File size: 9,226 Bytes
00fb7d5
 
 
e9b6e3f
00fb7d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e9b6e3f
 
190898f
e9b6e3f
 
 
00fb7d5
 
 
d14299c
00fb7d5
d14299c
00fb7d5
 
d14299c
 
00fb7d5
 
d14299c
00fb7d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
import logging, requests, os, io, glob, time
import json

from transformers import  T5TokenizerFast
from transformers import BertTokenizer
from transformers import PreTrainedModel
import torch

from fastai.text import *
import itertools
from typing import Optional, Dict, Union

from nltk import sent_tokenize

from transformers import(
    AutoModelForSeq2SeqLM,

    PreTrainedModel,
    PreTrainedTokenizer,
)
from transformers import AutoTokenizer
import torch


class QGPipeline:

    def __init__(
            self,
            model: PreTrainedModel,
            tokenizer: PreTrainedTokenizer,
            ans_model: PreTrainedModel,
            ans_tokenizer: PreTrainedTokenizer,
            qg_format: str,
            use_cuda: bool
    ):
        self.model = model
        self.tokenizer = tokenizer

        self.ans_model = ans_model
        self.ans_tokenizer = ans_tokenizer

        self.qg_format = qg_format

        self.device = "cuda" if torch.cuda.is_available() and use_cuda else "cpu"
        self.model.to(self.device)

        if self.ans_model is not self.model:
            self.ans_model.to(self.device)

        assert self.model.__class__.__name__ in ["MT5ForConditionalGeneration"]

        self.model_type = "mt5"

    def __call__(self, inputs: str):
        inputs = " ".join(inputs.split())
        sents, answers = self._extract_answers(inputs)
        flat_answers = list(itertools.chain(*answers))

        if len(flat_answers) == 0:
            return []

        qg_examples = self._prepare_inputs_for_qg_from_answers_hl(sents, answers)

        qg_inputs = [example['source_text'] for example in qg_examples]
        questions = self._generate_questions(qg_inputs)
        output = [{'answer': example['answer'], 'question': que} for example, que in zip(qg_examples, questions)]
        return output

    def _generate_questions(self, inputs):
        inputs = self._tokenize(inputs, padding=True, truncation=True)

        outs = self.model.generate(
            input_ids=inputs['input_ids'].to(self.device),
            attention_mask=inputs['attention_mask'].to(self.device),
            max_length=80,
            num_beams=4,
        )

        questions = [self.tokenizer.decode(ids, skip_special_tokens=True) for ids in outs]
        return questions

    def _extract_answers(self, context):
        sents, inputs = self._prepare_inputs_for_ans_extraction(context)

        inputs = self._tokenize(inputs, padding=True, truncation=True)

        outs = self.ans_model.generate(
            input_ids=inputs['input_ids'].to(self.device),
            attention_mask=inputs['attention_mask'].to(self.device),
            max_length=80,
        )


        dec = [self.ans_tokenizer.decode(ids, skip_special_tokens=True) for ids in outs]

        answers = [item.split('<sep>') for item in dec]

        answers = [i[:-1] for i in answers]
        answ_ = []
        for i in answers:
            l = []
            for b in i:
                l.append(b.replace("<pad>", ""))
            answ_.append(l)
        print(answers)
        return sents, answ_

    def _tokenize(self,
                  inputs,
                  padding=True,
                  truncation=True,
                  add_special_tokens=True,
                  max_length=512
                  ):
        inputs = self.tokenizer.batch_encode_plus(
            inputs,
            max_length=max_length,
            add_special_tokens=add_special_tokens,
            truncation=truncation,
            padding="max_length" if padding else False,
            pad_to_max_length=padding,
            return_tensors="pt"
        )

        return inputs

    def _prepare_inputs_for_ans_extraction(self, text):
        sents = sent_tokenize(text)

        inputs = []
        for i in range(len(sents)):
            source_text = "extract answers:"
            for j, sent in enumerate(sents):
                if i == j:
                    sent = "<hl> %s <hl>" % sent
                source_text = "%s %s" % (source_text, sent)
                source_text = source_text.strip()

            if self.model_type == "mt5":
                source_text = source_text + " </s>"

            inputs.append(source_text)



        return sents, inputs

    def _prepare_inputs_for_qg_from_answers_hl(self, sents, answers):
        inputs = []
        for i, answer in enumerate(answers):
            if len(answer) == 0: continue
            for answer_text in answer:
                sent = sents[i]
                sents_copy = sents[:]

