# 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. """The Loading scripts for ImageRewardDB.""" import pandas as pd import json import os import datasets from huggingface_hub import hf_hub_url # Find for instance the citation on arxiv or on the dataset repo/website _CITATION = """\ @misc{xu2023imagereward, title={ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation}, author={Jiazheng Xu and Xiao Liu and Yuchen Wu and Yuxuan Tong and Qinkai Li and Ming Ding and Jie Tang and Yuxiao Dong}, year={2023}, eprint={2304.05977}, archivePrefix={arXiv}, primaryClass={cs.CV} } """ # You can copy an official description _DESCRIPTION = """\ ImageRewardDB is a comprehensive text-to-image comparison dataset, focusing on text-to-image human preference. \ It consists of 137k pairs of expert comparisons, based on text prompts and corresponding model outputs from DiffusionDB. \ To build the ImageRewadDB, we design a pipeline tailored for it, establishing criteria for quantitative assessment and \ annotator training, optimizing labeling experience, and ensuring quality validation. \ """ _HOMEPAGE = "https://huggingface.co/datasets/THUDM/ImageRewardDB" _VERSION = datasets.Version("1.0.0") _LICENSE = "Apache License 2.0" # The HuggingFace Datasets library doesn't host the datasets but only points to the original files. # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method) _REPO_ID = "THUDM/ImageRewardDB" _URLS = {} _PART_IDS = { "train": 32, "validation": 2, "test": 2 } for name in list(_PART_IDS.keys()): _URLS[name] = {} for i in range(1, _PART_IDS[name]+1): _URLS[name][i] = hf_hub_url( _REPO_ID, filename=f"images/{name}/{name}_{i}.zip", repo_type="dataset" ) _URLS[name]["metadata"] = hf_hub_url( _REPO_ID, filename=f"metadata-{name}.parquet", repo_type="dataset" ) class ImageRewardDBConfig(datasets.BuilderConfig): '''BuilderConfig for ImageRewardDB''' def __init__(self, part_ids, **kwargs): '''BuilderConfig for ImageRewardDB Args: part_ids([int]): A list of part_ids. **kwargs: keyword arguments forwarded to super ''' super(ImageRewardDBConfig, self).__init__(version=_VERSION, **kwargs) self.part_ids = part_ids class ImageRewardDB(datasets.GeneratorBasedBuilder): """A dataset of 137k expert comparisons to date, demonstrating the text-to-image human preference.""" # This is an example of a dataset with multiple configurations. # If you don't want/need to define several sub-sets in your dataset, # just remove the BUILDER_CONFIG_CLASS and the BUILDER_CONFIGS attributes. # If you need to make complex sub-parts in the datasets with configurable options # You can create your own builder configuration class to store attribute, inheriting from datasets.BuilderConfig # BUILDER_CONFIG_CLASS = MyBuilderConfig # You will be able to load one or the other configurations in the following list with # data = datasets.load_dataset('my_dataset', 'first_domain') # data = datasets.load_dataset('my_dataset', 'second_domain') BUILDER_CONFIGS = [] for num_k in [1,2,4,8]: part_ids = { "train": 4*num_k, "validation": 2, "test": 2 } BUILDER_CONFIGS.append( ImageRewardDBConfig(name=f"{num_k}k_group", part_ids=part_ids, description=f"This is a {num_k}k-scale groups of ImageRewardDB") ) BUILDER_CONFIGS.append( ImageRewardDBConfig(name=f"{num_k}k", part_ids=part_ids, description=f"This is a {num_k}k-scale ImageRewardDB") ) BUILDER_CONFIGS.append( ImageRewardDBConfig(name=f"{num_k}k_pair", part_ids=part_ids, description=f"This is a {num_k}k-scale pairs of ImageRewardDB") ) DEFAULT_CONFIG_NAME = "8k" # It's not mandatory to have a default configuration. Just use one if it make sense. def _info(self): if "group" in self.config.name: features = datasets.Features( { "prompt_id": datasets.Value("string"), "prompt": datasets.Value("string"), "classification": datasets.Value("string"), "image": datasets.Sequence(datasets.Image()), "rank": datasets.Sequence(datasets.Value("int8")), "overall_rating": datasets.Sequence(datasets.Value("int8")), "image_text_alignment_rating": datasets.Sequence(datasets.Value("int8")), "fidelity_rating": datasets.Sequence(datasets.Value("int8")) } ) elif "pair" in self.config.name: features = datasets.Features( { "prompt_id": datasets.Value("string"), "prompt": datasets.Value("string"), "classification": datasets.Value("string"), "img_better": datasets.Image(), "img_worse": datasets.Image() } ) else: features = datasets.Features( { "image": datasets.Image(), "prompt_id": datasets.Value("string"), "prompt": datasets.Value("string"), "classification": datasets.Value("string"), "image_amount_in_total": datasets.Value("int8"), "rank": datasets.Value("int8"), "overall_rating": datasets.Value("int8"), "image_text_alignment_rating": datasets.Value("int8"), "fidelity_rating": datasets.Value("int8") } ) return datasets.DatasetInfo( # This is the description that will appear on the datasets page. description=_DESCRIPTION, # This defines the different columns of the dataset and their types features=features, # Here we define them above because they are different between the two configurations # If there's a common (input, target) tuple from the features, uncomment supervised_keys line