1. Data API
- tf.data 모듈안에 포함된 클래스 확인
- tf.data: https://www.tensorflow.org/api_docs/python/tf/data
Module: tf.data | TensorFlow v2.12.0
tf.data.Dataset API for input pipelines.
www.tensorflow.org
- tf.data.datasets
- tf.data.datasets
import tensorflow as tf
import tensorflow_datasets as tfds
# 데이터셋 확인
builders = tfds.list_builders()
print(builders)
# 출력 결과
['abstract_reasoning',
'accentdb',
'aeslc',
'aflw2k3d',
'ag_news_subset',
...
'yelp_polarity_reviews',
'yes_no',
'youtube_vis']# mnist 데이터 생성
data, info = tfds.load('mnist', with_info = True)
train_data, test_data = data['train'], data['test']
print(info)
# 출력 결과
tfds.core.DatasetInfo(
name='mnist',
full_name='mnist/3.0.1',
description="""
The MNIST database of handwritten digits.
""",
homepage='http://yann.lecun.com/exdb/mnist/',
data_path='C:\\Users\\YONG\\tensorflow_datasets\\mnist\\3.0.1',
file_format=tfrecord,
download_size=11.06 MiB,
dataset_size=21.00 MiB,
features=FeaturesDict({
'image': Image(shape=(28, 28, 1), dtype=uint8),
'label': ClassLabel(shape=(), dtype=int64, num_classes=10),
}),
supervised_keys=('image', 'label'),
disable_shuffling=False,
splits={
'test': <SplitInfo num_examples=10000, num_shards=1>,
'train': <SplitInfo num_examples=60000, num_shards=1>,
},
citation="""@article{lecun2010mnist,
title={MNIST handwritten digit database},
author={LeCun, Yann and Cortes, Corinna and Burges, CJ},
journal={ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist},
volume={2},
year={2010}
}""",
)
- tf.data
- 생성
- from_tensor_slices(): 개별 또는 다중 넘파이를 받고, 배치를 지원
- from_tensors(): 배치를 지원하지 않음
- froom_generator(): 생성자 하무에서 입력을 취함
- 변환
- batch(): 순차적으로 지정한 배치사이즈로 데이터셋을 분할
- repeat(): 데이터를 복제
- shuffle(): 데이터를 무작위로 섞음
- map(): 데이터에 함수를 적용
- filter(): 데이터를 거르고자 할 때 사용
- 반복
- next_batch = iterator.get_next() 사용
- from_tensor_slices
import numpy as np
num_items = 20
num_list = np.arange(num_items)
num_list_dataset = tf.data.Dataset.from_tensor_slices(num_list)
num_list_dataset
# 출력 결과
# shape은 아직 없는 상태
<TensorSliceDataset element_spec=TensorSpec(shape=(), dtype=tf.int32, name=None)>for item in num_list_dataset:
print(item)
# 출력 결과
# tensor 20개 생성
tf.Tensor(0, shape=(), dtype=int32)
tf.Tensor(1, shape=(), dtype=int32)
tf.Tensor(2, shape=(), dtype=int32)
tf.Tensor(3, shape=(), dtype=int32)
tf.Tensor(4, shape=(), dtype=int32)
tf.Tensor(5, shape=(), dtype=int32)
tf.Tensor(6, shape=(), dtype=int32)
tf.Tensor(7, shape=(), dtype=int32)
tf.Tensor(8, shape=(), dtype=int32)
tf.Tensor(9, shape=(), dtype=int32)
tf.Tensor(10, shape=(), dtype=int32)
tf.Tensor(11, shape=(), dtype=int32)
tf.Tensor(12, shape=(), dtype=int32)
tf.Tensor(13, shape=(), dtype=int32)
tf.Tensor(14, shape=(), dtype=int32)
tf.Tensor(15, shape=(), dtype=int32)
tf.Tensor(16, shape=(), dtype=int32)
tf.Tensor(17, shape=(), dtype=int32)
tf.Tensor(18, shape=(), dtype=int32)
tf.Tensor(19, shape=(), dtype=int32)
- from_generator()
- 해당 클래스 메서드를 사용하면 생성자에서 데이터셋 생성 가능
- output_types, output_shapes 인수로 출력 자료형과 크기를 지정해주어야 함
import itertools
# i는 1씩 증가하고 i의 개수만큼 1을 배열에 추가
def gen():
for i in itertools.count(1):
yield(i, [1] * i)
# 위에서 만든 generator
# 출력 형식은 int64
# 출력 형태는 TensorShape([])
dataset = tf.data.Dataset.from_generator(
gen,
(tf.int64, tf.int64),
(tf.TensorShape([]), tf.TensorShape([None]))
)
list(dataset.take(3).as_numpy_iterator())
# 출력 결과
[(1, array([1], dtype=int64)),
(2, array([1, 1], dtype=int64)),
(3, array([1, 1, 1], dtype=int64))]# stop이 없이 위의 코드와 같이 gen을 돌리면 무한히 돌아감
def gen(stop):
for i in itertools.count(1):
if i < stop:
yield(i, [1] * i)
dataset = tf.data.Dataset.from_generator(
gen, args = [10],
output_types = (tf.int64, tf.int64),
output_shapes = (tf.TensorShape([]), tf.TensorShape([None]))
)
list(dataset.take(5).as_numpy_iterator())
# 출력 결과
[(1, array([1], dtype=int64)),
(2, array([1, 1], dtype=int64)),
(3, array([1, 1, 1], dtype=int64)),
(4, array([1, 1, 1, 1], dtype=int64)),
(5, array([1, 1, 1, 1, 1], dtype=int64))]
- batch, repeat
- batch(): 배치 사이즈 크기
- repeat(): 반복 횟수
# 배치사이즈 7, 3번 반복
dataset = num_list_dataset.repeat(3).batch(7)for item in dataset:
