import os from typing import Dict, List, Tuple import datasets import jsonlines as jl import pandas as pd from seacrowd.utils import schemas from seacrowd.utils.configs import SEACrowdConfig from seacrowd.utils.constants import Licenses, Tasks _CITATION = """\ @inproceedings{thapliyal-etal-2022-crossmodal, title = "Crossmodal-3600: A Massively Multilingual Multimodal Evaluation Dataset", author = "Thapliyal, Ashish V. and Pont Tuset, Jordi and Chen, Xi and Soricut, Radu", editor = "Goldberg, Yoav and Kozareva, Zornitsa and Zhang, Yue", booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing", month = dec, year = "2022", address = "Abu Dhabi, United Arab Emirates", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2022.emnlp-main.45", doi = "10.18653/v1/2022.emnlp-main.45", pages = "715--729", } """ _DATASETNAME = "cc3m_35l" _DESCRIPTION = """\ CC3M-35L is created by translating Conceptual Captions 3M (Sharma et al., 2018), originally in English, to the other 34 languages using Google's machine translation API. """ _HOMEPAGE = "https://google.github.io/crossmodal-3600/" _LICENSE = Licenses.CC_BY_4_0.value # the image URLs are contained in tsv file together with the original captions which can be downloaded locally using google account. # those tsv file originally can be found and downloaded from this page https://ai.google.com/research/ConceptualCaptions/download # there are no direct image folder ready, so it needs to be downloaded one by one # some warnings may occur when downloading due to reasons such as security certificate and others _URLS = { "trans_train": "https://storage.googleapis.com/crossmodal-3600/cc3m_mt_train.jsonl.gz", "trans_dev": "https://storage.googleapis.com/crossmodal-3600/cc3m_mt_dev.jsonl.gz", } _SUPPORTED_TASKS = [Tasks.IMAGE_CAPTIONING] _SOURCE_VERSION = "1.0.0" _SEACROWD_VERSION = "2024.06.20" _LANGUAGES = ["fil", "ind", "tha", "vie"] _LOCAL = True class CC3M35L(datasets.GeneratorBasedBuilder): """ CC3M-35L is created by translating Conceptual Captions 3M (Sharma et al., 2018), originally in English, to the other 34 languages using Google's machine translation API. """ SOURCE_VERSION = datasets.Version(_SOURCE_VERSION) SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION) BUILDER_CONFIGS = [SEACrowdConfig(name=f"cc3m_35l_{lang}_source", version=datasets.Version(_SOURCE_VERSION), description=f"cc3m_35l_{lang} source schema", schema="source", subset_id=f"cc3m_35l_{lang}",) for lang in _LANGUAGES] + [ SEACrowdConfig( name=f"{_DATASETNAME}_{lang}_seacrowd_imtext", version=datasets.Version(_SEACROWD_VERSION), description=f"{_DATASETNAME}_{lang} SEACrowd schema", schema="seacrowd_imtext", subset_id=f"{_DATASETNAME}_{lang}", ) for lang in _LANGUAGES ] DEFAULT_CONFIG_NAME = "cc3m_35l_id_source" def _info(self) -> datasets.DatasetInfo: if self.config.schema == "source": features = datasets.Features( { "id": datasets.Value("string"), "image_paths": datasets.Value("string"), "src_lang": datasets.Value("string"), "caption_tokenized": datasets.Value("string"), "trg_lang": datasets.Value("string"), "translation_tokenized": datasets.Value("string"), "backtranslation_tokenized": datasets.Value("string"), } ) elif self.config.schema == "seacrowd_imtext": features = schemas.image_text_features() return datasets.DatasetInfo( description=_DESCRIPTION, features=features, homepage=_HOMEPAGE, license=_LICENSE, citation=_CITATION, ) def fill_img_path(self, df: pd.DataFrame, line: dict): exceptions = [] selected_row = df.query('caption==@line["caption_tokenized"]') # it may return several rows, skip of empty if not selected_row.empty: # for each row, download the image, use its path and put the translation for idx, row