from pathlib import Path from typing import Set from datasets import DatasetBuilder, GeneratorBasedBuilder, DatasetInfo, Features, Image, ClassLabel, Array3D, DownloadManager, SplitGenerator, BuilderConfig, Version import numpy as np import datasets VERSION = "v1_240507" HF_VERSION = "1.0.0" # Available Dataset View Names full_dataset_name = "full-dataset" semantic_segmentation_name = "semantic-segmentation" instance_segmentation_name = "instance-segmentation" animal_category_anomoalies_name = "animal-category-anomalies" re_id_best_name = "chicken-re-id-best-visibility" #re_id_good_name = "chicken-re-id-good-visibility" #re_id_bad_name = "chicken-re-id-bad-visibility" re_id_full_name = "chicken-re-id-all-visibility" # Example usage # from datasets import load_dataset # dataset = datasets.load_dataset( # "dariakern/Chicks4FreeID", # "chicken-re-id-best-visibility", # as_supervised=True, # trust_remote_code=True # ) ##### ONTOLOTGY ###### ontologies = { "v1_240507": {'tools': [{'classifications': [{'instructions': 'coop', 'options': [{'label': '1'}, {'label': '2'}, {'label': '3'}, {'label': '4'}, {'label': '5'}, {'label': '6'}, {'label': '7'}, {'label': '8'}, {'label': '9'}, {'label': '10'}, {'label': '11'},], 'required': True, 'type': 'radio'}, {'instructions': 'identity', 'options': [{'label': 'Beate'}, {'label': 'Borghild'}, {'label': 'Eleonore'}, {'label': 'Mona'}, {'label': 'Henriette'}, {'label': 'Margit'}, {'label': 'Millie'}, {'label': 'Sigrun'}, {'label': 'Kristina'}, {'label': 'Unknown'}, {'label': 'Tina'}, {'label': 'Gretel'}, {'label': 'Lena'}, {'label': 'Yolkoono'}, {'label': 'Skimmy'}, {'label': 'Mavi'}, {'label': 'Mirmir'}, {'label': 'Nugget'}, {'label': 'Fernanda'}, {'label': 'Isolde'}, {'label': 'Mechthild'}, {'label': 'Brunhilde'}, {'label': 'Spiderman'}, {'label': 'Brownie'}, {'label': 'Camy'}, {'label': 'Samy'}, {'label': 'Yin'}, {'label': 'Yuriko'}, {'label': 'Renate'}, {'label': 'Regina'}, {'label': 'Monika'}, {'label': 'Heidi'}, {'label': 'Erna'}, {'label': 'Marina'}, {'label': 'Kathrin'}, {'label': 'Isabella'}, {'label': 'Amalia'}, {'label': 'Edeltraut'}, {'label': 'Erdmute'}, {'label': 'Oktavia'}, {'label': 'Siglinde'}, {'label': 'Ulrike'}, {'label': 'Hermine'}, {'label': 'Matilda'}, {'label': 'Chantal'}, {'label': 'Chayenne'}, {'label': 'Jaqueline'}, {'label': 'Mandy'}, {'label': 'Henny'}, {'label': 'Shady'}, {'label': 'Shorty'}], 'required': True, 'type': 'radio'}, {'instructions': 'visibility', 'options': [{'label': 'best'}, {'label': 'good'}, {'label': 'bad'}], 'required': True, 'type': 'radio'}], 'color': '#1e1cff', 'name': 'chicken', 'required': False, 'tool': 'superpixel'}, {'color': '#FF34FF', 'name': 'background', 'required': False, 'tool': 'superpixel'}, {'classifications': [{'instructions': 'coop', 'options': [{'label': '1'}, {'label': '2'}, {'label': '3'}, {'label': '4'}, {'label': '5'}, {'label': '6'}, {'label': '7'}, {'label': '8'}, {'label': '9'}, {'label': '10'}, {'label': '11'}], 'required': True, 'type': 'radio'}, {'instructions': 'identity', 'options': [{'label': 'Evelyn'}, {'label': 'Marley'}], 'required': True, 'type': 'radio'}, {'instructions': 'visibility', 'options': [{'label': 'best'}, {'label': 'good'}, {'label': 'bad'}], 'required': True, 'type': 'radio'}], 'color': '#FF4A46', 'name': 'duck', 'required': False, 'tool': 'superpixel'}, {'classifications': [{'instructions': 'coop', 'options': [{'label': '1'}, {'label': '2'}, {'label': '3'}, {'label': '4'}, {'label': '5'}, {'label': '6'}, {'label': '7'}, {'label': '8'}, {'label': '9'}, {'label': '10'}, {'label': '11'}], 'required': True, 'type': 'radio'}, {'instructions': 'identity', 'options': [{'label': 'Elvis'}, {'label': 'Jackson'}], 'required': True, 'type': 'radio'}, {'instructions': 'visibility', 'options': [{'label': 'best'}, {'label': 'good'}, {'label': 'bad'}], 'required': True, 'type': 'radio'}], 'color': '#ff0000', 'name': 'rooster', 'required': False, 'tool': 'superpixel'}]} } ontologies["v1_240507_SMALL"] = ontologies["v1_240507"] class Ontology: ontology: dict = None def __init__(self, version_name: str): self.ontology: dict = ontologies[version_name] def names(self, class_name, tool_name=None, drop_unkown=False): """ Returns a list of all possible names for a given category (accross all tools) """ if class_name == "animal_category": return sorted(list({tool["name"] for tool in