TensorFlow2.0教程-使用RNN实现文本分类

原文地址：https://blog.csdn.net/qq_31456593/article/details/89923645

Tensorflow 2.0 教程持续更新：https://blog.csdn.net/qq_31456593/article/details/88606284

本教程主要由tensorflow2.0官方教程的个人学习复现笔记整理而来，并借鉴了一些keras构造神经网络的方法，中文讲解，方便喜欢阅读中文教程的朋友，tensorflow官方教程：https://www.tensorflow.org

完整tensorflow2.0教程代码请看https://github.com/czy36mengfei/tensorflow2_tutorials_chinese (欢迎star)

Tensorflow2.0部分教程内容：

TensorFlow 2.0 教程- Keras 快速入门
TensorFlow 2.0 教程-keras 函数api
TensorFlow 2.0 教程-使用keras训练模型
TensorFlow 2.0 教程-用keras构建自己的网络层
TensorFlow 2.0 教程-keras模型保存和序列化

1.使用tensorflow_datasets 构造输入数据

!pip install -q tensorflow_datasets

[31mspacy 2.0.18 has requirement numpy>=1.15.0, but you'll have numpy 1.14.3 which is incompatible.[0m
[31mplotnine 0.5.1 has requirement matplotlib>=3.0.0, but you'll have matplotlib 2.2.2 which is incompatible.[0m
[31mplotnine 0.5.1 has requirement pandas>=0.23.4, but you'll have pandas 0.23.0 which is incompatible.[0m
[31mneo4j-driver 1.6.2 has requirement neotime==1.0.0, but you'll have neotime 1.7.2 which is incompatible.[0m
[31mmizani 0.5.3 has requirement pandas>=0.23.4, but you'll have pandas 0.23.0 which is incompatible.[0m
[31mfastai 0.7.0 has requirement torch<0.4, but you'll have torch 0.4.1 which is incompatible.[0m
[33mYou are using pip version 10.0.1, however version 19.1 is available.
You should consider upgrading via the 'pip install --upgrade pip' command.[0m

import tensorflow_datasets as tfds
dataset, info = tfds.load('imdb_reviews/subwords8k', with_info=True,as_supervised=True)

获取训练集、测试集

train_dataset, test_dataset = dataset['train'], dataset['test']

获取tokenizer对象，用进行字符处理级id转换（这里先转换成subword，再转换为id）等操作

tokenizer = info.features['text'].encoder
print('vocabulary size: ', tokenizer.vocab_size)

vocabulary size:  8185

token对象测试

sample_string = 'Hello word , Tensorflow'
tokenized_string = tokenizer.encode(sample_string)
print('tokened id: ', tokenized_string)# 解码会原字符串
src_string = tokenizer.decode(tokenized_string)
print('original string: ', src_string)

tokened id:  [4025, 222, 2621, 1199, 6307, 2327, 2934]
original string:  Hello word , Tensorflow

解出每个subword

for t in tokenized_string:print(str(t)+'->['+tokenizer.decode([t])+ ']')

4025->[Hell]
222->[o ]
2621->[word]
1199->[ , ]
6307->[Ten]
2327->[sor]
2934->[flow]

构建批次训练集

BUFFER_SIZE=10000
BATCH_SIZE = 64train_dataset = train_dataset.shuffle(BUFFER_SIZE)
train_dataset = train_dataset.padded_batch(BATCH_SIZE, train_dataset.output_shapes)test_dataset = test_dataset.padded_batch(BATCH_SIZE, test_dataset.output_shapes)

