|TensorFlow.org에서 보기||Google Colab에서 실행||GitHub에서 소스 보기||노트북 다운로드||TF Hub 모델 보기|
이 노트북은 리뷰의 텍스트를 사용하여 영화 리뷰를 긍정적 또는 부정적으로 분류합니다. 이진( 또는 2-클래스 분류인 이 예는 광범위하게 적용할 수 있는 중요한 머신러닝 응용 사례입니다.
이 튜토리얼은 TensorFlow Hub 및 Keras를 사용한 전이 학습의 기본적인 응용을 보여줍니다.
pip install tensorflow-hub
pip install tensorflow-datasets
import os import numpy as np import tensorflow as tf import tensorflow_hub as hub import tensorflow_datasets as tfds print("Version: ", tf.__version__) print("Eager mode: ", tf.executing_eagerly()) print("Hub version: ", hub.__version__) print("GPU is", "available" if tf.config.list_physical_devices("GPU") else "NOT AVAILABLE")
Version: 2.6.0 Eager mode: True Hub version: 0.12.0 GPU is available
IMDB 데이터셋 다운로드
# Split the training set into 60% and 40% to end up with 15,000 examples # for training, 10,000 examples for validation and 25,000 examples for testing. train_data, validation_data, test_data = tfds.load( name="imdb_reviews", split=('train[:60%]', 'train[60%:]', 'test'), as_supervised=True)
잠시 데이터 형태를 알아 보죠. 이 데이터셋의 샘플은 전처리된 정수 배열입니다. 이 정수는 영화 리뷰에 나오는 단어를 나타냅니다. 레이블(label)은 정수 0 또는 1입니다. 0은 부정적인 리뷰이고 1은 긍정적인 리뷰입니다.
처음 10개의 샘플을 출력해 보겠습니다.
train_examples_batch, train_labels_batch = next(iter(train_data.batch(10))) train_examples_batch
<tf.Tensor: shape=(10,), 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.', b"This is a film which should be seen by anybody interested in, effected by, or suffering from an eating disorder. It is an amazingly accurate and sensitive portrayal of bulimia in a teenage girl, its causes and its symptoms. The girl is played by one of the most brilliant young actresses working in cinema today, Alison Lohman, who was later so spectacular in 'Where the Truth Lies'. I would recommend that this film be shown in all schools, as you will never see a better on this subject. Alison Lohman is absolutely outstanding, and one marvels at her ability to convey the anguish of a girl suffering from this compulsive disorder. If barometers tell us the air pressure, Alison Lohman tells us the emotional pressure with the same degree of accuracy. Her emotional range is so precise, each scene could be measured microscopically for its gradations of trauma, on a scale of rising hysteria and desperation which reaches unbearable intensity. Mare Winningham is the perfect choice to play her mother, and does so with immense sympathy and a range of emotions just as finely tuned as Lohman's. Together, they make a pair of sensitive emotional oscillators vibrating in resonance with one another. This film is really an astonishing achievement, and director Katt Shea should be proud of it. The only reason for not seeing it is if you are not interested in people. But even if you like nature films best, this is after all animal behaviour at the sharp edge. Bulimia is an extreme version of how a tormented soul can destroy her own body in a frenzy of despair. And if we don't sympathise with people suffering from the depths of despair, then we are dead inside.", b'Okay, you have:<br /><br />Penelope Keith as Miss Herringbone-Tweed, B.B.E. (Backbone of England.) She\'s killed off in the first scene - that\'s right, folks; this show has no backbone!<br /><br />Peter O\'Toole as Ol\' Colonel Cricket from The First War and now the emblazered Lord of the Manor.<br /><br />Joanna Lumley as the ensweatered Lady of the Manor, 20 years younger than the colonel and 20 years past her own prime but still glamourous (Brit spelling, not mine) enough to have a toy-boy on the side. It\'s alright, they have Col. Cricket\'s full knowledge and consent (they guy even comes \'round for Christmas!) Still, she\'s considerate of the colonel enough to have said toy-boy her own age (what a gal!)<br /><br />David McCallum as said toy-boy, equally as pointlessly glamourous as his squeeze. Pilcher couldn\'t come up with any cover for him within the story, so she gave him a hush-hush job at the Circus.<br /><br />and finally:<br /><br />Susan Hampshire as Miss Polonia Teacups, Venerable Headmistress of the Venerable Girls\' Boarding-School, serving tea in her office with a dash of deep, poignant advice for life in the outside world just before graduation. Her best bit of advice: "I\'ve only been to Nancherrow (the local Stately Home of England) once. I thought it was very beautiful but, somehow, not part of the real world." Well, we can\'t say they didn\'t warn us.