|
1
|
|
|
"""Representing suggested subjects.""" |
|
2
|
|
|
|
|
3
|
|
|
import abc |
|
4
|
|
|
import collections |
|
5
|
|
|
import numpy as np |
|
6
|
|
|
|
|
7
|
|
|
|
|
8
|
|
|
SubjectSuggestion = collections.namedtuple( |
|
9
|
|
|
'SubjectSuggestion', 'uri label notation score') |
|
10
|
|
|
WeightedSuggestion = collections.namedtuple( |
|
11
|
|
|
'WeightedSuggestion', 'hits weight subjects') |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class SuggestionFilter: |
|
15
|
|
|
"""A reusable filter for filtering SubjectSuggestion objects.""" |
|
16
|
|
|
|
|
17
|
|
|
def __init__(self, subject_index, limit=None, threshold=0.0): |
|
18
|
|
|
self._subject_index = subject_index |
|
19
|
|
|
self._limit = limit |
|
20
|
|
|
self._threshold = threshold |
|
21
|
|
|
|
|
22
|
|
|
def __call__(self, orighits): |
|
23
|
|
|
return LazySuggestionResult( |
|
24
|
|
|
lambda: orighits.filter(self._subject_index, |
|
25
|
|
|
self._limit, |
|
26
|
|
|
self._threshold)) |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
class SuggestionResult(metaclass=abc.ABCMeta): |
|
30
|
|
|
"""Abstract base class for a set of hits returned by an analysis |
|
31
|
|
|
operation.""" |
|
32
|
|
|
|
|
33
|
|
|
@abc.abstractmethod |
|
34
|
|
|
def as_list(self, subject_index): |
|
35
|
|
|
"""Return the hits as an ordered sequence of SubjectSuggestion objects, |
|
36
|
|
|
highest scores first.""" |
|
37
|
|
|
pass # pragma: no cover |
|
38
|
|
|
|
|
39
|
|
|
@abc.abstractmethod |
|
40
|
|
|
def as_vector(self, subject_index): |
|
41
|
|
|
"""Return the hits as a one-dimensional score vector |
|
42
|
|
|
where the indexes match the given subject index.""" |
|
43
|
|
|
pass # pragma: no cover |
|
44
|
|
|
|
|
45
|
|
|
@abc.abstractmethod |
|
46
|
|
|
def filter(self, subject_index, limit=None, threshold=0.0): |
|
47
|
|
|
"""Return a subset of the hits, filtered by the given limit and |
|
48
|
|
|
score threshold, as another SuggestionResult object.""" |
|
49
|
|
|
pass # pragma: no cover |
|
50
|
|
|
|
|
51
|
|
|
@abc.abstractmethod |
|
52
|
|
|
def __len__(self): |
|
53
|
|
|
"""Return the number of hits with non-zero scores.""" |
|
54
|
|
|
pass # pragma: no cover |
|
55
|
|
|
|
|
56
|
|
|
|
|
57
|
|
|
class LazySuggestionResult(SuggestionResult): |
|
58
|
|
|
"""SuggestionResult implementation that wraps another SuggestionResult which |
|
59
|
|
|
is initialized lazily only when it is actually accessed. Method calls |
|
60
|
|
|
will be proxied to the wrapped SuggestionResult.""" |
|
61
|
|
|
|
|
62
|
|
|
def __init__(self, construct): |
|
63
|
|
|
"""Create the proxy object. The given construct function will be |
|
64
|
|
|
called to create the actual SuggestionResult when it is needed.""" |
|
65
|
|
|
self._construct = construct |
|
66
|
|
|
self._object = None |
|
67
|
|
|
|
|
68
|
|
|
def _initialize(self): |
|
69
|
|
|
if self._object is None: |
|
70
|
|
|
self._object = self._construct() |
|
71
|
|
|
|
|
72
|
|
|
def as_list(self, subject_index): |
|
73
|
|
|
self._initialize() |
|
74
|
|
|
return self._object.as_list(subject_index) |
|
75
|
|
|
|
|
76
|
|
|
def as_vector(self, subject_index): |
|
77
|
|
|
self._initialize() |
|
78
|
|
|
return self._object.as_vector(subject_index) |
|
79
|
|
|
|
|
80
|
|
|
def filter(self, subject_index, limit=None, threshold=0.0): |
|
81
|
|
|
self._initialize() |
|
82
|
|
|
return self._object.filter(subject_index, limit, threshold) |
|
83
|
|
|
|
|
84
|
|
|
def __len__(self): |
|
85
|
|
|
self._initialize() |
|
86
|
|
|
return len(self._object) |
|
87
|
|
|
|
|
88
|
|
|
|
|
89
|
|
|
class VectorSuggestionResult(SuggestionResult): |
|
90
|
|
|
"""SuggestionResult implementation based primarily on NumPy vectors.""" |
|
91
|
|
|
|
|
92
|
|
|
def __init__(self, vector): |
|
93
|
|
|
self._vector = vector.astype(np.float32) |
|
94
|
|
|
self._subject_order = None |
|
95
|
|
|
self._lsr = None |
|
96
|
|
|
|
|
97
|
|
|
def _vector_to_list_suggestion(self, subject_index): |
|
98
|
|
|
hits = [] |
|
99
|
|
|
for subject_id in self.subject_order: |
|
100
