BioReason-0813/model/help_funcs.py

import torch
from nltk.translate.bleu_score import corpus_bleu
from nltk.translate.meteor_score import meteor_score
from rouge_score import rouge_scorer
from tqdm import tqdm
import numpy as np


def caption_evaluate(predictions, targets, tokenizer, text_trunc_length):
    targets = [t.strip() for t in targets]
    meteor_scores = []
    references = []
    hypotheses = []
    for gt, out in tqdm(zip(targets, predictions)):
        gt_tokens = tokenizer.tokenize(gt, truncation=True, max_length=text_trunc_length,
                                            padding='max_length')
        ## added for galactica
        gt_tokens = list(filter(('<pad>').__ne__, gt_tokens))
        gt_tokens = list(filter(('[PAD]').__ne__, gt_tokens))
        gt_tokens = list(filter(('[CLS]').__ne__, gt_tokens))
        gt_tokens = list(filter(('[SEP]').__ne__, gt_tokens))

        out_tokens = tokenizer.tokenize(out, truncation=True, max_length=text_trunc_length,
                                            padding='max_length')
        out_tokens = list(filter(('<pad>').__ne__, out_tokens))
        gt_tokens = list(filter(('[PAD]').__ne__, gt_tokens))
        out_tokens = list(filter(('[CLS]').__ne__, out_tokens))
        out_tokens = list(filter(('[SEP]').__ne__, out_tokens))

        references.append([gt_tokens])
        hypotheses.append(out_tokens)

        mscore = meteor_score([gt_tokens], out_tokens)
        meteor_scores.append(mscore)

    bleu2 = corpus_bleu(references, hypotheses, weights=(.5,.5))
    bleu4 = corpus_bleu(references, hypotheses, weights=(.25,.25,.25,.25))
    bleu2 *= 100
    bleu4 *= 100

    print('BLEU-2 score:', bleu2)
    print('BLEU-4 score:', bleu4)
    _meteor_score = np.mean(meteor_scores)
    _meteor_score *= 100
    print('Average Meteor score:', _meteor_score)

    scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'])

    rouge_scores = []

    references = []
    hypotheses = []

    for gt, out in tqdm(zip(targets, predictions)):
        rs = scorer.score(out, gt)
        rouge_scores.append(rs)

    print('ROUGE score:')
    rouge_1 = np.mean([rs['rouge1'].fmeasure for rs in rouge_scores]) * 100
    rouge_2 = np.mean([rs['rouge2'].fmeasure for rs in rouge_scores]) * 100
    rouge_l = np.mean([rs['rougeL'].fmeasure for rs in rouge_scores]) * 100
    print('rouge1:', rouge_1)
    print('rouge2:', rouge_2)
    print('rougeL:', rouge_l)
    return bleu2, bleu4, rouge_1, rouge_2, rouge_l, _meteor_score


class AttrDict(dict):
    def __init__(self, *args, **kwargs):
        super(AttrDict, self).__init__(*args, **kwargs)
        self.__dict__ = self


def pad_and_concat(tensor_list, fill_value=0):
    '''
    concat the first dimension and pad the second dimension
    tensor_list: [[B (diff), N_num, *], ...]
    '''
    device = tensor_list[0].device
    dtype=tensor_list[0].dtype
    max_dim1 = max(t.shape[1] for t in tensor_list)
    sum_dim0 = sum(t.shape[0] for t in tensor_list)
    if len(tensor_list[0].shape) == 3:
        out = torch.full((sum_dim0, max_dim1, tensor_list[0].shape[-1]), fill_value=fill_value, device=device, dtype=dtype)
        i = 0
        for t in tensor_list:
            out[i:i+t.shape[0], :t.shape[1]] = t
            i += t.shape[0]
        return out
    elif len(tensor_list[0].shape) == 2:
        out = torch.full((sum_dim0, max_dim1), fill_value=fill_value, device=device, dtype=dtype)
        i = 0
        for t in tensor_list:
            out[i:i+t.shape[0], :t.shape[1]] = t
            i += t.shape[0]
        return out
    raise NotImplementedError()


def hf_enable_gradient_checkpointing(hf_model):
    if hasattr(hf_model, "enable_input_require_grads"):
        hf_model.enable_input_require_grads()
    else:

        def make_inputs_require_grad(module, input, output):
            output.requires_grad_(True)

        hf_model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)

    # enable gradient checkpointing for memory efficiency
    hf_model.gradient_checkpointing_enable()
    return hf_model