<br /><br />Ah, Susan - time was, your character would have been running the whole show. They don\'t write \'em like that any more. Our loss, not yours.<br /><br />So - with a cast and setting like this, you have the re-makings of "Brideshead Revisited," right?<br /><br />Wrong! They took these 1-dimensional supporting roles because they paid so well. After all, acting is one of the oldest temp-jobs there is (YOU name another!)<br /><br />First warning sign: lots and lots of backlighting. They get around it by shooting outdoors - "hey, it\'s just the sunlight!"<br /><br />Second warning sign: Leading Lady cries a lot. When not crying, her eyes are moist. That\'s the law of romance novels: Leading Lady is "dewy-eyed."<br /><br />Henceforth, Leading Lady shall be known as L.L.<br /><br />Third warning sign: L.L. actually has stars in her eyes when she\'s in love. Still, I\'ll give Emily Mortimer an award just for having to act with that spotlight in her eyes (I wonder . did they use contacts?)<br /><br />And lastly, fourth warning sign: no on-screen female character is "Mrs." She\'s either "Miss" or "Lady."<br /><br />When all was said and done, I still couldn\'t tell you who was pursuing whom and why. I couldn\'t even tell you what was said and done.<br /><br />To sum up: they all live through World War II without anything happening to them at all.<br /><br />OK, at the end, L.L. finds she\'s lost her parents to the Japanese prison camps and baby sis comes home catatonic. Meanwhile (there\'s always a "meanwhile,") some young guy L.L. had a crush on (when, I don\'t know) comes home from some wartime tough spot and is found living on the street by Lady of the Manor (must be some street if SHE\'s going to find him there.) Both war casualties are whisked away to recover at Nancherrow (SOMEBODY has to be "whisked away" SOMEWHERE in these romance stories!)<br /><br />Great drama.', b'The film is based on a genuine 1950s novel.<br /><br />Journalist Colin McInnes wrote a set of three "London novels": "Absolute Beginners", "City of Spades" and "Mr Love and Justice". I have read all three. The first two are excellent. The last, perhaps an experiment that did not come off. But McInnes\'s work is highly acclaimed; and rightly so. This musical is the novelist\'s ultimate nightmare - to see the fruits of one\'s mind being turned into a glitzy, badly-acted, soporific one-dimensional apology of a film that says it captures the spirit of 1950s London, and does nothing of the sort.<br /><br />Thank goodness Colin McInnes wasn\'t alive to witness it.', b'I really love the sexy action and sci-fi films of the sixties and its because of the actress\'s that appeared in them. They found the sexiest women to be in these films and it didn\'t matter if they could act (Remember "Candy"?). The reason I was disappointed by this film was because it wasn\'t nostalgic enough. The story here has a European sci-fi film called "Dragonfly" being made and the director is fired. So the producers decide to let a young aspiring filmmaker (Jeremy Davies) to complete the picture. They\'re is one real beautiful woman in the film who plays Dragonfly but she\'s barely in it. Film is written and directed by Roman Coppola who uses some of his fathers exploits from his early days and puts it into the script. I wish the film could have been an homage to those early films. They could have lots of cameos by actors who appeared in them. There is one actor in this film who was popular from the sixties and its John Phillip Law (Barbarella). Gerard Depardieu, Giancarlo Giannini and Dean Stockwell appear as well. I guess I\'m going to have to continue waiting for a director to make a good homage to the films of the sixties. If any are reading this, "Make it as sexy as you can"! I\'ll be waiting!', b'Sure, this one isn\'t really a blockbuster, nor does it target such a position. "Dieter" is the first name of a quite popular German musician, who is either loved or hated for his kind of acting and thats exactly what this movie is about. It is based on the autobiography "Dieter Bohlen" wrote a few years ago but isn\'t meant to be accurate on that. The movie is filled with some sexual offensive content (at least for American standard) which is either amusing (not for the other "actors" of course) or dumb - it depends on your individual kind of humor or on you being a "Bohlen"-Fan or not. Technically speaking there isn\'t much to criticize. Speaking of me I find this movie to be an OK-movie.'], dtype=object)>
처음 10개의 레이블도 출력해 보겠습니다.
