|
1
|
|
|
"""Representing suggested subjects.""" |
|
2
|
|
|
|
|
3
|
|
|
import collections |
|
4
|
|
|
import itertools |
|
5
|
|
|
|
|
6
|
|
|
import numpy as np |
|
7
|
|
|
from scipy.sparse import csr_array, dok_array |
|
8
|
|
|
|
|
9
|
|
|
SubjectSuggestion = collections.namedtuple("SubjectSuggestion", "subject_id score") |
|
10
|
|
|
WeightedSuggestionsBatch = collections.namedtuple( |
|
11
|
|
|
"WeightedSuggestionsBatch", "hit_sets weight subjects" |
|
12
|
|
|
) |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
def vector_to_suggestions(vector, limit): |
|
16
|
|
|
limit = min(len(vector), limit) |
|
17
|
|
|
topk_idx = np.argpartition(vector, -limit)[-limit:] |
|
18
|
|
|
return ( |
|
19
|
|
|
SubjectSuggestion(subject_id=idx, score=float(vector[idx])) for idx in topk_idx |
|
20
|
|
|
) |
|
21
|
|
|
|
|
22
|
|
|
|
|
23
|
|
|
def filter_suggestion(preds, limit=None, threshold=0.0): |
|
24
|
|
|
"""filter a 2D sparse suggestion array (csr_array), retaining only the |
|
25
|
|
|
top K suggestions with a score above or equal to the threshold for each |
|
26
|
|
|
individual prediction; the rest will be left as zeros""" |
|
27
|
|
|
|
|
28
|
|
|
filtered = dok_array(preds.shape, dtype=np.float32) |
|
29
|
|
|
for row in range(preds.shape[0]): |
|
30
|
|
|
arow = preds.getrow(row) |
|
31
|
|
|
top_k = arow.data.argsort()[::-1] |
|
32
|
|
|
if limit is not None: |
|
33
|
|
|
top_k = top_k[:limit] |
|
34
|
|
|
for idx in top_k: |
|
35
|
|
|
val = arow.data[idx] |
|
36
|
|
|
if val < threshold: |
|
37
|
|
|
break |
|
38
|
|
|
filtered[row, arow.indices[idx]] = val |
|
39
|
|
|
return filtered.tocsr() |
|
40
|
|
|
|
|
41
|
|
|
|
|
42
|
|
|
class SuggestionResult: |
|
43
|
|
|
"""Suggestions for a single document, backed by a row of a sparse array.""" |
|
44
|
|
|
|
|
45
|
|
|
def __init__(self, array, idx): |
|
46
|
|
|
self._array = array |
|
47
|
|
|
self._idx = idx |
|
48
|
|
|
|
|
49
|
|
|
def __iter__(self): |
|
50
|
|
|
_, cols = self._array[[self._idx], :].nonzero() |
|
51
|
|
|
suggestions = [ |
|
52
|
|
|
SubjectSuggestion(subject_id=col, score=float(self._array[self._idx, col])) |
|
53
|
|
|
for col in cols |
|
54
|
|
|
] |
|
55
|
|
|
return iter( |
|
56
|
|
|
sorted(suggestions, key=lambda suggestion: suggestion.score, reverse=True) |
|
57
|
|
|
) |
|
58
|
|
|
|
|
59
|
|
|
def as_vector(self): |
|
60
|
|
|
return self._array[[self._idx], :].toarray()[0] |
|
61
|
|
|
|
|
62
|
|
|
def __len__(self): |
|
63
|
|
|
_, cols = self._array[[self._idx], :].nonzero() |
|
64
|
|
|
return len(cols) |
|
65
|
|
|
|
|
66
|
|
|
|
|
67
|
|
|
class SuggestionBatch: |
|
68
|
|
|
"""Subject suggestions for a batch of documents.""" |
|
69
|
|
|
|
|
70
|
|
|
def __init__(self, array): |
|
71
|
|
|
"""Create a new SuggestionBatch from a csr_array""" |
|
72
