# Copyright 2022 The HuggingFace Datasets Authors. # # 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. """Caltech 101 loading script""" from __future__ import annotations from pathlib import Path import datasets import numpy as np import scipy.io from datasets.tasks import ImageClassification _CITATION = """\ @article{FeiFei2004LearningGV, title={Learning Generative Visual Models from Few Training Examples: An Incremental Bayesian Approach Tested on 101 Object Categories}, author={Li Fei-Fei and Rob Fergus and Pietro Perona}, journal={Computer Vision and Pattern Recognition Workshop}, year={2004}, } """ _DESCRIPTION = """\ Pictures of objects belonging to 101 categories. About 40 to 800 images per category. Most categories have about 50 images. Collected in September 2003 by Fei-Fei Li, Marco Andreetto, and Marc'Aurelio Ranzato. The size of each image is roughly 300 x 200 pixels. """ _HOMEPAGE = "https://data.caltech.edu/records/20086" _LICENSE = "CC BY 4.0" _DATA_URL = "caltech-101.zip" _NAMES = [ "accordion", "airplanes", "anchor", "ant", "background_google", "barrel", "bass", "beaver", "binocular", "bonsai", "brain", "brontosaurus", "buddha", "butterfly", "camera", "cannon", "car_side", "ceiling_fan", "cellphone", "chair", "chandelier", "cougar_body", "cougar_face", "crab", "crayfish", "crocodile", "crocodile_head", "cup", "dalmatian", "dollar_bill", "dolphin", "dragonfly", "electric_guitar", "elephant", "emu", "euphonium", "ewer", "faces", "faces_easy", "ferry", "flamingo", "flamingo_head", "garfield", "gerenuk", "gramophone", "grand_piano", "hawksbill", "headphone", "hedgehog", "helicopter", "ibis", "inline_skate", "joshua_tree", "kangaroo", "ketch", "lamp", "laptop", "leopards", "llama", "lobster", "lotus", "mandolin", "mayfly", "menorah", "metronome", "minaret", "motorbikes", "nautilus", "octopus", "okapi", "pagoda", "panda", "pigeon", "pizza", "platypus", "pyramid", "revolver", "rhino", "rooster", "saxophone", "schooner", "scissors", "scorpion", "sea_horse", "snoopy", "soccer_ball", "stapler", "starfish", "stegosaurus", "stop_sign", "strawberry", "sunflower", "tick", "trilobite", "umbrella", "watch", "water_lilly", "wheelchair", "wild_cat", "windsor_chair", "wrench", "yin_yang", ] # For some reason, the category names in "101_ObjectCategories" and # "Annotations" do not always match. This is a manual map between the # two. Defaults to using same name, since most names are fine. _ANNOTATION_NAMES_MAP = { "Faces": "Faces_2", "Faces_easy": "Faces_3", "Motorbikes": "Motorbikes_16", "airplanes": "Airplanes_Side_2", } _TRAIN_POINTS_PER_CLASS = 30 class Caltech101(datasets.GeneratorBasedBuilder): """Caltech 101 dataset.""" VERSION = datasets.Version("1.0.0") _BUILDER_CONFIG_WITH_BACKGROUND = datasets.BuilderConfig( name="with_background_category", version=VERSION, description="Dataset containing the 101 categories and the additonnal background one. " "No annotations.", ) _BUILDER_CONFIG_WITHOUT_BACKGROUND = datasets.BuilderConfig( name="without_background_category", version=VERSION, description="Dataset containing only the 101 categories and their annotations.", ) BUILDER_CONFIGS = [ _BUILDER_CONFIG_WITH_BACKGROUND, _BUILDER_CONFIG_WITHOUT_BACKGROUND, ] def _info(self): if self.config.name == self._BUILDER_CONFIG_WITHOUT_BACKGROUND.name: features = datasets.Features( { "image": datasets.Image(), "label": datasets.features.ClassLabel(names=_NAMES), "annotation": { "obj_contour": datasets.features.Array2D( shape=(2, None), dtype="float64" ), "box_coord": datasets.features.Array2D( shape=(1, 