in selected_row.iterrows(): row["trans_caption"] = line["translation_tokenized"] row["backtrans_caption"] = line["backtranslation_tokenized"] # if the image cannot be downloaded for some reason, skip it # may cause difference in the total data each run try: row["img_path"] = datasets.DownloadManager().download(row["img_url"]) except: exceptions.append(idx) return selected_row, exceptions def is_target(self, line: dict, trg_lang: str): if line["trg_lang"] == trg_lang: return line def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]: """Returns SplitGenerators.""" dev_path = dl_manager.download_and_extract(_URLS["trans_dev"]) train_path = dl_manager.download_and_extract(_URLS["trans_train"]) if self.config.data_dir is None: raise ValueError("This is a local dataset. Please pass the data_dir kwarg to load_dataset.") else: data_dir = self.config.data_dir # read tsv from local train and validation files gcc_val = os.path.join(data_dir, "Validation_GCC-1.1.0-Validation.tsv") gcc_train = os.path.join(data_dir, "Train_GCC-training.tsv") # make it into pandas dataframe colnames = ["caption", "img_url"] gcc_val_df = pd.read_csv(gcc_val, sep="\t", header=None, names=colnames) gcc_train_df = pd.read_csv(gcc_train, sep="\t", header=None, names=colnames) # add new column to keep the downloaded image path gcc_val_df["img_path"] = None gcc_train_df["img_path"] = None # add new column to keep the translated caption gcc_val_df["trans_caption"] = None gcc_train_df["trans_caption"] = None gcc_val_df["backtrans_caption"] = None gcc_train_df["backtrans_caption"] = None # match the original captions in the translated set to the dataframe caption # download the images from the URL and use it as the filepath train_exceptions = [] val_exceptions = [] current_lang = self.config.subset_id.split("_")[2] val_caption_targets = [] train_caption_targets = [] # filter validation data with jl.open(os.path.join(dev_path), mode="r") as j: val_caption_targets = [line for line in j if line["trg_lang"] == current_lang] #for line in val_caption_targets[:100]: # this was for debugging only for line in val_caption_targets: res = self.fill_img_path(gcc_train_df, line) val_exceptions.extend(res[1]) gcc_val_df.update(res[0]) # clean the memory val_caption_targets = [] # filter train data with jl.open(os.path.join(train_path), mode="r") as j: train_caption_targets = [line for line in j if line["trg_lang"] == current_lang] #for line in train_caption_targets[:100]: # this was for debugging only for line in train_caption_targets: res = self.fill_img_path(gcc_val_df, line) train_exceptions.extend(res[1]) gcc_train_df.update(res[0]) # clean the memory train_caption_targets = [] return [ datasets.SplitGenerator( name=datasets.Split.TRAIN, gen_kwargs={ "filepath": gcc_train_df, "exceptions": train_exceptions, }, ), datasets.SplitGenerator( name=datasets.Split.VALIDATION, gen_kwargs={ "filepath": gcc_val_df, "exceptions": val_exceptions, }, ), ] def _generate_examples(self, filepath: dict, exceptions: list) -> Tuple[int, Dict]: """Yields examples as (key, example) tuples.""" for idx, row in filepath.iterrows(): if idx not in exceptions: if self.config.schema == "source": yield idx, { "id": str(idx), "image_paths": row["img_path"], "src_lang": "en", "caption_tokenized": row["caption"], "trg_lang": self.config.subset_id.split("_")[2], "translation_tokenized": row["trans_caption"], "backtranslation_tokenized": row["backtrans_caption"], } elif self.config.schema == "seacrowd_imtext": yield idx, { "id": str(idx), "image_paths": [row["img_path"]], "texts": row["trans_caption"], "metadata": { "context": None, "labels": None, }, } else: raise ValueError(f"Invalid config: {self.config.name}")