minicheck_web/inference.py

# Adapt code from https://github.com/yuh-zha/AlignScore/tree/main

import sys
sys.path.append("..")

import nltk
nltk.download('punkt')

from nltk.tokenize import sent_tokenize
import torch
from transformers import AutoConfig, AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
import torch.nn as nn
from tqdm import tqdm
import torch.nn.functional as F
import os


def sent_tokenize_with_newlines(text):
    blocks = text.split('\n')
    
    tokenized_blocks = [sent_tokenize(block) for block in blocks]
    tokenized_text = []
    for block in tokenized_blocks:
        tokenized_text.extend(block)
        tokenized_text.append('\n')  

    return tokenized_text[:-1]  


class Inferencer():
    def __init__(self, path, chunk_size, max_input_length, batch_size) -> None:
        
        self.path = path
        self.device = "cuda:0" if torch.cuda.is_available() else "cpu"

        self.model = AutoModelForSeq2SeqLM.from_pretrained(path).to(self.device)
        self.tokenizer = AutoTokenizer.from_pretrained(path)

        self.chunk_size=500 if chunk_size is None else chunk_size
        self.max_input_length=2048 if max_input_length is None else max_input_length
        self.max_output_length = 256
        
        self.model.eval()
        self.batch_size = batch_size
        self.softmax = nn.Softmax(dim=-1)

    def inference_example_batch(self, doc: list, claim: list):
        """
        inference a example,
        doc: list
        claim: list
        using self.inference to batch the process
        """

        assert len(doc) == len(claim), "doc must has the same length with claimthesis!"

        max_support_probs = []
        used_chunks = []
        support_prob_per_chunk = []
        
        for one_doc, one_claim in tqdm(zip(doc, claim), desc="Evaluating", total=len(doc)):
            output = self.inference_per_example(one_doc, one_claim)
            max_support_probs.append(output['max_support_prob'])
            used_chunks.append(output['used_chunks'])
            support_prob_per_chunk.append(output['support_prob_per_chunk'])
        
        return {
            'max_support_probs': max_support_probs,
            'used_chunks': used_chunks,
            'support_prob_per_chunk': support_prob_per_chunk
        }

    def inference_per_example(self, doc:str, claim: str):
        """
        inference a example,
        doc: string
        claim: string
        using self.inference to batch the process
        """
        def chunks(lst, n):
            """Yield successive chunks from lst with each having approximately n tokens.

            For flan-t5, we split using the white space;
            For roberta and deberta, we split using the tokenization.
            """
            current_chunk = []
            current_word_count = 0
            for sentence in lst:
                sentence_word_count = len(sentence.split())
                if current_word_count + sentence_word_count > n:
                    yield ' '.join(current_chunk)
                    current_chunk = [sentence]
                    current_word_count = sentence_word_count
                else:
                    current_chunk.append(sentence)
                    current_word_count += sentence_word_count
            if current_chunk:
                yield ' '.join(current_chunk)

        doc_sents = sent_tokenize_with_newlines(doc)
        doc_sents = doc_sents or ['']

        doc_chunks = [chunk.replace(" \n ", '\n').strip() for chunk in chunks(doc_sents, self.chunk_size)]

        '''
        [chunk_1, chunk_2, chunk_3, chunk_4, ...]
        [claim]
        '''
        claim_repeat = [claim] * len(doc_chunks)
        
        output = self.inference(doc_chunks, claim_repeat)
        
        return output

    def inference(self, doc, claim):
        """
        inference a list of doc and claim

        Standard aggregation (max) over chunks of doc

        Note: We do not have any post-processing steps for 'claim'
        and directly check 'doc' against 'claim'. If there are multiple 
        sentences in 'claim'. Sentences are not splitted and are checked 
        as a single piece of text.
        
        If there are multiple sentences in 'claim', we suggest users to 
        split 'claim' into sentences beforehand and prepares data like 
        (doc, claim_1), (doc, claim_2), ... for a multi-sentence 'claim'.

        **We leave the user to decide how to aggregate the results from multiple sentences.**

        Note: AggreFact-CNN is the only dataset that contains three-sentence 
        summaries and have annotations on the whole summaries, so we do not 
        split the sentences in each 'claim' during prediciotn for simplicity. 
        Therefore, for this dataset, our result is based on treating the whole 
        summary as a single piece of text (one 'claim').

        In general, sentence-level prediciton performance is better than that on 
        the full-response-level.
        """

        if isinstance(doc, str) and isinstance(claim, str):
            doc = [doc]
            claim = [claim]
        
        batch_input, _, batch_org_chunks = self.batch_tokenize(doc, claim)

        label_probs_list = []
        used_chunks = []

        for mini_batch_input, batch_org_chunk in zip(batch_input, batch_org_chunks):

            mini_batch_input = {k: v.to(self.device) for k, v in mini_batch_input.items()}

            with torch.no_grad():
                    
                decoder_input_ids = torch.zeros((mini_batch_input['input_ids'].size(0), 1), dtype=torch.long).to(self.device)
                outputs = self.model(input_ids=mini_batch_input['input_ids'], attention_mask=mini_batch_input['attention_mask'], decoder_input_ids=decoder_input_ids)
                logits = outputs.logits.squeeze(1)

                # 3 for no support and 209 for support
                label_logits = logits[:, torch.tensor([3, 209])].cpu()
                label_probs = torch.nn.functional.softmax(label_logits, dim=-1)

                label_probs_list.append(label_probs)
                used_chunks.extend(batch_org_chunk)

        label_probs = torch.cat(label_probs_list)
        support_prob_per_chunk = label_probs[:, 1].cpu().numpy()
        max_support_prob = label_probs[:, 1].max().item()
        
        return {
            'max_support_prob': max_support_prob,
            'used_chunks': used_chunks,
            'support_prob_per_chunk': support_prob_per_chunk
        }

    def batch_tokenize(self, doc, claim):
        """
        input doc and claims are lists
        """
        assert isinstance(doc, list) and isinstance(claim, list)
        assert len(doc) == len(claim), "doc and claim should be in the same length."

        original_text = [self.tokenizer.eos_token.join([one_doc, one_claim]) for one_doc, one_claim in zip(doc, claim)]

        batch_input = []
        batch_concat_text = []
        batch_org_chunks = []
        for mini_batch in self.chunks(original_text, self.batch_size):
            model_inputs = self.tokenizer(
                ['predict: ' + text for text in mini_batch], 
                max_length=self.max_input_length, 
                truncation=True, 
                padding=True, 
                return_tensors="pt"
            ) 

            batch_input.append(model_inputs) 
            batch_concat_text.append(mini_batch)  
            batch_org_chunks.append([item[:item.find('</s>')] for item in mini_batch]) 

        return batch_input, batch_concat_text, batch_org_chunks

    def chunks(self, lst, n):
        """Yield successive n-sized chunks from lst."""
        for i in range(0, len(lst), n):
            yield lst[i:i + n]
    
    def fact_check(self, doc, claim):

        outputs = self.inference_example_batch(doc, claim)
        return outputs['max_support_probs'], outputs['used_chunks'], outputs['support_prob_per_chunk']