below and # specify them. They'll be used if as_supervised=True in builder.as_dataset. # supervised_keys=("sentence", "label"), # Homepage of the dataset for documentation homepage=_HOMEPAGE, # License for the dataset if available license=_LICENSE, # Citation for the dataset citation=_CITATION, ) def _split_generators(self, dl_manager): # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name # dl_manager is a datasets.download.DownloadManager that can be used to download and extract URLS # It can accept any type or nested list/dict and will give back the same structure with the url replaced with path to local files. # By default the archives will be extracted and a path to a cached folder where they are extracted is returned instead of the archive data_dirs = {name: [] for name in list(_PART_IDS.keys())} json_paths = {name: [] for name in list(_PART_IDS.keys())} metadata_paths = {name: [] for name in list(_PART_IDS.keys())} for key in list(self.config.part_ids.keys()): for i in range(1, self.config.part_ids[key]+1): data_dir = dl_manager.download_and_extract(_URLS[key][i]) data_dirs[key].append(data_dir) json_paths[key].append(os.path.join(data_dir, f"{key}_{i}.json")) metadata_paths[key] = dl_manager.download(_URLS[key]["metadata"]) return [ datasets.SplitGenerator( name=datasets.Split.TRAIN, # These kwargs will be passed to _generate_examples gen_kwargs={ "split": "train", "data_dirs": data_dirs["train"], "json_paths": json_paths["train"], "metadata_path": metadata_paths["train"] }, ), datasets.SplitGenerator( name=datasets.Split.VALIDATION, # These kwargs will be passed to _generate_examples gen_kwargs={ "split": "validation", "data_dirs": data_dirs["validation"], "json_paths": json_paths["validation"], "metadata_path": metadata_paths["validation"] }, ), datasets.SplitGenerator( name=datasets.Split.TEST, # These kwargs will be passed to _generate_examples gen_kwargs={ "split": "test", "data_dirs": data_dirs["test"], "json_paths": json_paths["test"], "metadata_path": metadata_paths["test"] }, ), ] # method parameters are unpacked from `gen_kwargs` as given in `_split_generators` def _generate_examples(self, split, data_dirs, json_paths, metadata_path): # The `key` is for legacy reasons (tfds) and is not important in itself, but must be unique for each example. num_data_dirs = len(data_dirs) assert num_data_dirs == len(json_paths) #Iterate throug all extracted zip folders for images # metadata_table = pd.read_parquet(metadata_path) for index, json_path in enumerate(json_paths): json_data = json.load(open(json_path, "r", encoding="utf-8")) if "group" in self.config.name or "pair" in self.config.name: group_num = 0 image_path = [] rank = [] overall_rating, image_text_alignment_rating, fidelity_rating = [], [], [] for sample in json_data: if group_num == 0: image_path.clear() rank.clear() overall_rating.clear() image_text_alignment_rating.clear() fidelity_rating.clear() prompt_id = sample["prompt_id"] prompt = sample["prompt"] classification = sample["classification"] image_amount_in_total = sample["image_amount_in_total"] # image_path.append(sample["image_path"]) image_path.append(os.path.join(data_dirs[index], str(sample["image_path"]).split("/")[-1])) rank.append(sample["rank"]) overall_rating.append(sample["overall_rating"]) image_text_alignment_rating.append(sample["image_text_alignment_rating"]) fidelity_rating.append(sample["fidelity_rating"]) group_num += 1 if group_num == image_amount_in_total: group_num = 0 if "group" in self.config.name: yield prompt_id, ({ "prompt_id": prompt_id, "prompt": prompt, "classification": classification, "image": [{ "path": image_path[idx], "bytes": open(image_path[idx], "rb").read() } for idx in range(image_amount_in_total)], "rank": rank, "overall_rating": overall_rating, "image_text_alignment_rating": image_text_alignment_rating, "fidelity_rating": fidelity_rating, }) else: for idx in range(image_amount_in_total): for idy in range(idx+1, image_amount_in_total): if rank[idx] < rank[idy]: yield prompt_id, ({ "prompt_id": prompt_id, "prompt": prompt, "classification": classification, "img_better": { "path": image_path[idx], "bytes": open(image_path[idx], "rb").read() }, "img_worse": { "path": image_path[idy], "bytes": open(image_path[idy], "rb").read() } }) elif rank[idx] > rank[idy]: yield prompt_id, ({ "prompt_id": prompt_id, "prompt": prompt, "classification": classification, "img_better": { "path": image_path[idy], "bytes": open(image_path[idy], "rb").read() }, "img_worse": { "path": image_path[idx], "bytes": open(image_path[idx], "rb").read() } }) else: for example in json_data: image_path = os.path.join(data_dirs[index], str(example["image_path"]).split("/")[-1]) yield example["image_path"], { "image": { "path": image_path, "bytes": open(image_path, "rb").read() }, "prompt_id": example["prompt_id"], "prompt": example["prompt"], "classification": example["classification"], "image_amount_in_total": example["image_amount_in_total"], "rank": example["rank"], "overall_rating": example["overall_rating"], "image_text_alignment_rating": example["image_text_alignment_rating"], "fidelity_rating": example["fidelity_rating"] }