print(item)
# 출력 결과
# 배치 사이즈가 7이므로 7개씩 나뉨
# 그렇게 3번 반복
tf.Tensor([0 1 2 3 4 5 6], shape=(7,), dtype=int32)
tf.Tensor([ 7 8 9 10 11 12 13], shape=(7,), dtype=int32)
tf.Tensor([14 15 16 17 18 19 0], shape=(7,), dtype=int32)
tf.Tensor([1 2 3 4 5 6 7], shape=(7,), dtype=int32)
tf.Tensor([ 8 9 10 11 12 13 14], shape=(7,), dtype=int32)
tf.Tensor([15 16 17 18 19 0 1], shape=(7,), dtype=int32)
tf.Tensor([2 3 4 5 6 7 8], shape=(7,), dtype=int32)
tf.Tensor([ 9 10 11 12 13 14 15], shape=(7,), dtype=int32)
tf.Tensor([16 17 18 19], shape=(4,), dtype=int32)# 뒤에 남는 수 없이 정확한 배치 사이즈로 나누어 떨어지도록 하고 싶으면
# drop_remainder = True 옵션 설정
dataset = num_list_dataset.repeat(3).batch(7, drop_remainder = True)
for item in dataset:
print(item)
# 출력 결과
# 마지막에 4개만 있던 데이터 사라짐
tf.Tensor([0 1 2 3 4 5 6], shape=(7,), dtype=int32)
tf.Tensor([ 7 8 9 10 11 12 13], shape=(7,), dtype=int32)
tf.Tensor([14 15 16 17 18 19 0], shape=(7,), dtype=int32)
tf.Tensor([1 2 3 4 5 6 7], shape=(7,), dtype=int32)
tf.Tensor([ 8 9 10 11 12 13 14], shape=(7,), dtype=int32)
tf.Tensor([15 16 17 18 19 0 1], shape=(7,), dtype=int32)
tf.Tensor([2 3 4 5 6 7 8], shape=(7,), dtype=int32)
tf.Tensor([ 9 10 11 12 13 14 15], shape=(7,), dtype=int32)
- map, filter
- 전처리 단계레서 시행하여 원하지 않는 데이터를 거를 수 있음
- tf.Tensor 자료형을 다룸
# map 함수 적용
from tensorflow.data import Dataset
# [1, 2, 3, 4, 5]의 리스트
dataset = Dataset.range(1, 6)
# 리스트 각 값에 2씩 곱하는 과정을 map 함수로 적용
dataset = dataset.map(lambda x: x * 2)
list(dataset.as_numpy_iterator())
# 출력 결과
[2, 4, 6, 8, 10]
# as_numpy_iterator()형태로 출력하지 않고 그대로 출력하는 경우
dataset = Dataset.range(5)
result = dataset.map(lambda x: x + 1)
result
# 출력 결과
<MapDataset element_spec=TensorSpec(shape=(), dtype=tf.int64, name=None)># map 함수를 사용해 원하는 데이터만 전처리하여 가져올 수 있음
elements = [(1, 'one'), (2, 'two'), (3, 'three')]
dataset = Dataset.from_generator(lambda: elements, (tf.int32, tf.string))
result = dataset.map(lambda x_int, y_str: x_int)
list(result.as_numpy_iterator())
# 출력 결과
[1, 2, 3]dataset = Dataset.range(3)
# 1. 기본적인 선언
def g(x):
return tf.constant(10.5), tf.constant(['One', 'Two', 'Three'])
result = dataset.map(g)
# 각 원소의 스펙 확인
result.element_spec
# 출력 결과
(TensorSpec(shape=(), dtype=tf.float32, name=None),
TensorSpec(shape=(3,), dtype=tf.string, name=None))
# 2. tf.constant로 텐서플로우 타입을 명시하지 않아도 기본적으로 적용됨
def h(x):
return 10.5, ['One', 'Two', 'Three'], np.array([1., 2.], dtype = np.float64)
result = dataset.map(h)
result.element_spec
# 출력 결과
(TensorSpec(shape=(), dtype=tf.float32, name=None),
TensorSpec(shape=(3,), dtype=tf.string, name=None),
TensorSpec(shape=(2,), dtype=tf.float64, name=None))
# 3. 내부에 데이터 리스트 형태 추가
def i(x):
return (10.5, [12.5, 11.1]), "One", "Two"
result = dataset.map(i)
result.element_spec
# 출력 결과
((TensorSpec(shape=(), dtype=tf.float32, name=None),
TensorSpec(shape=(2,), dtype=tf.float32, name=None)),
TensorSpec(shape=(), dtype=tf.string, name=None),
TensorSpec(shape=(), dtype=tf.string, name=None))# 1. 필터로 조건 지정
dataset = Dataset.from_tensor_slices([1, 2, 3])
dataset = dataset.filter(lambda x: x < 3)
list(dataset.as_numpy_iterator())
# 출력 결과
[1, 2]
# 2. 필터를 함수로 지정가능
# 1이랑 같은 것만 필터링
def filter_fn(x):
return tf.math.equal(x, 1)
dataset = dataset.filter(filter_fn)
list(dataset.as_numpy_iterator())
- shuffle, take
# 데이터 가져오기
dataset, info = tfds.load('imdb_reviews', with_info = True, as_supervised = True)
train_dataset = dataset['train']
# 5개로 구분하여 셔플하고, 2개를 가져오기
train_dataset = train_dataset.batch(5).shuffle(5).take(2)
for data in train_dataset:
print(data)
# 출력 결과(영화 리뷰 데이터를 5개 가져와 섞은 뒤 그 중 2개를 출력한 것)
(<tf.Tensor: shape=(5,), dtype=string, numpy=
array([b'It was disgusting and painful. What a waste of a cast! I swear, the audience (1/2 full) laughed TWICE in 90 minutes. This is not a lie. Do not even rent it.<br /><br />Zeta Jones was just too mean to be believable.<br /><br />Cusack was OK. Just OK. I felt sorry for him (the actor) in case people remember this mess.<br /><br />Roberts was the same as she always is. Charming and sweet, but with no purpose. The "romance" with John was completely unbelievable.',