|
|
|
score = self._vector[subject_id] |
|
101
|
|
|
if score <= 0.0: |
|
102
|
|
|
continue # we can skip the remaining ones |
|
103
|
|
|
subject = subject_index[subject_id] |
|
104
|
|
|
hits.append( |
|
105
|
|
|
SubjectSuggestion( |
|
106
|
|
|
uri=subject[0], |
|
107
|
|
|
label=subject[1], |
|
108
|
|
|
notation=subject[2], |
|
109
|
|
|
score=float(score))) |
|
110
|
|
|
return ListSuggestionResult(hits) |
|
111
|
|
|
|
|
112
|
|
|
@property |
|
113
|
|
|
def subject_order(self): |
|
114
|
|
|
if self._subject_order is None: |
|
115
|
|
|
self._subject_order = np.argsort(self._vector)[::-1] |
|
116
|
|
|
return self._subject_order |
|
117
|
|
|
|
|
118
|
|
|
def as_list(self, subject_index): |
|
119
|
|
|
if self._lsr is None: |
|
120
|
|
|
self._lsr = self._vector_to_list_suggestion(subject_index) |
|
121
|
|
|
return self._lsr.as_list(subject_index) |
|
122
|
|
|
|
|
123
|
|
|
def as_vector(self, subject_index): |
|
124
|
|
|
return self._vector |
|
125
|
|
|
|
|
126
|
|
|
def filter(self, subject_index, limit=None, threshold=0.0): |
|
127
|
|
|
mask = (self._vector > threshold) |
|
128
|
|
|
deprecated_ids = subject_index.deprecated_ids() |
|
129
|
|
|
if limit is not None: |
|
130
|
|
|
limit_mask = np.zeros_like(self._vector, dtype=np.bool) |
|
131
|
|
|
top_k_subjects = [subj for subj in self.subject_order |
|
132
|
|
|
if subj not in deprecated_ids][:limit] |
|
133
|
|
|
limit_mask[top_k_subjects] = True |
|
134
|
|
|
mask = mask & limit_mask |
|
135
|
|
|
else: |
|
136
|
|
|
deprecated_mask = np.ones_like(self._vector, dtype=np.bool) |
|
137
|
|
|
deprecated_mask[deprecated_ids] = False |
|
138
|
|
|
mask = mask & deprecated_mask |
|
139
|
|
|
vsr = VectorSuggestionResult(self._vector * mask) |
|
140
|
|
|
return ListSuggestionResult(vsr.as_list(subject_index)) |
|
141
|
|
|
|
|
142
|
|
|
def __len__(self): |
|
143
|
|
|
return (self._vector > 0.0).sum() |
|
144
|
|
|
|
|
145
|
|
|
|
|
146
|
|
|
class ListSuggestionResult(SuggestionResult): |
|
147
|
|
|
"""SuggestionResult implementation based primarily on lists of hits.""" |
|
148
|
|
|
|
|
149
|
|
|
def __init__(self, hits): |
|
150
|
|
|
self._list = [hit for hit in hits if hit.score > 0.0] |
|
151
|
|
|
self._vector = None |
|
152
|
|
|
|
|
153
|
|
|
@classmethod |
|
154
|
|
|
def create_from_index(cls, hits, subject_index): |
|
155
|
|
|
subject_suggestions = [] |
|
156
|
|
|
for hit in hits: |
|
157
|
|
|
subject_id = subject_index.by_uri(hit.uri) |
|
158
|
|
|
if subject_id is None: |
|
159
|
|
|
continue |
|
160
|
|
|
subject = subject_index[subject_id] |
|
161
|
|
|
subject_suggestions.append( |
|
162
|
|
|
SubjectSuggestion(uri=hit.uri, |
|
163
|
|
|
label=subject[1], |
|
164
|
|
|
notation=subject[2], |
|
165
|
|
|
score=hit.score)) |
|
166
|
|
|
return ListSuggestionResult(subject_suggestions) |
|
167
|
|
|
|
|
168
|
|
|
def _list_to_vector(self, subject_index): |
|
169
|
|
|
vector = np.zeros(len(subject_index), dtype=np.float32) |
|
170
|
|
|
for hit in self._list: |
|
171
|
|
|
subject_id = subject_index.by_uri(hit.uri) |
|
172
|
|
|
if subject_id is not None: |
|
173
|
|
|
vector[subject_id] = hit.score |
|
174
|
|
|
return vector |
|
175
|
|
|
|
|
176
|
|
|
def as_list(self, subject_index): |
|
177
|
|
|
return self._list |
|
178
|
|
|
|
|
179
|
|
|
def as_vector(self, subject_index): |
|
180
|
|
|
if self._vector is None: |
|
181
|
|
|
self._vector = self._list_to_vector(subject_index) |
|
182
|
|
|
return self._vector |
|
183
|
|
|
|
|
184
|
|
|
def filter(self, subject_index, limit=None, threshold=0.0): |
|
185
|
|
|
hits = sorted(self._list, key=lambda hit: hit.score, reverse=True) |
|
186
|
|
|
filtered_hits = [hit for hit in hits |
|
187
|
|
|
if hit.score >= threshold and hit.score > 0.0 and |
|
188
|
|
|
hit.label is not None] |
|
189
|
|
|
if limit is not None: |
|
190
|
|
|
filtered_hits = filtered_hits[:limit] |
|
191
|
|
|
return ListSuggestionResult(filtered_hits) |
|
192
|
|
|
|
|
193
|
|
|
def __len__(self): |
|
194
|
|
|
return len(self._list) |
|
195
|
|
|
|
NatLibFi /
Annif