|
|
|
assert isinstance(array, csr_array) |
|
73
|
|
|
self.array = array |
|
74
|
|
|
|
|
75
|
|
|
@classmethod |
|
76
|
|
|
def from_sequence(cls, suggestion_results, subject_index, limit=None): |
|
77
|
|
|
"""Create a new SuggestionBatch from a sequence where each item is |
|
78
|
|
|
a sequence of SubjectSuggestion objects.""" |
|
79
|
|
|
|
|
80
|
|
|
deprecated = set(subject_index.deprecated_ids()) |
|
81
|
|
|
|
|
82
|
|
|
ar = dok_array((len(suggestion_results), len(subject_index)), dtype=np.float32) |
|
83
|
|
|
for idx, result in enumerate(suggestion_results): |
|
84
|
|
|
for suggestion in itertools.islice(result, limit): |
|
85
|
|
|
if suggestion.subject_id in deprecated or suggestion.score <= 0.0: |
|
86
|
|
|
continue |
|
87
|
|
|
ar[idx, suggestion.subject_id] = min(suggestion.score, 1.0) |
|
88
|
|
|
return cls(ar.tocsr()) |
|
89
|
|
|
|
|
90
|
|
|
@classmethod |
|
91
|
|
|
def from_averaged(cls, batches, weights): |
|
92
|
|
|
"""Create a new SuggestionBatch where the subject scores are the |
|
93
|
|
|
weighted average of scores in several SuggestionBatches""" |
|
94
|
|
|
|
|
95
|
|
|
avg_array = sum( |
|
96
|
|
|
[batch.array * weight for batch, weight in zip(batches, weights)] |
|
97
|
|
|
) / sum(weights) |
|
98
|
|
|
return SuggestionBatch(avg_array) |
|
99
|
|
|
|
|
100
|
|
|
def filter(self, limit=None, threshold=0.0): |
|
101
|
|
|
"""Return a subset of the hits, filtered by the given limit and |
|
102
|
|
|
score threshold, as another SuggestionBatch object.""" |
|
103
|
|
|
|
|
104
|
|
|
return SuggestionBatch(filter_suggestion(self.array, limit, threshold)) |
|
105
|
|
|
|
|
106
|
|
|
def __getitem__(self, idx): |
|
107
|
|
|
if idx < 0 or idx >= len(self): |
|
108
|
|
|
raise IndexError |
|
109
|
|
|
return SuggestionResult(self.array, idx) |
|
110
|
|
|
|
|
111
|
|
|
def __len__(self): |
|
112
|
|
|
return self.array.shape[0] |
|
113
|
|
|
|
|
114
|
|
|
|
|
115
|
|
|
class SuggestionResults: |
|
116
|
|
|
"""Subject suggestions for a potentially very large number of documents.""" |
|
117
|
|
|
|
|
118
|
|
|
def __init__(self, batches): |
|
119
|
|
|
"""Initialize a new SuggestionResults from an iterable that provides |
|
120
|
|
|
SuggestionBatch objects.""" |
|
121
|
|
|
|
|
122
|
|
|
self.batches = batches |
|
123
|
|
|
|
|
124
|
|
|
def filter(self, limit=None, threshold=0.0): |
|
125
|
|
|
"""Return a view of these suggestions, filtered by the given limit |
|
126
|
|
|
and/or threshold, as another SuggestionResults object.""" |
|
127
|
|
|
|
|
128
|
|
|
return SuggestionResults( |
|
129
|
|
|
(batch.filter(limit, threshold) for batch in self.batches) |
|
130
|
|
|
) |
|
131
|
|
|
|
|
132
|
|
|
def __iter__(self): |
|
133
|
|
|
return iter(itertools.chain.from_iterable(self.batches)) |
|
134
|
|
|
|
NatLibFi /
Annif