<tf.Tensor: shape=(10,), dtype=int64, numpy=array([0, 0, 0, 1, 1, 1, 0, 0, 0, 0])>
신경망은 층을 쌓아서 만듭니다. 여기에는 세 개의 중요한 구조적 결정이 필요합니다:
- 어떻게 텍스트를 표현할 것인가?
- 모델에서 얼마나 많은 층을 사용할 것인가?
- 각 층에서 얼마나 많은 은닉 유닛(hidden unit)을 사용할 것인가?
이 예제의 입력 데이터는 문장으로 구성됩니다. 예측할 레이블은 0 또는 1입니다.
텍스트를 표현하는 한 가지 방법은 문장을 임베딩 벡터로 변환하는 것입니다. 사전 훈련 된 텍스트 임베딩을 첫 번째 레이어로 사용할 수 있으며, 두 가지 이점이 있습니다.
- 텍스트 전처리에 대해 걱정할 필요가 없습니다.
- 전이 학습에 따른 이점이 있습니다.
- 임베딩은 고정 크기이기 때문에 처리 과정이 단순해집니다.
이 튜토리얼에서 사용할 수 있는 TFHub의 다른 많은 사전 훈련된 텍스트 임베딩이 있습니다.
- google/nnlm-en-dim128/2 - google/nnlm-en-dim50/2와 동일한 데이터에 동일한 NNLM 아키텍처로 훈련하지만 임베딩 차원이 더 큽니다. 보다 큰 차원의 임베딩은 작업을 개선할 수 있지만 모델을 훈련하는 데 더 오래 걸릴 수 있습니다.
- google/nnlm-en-dim128-with-normalization/2 - google/nnlm-en-dim128/2와 동일하지만 구두점 제거와 같은 추가적인 텍스트 정규화가 있습니다. 이는 작업의 텍스트에 추가 문자나 구두점이 포함된 경우 도움이 될 수 있습니다.
- google/universal-sentence-encoder/4 - DAN(deep averaging network) 인코더로 훈련된 512 차원 임베딩을 생성하는 훨씬 더 큰 모델입니다.
그 밖에도 많이 있습니다! TFHub에서 더 많은 텍스트 임베딩 모델을 찾아보세요.
먼저 문장을 임베딩시키기 위해 텐서플로 허브 모델을 사용하는 케라스 층을 만들어 보죠. 그다음 몇 개의 샘플을 입력하여 테스트해 보겠습니다. 입력 텍스트의 길이에 상관없이 임베딩의 출력 크기는
(num_examples, embedding_dimension)가 됩니다.