b'This is a straight-to-video movie, so it should go without saying that it\'s not going to rival the first Lion King, but that said, this was downright good.<br /><br />My kids loved this, but that\'s a given, they love anything that\'s a cartoon. The big shock was that *I* liked it too, it was laugh out loud funny at some parts (even the fart jokes*), had lots of rather creative tie-ins with the first movie, and even some jokes that you had to be older to understand (but without being risqu\xc3\xa9 like in Shrek ["do you think he\'s compensating for something?"]).<br /><br />A special note on the fart jokes, I was surprised to find that none of the jokes were just toilet noises (in fact there were almost no noises/imagery at all, the references were actually rather subtle), they actually had a setup/punchline/etc, and were almost in good taste. I\'d like my kids to think that there\'s more to humor than going to the bathroom, and this movie is fine in those regards.<br /><br />Hmm what else? The music was so-so, not nearly as creative as in the first or second movie, but plenty of fun for the kids. No painfully corny moments, which was a blessing for me. A little action but nothing too scary (the Secret of NIMH gave my kids nightmares, not sure a G rating was appropriate for that one...)<br /><br />All in all I\'d say this is a great movie for kids of any age, one that\'s 100% safe to let them watch (I try not to be overly sensitive but I\'ve had to jump up and turn off the TV during a few movies that were less kid-appropriate than expected) - but you\'re safe to leave the room during this one. I\'d say stick around anyway though, you might find that you enjoy it too :)',
b'Finally, Timon and Pumbaa in their own film...<br /><br />\'The Lion King 1 1/2: Hakuna Matata\' is an irreverent new take on a classic tale. Which classic tale, you ask? Why, \'The Lion King\' of course!<br /><br />Yep, if there\'s one thing that Disney is never short of, it\'s narcissism.<br /><br />But that doesn\'t mean that this isn\'t a good film. It\'s basically the events of \'The Lion King\' as told from Timon and Pumbaa\'s perspective. And it\'s because of this that you\'ll have to know the story of \'The Lion King\' by heart to see where they\'re coming from.<br /><br />Anyway, at one level I was watching this and thinking "Oh my god this is so lame..." and on another level I was having a ball. Much of the humour is predictable - I mean, when Pumbaa makes up two beds, a big one for himself and a small one for Timon, within the first nanosecond we all know that Timon is going to take the big one. But that doesn\'t stop it from being hilarious, which, IMO, is \'Hakuna Matata\' in a nutshell. It\'s not what happens, it\'s how.<br /><br />And a note of warning: there are also some fart jokes. Seriously, did you expect anything else in a film where Pumbaa takes centre stage? But as fart jokes go, these are especially good, and should satisfy even the most particular connoisseur.<br /><br />The returning voice talent is great. I\'m kinda surprised that some of the actors were willing to return, what with most of them only having two or three lines (if they\'re lucky). Whoopi Goldberg is particularly welcome.<br /><br />The music is also great. From \'Digga Tunnah\' at the start to \'That\'s all I need\', an adaption of \'Warthog Rhapsody\' (a song that was cut from \'The Lion King\' and is frankly much improved in this incarnation), the music leaves me with nothing to complain about whatsoever.<br /><br />In the end, Timon and Pumbaa are awesome characters, and while it may be argued that \'Hakuna Matata\' is simply an excuse to see them in various fun and assorted compromising situations then so be it. It\'s rare to find characters that you just want to spend time with.<br /><br />Am I starting to sound creepy?<br /><br />Either way, \'The Lion King 1 1/2\' is great if you\'ve seen \'The Lion King\' far too many times. Especially if you are right now thinking "Don\'t be silly, there\'s no such thing as seeing \'The Lion King\' too many times!"',
