 Voir sur TensorFlow.org |  Exécuter dans Google Colab |  Afficher sur GitHub |  Télécharger le cahier |  Voir les modèles de hub TF |
Ce cahier classe les critiques de films comme positives ou négatives en utilisant le texte de la critique. Il s'agit d'un exemple de classification binaire - ou à deux classes -, un type de problème d'apprentissage automatique important et largement applicable.
Le didacticiel présente l'application de base de l'apprentissage par transfert avec TensorFlow Hub et Keras.
Il utilise l'ensemble de données IMDB qui contient le texte de 50 000 critiques de films de l' Internet Movie Database . Ceux-ci sont divisés en 25 000 avis pour la formation et 25 000 avis pour les tests. Les ensembles de formation et de test sont équilibrés , ce qui signifie qu'ils contiennent un nombre égal d'avis positifs et négatifs.
Ce bloc-notes utilise tf.keras , une API de haut niveau pour créer et former des modèles dans TensorFlow, et tensorflow_hub , une bibliothèque permettant de charger des modèles formés à partir de TFHub en une seule ligne de code. Pour un didacticiel de classification de texte plus avancé à l'aide tf.keras , consultez le Guide de classification de texte MLCC .
pip install tensorflow-hubpip 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: 2.8.0-rc1 Eager mode: True Hub version: 0.12.0 GPU is available
Télécharger le jeu de données IMDB
L'ensemble de données IMDB est disponible sur les revues imdb ou sur les ensembles de données TensorFlow . Le code suivant télécharge l'ensemble de données IMDB sur votre machine (ou le runtime colab) :
# 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)
Explorer les données
Prenons un moment pour comprendre le format des données. Chaque exemple est une phrase représentant la critique du film et une étiquette correspondante. La phrase n'est en aucun cas prétraitée. L'étiquette est une valeur entière de 0 ou 1, où 0 est un avis négatif et 1 est un avis positif.
Imprimons les 10 premiers exemples.
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)>
Imprimons également les 10 premières étiquettes.
train_labels_batch
<tf.Tensor: shape=(10,), dtype=int64, numpy=array([0, 0, 0, 1, 1, 1, 0, 0, 0, 0])>
Construire le modèle
Le réseau de neurones est créé en empilant des couches, ce qui nécessite trois décisions architecturales principales :
- Comment représenter le texte ?
- Combien de couches utiliser dans le modèle ?
- Combien d' unités cachées utiliser pour chaque couche ?
Dans cet exemple, les données d'entrée sont constituées de phrases. Les libellés à prédire sont 0 ou 1.
Une façon de représenter le texte consiste à convertir des phrases en vecteurs d'intégration. Utilisez une incorporation de texte pré-entraînée comme premier calque, ce qui aura trois avantages :
- Vous n'avez pas à vous soucier du prétraitement du texte,
- Bénéficiez d'un apprentissage par transfert,
- l'intégration a une taille fixe, elle est donc plus simple à traiter.
Pour cet exemple, vous utilisez un modèle d'intégration de texte pré-entraîné de TensorFlow Hub appelé google/nnlm-en-dim50/2 .
Il existe de nombreuses autres intégrations de texte pré-formées de TFHub qui peuvent être utilisées dans ce didacticiel :
- google/nnlm-en-dim128/2 - formé avec la même architecture NNLM sur les mêmes données que google/nnlm-en-dim50/2 , mais avec une dimension d'intégration plus grande. Des plongements dimensionnels plus grands peuvent améliorer votre tâche, mais l'entraînement de votre modèle peut prendre plus de temps.
- google/nnlm-en-dim128-with-normalization/2 - identique à google/nnlm-en-dim128/2 , mais avec une normalisation de texte supplémentaire telle que la suppression de la ponctuation. Cela peut être utile si le texte de votre tâche contient des caractères supplémentaires ou des signes de ponctuation.
- google/universal-sentence-encoder/4 - un modèle beaucoup plus grand produisant 512 incorporations dimensionnelles formées avec un encodeur de réseau de moyenne profonde (DAN).
Et beaucoup plus! Trouvez plus de modèles d'intégration de texte sur TFHub.
Commençons par créer une couche Keras qui utilise un modèle TensorFlow Hub pour intégrer les phrases, et essayons-la sur quelques exemples d'entrée. Notez que quelle que soit la longueur du texte d'entrée, la forme de sortie des incorporations est : (num_examples, embedding_dimension) .
