op_search.py

# -*- coding: utf-8 -*-
'''
Custom theano class to query the search engine.
'''

import numpy as np
import theano
from theano import gof
from theano import tensor
import parameters as prm
import utils
import average_precision
import random


class Search(theano.Op):
    __props__ = ()

    def __init__(self,options):
        self.options = options
        self.options['reformulated_queries'] = {}


    def make_node(self, x1, x2, x3, x4):
        assert hasattr(self, '_props'), "Your version of theano is too old to support __props__."
        x1 = tensor.as_tensor_variable(x1)
        x2 = tensor.as_tensor_variable(x2)
        x3 = tensor.as_tensor_variable(x3)
        x4 = tensor.as_tensor_variable(x4)
        out = [tensor.fmatrix().type(), tensor.itensor3().type(), tensor.imatrix().type(), tensor.fmatrix().type()]

        return theano.Apply(self, [x1, x2, x3, x4], out)


    def perform(self, node, inputs, output_storage):
        q_m = inputs[0]
        D_truth = inputs[1]
        n_iter = int(inputs[2])
        is_train = int(inputs[3])

        #outputs
        metrics = np.zeros((len(q_m), len(prm.metrics_map)), np.float32)

        if is_train:
            max_feedback_docs = prm.max_feedback_docs_train
        else:
            max_feedback_docs = prm.max_feedback_docs

        D_i = -2 * np.ones((len(q_m), max_feedback_docs, prm.max_words_input), np.int32)
        D_gt_m = np.zeros((len(q_m), prm.max_candidates), np.float32)
        D_id = np.zeros((len(q_m), prm.max_candidates), np.int32)

        # no need to retrieve extra terms in the last iteration
        if n_iter == prm.n_iterations - 1:
            extra_terms = False
        else:
            extra_terms = True

        # allow the search engine to cache queries only in the first iteration.
        if n_iter == 0:
            save_cache = prm.use_cache
        else:
            save_cache = False
    
        max_cand = prm.max_candidates

        qs = []
        for i, q_lst in enumerate(self.options['current_queries']):
            q = []
            for j, word in enumerate(q_lst):
                if q_m[i,j] == 1:
                    q.append(str(word))
            q = ' '.join(q)

            if len(q) == 0:
                q = 'dummy'
            qs.append(q)

        # only used to print the reformulated queries.
        self.options['reformulated_queries'][n_iter] = qs

        # always return one more candidate because one of them might be the input doc.
        candss = self.options['engine'].get_candidates(qs, max_cand, prm.max_feedback_docs, save_cache, extra_terms)

        for i, cands in enumerate(candss):

            D_truth_dic = {}
            for d_truth in D_truth[i]:
                if d_truth > -1:
                    D_truth_dic[d_truth] = 0

            D_id[i,:len(cands.keys())] = cands.keys()

            j = 0
            m = 0
            cand_ids = []

            selected_docs = np.arange(prm.max_feedback_docs)

            if is_train:
                selected_docs = np.random.choice(selected_docs, size=prm.max_feedback_docs_train, replace=False)

            for k, (cand_id, (words_idx, words)) in enumerate(cands.items()):

                cand_ids.append(cand_id)

            
                # no need to add candidate words in the last iteration.
                if n_iter < prm.n_iterations - 1:
                    # only add docs selected by sampling (if training).
                    if k in selected_docs:
                        
                        words = words[:prm.max_terms_per_doc]
                        words_idx = words_idx[:prm.max_terms_per_doc]

                        D_i[i,m,:len(words_idx)] = words_idx

                        # append empty strings, so the list size becomes <dim>.
                        words = words + max(0, prm.max_words_input - len(words)) * ['']

                        # append new words to the list of current queries.
                        self.options['current_queries'][i] += words

                        m += 1

                if cand_id in D_truth_dic:
                    D_gt_m[i,j] = 1.
                
                j += 1

            cands_set = set(cands.keys())

            if qs[i].lower() in self.options['engine'].title_id_map:
                input_doc_id = self.options['engine'].title_id_map[qs[i].lower()]
                # Remove input doc from returned docs.
                # This operation does not raise an error if the element is not there.
                cands_set.discard(input_doc_id) 

            intersec = len(set(D_truth_dic.keys()) & cands_set)
            recall = intersec / max(1., float(len(D_truth_dic)))
            precision = intersec / max(1., float(prm.max_candidates))
            metrics[i,prm.metrics_map['RECALL']] = recall
            metrics[i,prm.metrics_map['PRECISION']] = precision
            metrics[i,prm.metrics_map['F1']] = 2 * recall * precision / max(0.01, recall + precision)
            avg_precision = average_precision.compute(D_truth_dic.keys(), cand_ids)
            metrics[i,prm.metrics_map['MAP']] = avg_precision
            metrics[i,prm.metrics_map['LOG-GMAP']] = np.log(avg_precision + 1e-5)

        output_storage[0][0] = metrics
        output_storage[1][0] = D_i
        output_storage[2][0] = D_id
        output_storage[3][0] = D_gt_m

    def grad(self, inputs, output_grads):
        return [tensor.zeros_like(ii, dtype=theano.config.floatX) for ii in inputs]