b'Indian Directors have it tough, They have to compete with movies like "Laggan" where 11 henpecked,Castrated males defend their village and half of them are certifiable idiots. "Devdas", a hapless, fedar- festooned foreign return drinking to oblivion, with characters running in endless corridors oblivious to any one\'s feelings or sentiments-alas they live in an ornate squalor of red tapestry and pageantry. But to make a good movie, you have to tight-rope walk to appease the frontbenchers who are the quentessential gapers who are mesmerized with Split skirts and Dishum-Dishum fights preferably involving a nitwit "Bollywood" leading actor who is marginally handsome. So you can connect with a director who wants to tell a tale of Leonine village head who in own words "defending his Village" this is considered a violent movie or too masculine for a male audience. There are very few actors who can convey the anger and pathos like Nana Patekar (Narasimhan). Nana Patekar lets you in his courtyard and watch him beret and mock the Politician when his loyal admirers burst in laughter with every word of satire thrown at him, meanwhile his daughter is bathing his Grandson.This is as authentic a scene you can get in rural India. Nana Patekar is the essential actor who belongs to the old school of acting which is a disappearing breed in Hindi Films. The violence depicted is an intricate part of storytelling with Song&Dances thrown in for the gawkers without whom movies won\'t sell, a sad but true state of affairs. Faster this changes better for "Bollywood". All said and done this is one good Movie.',
b"Nathan Detroit runs illegal craps games for high rollers in NYC, but the heat is on and he can't find a secure location. He bets chronic gambler Sky Masterson that Sky can't make a prim missionary, Sarah Brown, go out to dinner with him. Sky takes up the challenge, but both men have some surprises in store \xc2\x85<br /><br />This is one of those expensive fifties MGM musicals in splashy colour, with big sets, loud music, larger-than-life roles and performances to match; Broadway photographed for the big screen if you like that sort of thing, which I don't. My main problem with these type of movies is simply the music. I like all kinds of music, from Albinoni to ZZ Top, but Broadway show tunes in swing time with never-ending pah-pah-tah-dah trumpet flourishes at the end of every fourth bar aren't my cup of tea. This was written by the tag team of Frank Loesser, Mankiewicz, Jo Swerling and Abe Burrows (based on a couple of Damon Runyon stories), and while the plot is quite affable the songs are weak. Blaine's two numbers for example are identical, unnecessary, don't advance the plot and grate on the ears (and are also flagrantly misogynistic if that sort of thing bothers you). There are only two memorable tunes, Luck Be A Lady (sung by Brando, not Sinatra as you might expect) and Sit Down, You're Rockin' The Boat (nicely performed by Kaye) but you have to sit through two hours to get to them. The movie's trump card is a young Brando giving a thoughtful, laid-back performance; he also sings quite well and even dances a little, and is evenly matched with the always interesting Simmons. The sequence where the two of them escape to Havana for the night is a welcome respite from all the noise, bustle and vowel-murdering of Noo Yawk. Fans of musicals may dig this, but in my view a musical has to do something more than just film the stage show."],
dtype=object)>, <tf.Tensor: shape=(5,), dtype=int64, numpy=array([0, 1, 1, 1, 0], dtype=int64)>)
(<tf.Tensor: shape=(5,), dtype=string, numpy=
array([b"This was an absolutely terrible movie. Don't be lured in by Christopher Walken or Michael Ironside. Both are great actors, but this must simply be their worst role in history. Even their great acting could not redeem this movie's ridiculous storyline. This movie is an early nineties US propaganda piece. The most pathetic scenes were those when the Columbian rebels were making their cases for revolutions. Maria Conchita Alonso appeared phony, and her pseudo-love affair with Walken was nothing but a pathetic emotional plug in a movie that was devoid of any real meaning. I am disappointed that there are movies like this, ruining actor's like Christopher Walken's good name. I could barely sit through it.",
