[Tensorflow] 딥러닝을 위한 공개 데이터셋 Tensorflow Datasets

딥러닝, 머신러닝을 공부하다 보면 예제로 활용할 수 있는 데이터를 마련하거나 찾기가 어려워서 곤란할 때가 있습니다. 특히 라벨링이 된 이미지, 영상, 음성 등의 데이터의 경우 자체적으로 마련하기가 쉽지 않습니다. 

이번 포스팅에서는 딥러닝, 기계학습을 하는데 활용할 수 있도록 공개된 데이터셋을 TensorFlow Datasets 에서 다운로드하고 fetch 하는 방법을 소개하겠습니다. 

TensorFlow 데이터셋은 아래의 두 곳에서 다운로드 할 수 있습니다. 

(Many many thanks to TensorFlow team!!! ^__^)

• TensorFlow Datasets : https://www.tensorflow.org/datasets

• TensorFlow Datasets on GitHub : https://github.com/tensorflow/datasets

(1) TensorFlow 2.0 과 TensorFlow Datasets 라이브러리 설치

cmd 명령 프롬프트 창에서 pip install 로 TensorFlow 2.0 과 TensorFlow DataSets 라이브러리를 설치합니다. 

(* CPU 를 사용할 경우)

$ pip install --upgrade pip $ pip install tensorflow $ pip install tensorflow_datasets

(* GPU를 사용할 경우)

$ pip install tensorflow-gpu

(2) tensorflow 와 tensorflow_datasets 라이브러리 import 후 Dataset 로딩하기

TensorFlow v2와 tensorflow_datasets 라이브르러를 import 하겠습니다. 

import tensorflow.compat.v2 as tf import tensorflow_datasets as tfds

TensorFlow v2 부터는 PyTorch처럼 Eager 모드를 지원합니다. Eager모드와 Graph 모드를 활성화시키겠습니다. 

# tfds works in both Eager and Graph modes tf.enable_v2_behavior()

TensorFlow Datasets에 등록된 모든 공개 데이터셋 리스트를 조회해보겠습니다. 양이 너무 많아서 중간은 생략했는데요, 아래에 카테고리별로 리스트를 다시 정리해보았습니다. 

# Tensorflow Datasets Lists tfds.list_builders() ['abstract_reasoning', 'aeslc', 'aflw2k3d', 'amazon_us_reviews', 'arc', 'bair_robot_pushing_small', : -- 너무 많아서 중간 생략 ^^; -- : 'wmt18_translate', 'wmt19_translate', 'wmt_t2t_translate', 'wmt_translate', 'xnli', 'xsum', 'yelp_polarity_reviews']

Audio, Image, Object Detection, Structured, Summarization, Text, Translate, Video 의 8개 범주로 데이터셋이 구분되어 정리되어 있습니다. (* link: https://www.tensorflow.org/datasets/catalog/overview)

Audio	Image
groove librispeech libritts ljspeech nsynth savee speech_commands	abstract_reasoning aflw2k3d arc beans bigearthnet binarized_mnist binary_alpha_digits caltech101 caltech_birds2010 caltech_birds2011 cars196 cassava cats_vs_dogs celeb_a celeb_a_hq cifar10 cifar100 cifar10_1 cifar10_corrupted citrus_leaves cityscapes clevr cmaterdb coil100 colorectal_histology colorectal_histology_large curated_breast_imaging_ddsm cycle_gan deep_weeds diabetic_retinopathy_detection div2k dmlab downsampled_imagenet dsprites dtd duke_ultrasound emnist eurosat fashion_mnist flic food101 geirhos_conflict_stimuli horses_or_humans i_naturalist2017 image_label_folder imagenet2012 imagenet2012_corrupted imagenet_resized imagenette imagewang kmnist lfw lost_and_found lsun malaria mnist mnist_corrupted omniglot oxford_flowers102 oxford_iiit_pet patch_camelyon pet_finder places365_small plant_leaves plant_village plantae_k quickdraw_bitmap resisc45 rock_paper_scissors scene_parse150 shapes3d smallnorb so2sat stanford_dogs stanford_online_products sun397 svhn_cropped tf_flowers the300w_lp uc_merced vgg_face2 visual_domain_decathlon
Object Detection
coco kitti open_images_v4 voc wider_face
Structured
amazon_us_reviews forest_fires german_credit_numeric higgs iris rock_you titanic
Summarization
aeslc big_patent billsum cnn_dailymail gigaword multi_news newsroom opinosis reddit_tifu scientific_papers wikihow xsum
Text
blimp c4 cfq civil_comments cos_e definite_pronoun_resolution eraser_multi_rc esnli gap glue imdb_reviews librispeech_lm lm1b math_dataset movie_rationales multi_nli multi_nli_mismatch natural_questions qa4mre scan scicite snli squad super_glue tiny_shakespeare trivia_qa wikipedia xnli yelp_polarity_reviews
Translate
flores para_crawl ted_hrlr_translate ted_multi_translate wmt14_translate wmt15_translate wmt16_translate wmt17_translate wmt18_translate wmt19_translate wmt_t2t_translate
Video
bair_robot_pushing_small moving_mnist robonet starcraft_video ucf101