self.ontology["tools"]} - {"background"})) result = [] for tool in self.ontology["tools"]: if "classifications" in tool: for classification in tool["classifications"]: if classification["instructions"] == class_name and (tool_name is None or tool_name == tool["name"]): result.extend([option["label"] for option in classification["options"] if not (drop_unkown and option["label"] == "Unknown") and option["label"] not in result]) return list(result) def get_color_map(self): """ Returns a dictionary mapping class names to their respective colors """ return {tool["name"]: tool["color"] for tool in self.ontology["tools"]} ontology = Ontology(VERSION) # Feature Names IMAGE = "image" image_feature = {IMAGE: Image()} SEGMENTATION_MAKS = "segmentation_mask" segmentation_mask_feature = {SEGMENTATION_MAKS: Image()} INSTANCE_MASK = "instance_mask" instance_mask_feature = {INSTANCE_MASK: Image()} CROP = "crop" crop_feature = {CROP: Image()} ID = "identity" identity_feature = {ID: ClassLabel(names=ontology.names(ID))} chicken_only_identitiy_feature = {ID: ClassLabel(names=ontology.names(ID, "chicken", drop_unkown=True))} VISIBILITY = "visibility" visibility_feature = {VISIBILITY: ClassLabel(names=ontology.names(VISIBILITY))} COOP = "coop" coop_feature = {COOP: ClassLabel(names=ontology.names(COOP))} CATEGORY = "animal_category" animal_category_feature = {CATEGORY: ClassLabel(names=ontology.names(CATEGORY))} INSTANCES = "instances" instance_features = { **crop_feature, **instance_mask_feature, **identity_feature, **visibility_feature, **animal_category_feature, } all_features = { **image_feature, **segmentation_mask_feature, **coop_feature, INSTANCES: [instance_features], } def name_to_dict(filename: str): """ Converts a filename to a dictionary object by splitting the filename by underscores and using the even indices as keys and the odd indices as values. """ return {filename.split('_')[i]: filename.split('_')[i + 1] for i in range(0, len(filename.split('_')) - 1, 2)} class ChicksDataset(GeneratorBasedBuilder): BUILDER_CONFIGS = [ BuilderConfig(name=full_dataset_name, version=Version(HF_VERSION), description="The complete dataset including all features and image types. Includes all coops, visibility ratings, identities, and animal categories, as well as segmentation masks and instance masks."), BuilderConfig(name=semantic_segmentation_name, version=Version(HF_VERSION), description="Includes images and color-coded segmentation masks."), BuilderConfig(name=instance_segmentation_name, version=Version(HF_VERSION), description="Includes images and a corresponding sequence of binary instance segmentation masks for each instance on the image."), BuilderConfig(name=animal_category_anomoalies_name, version=Version(HF_VERSION), description="Includes images of mostly chicken, but also some roosters and ducks, which make up the anomalies in the dataset."), BuilderConfig(name=re_id_best_name, version=Version(HF_VERSION), description="Includes crops of chickens which have the best visibility rating for re-identification."), #BuilderConfig(name=re_id_good_name, version=Version(HF_VERSION), description="Includes crops of chickens which have neither the best nor the worst visibility rating for re-identification."), #BuilderConfig(name=re_id_bad_name, version=Version(HF_VERSION), description="Includes crops of chickens which have the worst (bad) visibility rating for re-identification."), BuilderConfig(name=re_id_full_name, version=Version(HF_VERSION), description="Includes crops of chickens with all visibilities for re-identification without any filtering on visibility rating."), ] def _info(self, *args, **kwargs): if self.config.name == full_dataset_name: return DatasetInfo( features=Features(all_features), ) elif self.config.name in [ re_id_full_name, re_id_best_name, # re_id_good_name, re_id_bad_name ]: return DatasetInfo( features=Features({ **crop_feature, **chicken_only_identitiy_feature, }), supervised_keys=( CROP, ID, ), ) elif self.config.name == semantic_segmentation_name: return DatasetInfo( features=Features({ **image_feature, **segmentation_mask_feature, }), supervised_keys=( IMAGE, SEGMENTATION_MAKS, ) ) elif self.config.name == instance_segmentation_name: return DatasetInfo( features=Features({ **image_feature, INSTANCES: [instance_mask_feature], }), supervised_keys=( IMAGE, INSTANCES, # TODO use nested