embedding = "https://tfhub.dev/google/nnlm-en-dim50/2" hub_layer = hub.KerasLayer(embedding, input_shape=, dtype=tf.string, trainable=True) hub_layer(train_examples_batch[:3])
<tf.Tensor: shape=(3, 50), dtype=float32, numpy= array([[ 0.5423194 , -0.01190171, 0.06337537, 0.0686297 , -0.16776839, -0.10581177, 0.168653 , -0.04998823, -0.31148052, 0.07910344, 0.15442258, 0.01488661, 0.03930155, 0.19772716, -0.12215477, -0.04120982, -0.27041087, -0.21922147, 0.26517656, -0.80739075, 0.25833526, -0.31004202, 0.2868321 , 0.19433866, -0.29036498, 0.0386285 , -0.78444123, -0.04793238, 0.41102988, -0.36388886, -0.58034706, 0.30269453, 0.36308962, -0.15227163, -0.4439151 , 0.19462997, 0.19528405, 0.05666233, 0.2890704 , -0.28468323, -0.00531206, 0.0571938 , -0.3201319 , -0.04418665, -0.08550781, -0.55847436, -0.2333639 , -0.20782956, -0.03543065, -0.17533456], [ 0.56338924, -0.12339553, -0.10862677, 0.7753425 , -0.07667087, -0.15752274, 0.01872334, -0.08169781, -0.3521876 , 0.46373403, -0.08492758, 0.07166861, -0.00670818, 0.12686071, -0.19326551, -0.5262643 , -0.32958236, 0.14394784, 0.09043556, -0.54175544, 0.02468163, -0.15456744, 0.68333143, 0.09068333, -0.45327246, 0.23180094, -0.8615696 , 0.3448039 , 0.12838459, -0.58759046, -0.40712303, 0.23061076, 0.48426905, -0.2712814 , -0.5380918 , 0.47016335, 0.2257274 , -0.00830665, 0.28462422, -0.30498496, 0.04400366, 0.25025868, 0.14867125, 0.4071703 , -0.15422425, -0.06878027, -0.40825695, -0.31492147, 0.09283663, -0.20183429], [ 0.7456156 , 0.21256858, 0.1440033 , 0.52338624, 0.11032254, 0.00902788, -0.36678016, -0.08938274, -0.24165548, 0.33384597, -0.111946 , -0.01460045, -0.00716449, 0.19562715, 0.00685217, -0.24886714, -0.42796353, 0.1862 , -0.05241097, -0.664625 , 0.13449019, -0.22205493, 0.08633009, 0.43685383, 0.2972681 , 0.36140728, -0.71968895, 0.05291242, -0.1431612 , -0.15733941, -0.15056324, -0.05988007, -0.08178931, -0.15569413, -0.09303784, -0.18971168, 0.0762079 , -0.02541647, -0.27134502, -0.3392682 , -0.10296471, -0.27275252, -0.34078008, 0.20083308, -0.26644838, 0.00655449, -0.05141485, -0.04261916, -0.4541363 , 0.20023566]], dtype=float32)>
이제 전체 모델을 만들어 보겠습니다:
model = tf.keras.Sequential() model.add(hub_layer) model.add(tf.keras.layers.Dense(16, activation='relu')) model.add(tf.keras.layers.Dense(1)) model.summary()
Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= keras_layer (KerasLayer) (None, 50) 48190600 _________________________________________________________________ dense (Dense) (None, 16) 816 _________________________________________________________________ dense_1 (Dense) (None, 1) 17 ================================================================= Total params: 48,191,433 Trainable params: 48,191,433 Non-trainable params: 0 _________________________________________________________________
순서대로 층을 쌓아 분류기를 만듭니다:
- 첫 번째 레이어는 TensorFlow Hub 레이어입니다. 이 레이어는 사전 훈련된 저장된 모델을 사용하여 문장을 임베딩 벡터에 매핑합니다. 사용 중인 사전 훈련된 텍스트 임베딩 모델(google/nnlm-en-dim50/2)은 문장을 토큰으로 분할하고 각 토큰을 임베딩한 다음 임베딩을 결합합니다. 결과적인 차원은
(num_examples, embedding_dimension)입니다. 이 NNLM 모델의 경우에는
- 이 고정 크기의 출력 벡터는 16개의 은닉 유닛(hidden unit)을 가진 완전 연결 층(
- 마지막 층은 하나의 출력 노드를 가진 완전 연결 층입니다.
sigmoid활성화 함수를 사용하므로 확률 또는 신뢰도 수준을 표현하는 0~1 사이의 실수가 출력됩니다.
이제 모델을 컴파일합니다.
손실 함수와 옵티마이저
모델에는 훈련을 위한 손실 함수와 옵티마이저가 필요합니다. 이진 분류 문제이고 모델이 로짓(선형 활성화가 있는 단일 단위 레이어)을 출력하므로
binary_crossentropy 손실 함수를 사용합니다.
다른 손실 함수를 선택할 수 없는 것은 아닙니다. 예를 들어
mean_squared_error를 선택할 수 있습니다. 하지만 일반적으로
binary_crossentropy가 확률을 다루는데 적합합니다. 이 함수는 확률 분포 간의 거리를 측정합니다. 여기에서는 정답인 타깃 분포와 예측 분포 사이의 거리입니다.