                answer_text = answer_text.strip()

                try:

                    ans_start_idx = sent.index(answer_text)

                    sent = f"{sent[:ans_start_idx]} <hl> {answer_text} <hl> {sent[ans_start_idx + len(answer_text):]}"
                    sents_copy[i] = sent

                    source_text = " ".join(sents_copy)
                    source_text = f"generate question: {source_text}"
                    if self.model_type == "mt5":
                        source_text = source_text + " </s>"
                except:

                    continue

                inputs.append({"answer": answer_text, "source_text": source_text})

        return inputs


class MultiTaskQAQGPipeline(QGPipeline):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

    def __call__(self, inputs: Union[Dict, str]):
        if type(inputs) is str:
            # do qg
            return super().__call__(inputs)
        else:
            # do qa
            return self._extract_answer(inputs["question"], inputs["context"])

    def _prepare_inputs_for_qa(self, question, context):
        source_text = f"question: {question}  context: {context}"
        if self.model_type == "mt5":
            source_text = source_text + " </s>"
        return source_text

    def _extract_answer(self, question, context):
        source_text = self._prepare_inputs_for_qa(question, context)
        inputs = self._tokenize([source_text], padding=False)
        outs = self.model.generate(
            input_ids=inputs['input_ids'].to(self.device),
            attention_mask=inputs['attention_mask'].to(self.device),
            max_length=80,
        )

        answer = self.tokenizer.decode(outs[0], skip_special_tokens=True)

        return answer


SUPPORTED_TASKS = {
    "multitask-qa-qg": {
        "impl": MultiTaskQAQGPipeline,
        "default": {
            "model": "ozcangundes/mt5-multitask-qa-qg-turkish",
        }
    },
}


def pipelinex(
        task: str,
        model: Optional = None,
        tokenizer: Optional[Union[str, PreTrainedTokenizer]] = None,
        qg_format: Optional[str] = "highlight",
        ans_model: Optional = None,
        ans_tokenizer: Optional[Union[str, PreTrainedTokenizer]] = None,
        use_cuda: Optional[bool] = True,
        **kwargs,
):
    # Retrieve the task
    if task not in SUPPORTED_TASKS:
        raise KeyError("Unknown task {}, available tasks are {}".format(task, list(SUPPORTED_TASKS.keys())))

    targeted_task = SUPPORTED_TASKS[task]
    task_class = targeted_task["impl"]

    # Use default model/config/tokenizer for the task if no model is provided
    if model is None:
        model = targeted_task["default"]["model"]

    # Try to infer tokenizer from model or config name (if provided as str)
    if tokenizer is None:
        if isinstance(model, str):
            tokenizer = model
        else:
            # Impossible to guest what is the right tokenizer here
            raise Exception(
                "Impossible to guess which tokenizer to use. "
                "Please provided a PretrainedTokenizer class or a path/identifier to a pretrained tokenizer."
            )

    # Instantiate tokenizer if needed
    if isinstance(tokenizer, (str, tuple)):
        if isinstance(tokenizer, tuple):
            # For tuple we have (tokenizer name, {kwargs})
            tokenizer = AutoTokenizer.from_pretrained(tokenizer[0], **tokenizer[1])
        else:
            tokenizer = AutoTokenizer.from_pretrained(tokenizer)

    # Instantiate model if needed
    if isinstance(model, str):
        model = AutoModelForSeq2SeqLM.from_pretrained(model)
    print(ans_model)
    return task_class(model=model, tokenizer=tokenizer, ans_model=model, ans_tokenizer=tokenizer, qg_format=qg_format,
                      use_cuda=use_cuda)

################################################################################################




# loads the model into memory from disk and returns it
def model_fn(model_dir):
  # Load model from HuggingFace Hub
  tokenizer = T5TokenizerFast.from_pretrained(model_dir, extra_ids=0)
  model = AutoModelForSeq2SeqLM.from_pretrained(model_dir)
  return model, tokenizer



# Perform prediction on the deserialized object, with the loaded model
def predict_fn(data, model_tokenizer):
    
    model,tokenizer = model_tokenizer 
    
    multimodel = pipelinex("multitask-qa-qg",tokenizer=tokenizer,model=model)
    answers = multimodel(data)
 
    
    return answers