4), dtype="int64" ), }, } ) else: features = datasets.Features( { "image": datasets.Image(), "label": datasets.features.ClassLabel(names=_NAMES), } ) return datasets.DatasetInfo( description=_DESCRIPTION, features=features, homepage=_HOMEPAGE, license=_LICENSE, citation=_CITATION, ) def _split_generators(self, dl_manager): data_root_dir = dl_manager.download_and_extract(_DATA_URL) img_folder_compress_path = [ file for file in dl_manager.iter_files(data_root_dir) if Path(file).name == "101_ObjectCategories.tar.gz" ][0] annotations_folder_compress_path = [ file for file in dl_manager.iter_files(data_root_dir) if Path(file).name == "Annotations.tar" ][0] img_dir = dl_manager.extract(img_folder_compress_path) annotation_dir = dl_manager.extract(annotations_folder_compress_path) return [ datasets.SplitGenerator( name=datasets.Split.TRAIN, gen_kwargs={ "img_dir": Path(img_dir) / "101_ObjectCategories", "annotation_dir": Path(annotation_dir) / "Annotations", "split": "train", "config_name": self.config.name, }, ), datasets.SplitGenerator( name=datasets.Split.TEST, gen_kwargs={ "img_dir": Path(img_dir) / "101_ObjectCategories", "annotation_dir": Path(annotation_dir) / "Annotations", "split": "test", "config_name": self.config.name, }, ), ] def _generate_examples(self, img_dir, annotation_dir, split, config_name): # Same stratagy as the one proposed in TF datasets: 30 random examples from each class are added to the train # split, and the remainder are added to the test split. # Source: https://github.com/tensorflow/datasets/blob/1106d587f97c4fca68c5b593dc7dc48c790ffa8c/tensorflow_datasets/image_classification/caltech.py#L88-L140 is_train_split = split == "train" # Sets random seed so the random partitioning of files is the same when # called for the train and test splits. numpy_original_state = np.random.get_state() np.random.seed(1234) for class_dir in img_dir.iterdir(): class_name = class_dir.name index_codes = [ image_path.name.split("_")[1][: -len(".jpg")] for image_path in class_dir.iterdir() if image_path.name.endswith(".jpg") ] # _TRAIN_POINTS_PER_CLASS datapoints are sampled for the train split, # the others constitute the test split. if _TRAIN_POINTS_PER_CLASS > len(index_codes): raise ValueError( f"Fewer than {_TRAIN_POINTS_PER_CLASS} ({len(index_codes)}) points in class {class_dir.name}" ) train_indices = np.random.choice( index_codes, _TRAIN_POINTS_PER_CLASS, replace=False ) test_indices = set(index_codes).difference(train_indices) indices_to_emit = train_indices if is_train_split else test_indices if ( class_name == "BACKGROUND_Google" and config_name == self._BUILDER_CONFIG_WITHOUT_BACKGROUND.name ): print("skip BACKGROUND_Google") continue for indice in indices_to_emit: record = { "image": str(class_dir / f"image_{indice}.jpg"), "label": class_dir.name.lower(), } if config_name == self._BUILDER_CONFIG_WITHOUT_BACKGROUND.name: if class_name in _ANNOTATION_NAMES_MAP: annotations_class_name = _ANNOTATION_NAMES_MAP[class_name] else: annotations_class_name = class_name data = scipy.io.loadmat( str( annotation_dir / annotations_class_name / f"annotation_{indice}.mat" ) ) # raise ValueError(data["obj_contour"].dtype, data["box_coord"]) record["annotation"] = { "obj_contour": data["obj_contour"], "box_coord": data["box_coord"], } yield f"{class_dir.name.lower()}/{f'image_{indice}.jpg'}", record # Resets the seeds to their previous states. np.random.set_state(numpy_original_state)