b'I have been known to fall asleep during films, but this is usually due to a combination of things including, really tired, being warm and comfortable on the sette and having just eaten a lot. However on this occasion I fell asleep because the film was rubbish. The plot development was constant. Constantly slow and boring. Things seemed to happen, but with no explanation of what was causing them or why. I admit, I may have missed part of the film, but i watched the majority of it and everything just seemed to happen of its own accord without any real concern for anything else. I cant recommend this film at all.',
b'Mann photographs the Alberta Rocky Mountains in a superb fashion, and Jimmy Stewart and Walter Brennan give enjoyable performances as they always seem to do. <br /><br />But come on Hollywood - a Mountie telling the people of Dawson City, Yukon to elect themselves a marshal (yes a marshal!) and to enforce the law themselves, then gunfighters battling it out on the streets for control of the town? <br /><br />Nothing even remotely resembling that happened on the Canadian side of the border during the Klondike gold rush. Mr. Mann and company appear to have mistaken Dawson City for Deadwood, the Canadian North for the American Wild West.<br /><br />Canadian viewers be prepared for a Reefer Madness type of enjoyable howl with this ludicrous plot, or, to shake your head in disgust.',
b'This is the kind of film for a snowy Sunday afternoon when the rest of the world can go ahead with its own business as you descend into a big arm-chair and mellow for a couple of hours. Wonderful performances from Cher and Nicolas Cage (as always) gently row the plot along. There are no rapids to cross, no dangerous waters, just a warm and witty paddle through New York life at its best. A family film in every sense and one that deserves the praise it received.',
b'As others have mentioned, all the women that go nude in this film are mostly absolutely gorgeous. The plot very ably shows the hypocrisy of the female libido. When men are around they want to be pursued, but when no "men" are around, they become the pursuers of a 14 year old boy. And the boy becomes a man really fast (we should all be so lucky at this age!). He then gets up the courage to pursue his true love.'],
dtype=object)>, <tf.Tensor: shape=(5,), dtype=int64, numpy=array([0, 0, 0, 1, 1], dtype=int64)>)
- get_next()
dataset = Dataset.range(2)
for element in dataset:
print(element)
# 출력 결과
tf.Tensor(0, shape=(), dtype=int64)
tf.Tensor(1, shape=(), dtype=int64)dataset = Dataset.range(2)
iterator = iter(dataset)
print(dataset)
# 다음 데이터에 접근
print(iterator.get_next())
print(iterator.get_next())
# 출력 결과
<RangeDataset element_spec=TensorSpec(shape=(), dtype=tf.int64, name=None)>
tf.Tensor(0, shape=(), dtype=int64)
tf.Tensor(1, shape=(), dtype=int64)a = np.random.randint(0, 10, size = (2, 3))
print(a)
dataset = Dataset.from_tensor_slices(a)
iterator = iter(dataset)
print(iterator.get_next())
print(iterator.get_next())
# 출력 결과
# a의 원래 2행짜리 데이터에서 get_next()가 실행될 때마다 다음 행에 접근
[[0 7 2]
[6 1 4]]
tf.Tensor([0 7 2], shape=(3,), dtype=int32)
tf.Tensor([6 1 4], shape=(3,), dtype=int32)
2. tf.dataset을 이용한 Fashion-MNIST 분류
- modules import
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras.layers import Dense, Input, Flatten, Dropout, Activation, BatchNormalization
from tensorflow.keras.models import Model
from tensorflow.keras.datasets.fashion_mnist import load_data
- 데이터 로드
(x_train, y_train), (x_test, y_test) = load_data()
# 데이터 형태 확인
print(x_train.shape)
print(y_train.shape)
print(x_test.shape)
print(y_test.shape)
# 출력 결과
(60000, 28, 28)
(60000,)
(10000, 28, 28)
(10000,)
- 데이터 전처리
x_train = x_train / 255.
