|
1
|
|
|
"""TODO""" |
|
2
|
|
|
|
|
3
|
|
|
from __future__ import annotations |
|
4
|
|
|
|
|
5
|
|
|
import json |
|
6
|
|
|
import os |
|
7
|
|
|
from typing import TYPE_CHECKING, Any |
|
8
|
|
|
|
|
9
|
|
|
# import tiktoken |
|
10
|
|
|
from transformers import pipeline |
|
11
|
|
|
|
|
12
|
|
|
import annif.eval |
|
13
|
|
|
import annif.parallel |
|
14
|
|
|
import annif.util |
|
15
|
|
|
from annif.exception import NotSupportedException |
|
16
|
|
|
from annif.suggestion import SubjectSuggestion, SuggestionBatch |
|
17
|
|
|
|
|
18
|
|
|
from . import backend |
|
19
|
|
|
|
|
20
|
|
|
# from openai import AsyncAzureOpenAI |
|
21
|
|
|
|
|
22
|
|
|
|
|
23
|
|
|
if TYPE_CHECKING: |
|
24
|
|
|
from datetime import datetime |
|
25
|
|
|
|
|
26
|
|
|
from annif.corpus.document import DocumentCorpus |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
class RescorerBackend(backend.AnnifBackend): |
|
30
|
|
|
# """TODO backend that combines results from multiple projects""" |
|
31
|
|
|
|
|
32
|
|
|
name = "rescorer" |
|
33
|
|
|
|
|
34
|
|
|
def _get_sources_attribute(self, attr: str) -> list[bool | None]: |
|
35
|
|
|
params = self._get_backend_params(None) |
|
36
|
|
|
sources = annif.util.parse_sources(params["sources"]) |
|
37
|
|
|
return [ |
|
38
|
|
|
getattr(self.project.registry.get_project(project_id), attr) |
|
39
|
|
|
for project_id, _ in sources |
|
40
|
|
|
] |
|
41
|
|
|
|
|
42
|
|
|
def initialize(self, parallel: bool = False) -> None: |
|
43
|
|
|
# initialize all the source projects |
|
44
|
|
|
params = self._get_backend_params(None) |
|
45
|
|
|
for project_id, _ in annif.util.parse_sources(params["sources"]): |
|
46
|
|
|
project = self.project.registry.get_project(project_id) |
|
47
|
|
|
project.initialize(parallel) |
|
48
|
|
|
|
|
49
|
|
|
self.classifier = pipeline( |
|
50
|
|
|
"zero-shot-classification", model=params.get("model"), |
|
51
|
|
|
from_pt=True, |
|
52
|
|
|
multi_label=True, |
|
53
|
|
|
) |
|
54
|
|
|
|
|
55
|
|
|
def _suggest_with_sources( |
|
56
|
|
|
self, texts: list[str], sources: list[tuple[str, float]] |
|
57
|
|
|
) -> dict[str, SuggestionBatch]: |
|
58
|
|
|
return { |
|
59
|
|
|
project_id: self.project.registry.get_project(project_id).suggest(texts) |
|
60
|
|
|
for project_id, _ in sources |
|
61
|
|
|
} |
|
62
|
|
|
|
|
63
|
|
|
@property |
|
64
|
|
|
def is_trained(self) -> bool: |
|
65
|
|
|
sources_trained = self._get_sources_attribute("is_trained") |
|
66
|
|
|
return all(sources_trained) |
|
67
|
|
|
|
|
68
|
|
|
@property |
|
69
|
|
|
def modification_time(self) -> datetime | None: |
|
70
|
|
|
mtimes = self._get_sources_attribute("modification_time") |
|
71
|
|
|
return max(filter(None, mtimes), default=None) |
|
72
|
|
|
|
|
73
|
|
|
def _train(self, corpus: DocumentCorpus, params: dict[str, Any], jobs: int = 0): |
|
74
|
|
|
raise NotSupportedException("Training rescorer backend is not possible.") |
|
75
|
|
|
|
|
76
|
|
|
def _suggest_batch( |
|
77
|
|
|
self, texts: list[str], params: dict[str, Any] |
|
78
|
|
|
) -> SuggestionBatch: |
|
79
|
|