reference to instance_mask_feature ) ) elif self.config.name == animal_category_anomoalies_name: return DatasetInfo( features=Features({ **crop_feature, **animal_category_feature, }), supervised_keys=( CROP, CATEGORY ) ) def _split_generators(self, dl_manager: DownloadManager): URL = f"https://huggingface.co/datasets/dariakern/Chicks4FreeID/resolve/main/{VERSION}.zip?download=true" base_path = Path(dl_manager.download_and_extract(URL)) # Only offer train test split for chicken-re-id task if self.config.name in [ re_id_full_name, re_id_best_name ]: from sklearn.model_selection import train_test_split # all crop files (only chicken, remove unknowns) all_crops = sorted([ crop_file for crop_file in base_path.rglob(f"**/{VERSION}/reId/chicken/**/*crop_*.png") if "Unknown" not in crop_file.parts ]) # all identity targets (labels) identities = [name_to_dict(crop.stem)[ID] for crop in all_crops] if VERSION == "v1_240507_SMALL": train_crops, test_crops = all_crops, all_crops else: # Splitting the dataset into train and test using stratified train_test_split train_crops, test_crops, _, _ = train_test_split( all_crops, identities, test_size=0.2, stratify=identities, shuffle=True, random_state=42 ) return [ SplitGenerator( gen_kwargs={"base_path": base_path, "split": set(train_crops)}, name=datasets.Split.TRAIN, ), SplitGenerator( gen_kwargs={"base_path": base_path, "split": set(test_crops)}, name=datasets.Split.TEST, ) ] else: return [ SplitGenerator( name=datasets.Split.TRAIN, gen_kwargs={"base_path": base_path, "split": None}) ] def _generate_all(self, base_path: Path, split: Set[Path]=None): """ Generates all examples for the dataset, including all features. Args: base_path (Path): The base path to the dataset split (Set[Path]): The paths to all instance crops to include in the current dataset """ img_dir = base_path / f"{VERSION}/images" mask_dir = base_path / f"{VERSION}/masks" reid_dir = base_path / f"{VERSION}/reId" # Collecting images, segmentation masks, and instance masks for img_file in img_dir.iterdir(): image_id = img_file.stem image_path = img_file segmentation_mask_path = mask_dir / f"{image_id}_segmentationMask.png" instance_masks = list(mask_dir.rglob(f"{image_id}_instanceMask_*.png")) instance_crops = list(reid_dir.rglob(f"**/{image_id}_crop_*.png")) # Check if all crops have a corresponding instance mask assert len(instance_masks) == len(instance_crops) and len(instance_masks) > 0 # Remove any instance_crops that are not in crops_split if split is not None: instance_crops = [crop for crop in instance_crops if crop in split] instance_data = [] infos = {} for instance_mask_path, crop_path in zip(instance_masks, instance_crops): infos = name_to_dict(crop_path.stem) instance_data.append({ INSTANCE_MASK: str(instance_mask_path), CROP: str(crop_path), VISIBILITY: infos[VISIBILITY], ID: infos[ID], CATEGORY: crop_path.relative_to(reid_dir).parts[0], }) if instance_data: yield image_id, { IMAGE: str(image_path), SEGMENTATION_MAKS: str(segmentation_mask_path), COOP: infos[COOP], INSTANCES: instance_data, } def _generate_examples(self, **kwargs): if self.config.name in [full_dataset_name]: yield from self._generate_all(**kwargs) elif self.config.name == semantic_segmentation_name: for image_id, example in self._generate_all(**kwargs): yield image_id, { IMAGE: example[IMAGE], SEGMENTATION_MAKS: example[SEGMENTATION_MAKS], } elif self.config.name == instance_segmentation_name: for image_id, example in self._generate_all(**kwargs): yield image_id, { IMAGE: example[IMAGE], INSTANCES: [ { INSTANCE_MASK: instance[INSTANCE_MASK] } for instance in example[INSTANCES] ] } elif self.config.name == animal_category_anomoalies_name: for image_id, example in self._generate_all(**kwargs): for instance in example[INSTANCES]: instance_id = Path(instance[CROP]).stem yield instance_id, { CROP: instance[CROP], CATEGORY: instance[CATEGORY], } elif self.config.name in [ re_id_best_name, re_id_full_name, # re_id_good_name, re_id_bad_name ]: for image_id, example in self._generate_all(**kwargs): for instance in example[INSTANCES]: # Conditions for filtering use_all = self.config.name == re_id_full_name selected_visibility = instance[VISIBILITY] == self.config.name.split("-")[-2] if use_all or selected_visibility: instance_id = Path(instance[CROP]).stem yield instance_id, { CROP: instance[CROP], ID: instance[ID], }