x_test = x_test / 255.
- tf.data 이용
train_ds = Dataset.from_tensor_slices((x_train, y_train))
train_ds = train_ds.shuffle(1000)
train_ds = train_ds.batch(32)
test_ds = Dataset.from_tensor_slices((x_test, y_test))
test_ds = test_ds.batch(32)
- 데이터 확인
class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',
'Sandal', 'Shirt', 'Sneakers', 'Bag', 'Ankle boot']
for image, label in train_ds.take(2):
plt.title("{}".format(class_names[label[0]]))
plt.imshow(image[0, :, :], cmap = 'gray')
plt.show()
- 모델 생성
- 임의의 모델
def build_model():
input = Input(shape = (28, 28), name = 'input')
flatten = Flatten(input_shape = [28, 28], name = 'flatten')(input)
hidden1 = Dense(256, kernel_initializer = 'he_normal', name = 'hidden1')(flatten)
hidden1 = BatchNormalization()(hidden1)
hidden1 = Activation('relu')(hidden1)
dropout1 = Dropout(0.5)(hidden1)
hidden2 = Dense(100, kernel_initializer = 'he_normal', name = 'hidden2')(dropout1)
hidden2 = BatchNormalization()(hidden2)
hidden2 = Activation('relu')(hidden2)
dropout2 = Dropout(0.5)(hidden2)
hidden3 = Dense(100, kernel_initializer = 'he_normal', name = 'hidden3')(dropout2)
hidden3 = BatchNormalization()(hidden3)
hidden3 = Activation('relu')(hidden3)
dropout3 = Dropout(0.5)(hidden3)
hidden4 = Dense(50, kernel_initializer = 'he_normal', name = 'hidden4')(dropout3)
hidden4 = BatchNormalization()(hidden4)
hidden4 = Activation('relu')(hidden4)
dropout4 = Dropout(0.5)(hidden4)
output = Dense(10, activation = 'softmax', name = 'output')(dropout4)
model = Model(inputs = [input], outputs = [output])
return modelmodel = build_model()
model.summary()
# 출력 결과
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 28, 28)] 0
flatten (Flatten) (None, 784) 0
hidden1 (Dense) (None, 256) 200960
batch_normalization (BatchN (None, 256) 1024
ormalization)
activation (Activation) (None, 256) 0
dropout (Dropout) (None, 256) 0
hidden2 (Dense) (None, 100) 25700
batch_normalization_1 (Batc (None, 100) 400
hNormalization)
activation_1 (Activation) (None, 100) 0
dropout_1 (Dropout) (None, 100) 0
hidden3 (Dense) (None, 100) 10100
batch_normalization_2 (Batc (None, 100) 400
hNormalization)
activation_2 (Activation) (None, 100) 0
dropout_2 (Dropout) (None, 100) 0
hidden4 (Dense) (None, 50) 5050
batch_normalization_3 (Batc (None, 50) 200
hNormalization)
activation_3 (Activation) (None, 50) 0
dropout_3 (Dropout) (None, 50) 0
output (Dense) (None, 10) 510
=================================================================
Total params: 244,344
Trainable params: 243,332
Non-trainable params: 1,012
_________________________________________________________________
- 모델 컴파일
- 평가(metrics)방식의 다른 방법
- tf.keras.metrics.Mean
- tf.keras.metrics.SparseCategoricalAccuracy
- 위 두 방식을 이용하여 loss 값을 좀 더 smooth하게 만들기(평균을 내는 방식)
loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
optimizer = tf.keras.optimizers.Adam()
train_loss = tf.keras.metrics.Mean(name = 'train_loss')
train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name = 'train_accuracy')
