RNN与情感分类问题实战-加载IMDB数据集

目录

  • Sentiment Analysis
  • Two approaches
    • Single layer
    • Multi-layers


Sentiment Analysis

RNN与情感分类问题实战-加载IMDB数据集

Two approaches

  • SimpleRNNCell

    • single layer

    • multi-layers

  • RNNCell

Single layer

import os
import tensorflow as tf
import numpy as np
from tensorflow import keras
from tensorflow.keras import layers

tf.random.set_seed(22)
np.random.seed(22)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
assert tf.__version__.startswith('2.')

batchsz = 128

# the most frequest words
total_words = 10000
max_review_len = 80
embedding_len = 100
(x_train,
 y_train), (x_test,
            y_test) = keras.datasets.imdb.load_data(num_words=total_words)
# x_train:[b, 80]
# x_test: [b, 80]
x_train = keras.preprocessing.sequence.pad_sequences(x_train,
                                                     maxlen=max_review_len)
x_test = keras.preprocessing.sequence.pad_sequences(x_test,
                                                    maxlen=max_review_len)

db_train = tf.data.Dataset.from_tensor_slices((x_train, y_train))
db_train = db_train.shuffle(1000).batch(batchsz, drop_remainder=True)
db_test = tf.data.Dataset.from_tensor_slices((x_test, y_test))
db_test = db_test.batch(batchsz, drop_remainder=True)
print('x_train shape:', x_train.shape, tf.reduce_max(y_train),
      tf.reduce_min(y_train))
print('x_test shape:', x_test.shape)


class MyRNN(keras.Model):
    def __init__(self, units):
        super(MyRNN, self).__init__()

        # [b, 64]
        self.state0 = [tf.zeros([batchsz, units])]
        self.state1 = [tf.zeros([batchsz, units])]

        # transform text to embedding representation
        # [b, 80] => [b, 80, 100]
        self.embedding = layers.Embedding(total_words,
                                          embedding_len,
                                          input_length=max_review_len)

        # [b, 80, 100] , h_dim: 64
        # RNN: cell1 ,cell2, cell3
        # SimpleRNN,units=64表示100个向量转成64个初始的状态
        self.rnn_cell0 = layers.SimpleRNNCell(units, dropout=0.5)
        self.rnn_cell1 = layers.SimpleRNNCell(units, dropout=0.5)

        # fc, [b, 80, 100] => [b, 64] => [b, 1]
        self.outlayer = layers.Dense(1)

    def call(self, inputs, training=None):
        """
        net(x) net(x, training=True) :train mode
        net(x, training=False): test
        :param inputs: [b, 80]
        :param training:
        :return:
        """
        # [b, 80]
        x = inputs
        # embedding: [b, 80] => [b, 80, 100]
        x = self.embedding(x)
        # rnn cell compute
        # [b, 80, 100] => [b, 64]
        state0 = self.state0
        state1 = self.state1
        for word in tf.unstack(x, axis=1):  # word: [b, 100]
            # h1 = x*wxh+h0*whh
            # out0: [b, 64]
            out0, state0 = self.rnn_cell0(word, state0, training)
            # out1: [b, 64]
            out1, state1 = self.rnn_cell1(out0, state1, training)

        # out: [b, 64] => [b, 1]
        x = self.outlayer(out1)
        # p(y is pos|x)
        prob = tf.sigmoid(x)

        return prob


def main():
    units = 64
    epochs = 4

    model = MyRNN(units)
    model.compile(optimizer=keras.optimizers.Adam(0.001),
                  loss=tf.losses.BinaryCrossentropy(),
                  metrics=['accuracy'])
    model.fit(db_train, epochs=epochs, validation_data=db_test)

    model.evaluate(db_test)


if __name__ == '__main__':
    main()

Multi-layers

import os
import tensorflow as tf
import numpy as np
from tensorflow import keras
from tensorflow.keras import layers

tf.random.set_seed(22)
np.random.seed(22)
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
assert tf.__version__.startswith('2.')

batchsz = 128

# the most frequest words
total_words = 10000  # 编码10000个单词
max_review_len = 80  # 句子长度80
embedding_len = 100
(x_train,
 y_train), (x_test,
            y_test) = keras.datasets.imdb.load_data(num_words=total_words)
# x_train:[b, 80]
# x_test: [b, 80]
x_train = keras.preprocessing.sequence.pad_sequences(x_train,
                                                     maxlen=max_review_len)
x_test = keras.preprocessing.sequence.pad_sequences(x_test,
                                                    maxlen=max_review_len)

db_train = tf.data.Dataset.from_tensor_slices((x_train, y_train))
# drop_remainder,丢弃最后一个大小不合适的batch
db_train = db_train.shuffle(1000).batch(batchsz, drop_remainder=True)
db_test = tf.data.Dataset.from_tensor_slices((x_test, y_test))
db_test = db_test.batch(batchsz, drop_remainder=True)
print('x_train shape:', x_train.shape, tf.reduce_max(y_train),
      tf.reduce_min(y_train))
print('x_test shape:', x_test.shape)


class MyRNN(keras.Model):
    def __init__(self, units):
        super(MyRNN, self).__init__()

        # transform text to embedding representation
        # [b, 80] => [b, 80, 100]  # embedding_len=100表示一个单词为100的向量
        self.embedding = layers.Embedding(total_words,
                                          embedding_len,
                                          input_length=max_review_len)

        # [b, 80, 100] , h_dim: 64
        self.rnn = keras.Sequential([
            layers.SimpleRNN(units,
                             dropout=0.5,
                             return_sequences=True,
                             unroll=True),
            layers.SimpleRNN(units, dropout=0.5, unroll=True)
        ])

        # fc, [b, 80, 100] => [b, 64] => [b, 1] # 得到分类结果
        self.outlayer = layers.Dense(1)

    def call(self, inputs, training=None):
        """
        net(x) net(x, training=True) :train mode
        net(x, training=False): test
        :param inputs: [b, 80]
        :param training: 计算过程是train还是test
        :return:
        """
        # [b, 80]
        x = inputs
        # embedding: [b, 80] => [b, 80, 100]
        x = self.embedding(x)
        # rnn cell compute
        # x: [b, 80, 100] => [b, 64]
        x = self.rnn(x)

        # out: [b, 64] => [b, 1]
        x = self.outlayer(x)
        # p(y is pos|x)
        prob = tf.sigmoid(x)

        return prob


def main():
    units = 64
    epochs = 4

    model = MyRNN(units)
    model.compile(optimizer=keras.optimizers.Adam(0.001),
                  loss=tf.losses.BinaryCrossentropy(),
                  metrics=['accuracy'])
    model.fit(db_train, epochs=epochs, validation_data=db_test)

    model.evaluate(db_test)


if __name__ == '__main__':
    main()
上一篇:10.4-10.5随笔


下一篇:RNN网络考点精析