나중에 회귀 문제(예: 주택 가격 예측)를 살펴볼 때 평균 제곱 오차라고 하는 또 다른 손실 함수를 사용하는 방법을 살펴볼 것입니다.
이제 모델이 사용할 옵티마이저와 손실 함수를 설정해 보죠:
model.compile(optimizer='adam', loss=tf.keras.losses.BinaryCrossentropy(from_logits=True), metrics=['accuracy'])
512개 샘플의 미니 배치에서 10개 epoch 동안 모델을 훈련합니다. 이 동작은
y_train 텐서의 모든 샘플에 대한 10회 반복에 해당합니다. 훈련하는 동안 검증 세트의 10,000개 샘플에서 모델의 손실과 정확도를 모니터링합니다.
history = model.fit(train_data.shuffle(10000).batch(512), epochs=10, validation_data=validation_data.batch(512), verbose=1)
Epoch 1/10 30/30 [==============================] - 3s 53ms/step - loss: 0.6830 - accuracy: 0.5123 - val_loss: 0.6275 - val_accuracy: 0.5114 Epoch 2/10 30/30 [==============================] - 2s 47ms/step - loss: 0.5770 - accuracy: 0.6149 - val_loss: 0.5367 - val_accuracy: 0.6947 Epoch 3/10 30/30 [==============================] - 2s 48ms/step - loss: 0.4538 - accuracy: 0.7857 - val_loss: 0.4304 - val_accuracy: 0.7879 Epoch 4/10 30/30 [==============================] - 2s 45ms/step - loss: 0.3312 - accuracy: 0.8697 - val_loss: 0.3601 - val_accuracy: 0.8357 Epoch 5/10 30/30 [==============================] - 2s 45ms/step - loss: 0.2413 - accuracy: 0.9124 - val_loss: 0.3236 - val_accuracy: 0.8616 Epoch 6/10 30/30 [==============================] - 2s 49ms/step - loss: 0.1762 - accuracy: 0.9423 - val_loss: 0.3093 - val_accuracy: 0.8726 Epoch 7/10 30/30 [==============================] - 2s 45ms/step - loss: 0.1286 - accuracy: 0.9609 - val_loss: 0.3038 - val_accuracy: 0.8699 Epoch 8/10 30/30 [==============================] - 2s 45ms/step - loss: 0.0936 - accuracy: 0.9758 - val_loss: 0.3074 - val_accuracy: 0.8728 Epoch 9/10 30/30 [==============================] - 2s 46ms/step - loss: 0.0679 - accuracy: 0.9853 - val_loss: 0.3153 - val_accuracy: 0.8710 Epoch 10/10 30/30 [==============================] - 2s 46ms/step - loss: 0.0491 - accuracy: 0.9913 - val_loss: 0.3260 - val_accuracy: 0.8731
모델의 성능을 확인해 보죠. 두 개의 값이 반환됩니다. 손실(오차를 나타내는 숫자이므로 낮을수록 좋습니다)과 정확도입니다.
results = model.evaluate(test_data.batch(512), verbose=2) for name, value in zip(model.metrics_names, results): print("%s: %.3f" % (name, value))
49/49 - 3s - loss: 0.3497 - accuracy: 0.8561 loss: 0.350 accuracy: 0.856
이 예제는 매우 단순한 방식을 사용하므로 87% 정도의 정확도를 달성했습니다. 고급 방법을 사용한 모델은 95%에 가까운 정확도를 얻습니다.
- 문자열 입력으로 작업하는 보다 일반적인 방법과 훈련 중 정확도 및 손실의 진행 상황에 대한 보다 자세한 분석은 전처리된 텍스트를 사용한 텍스트 분류 튜토리얼을 참조하세요.
- TFHub에서 훈련된 모델을 사용하여 더 많은 텍스트 관련 튜토리얼을 시도해 보세요.
# MIT License # # Copyright (c) 2017 François Chollet # # Permission is hereby granted, free of charge, to any person obtaining a # copy of this software and associated documentation files (the "Software"), # to deal in the Software without restriction, including without limitation # the rights to use, copy, modify, merge, publish, distribute, sublicense, # and/or sell copies of the Software, and to permit persons to whom the # Software is furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL # THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER # DEALINGS IN THE SOFTWARE.