infer.py

from model.mrc_model import MRCQuestionAnswering
from transformers import AutoTokenizer, pipeline, RobertaForQuestionAnswering
import torch
from nltk import word_tokenize
from transformers.models.auto import MODEL_FOR_QUESTION_ANSWERING_MAPPING


def tokenize_function(example, tokenizer):
    question_word = word_tokenize(example["question"])
    context_word = word_tokenize(example["context"])

    question_sub_words_ids = [tokenizer.convert_tokens_to_ids(tokenizer.tokenize(w)) for w in question_word]
    context_sub_words_ids = [tokenizer.convert_tokens_to_ids(tokenizer.tokenize(w)) for w in context_word]
    valid = True
    if len([j for i in question_sub_words_ids + context_sub_words_ids for j in
            i]) > tokenizer.max_len_single_sentence - 1:
        valid = False

    question_sub_words_ids = [[tokenizer.bos_token_id]] + question_sub_words_ids + [[tokenizer.eos_token_id]]
    context_sub_words_ids = context_sub_words_ids + [[tokenizer.eos_token_id]]

    input_ids = [j for i in question_sub_words_ids + context_sub_words_ids for j in i]
    if len(input_ids) > tokenizer.max_len_single_sentence + 2:
        valid = False

    words_lengths = [len(item) for item in question_sub_words_ids + context_sub_words_ids]

    return {
        "input_ids": input_ids,
        "words_lengths": words_lengths,
        "valid": valid
    }


def data_collator(samples, tokenizer):
    if len(samples) == 0:
        return {}

    def collate_tokens(values, pad_idx, eos_idx=None, left_pad=False, move_eos_to_beginning=False):
        """Convert a list of 1d tensors into a padded 2d tensor."""
        size = max(v.size(0) for v in values)
        res = values[0].new(len(values), size).fill_(pad_idx)

        def copy_tensor(src, dst):
            assert dst.numel() == src.numel()
            if move_eos_to_beginning:
                assert src[-1] == eos_idx
                dst[0] = eos_idx
                dst[1:] = src[:-1]
            else:
                dst.copy_(src)

        for i, v in enumerate(values):
            copy_tensor(v, res[i][size - len(v):] if left_pad else res[i][:len(v)])
        return res

    input_ids = collate_tokens([torch.tensor(item['input_ids']) for item in samples], pad_idx=tokenizer.pad_token_id)
    attention_mask = torch.zeros_like(input_ids)
    for i in range(len(samples)):
        attention_mask[i][:len(samples[i]['input_ids'])] = 1
    words_lengths = collate_tokens([torch.tensor(item['words_lengths']) for item in samples], pad_idx=0)

    batch_samples = {
        'input_ids': input_ids,
        'attention_mask': attention_mask,
        'words_lengths': words_lengths,
    }

    return batch_samples


def extract_answer(inputs, outputs, tokenizer):
    plain_result = []
    for sample_input, start_logit, end_logit in zip(inputs, outputs.start_logits, outputs.end_logits):
        sample_words_length = sample_input['words_lengths']
        input_ids = sample_input['input_ids']
        # Get the most likely beginning of answer with the argmax of the score
        answer_start = sum(sample_words_length[:torch.argmax(start_logit)])
        # Get the most likely end of answer with the argmax of the score
        answer_end = sum(sample_words_length[:torch.argmax(end_logit) + 1])

        if answer_start <= answer_end:
            answer = tokenizer.convert_tokens_to_string(
                tokenizer.convert_ids_to_tokens(input_ids[answer_start:answer_end]))
            if answer == tokenizer.bos_token:
                answer = ''
        else:
            answer = ''

        score_start = torch.max(torch.softmax(start_logit, dim=-1)).cpu().detach().numpy().tolist()
        score_end = torch.max(torch.softmax(end_logit, dim=-1)).cpu().detach().numpy().tolist()
        plain_result.append({
            "answer": answer,
            "score_start": score_start,
            "score_end": score_end
        })
    return plain_result


if __name__ == "__main__":
    model_checkpoint = "nguyenvulebinh/vi-mrc-base"
    tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
    model = MRCQuestionAnswering.from_pretrained(model_checkpoint)

    print(model)

    QA_input = {
        'question': "Bình được công nhận với danh hiệu gì ?",
        'context': "Bình Nguyễn là một người đam mê với lĩnh vực xử lý ngôn ngữ tự nhiên . Anh nhận chứng chỉ Google Developer Expert năm 2020"
    }
    while True:
        if len(QA_input['question'].strip()) > 0:
            inputs = [tokenize_function(QA_input)]
            inputs_ids = data_collator(inputs)
            outputs = model(**inputs_ids)
            answer = extract_answer(inputs, outputs, tokenizer)[0]
            print("answer: {}. Score start: {}, Score end: {}".format(answer['answer'],
                                                                      answer['score_start'],
                                                                      answer['score_end']))

        else:
            QA_input['context'] = input('Context: ')
        QA_input['question'] = input('Question: ')