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)>
Construisons maintenant le modèle complet :
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
_________________________________________________________________
Les couches sont empilées séquentiellement pour construire le classifieur :
- La première couche est une couche TensorFlow Hub. Cette couche utilise un modèle enregistré pré-formé pour mapper une phrase dans son vecteur d'intégration. Le modèle d'intégration de texte pré-formé que vous utilisez ( google/nnlm-en-dim50/2 ) divise la phrase en jetons, intègre chaque jeton, puis combine l'intégration. Les dimensions résultantes sont :
(num_examples, embedding_dimension). Pour ce modèle NNLM,embedding_dimensionest 50. - Ce vecteur de sortie de longueur fixe est acheminé via une couche entièrement connectée (
Dense) avec 16 unités cachées. - La dernière couche est densément connectée à un seul nœud de sortie.
Compilons le modèle.
Fonction de perte et optimiseur
Un modèle a besoin d'une fonction de perte et d'un optimiseur pour la formation. Puisqu'il s'agit d'un problème de classification binaire et que le modèle génère des logits (une couche à une seule unité avec une activation linéaire), vous utiliserez la fonction de perte binary_crossentropy .
Ce n'est pas le seul choix pour une fonction de perte, vous pouvez, par exemple, choisir mean_squared_error . Mais, en général, binary_crossentropy est meilleur pour traiter les probabilités - il mesure la "distance" entre les distributions de probabilités, ou dans notre cas, entre la distribution de vérité terrain et les prédictions.
Plus tard, lorsque vous explorerez des problèmes de régression (par exemple, pour prédire le prix d'une maison), vous verrez comment utiliser une autre fonction de perte appelée erreur quadratique moyenne.
Maintenant, configurez le modèle pour utiliser un optimiseur et une fonction de perte :
model.compile(optimizer='adam',
loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
metrics=['accuracy'])
Former le modèle
Entraînez le modèle pour 10 époques en mini-lots de 512 échantillons. Il s'agit de 10 itérations sur tous les échantillons dans les tenseurs x_train et y_train . Pendant l'entraînement, surveillez la perte et la précision du modèle sur les 10 000 échantillons de l'ensemble de validation :
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 59ms/step - loss: 0.6707 - accuracy: 0.5317 - val_loss: 0.6150 - val_accuracy: 0.5882 Epoch 2/10 30/30 [==============================] - 2s 57ms/step - loss: 0.5382 - accuracy: 0.7012 - val_loss: 0.4972 - val_accuracy: 0.7450 Epoch 3/10 30/30 [==============================] - 2s 56ms/step - loss: 0.3976 - accuracy: 0.8246 - val_loss: 0.4023 - val_accuracy: 0.8149 Epoch 4/10 30/30 [==============================] - 2s 61ms/step - loss: 0.2879 - accuracy: 0.8894 - val_loss: 0.3503 - val_accuracy: 0.8385 Epoch 5/10 30/30 [==============================] - 2s 59ms/step - loss: 0.2120 - accuracy: 0.9243 - val_loss: 0.3248 - val_accuracy: 0.8530 Epoch 6/10 30/30 [==============================] - 2s 54ms/step - loss: 0.1560 - accuracy: 0.9513 - val_loss: 0.3148 - val_accuracy: 0.8663 Epoch 7/10 30/30 [==============================] - 2s 55ms/step - loss: 0.1147 - accuracy: 0.9680 - val_loss: 0.3168 - val_accuracy: 0.8675 Epoch 8/10 30/30 [==============================] - 2s 56ms/step - loss: 0.0847 - accuracy: 0.9791 - val_loss: 0.3199 - val_accuracy: 0.8670 Epoch 9/10 30/30 [==============================] - 2s 54ms/step - loss: 0.0617 - accuracy: 0.9880 - val_loss: 0.3272 - val_accuracy: 0.8649 Epoch 10/10 30/30 [==============================] - 2s 57ms/step - loss: 0.0449 - accuracy: 0.9931 - val_loss: 0.3379 - val_accuracy: 0.8651
Évaluer le modèle
Et voyons comment le modèle fonctionne. Deux valeurs seront renvoyées. Perte (un nombre qui représente notre erreur, les valeurs inférieures sont meilleures) et précision.
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 - 2s - loss: 0.3599 - accuracy: 0.8523 - 2s/epoch - 34ms/step loss: 0.360 accuracy: 0.852
Cette approche assez naïve permet d'atteindre une précision d'environ 87 %. Avec des approches plus poussées, le modèle devrait se rapprocher de 95 %.
Lectures complémentaires
- Pour une manière plus générale de travailler avec des entrées de chaîne et pour une analyse plus détaillée de la progression de la précision et de la perte pendant la formation, consultez le didacticiel Classification de texte avec texte prétraité .
- Essayez d'autres didacticiels liés au texte à l' aide de modèles formés de TFHub.
# 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.