test_loss = tf.keras.metrics.Mean(name = 'test_loss')
test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name = 'test_accuracy')
- 모델 학습
- tf.function으로 인해 학습이 시작되면 그래프를 생성하여 속도가 빠름
# tf.function 사용시 오토 그래프 생성으로 성능 향상
@tf.function
def train_step(images, labels):
with tf.GradientTape() as tape:
predictions = model(images)
loss = loss_object(labels, predictions)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
train_loss(loss)
train_accuracy(labels, predictions)
@tf.function
def test_step(images, labels):
predictions = model(images)
t_loss = loss_object(labels, predictions)
test_loss(t_loss)
test_accuracy(labels, predictions)
epochs = 20
for epoch in range(epochs):
for images, labels in train_ds:
train_step(images, labels)
for test_images, test_labels in test_ds:
test_step(test_images, test_labels)
template = "Epochs: {:3d}\tLoss: {:.4f}\tAccuracy: {:.4f}\tTest Loss: {:.4f}\tTest Accuracy: {:.4f}"
print(template.format(epoch + 1,
train_loss.result(),
train_accuracy.result() * 100,
test_loss.result(),
test_accuracy.result() * 100))
# 출력 결과
Epochs: 1 Loss: 0.3975 Accuracy: 85.4906 Test Loss: 0.3896 Test Accuracy: 85.6400
Epochs: 2 Loss: 0.3756 Accuracy: 86.2650 Test Loss: 0.3840 Test Accuracy: 85.9050
Epochs: 3 Loss: 0.3586 Accuracy: 86.8523 Test Loss: 0.3768 Test Accuracy: 86.2340
Epochs: 4 Loss: 0.3450 Accuracy: 87.3364 Test Loss: 0.3706 Test Accuracy: 86.4583
Epochs: 5 Loss: 0.3333 Accuracy: 87.7414 Test Loss: 0.3684 Test Accuracy: 86.6014
Epochs: 6 Loss: 0.3232 Accuracy: 88.0877 Test Loss: 0.3648 Test Accuracy: 86.7925
Epochs: 7 Loss: 0.3144 Accuracy: 88.3983 Test Loss: 0.3639 Test Accuracy: 86.8289
Epochs: 8 Loss: 0.3066 Accuracy: 88.6765 Test Loss: 0.3618 Test Accuracy: 87.0010
Epochs: 9 Loss: 0.2994 Accuracy: 88.9215 Test Loss: 0.3595 Test Accuracy: 87.1400
Epochs: 10 Loss: 0.2927 Accuracy: 89.1588 Test Loss: 0.3595 Test Accuracy: 87.1833
Epochs: 11 Loss: 0.2864 Accuracy: 89.3894 Test Loss: 0.3573 Test Accuracy: 87.3015
Epochs: 12 Loss: 0.2808 Accuracy: 89.5865 Test Loss: 0.3570 Test Accuracy: 87.3336
Epochs: 13 Loss: 0.2753 Accuracy: 89.7777 Test Loss: 0.3583 Test Accuracy: 87.4113
Epochs: 14 Loss: 0.2703 Accuracy: 89.9568 Test Loss: 0.3577 Test Accuracy: 87.4900
Epochs: 15 Loss: 0.2654 Accuracy: 90.1251 Test Loss: 0.3583 Test Accuracy: 87.5524
Epochs: 16 Loss: 0.2609 Accuracy: 90.2880 Test Loss: 0.3615 Test Accuracy: 87.5750
Epochs: 17 Loss: 0.2565 Accuracy: 90.4376 Test Loss: 0.3626 Test Accuracy: 87.6426
Epochs: 18 Loss: 0.2525 Accuracy: 90.5751 Test Loss: 0.3634 Test Accuracy: 87.6910
Epochs: 19 Loss: 0.2484 Accuracy: 90.7171 Test Loss: 0.3651 Test Accuracy: 87.7324
Epochs: 20 Loss: 0.2446 Accuracy: 90.8512 Test Loss: 0.3667 Test Accuracy: 87.7555
- 모델 학습: 2번째 방법(Keras)
from sklearn.model_selection import train_test_split
(x_train_full, y_train_full), (x_test, y_test) = load_data()
x_train, x_val, y_train, y_val = train_test_split(x_train_full, y_train_full, test_size = 0.3, random_state = 777)
x_train = x_train / 255.