(3) CIFAR 100 데이터셋을 로컬 디스크에 다운로드 하고 로딩하기         (download & load CIFAR 100 dataset)

딥러닝을 활용한 이미지 분류 학습에 많이 사용되는 예제 데이터셋인 CIFAR 100 Dataset 을 로컬 디스크에 다운로드해보겠습니다. 

[ CIFAR 10 이미지 시각화 (예시) ]

cifar_builder = tfds.builder("cifar100") cifar_builder.download_and_prepare()

CIFAR 100 데이터셋은 ./tensorflow_datasets/cifar100/3.0.0. 폴더 밑에 다운로드되어 있습니다. train, test 데이터셋 레코드와 label 데이터, dataset과 image에 대한 JSON 데이터셋이 로컬 디스크에 다운로드 되었습니다. 

(4) 데이터셋의 속성 정보 조회하기 (Datasets Attributes)

cifiar_builder.info 로 데이터셋의 속성을 조회해보면 아래와 같습니다. 아래의 JSON 속성 정보를 참고해서 필요한 정보를 참조할 수 있습니다. 

print(cifar_builder.info) [Out]: tfds.core.DatasetInfo( name='cifar100', version=3.0.0, description='This dataset is just like the CIFAR-10, except it has 100 classes containing 600 images each. There are 500 training images and 100 testing images per class. The 100 classes in the CIFAR-100 are grouped into 20 superclasses. Each image comes with a "fine" label (the class to which it belongs) and a "coarse" label (the superclass to which it belongs).', homepage='https://www.cs.toronto.edu/~kriz/cifar.html', features=FeaturesDict({ 'coarse_label': ClassLabel(shape=(), dtype=tf.int64, num_classes=20), 'image': Image(shape=(32, 32, 3), dtype=tf.uint8), 'label': ClassLabel(shape=(), dtype=tf.int64, num_classes=100), }), total_num_examples=60000, splits={ 'test': 10000, 'train': 50000, }, supervised_keys=('image', 'label'), citation="""@TECHREPORT{Krizhevsky09learningmultiple, author = {Alex Krizhevsky}, title = {Learning multiple layers of features from tiny images}, institution = {}, year = {2009} }""", redistribution_info=, )

CIFAR 100 데이터셋의 100개 라벨을 인쇄해보겠습니다. 위의 JSON 정보를 참고해서 features["label"].names 로 속성값을 조회할 수 있습니다. 

# label naems print(cifar_builder.info.features["label"].names) [Out]: ['apple', 'aquarium_fish', 'baby', 'bear', 'beaver', 'bed', 'bee', 'beetle', 'bicycle', 'bottle', 'bowl', 'boy', 'bridge', 'bus', 'butterfly', 'camel', 'can', 'castle', 'caterpillar', 'cattle', 'chair', 'chimpanzee', 'clock', 'cloud', 'cockroach', 'couch', 'crab', 'crocodile', 'cup', 'dinosaur', 'dolphin', 'elephant', 'flatfish', 'forest', 'fox', 'girl', 'hamster', 'house', 'kangaroo', 'keyboard', 'lamp', 'lawn_mower', 'leopard', 'lion', 'lizard', 'lobster', 'man', 'maple_tree', 'motorcycle', 'mountain', 'mouse', 'mushroom', 'oak_tree', 'orange', 'orchid', 'otter', 'palm_tree', 'pear', 'pickup_truck', 'pine_tree', 'plain', 'plate', 'poppy', 'porcupine', 'possum', 'rabbit', 'raccoon', 'ray', 'road', 'rocket', 'rose', 'sea', 'seal', 'shark', 'shrew', 'skunk', 'skyscraper', 'snail', 'snake', 'spider', 'squirrel', 'streetcar', 'sunflower', 'sweet_pepper', 'table', 'tank', 'telephone', 'television', 'tiger', 'tractor', 'train', 'trout', 'tulip', 'turtle', 'wardrobe', 'whale', 'willow_tree', 'wolf', 'woman', 'worm'] print(cifar_builder.info.features["coarse_label"].names) [Out]: ['aquatic_mammals', 'fish', 'flowers', 'food_containers', 'fruit_and_vegetables', 'household_electrical_devices', 'household_furniture', 'insects', 'large_carnivores', 'large_man-made_outdoor_things', 'large_natural_outdoor_scenes', 'large_omnivores_and_herbivores', 'medium_mammals', 'non-insect_invertebrates', 'people', 'reptiles', 'small_mammals', 'trees', 'vehicles_1', 'vehicles_2']