|
sources = annif.util.parse_sources(params["sources"]) |
|
80
|
|
|
new_scores_weight = float(params["new_scores_weight"]) |
|
81
|
|
|
# llm_probs_weight = float(params["llm_probs_weight"]) |
|
82
|
|
|
# encoding = tiktoken.encoding_for_model(model.rsplit("-", 1)[0]) |
|
83
|
|
|
|
|
84
|
|
|
batch_results = [] |
|
85
|
|
|
base_suggestion_batch = self._suggest_with_sources(texts, sources)[ |
|
86
|
|
|
sources[0][0] |
|
87
|
|
|
] |
|
88
|
|
|
|
|
89
|
|
|
from time import time |
|
90
|
|
|
start_t = time() |
|
91
|
|
|
for text, base_suggestions in zip(texts, base_suggestion_batch): |
|
92
|
|
|
base_labels = [ |
|
93
|
|
|
self.project.subjects[s.subject_id].labels["en"] |
|
94
|
|
|
for s in base_suggestions |
|
95
|
|
|
] |
|
96
|
|
|
|
|
97
|
|
|
# text = self._truncate_text(text, encoding) |
|
98
|
|
|
result = self.classifier(text, base_labels) |
|
99
|
|
|
print(result) |
|
100
|
|
|
# try: |
|
101
|
|
|
# llm_result = json.loads(answer) |
|
102
|
|
|
# except (TypeError, json.decoder.JSONDecodeError) as err: |
|
103
|
|
|
# print(err) |
|
104
|
|
|
# llm_result = dict() |
|
105
|
|
|
rescored_results = self._rescore_suggestions( |
|
106
|
|
|
result, |
|
107
|
|
|
base_labels, |
|
108
|
|
|
base_suggestions, |
|
109
|
|
|
new_scores_weight, |
|
110
|
|
|
) |
|
111
|
|
|
batch_results.append(rescored_results) |
|
112
|
|
|
print(f"Time: {time() - start_t:.2f} s") |
|
113
|
|
|
return SuggestionBatch.from_sequence(batch_results, self.project.subjects) |
|
114
|
|
|
|
|
115
|
|
|
# def _truncate_text(self, text, encoding): |
|
116
|
|
|
# """truncate text so it contains at most MAX_PROMPT_TOKENS according to the |
|
117
|
|
|
# OpenAI tokenizer""" |
|
118
|
|
|
|
|
119
|
|
|
# MAX_PROMPT_TOKENS = 14000 |
|
120
|
|
|
# tokens = encoding.encode(text) |
|
121
|
|
|
# return encoding.decode(tokens[:MAX_PROMPT_TOKENS]) |
|
122
|
|
|
|
|
123
|
|
|
def _rescore_suggestions( |
|
124
|
|
|
self, |
|
125
|
|
|
result, |
|
126
|
|
|
base_labels, |
|
127
|
|
|
base_suggestions, |
|
128
|
|
|
new_scores_weight, |
|
129
|
|
|
): |
|
130
|
|
|
suggestions = [] |
|
131
|
|
|
for blabel, bsuggestion in zip(base_labels, base_suggestions): |
|
132
|
|
|
try: |
|
133
|
|
|
ind = result["labels"].index(blabel) |
|
134
|
|
|
score = result["scores"][ind] |
|
135
|
|
|
except ValueError: |
|
136
|
|
|
print(f"Base label {blabel} not found in new labels") |
|
137
|
|
|
score = bsuggestion.score # use only base suggestion score |
|
138
|
|
|
subj_id = bsuggestion.subject_id |
|
139
|
|
|
|
|
140
|
|
|
base_scores_weight = 1.0 - new_scores_weight |
|
141
|
|
|
mean_score = ( |
|
142
|
|
|
base_scores_weight * bsuggestion.score |
|
143
|
|
|
+ new_scores_weight * score # * probability * llm_probs_weight |
|
144
|
|
|
) / ( |
|
145
|
|
|
base_scores_weight |
|
146
|
|
|
+ new_scores_weight # * probability * llm_probs_weight |
|
147
|
|
|
) # weighted mean of LLM and base scores! |
|
148
|
|
|
suggestions.append(SubjectSuggestion(subject_id=subj_id, score=mean_score)) |
|
149
|
|
|
return suggestions |
|
150
|
|
|
|
NatLibFi /
Annif