x_val = x_val / 255.
x_test = x_test / 255.
print(x_train.shape)
print(y_train.shape)
print(x_val.shape)
print(y_val.shape)
print(x_test.shape)
print(y_test.shape)
model = build_model()
model.compile(optimizer = 'sgd',
loss = 'sparse_categorical_crossentropy',
metrics = ['accuracy'])
model.summary()
# 출력 결과
(42000, 28, 28)
(42000,)
(18000, 28, 28)
(18000,)
(10000, 28, 28)
(10000,)
Model: "model_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 28, 28)] 0
flatten (Flatten) (None, 784) 0
hidden1 (Dense) (None, 256) 200960
batch_normalization_4 (Batc (None, 256) 1024
hNormalization)
activation_4 (Activation) (None, 256) 0
dropout_4 (Dropout) (None, 256) 0
hidden2 (Dense) (None, 100) 25700
batch_normalization_5 (Batc (None, 100) 400
hNormalization)
activation_5 (Activation) (None, 100) 0
dropout_5 (Dropout) (None, 100) 0
hidden3 (Dense) (None, 100) 10100
batch_normalization_6 (Batc (None, 100) 400
hNormalization)
activation_6 (Activation) (None, 100) 0
dropout_6 (Dropout) (None, 100) 0
hidden4 (Dense) (None, 50) 5050
batch_normalization_7 (Batc (None, 50) 200
hNormalization)
activation_7 (Activation) (None, 50) 0
dropout_7 (Dropout) (None, 50) 0
output (Dense) (None, 10) 510
=================================================================
Total params: 244,344
Trainable params: 243,332
Non-trainable params: 1,012
_________________________________________________________________from tensorflow.keras.callbacks import EarlyStopping
early_stopping_cb = EarlyStopping(patience = 3, monitor = 'val_loss',
restore_best_weights = True)
history = model.fit(x_train, y_train,
batch_size = 256,
epochs = 200,
shuffle = True,
validation_data = (x_val, y_val),
callbacks = [early_stopping_cb])
- 모델 평가
model.evaluate(x_test, y_test, batch_size = 100)
# 출력 결과
loss: 0.4427 - accuracy: 0.8464
[0.44270941615104675, 0.8464000225067139]
- 결과 확인
# 첫번째 테스트 데이터 결과
test_img = x_test[0, :, :]
plt.title(class_names[y_test[0]])
plt.imshow(test_img, cmap = 'gray')
plt.show()
pred = model.predict(test_img.reshape(1, 28, 28))
pred.shape
# 출력 결과
(1, 10)
pred
# 출력 결과
array([[8.9198991e-05, 3.5745958e-05, 7.4570953e-06, 1.5882608e-05,
8.0741156e-06, 3.3398017e-02, 4.0778108e-05, 1.1560775e-01,
7.1698561e-04, 8.5008013e-01]], dtype=float32)
# 가장 확률이 높은 것을 정답으로 출력
class_names[np.argmax(pred)]
# 출력 결과
'Ankle boot'
- Test Batch Dataset
test_batch = x_test[:32, :, :]
test_batch_y = y_test[:32]
print(test_batch.shape)
# 출력 결과
(32, 28, 28)preds = model.predict(test_batch)
preds.shape
# 출력 결과
(32, 10)pred_arg = np.argmax(preds, -1)
num_rows = 8
num_cols = 4
num_images = num_rows * num_cols
plt.figure(figsize = (16, 10))
for idx in range(1, 33, 1):
plt.subplot(num_rows, num_cols, idx)
plt.title('Predicted: {}, True: {}'.format(class_names[pred_arg[idx - 1]],
class_names[test_batch_y[idx - 1]]))
plt.imshow(test_batch[idx - 1], cmap = 'gray')
plt.show()
'Python > Deep Learning' 카테고리의 다른 글
| [딥러닝-케라스] 케라스 Fashion MNIST CNN 모델 (0) | 2023.05.14 |
|---|---|
| [딥러닝-케라스] 케라스 컨볼루션 신경망 (0) | 2023.05.13 |
| [딥러닝-케라스] 케라스의 다양한 학습 기술 (0) | 2023.05.11 |
| [딥러닝-케라스] 케라스 Fashion MNIST 모델 (0) | 2023.05.08 |
| [딥러닝-케라스] 케라스 자동차 연비 예측 모델 (0) | 2023.04.28 |