sescore.py

# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""SEScore: a text generation evaluation metric """

import evaluate
import datasets

import comet
from typing import Dict
import torch
from comet.encoders.base import Encoder
from comet.encoders.bert import BERTEncoder
from transformers import AutoModel, AutoTokenizer

class robertaEncoder(BERTEncoder):
    def __init__(self, pretrained_model: str) -> None:
        super(Encoder, self).__init__()
        self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model)
        self.model = AutoModel.from_pretrained(
            pretrained_model, add_pooling_layer=False
        )
        self.model.encoder.output_hidden_states = True

    @classmethod
    def from_pretrained(cls, pretrained_model: str) -> Encoder:
        return robertaEncoder(pretrained_model)

    def forward(
        self, input_ids: torch.Tensor, attention_mask: torch.Tensor, **kwargs
    ) -> Dict[str, torch.Tensor]:
        last_hidden_states, _, all_layers = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            output_hidden_states=True,
            return_dict=False,
        )
        return {
            "sentemb": last_hidden_states[:, 0, :],
            "wordemb": last_hidden_states,
            "all_layers": all_layers,
            "attention_mask": attention_mask,
        }


# TODO: Add BibTeX citation
_CITATION = """\
@inproceedings{xu-etal-2022-not,
  title={Not All Errors are Equal: Learning Text Generation Metrics using Stratified Error Synthesis},
  author={Xu, Wenda and Tuan, Yi-lin and Lu, Yujie and Saxon, Michael and Li, Lei and Wang, William Yang},
  booktitle ={Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing},
  month={dec},
  year={2022},
  url={https://arxiv.org/abs/2210.05035}
}
"""

_DESCRIPTION = """\
SEScore is an evaluation metric that trys to compute an overall score to measure text generation quality.
"""

_KWARGS_DESCRIPTION = """
Calculates how good are predictions given some references
Args:
    predictions: list of candidate outputs
    references: list of references 
Returns:
    {"mean_score": mean_score, "scores": scores}
    
Examples:
    >>> import evaluate
    >>> sescore = evaluate.load("xu1998hz/sescore")
    >>> score = sescore.compute(
            references=['sescore is a simple but effective next-generation text evaluation metric'],
            predictions=['sescore is simple effective text evaluation metric for next generation']
        )
"""

# TODO: Define external resources urls if needed
BAD_WORDS_URL = "http://url/to/external/resource/bad_words.txt"


@evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
class SEScore(evaluate.Metric):
    """SEScore"""

    def _info(self):
        # TODO: Specifies the evaluate.EvaluationModuleInfo object
        return evaluate.MetricInfo(
            # This is the description that will appear on the modules page.
            module_type="metric",
            description=_DESCRIPTION,
            citation=_CITATION,
            inputs_description=_KWARGS_DESCRIPTION,
            # This defines the format of each prediction and reference
            features=datasets.Features({
                'predictions': datasets.Value("string", id="sequence"),
                'references': datasets.Value("string", id="sequence"),
            }),
            # Homepage of the module for documentation
            homepage="http://module.homepage",
            # Additional links to the codebase or references
            codebase_urls=["http://github.com/path/to/codebase/of/new_module"],
            reference_urls=["http://path.to.reference.url/new_module"]
        )

    def _download_and_prepare(self, dl_manager):
        """download SEScore checkpoints to compute the scores"""
        # Download SEScore checkpoint
        from comet import load_from_checkpoint
        import os
        from huggingface_hub import snapshot_download
        # initialize roberta into str2encoder
        comet.encoders.str2encoder['RoBERTa'] = robertaEncoder
        destination = snapshot_download(repo_id="xu1998hz/sescore_english_mt", revision="main")
        self.scorer = load_from_checkpoint(f'{destination}/checkpoint/sescore_english_mt.ckpt')

    def _compute(self, predictions, references, gpus=None, progress_bar=False):
        if gpus is None:
            gpus = 1 if torch.cuda.is_available() else 0

        data = {"src": references, "mt": predictions}
        data = [dict(zip(data, t)) for t in zip(*data.values())]
        scores, mean_score = self.scorer.predict(data, gpus=gpus, progress_bar=progress_bar)
        return {"mean_score": mean_score, "scores": scores}