模型构建

下面，因为此处的句子是变长的，所以只能使用序列模型，而不能使用keras的函数api

# def get_model():
#     inputs = tf.keras.Input((1240,))
#     emb = tf.keras.layers.Embedding(tokenizer.vocab_size, 64)(inputs)
#     h1 = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(64))(emb)
#     h1 = tf.keras.layers.Dense(64, activation='relu')(h1)
#     outputs = tf.keras.layers.Dense(1, activation='sigmoid')(h1)
#     model = tf.keras.Model(inputs, outputs)
#     return modeldef get_model():model = tf.keras.Sequential([tf.keras.layers.Embedding(tokenizer.vocab_size, 64),tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(64)),tf.keras.layers.Dense(64, activation='relu'),tf.keras.layers.Dense(1, activation='sigmoid')])return model

model = get_model()
model.compile(loss='binary_crossentropy',optimizer='adam',metrics=['accuracy'])

模型训练

history = model.fit(train_dataset, epochs=10,validation_data=test_dataset)

Epoch 1/10
391/391 [==============================] - 827s 2s/step - loss: 0.5606 - accuracy: 0.7068 - val_loss: 0.0000e+00 - val_accuracy:
....
Epoch 10/10
391/391 [==============================] - 791s 2s/step - loss: 0.1333 - accuracy: 0.9548 - val_loss: 0.6117 - val_accuracy: 0.8199

# 查看训练过程
import matplotlib.pyplot as pltdef plot_graphs(history, string):plt.plot(history.history[string])plt.plot(history.history['val_'+string])plt.xlabel('epochs')plt.ylabel(string)plt.legend([string, 'val_'+string])plt.show()plot_graphs(history, 'accuracy')

plot_graphs(history, 'loss')

测试

test_loss, test_acc = model.evaluate(test_dataset)
print('test loss: ', test_loss)
print('test acc: ', test_acc)

    391/Unknown - 68s 174ms/step - loss: 0.6117 - accuracy: 0.8199test loss:  0.6117385012262008
test acc:  0.81988

上述模型不会mask掉序列的padding，所以如果在有padding的寻列上训练，测试没有padding的序列时可能有所偏差。

def pad_to_size(vec, size):zeros = [0] * (size-len(vec))vec.extend(zeros)return vecdef sample_predict(sentence, pad=False):tokened_sent = tokenizer.encode(sentence)if pad:tokened_sent = pad_to_size(tokened_sent, 64)pred = model.predict(tf.expand_dims(tokened_sent, 0))return pred

# 没有padding的情况
sample_pred_text = ('The movie was cool. The animation and the graphics ''were out of this world. I would recommend this movie.')
predictions = sample_predict(sample_pred_text, pad=False)
print(predictions)

[[0.2938048]]

# 有paddin的情况
sample_pred_text = ('The movie was cool. The animation and the graphics ''were out of this world. I would recommend this movie.')
predictions = sample_predict(sample_pred_text, pad=True)
print (predictions)

[[0.42541984]]

堆叠更多的lstm层

from tensorflow.keras import layers
model = keras.Sequential(
[layers.Embedding(tokenizer.vocab_size, 64),layers.Bidirectional(layers.LSTM(64, return_sequences=True)),layers.Bidirectional(layers.LSTM(32)),layers.Dense(64, activation='relu'),layers.Dense(1, activation='sigmoid')
])

model.compile(loss=tf.keras.losses.binary_crossentropy,optimizer=tf.keras.optimizers.Adam(),metrics=['accuracy'])

history=model.fit(train_dataset, epochs=6, validation_data=test_dataset)

Epoch 1/6
391/391 [==============================] - 1646s 4s/step - loss: 0.5270 - accuracy: 0.7414 - val_loss: 0.0000e+00 - val_accuracy: ....
Epoch 6/6
391/391 [==============================] - 1622s 4s/step - loss: 0.1619 - accuracy: 0.9430 - val_loss: 0.5484 - val_accuracy: 0.7808

plot_graphs(history, 'accuracy')

plot_graphs(history, 'loss')

res = model.evaluate(test_dataset)
print(res)

    391/Unknown - 125s 320ms/step - loss: 0.5484 - accuracy: 0.7808[0.5484032468570162, 0.78084]