(5) Train, Validation Set 분할하여 불러오기

# Train/ Validation Datasets train_cifar_dataset = cifar_builder.as_dataset(split=tfds.Split.TRAIN) val_cifar_dataset = cifar_builder.as_dataset(split=tfds.Split.TEST) print(train_cifar_dataset) [Out]: <DatasetV1Adapter shapes: {coarse_label: (), image: (32, 32, 3), label: ()}, types: {coarse_label: tf.int64, image: tf.uint8, label: tf.int64}> print(val_cifar_dataset) [Out]: <DatasetV1Adapter shapes: {coarse_label: (), image: (32, 32, 3), label: ()}, types: {coarse_label: tf.int64, image: tf.uint8, label: tf.int64}> # Number of classes: 100 num_classes = cifar_builder.info.features['label'].num_classes # Number of images: train 50,000 .  validation 10,000 num_train_imgs = cifar_builder.info.splits['train'].num_examples num_val_imgs = cifar_builder.info.splits['test'].num_examples print("Training dataset instance:", train_cifar_dataset) Training dataset instance: <DatasetV1Adapter shapes: {coarse_label: (), image: (32, 32, 3), label: ()}, types: {coarse_label: tf.int64, image: tf.uint8, label: tf.int64}>

(6) 데이터셋 전처리 (크기 조정, 증식, 배치 샘플링, 검증 데이터셋 생성)

* code reference: Hands-on Computer Vision with TensorFlow 2 by Eliot Andres & Benjamin Planche

(https://www.amazon.com/Hands-Computer-Vision-TensorFlow-processing-ebook/dp/B07SMQGX48)

import math input_shape = [224, 224, 3] batch_size = 32 num_epochs = 30 train_cifar_dataset = train_cifar_dataset.repeat(num_epochs).shuffle(10000)   def _prepare_data_fn(features, input_shape, augment=False):     """     Resize image to expected dimensions, and opt. apply some random transformations.     - param features: Data     - param input_shape: Shape expected by the models (images will be resized accordingly)     - param augment: Flag to apply some random augmentations to the images     - return: Augmented Images, Labels     """     input_shape = tf.convert_to_tensor(input_shape)          # Tensorflow-dataset returns batches as feature dictionaries, expected by Estimators.      # To train Keras models, it is mor straightforward to return the batch content as tuples.      image = features['image']     label = features['label']          # Convert the images to float type, also scaling their values from [0, 255] to [0., 1.]     image = tf.image.convert_image_dtype(image, tf.float32)          if augment:         # Randomly applied horizontal flip         image = tf.image.random_flip_left_right(image)                  # Random B/S changes         image = tf.image.random_brightness(image, max_delta=0.1)         image = tf.image.random_saturation(image, lower=0.5, upper=1.5)         image = tf.clip_by_value(image, 0.0, 1.0) # keeping pixel values in check                  # random resize and random crop back to expected size         random_scale_factor = tf.random.uniform([1], minval=1., maxval=1.4, dtype=tf.float32)         scaled_height = tf.cast(tf.cast(input_shape[0], tf.float32) * random_scale_factor, tf.int32)         scaled_width = tf.cast(tf.cast(input_shape[1], tf.float32) * random_scale_factor, tf.int32)                  scaled_shape = tf.squeeze(tf.stack([scaled_height, scaled_width]))         image = tf.image.resize(image, scaled_shape)         image = tf.image.random_crop(image, input_shape)     else:         image = tf.image.resize(image, input_shape[:2])              return image, label import functools prepare_data_fn_for_train = functools.partial(_prepare_data_fn,                                                input_shape=input_shape,                                               augment=True) train_cifar_dataset = train_cifar_dataset.map(prepare_data_fn_for_train, num_parallel_calls=4) # batch the samples train_cifar_dataset = train_cifar_dataset.batch(batch_size) train_cifar_dataset = train_cifar_dataset.prefetch(1) # validation dataset (not shuffling or augmenting it) prepare_data_fn_for_val = functools.partial(_prepare_data_fn,                                            input_shape=input_shape,                                            augment=False) val_cifar_dataset = (val_cifar_dataset                      .repeat()                     .map(prepare_data_fn_for_val, num_parallel_calls=4)                     .batch(batch_size)                     .prefetch(1)) train_steps_per_epoch = math.ceil(num_train_imgs / batch_size) val_steps_per_epoch = math.ceil(num_val_imgs / batch_size)

많은 도움이 되었기를 바랍니다. 

이번 포스팅이 도움이 되었다면 아래의 '공감~'를 꾹 눌러주세요. :-)