    choosers_predict_filters : list of str, optional
+        Filters applied to the choosers table before calculating
+        new data points.
+    alts_fit_filters : list of str, optional
+        Filters applied to the alternatives table before fitting the model.
+    alts_predict_filters : list of str, optional
+        Filters applied to the alternatives table before calculating
+        new data points.
+    interaction_predict_filters : list of str, optional
+        Filters applied to the merged choosers/alternatives table
+        before predicting agent choices.
+    choice_column : optional
+        Name of the column in the `alternatives` table that choosers
+        should choose. e.g. the 'building_id' column. If not provided
+        the alternatives index is used.
+    name : optional
+        Optional descriptive name for this model that may be used
+        in output.
+
+    """
+    def __init__(self, model_expression, sample_size, location_id_col=None,
+                 choosers_fit_filters=None, choosers_predict_filters=None,
+                 alts_fit_filters=None, alts_predict_filters=None,
+                 interaction_predict_filters=None,
+                 choice_column=None, name=None):
+        # LCMs never have a constant
+        self.model_expression = model_expression + ' - 1'
+        self.sample_size = sample_size
+        self.location_id_col = location_id_col
+        self.choosers_fit_filters = choosers_fit_filters
+        self.choosers_predict_filters = choosers_predict_filters
+        self.alts_fit_filters = alts_fit_filters
+        self.alts_predict_filters = alts_predict_filters
+        self.interaction_predict_filters = interaction_predict_filters
+        self.choice_column = choice_column
+        self.name = name or 'MNLLocationChoiceModel'
+
+        self._log_lks = None
+        self._model_columns = None
+        self.fit_results = None
+
+    def fit(self, choosers, alternatives, current_choice):
+        """
+        Fit and save model parameters based on given data.
+
+        Parameters
+        ----------
+        choosers : pandas.DataFrame
+            Table describing the agents making choices, e.g. households.
+        alternatives : pandas.DataFrame
+            Table describing the things from which agents are choosing,
+            e.g. buildings.
+        current_choice : pandas.Series
+            A Series describing the `alternatives` currently chosen
+            by the `choosers`. Should have an index matching `choosers`
+            and values matching the index of `alternatives`.
+
+        Returns
+        -------
+        null_ll : float
+            Null Log-liklihood
+        conv_ll : float
+            Log-liklihood at convergence
+        ll_ratio : float
+            Log-liklihood ratio
+
+        """
+        choosers = util.apply_filter_query(choosers, self.choosers_fit_filters)
+        current_choice = current_choice.loc[choosers.index]
+        alternatives = util.apply_filter_query(
+            alternatives, self.alts_fit_filters)
+        _, merged, chosen = interaction.mnl_interaction_dataset(
+            choosers, alternatives, self.sample_size, current_choice)
+        model_design = dmatrix(
+            self.model_expression, data=merged, return_type='dataframe')
+        self._model_columns = model_design.columns  # used for report
+        fit, results = mnl.mnl_estimate(
+            model_design.as_matrix(), chosen, self.sample_size)
+        self._log_lks = fit
+        self.fit_results = results
+        return fit
+
+    @property
+    def fitted(self):
+        """
+        True if model is ready for prediction.
+
+        """
+        return bool(self.fit_results)
+
+    def assert_fitted(self):
+        """
+        Raises `RuntimeError` if the model is not ready for prediction.
+
+        """
+        if not self.fitted:
+            raise RuntimeError('Model has not been fit.')
+
+    @property
+    def coefficients(self):
+        """
+        Model coefficients as a list.
+
+        """
+        self.assert_fitted()
+        return [x[0] for x in self.fit_results]
+
+    def report_fit(self):
+        """
+        Print a report of the fit results.
+
+        """
+        if not self.fitted:
+            print('Model not yet fit.')
+            return
+
+        print('Null Log-liklihood: {}'.format(self._log_lks[0]))
+        print('Log-liklihood at convergence: {}'.format(self._log_lks[1]))
+        print('Log-liklihood Ratio: {}\n'.format(self._log_lks[2]))
+
+        tbl = PrettyTable(
+            ['Component', 'Coefficient', 'Std. Error', 'T-Score'])
+        tbl.align['Component'] = 'l'
+        for c, x in zip(self._model_columns, self.fit_results):
+            tbl.add_row((c,) + x)
+
+        print(tbl)
+
+    def predict(self, choosers, alternatives):
+        """
+        Choose from among alternatives for a group of agents.
+
+        Parameters
+        ----------
+        choosers : pandas.DataFrame
+            Table describing the agents making choices, e.g. households.
+            Only the first item in this table is used for determining
+            agent probabilities of choosing alternatives.
+        alternatives : pandas.DataFrame
+            Table describing the things from which agents are choosing.
+
+        Returns
+        -------
+        choices : pandas.Series
+            Mapping of chooser ID to alternative ID. Some choosers
+            will map to a nan value when there are not enough alternatives
+            for all the choosers.
+
+        """
+        self.assert_fitted()
+
+        if self.location_id_col:
+            choosers = choosers[choosers[self.location_id_col].isnull()]
+        choosers = util.apply_filter_query(
+            choosers, self.choosers_predict_filters)
+        alternatives = util.apply_filter_query(
+            alternatives, self.alts_predict_filters)
+
+        # TODO: only using 1st item in choosers for determining probabilities.
+        # Need to expand options around this.
+        _, merged, _ = interaction.mnl_interaction_dataset(
+            choosers.head(1), alternatives, len(alternatives))
+        merged = util.apply_filter_query(
+            merged, self.interaction_predict_filters)
+        model_design = dmatrix(
+            self.model_expression, data=merged, return_type='dataframe')
+
+        # probabilities are returned from mnl_simulate as a 2d array
+        # and need to be flatted for use in unit_choice.
+        probabilities = mnl.mnl_simulate(
+            model_design.as_matrix(), self.coefficients,
+            numalts=len(merged), returnprobs=True).flatten()
+
+        # figure out exactly which things from which choices are drawn
+        alt_choices = (
+            merged[self.choice_column] if self.choice_column else merged.index)
+
+        return unit_choice(
+            choosers.index, alt_choices, probabilities)