TensorFlow 2.0 笔记（八）—— 循环神经网络

x = tf.random.normal([4, 80, 100])
xt0 = x[:, 0, :]
cell = tf.keras.layers.SimpleRNNCell(64)
out, xt1 = cell(xt0, [tf.zeros([4, 64])])
out.shape, xt1[0].shape
id(out), id(xt1[0])
cell.trainable_variables

x = tf.random.normal([4, 80, 100])
xt0 = x[:, 0, :]
cell = tf.keras.layers.SimpleRNNCell(64)
cell2 = tf.keras.layers.SimpleRNNCell(64)
state0 = [tf.zeros([4, 64])]
state1 = [tf.zeros([4, 64])]
out0, state0 = cell(xt0, state0)
out1, state1 = cell2(out0, state1)
out1.shape, state1[0].shape

• RNN Cell

import os
import tensorflow as tf
import numpy as np

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
assert tf.__version__.startswith('2.')
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0
tf.config.experimental.set_memory_growth(physical_devices[0], True)

tf.random.set_seed(22)
np.random.seed(22)

batchsz = 128
# the most frequest words
total_words = 10000
max_review_len = 80
embedding_len = 100
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.imdb.load_data(num_words=total_words)
# x_train: [b, 80]
# x_test:  [b, 80]
x_train = tf.keras.preprocessing.sequence.pad_sequences(x_train, maxlen=max_review_len)
x_test = tf.keras.preprocessing.sequence.pad_sequences(x_test, maxlen=max_review_len)

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(1000).batch(batchsz, drop_remainder=True)
test_data = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_data = test_data.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(tf.keras.Model):

    def __init__(self, units):
        super(MyRNN, self).__init__()

        self.state0 = [tf.zeros([batchsz, units])]
        self.state1 = [tf.zeros([batchsz, units])]
        # transform text to embedding representation
        # [b, 80] => [b, 80, 100]
        self.embedding = tf.keras.layers.Embedding(total_words, embedding_len,
                                                   input_length=max_review_len)
        # [b, 80, 100], h_dim: 64
        # RNN: cell1, cell2, cell3
        # SimpleRNN
        self.rnn_cell0 = tf.keras.layers.SimpleRNNCell(units, dropout=0.5)
        self.rnn_cell1 = tf.keras.layers.SimpleRNNCell(units, dropout=0.5)
        # fc, [b, 80, 100] => [b, 64] => [b, 1]
        self.out_layer = tf.keras.layers.Dense(1)

    def call(self, inputs, training=None, mask=None):
        """
        net(x) net(x, training=True): train mode
        net(x, training=False): test
        :param inputs: [b, 80]
        :param training:
        :param mask:
        :return:
        """
        # [b, 80}
        x = inputs
        # embedding: [b, 80] => [b, 80, 100]
        x = self.embedding(x)
        # run 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.out_layer(out1)
        # p(y is pos|x)
        prob = tf.sigmoid(x)

        return prob


def main():
    units = 64
    epochs = 4
    model = MyRNN(units)
    model.compile(optimizer=tf.keras.optimizers.Adam(0.001),
                  loss=tf.losses.BinaryCrossentropy(),
                  experimental_run_tf_function=False,
                  metrics=['accuracy'])
    model.fit(train_data, epochs=epochs, validation_data=test_data, validation_freq=1)

    model.evaluate(test_data)


if __name__ == '__main__':
    main()

• RNN layer

import os
import time
import tensorflow as tf
import numpy as np

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
assert tf.__version__.startswith('2.')
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0
tf.config.experimental.set_memory_growth(physical_devices[0], True)

tf.random.set_seed(22)
np.random.seed(22)

batchsz = 128
# the most frequest words
total_words = 10000
max_review_len = 80
embedding_len = 100
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.imdb.load_data(num_words=total_words)
# x_train: [b, 80]
# x_test:  [b, 80]
x_train = tf.keras.preprocessing.sequence.pad_sequences(x_train, maxlen=max_review_len)
x_test = tf.keras.preprocessing.sequence.pad_sequences(x_test, maxlen=max_review_len)

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(1000).batch(batchsz, drop_remainder=True)
test_data = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_data = test_data.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(tf.keras.Model):

    def __init__(self, units):
        super(MyRNN, self).__init__()

        # transform text to embedding representation
        # [b, 80] => [b, 80, 100]
        self.embedding = tf.keras.layers.Embedding(total_words, embedding_len,
                                                   input_length=max_review_len)
        # [b, 80, 100], h_dim: 64
        # RNN: cell1, cell2, cell3
        # SimpleRNN
        # unroll = True 可以加快RNN
        self.rnn = tf.keras.Sequential([
            tf.keras.layers.SimpleRNN(units, dropout=0.5, return_sequences=True),
            tf.keras.layers.SimpleRNN(units, dropout=0.5)
        ])
        # fc, [b, 80, 100] => [b, 64] => [b, 1]
        self.out_layer = tf.keras.layers.Dense(1)

    def call(self, inputs, training=None, mask=None):
        """
        net(x) net(x, training=True): train mode
        net(x, training=False): test
        :param inputs: [b, 80]
        :param training:
        :param mask:
        :return:
        """
        # [b, 80}
        x = inputs
        # embedding: [b, 80] => [b, 80, 100]
        x = self.embedding(x)
        # run cell compute
        # [b, 80, 100] => [b, 64]
        x = self.rnn(x)
        # out: [b, 64] => [b, 1]
        x = self.out_layer(x)
        # p(y is pos|x)
        prob = tf.sigmoid(x)

        return prob


def main():
    units = 64
    epochs = 4
    t0 = time.time()
    model = MyRNN(units)
    model.compile(optimizer=tf.keras.optimizers.Adam(0.001),
                  loss=tf.losses.BinaryCrossentropy(),
                  metrics=['accuracy'])
    model.fit(train_data, epochs=epochs, validation_data=test_data, validation_freq=1)

    model.evaluate(test_data)

    # accuracy: 0.8107 total time cost: 39.90
    t1 = time.time()
    print('total time cost:', t1 - t0)


if __name__ == '__main__':
    main()

• LSTM Cell

import os
import time
import tensorflow as tf
import numpy as np

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
assert tf.__version__.startswith('2.')
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0
tf.config.experimental.set_memory_growth(physical_devices[0], True)

tf.random.set_seed(22)
np.random.seed(22)

batchsz = 128
# the most frequest words
total_words = 10000
max_review_len = 80
embedding_len = 100
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.imdb.load_data(num_words=total_words)
# x_train: [b, 80]
# x_test:  [b, 80]
x_train = tf.keras.preprocessing.sequence.pad_sequences(x_train, maxlen=max_review_len)
x_test = tf.keras.preprocessing.sequence.pad_sequences(x_test, maxlen=max_review_len)

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(1000).batch(batchsz, drop_remainder=True)
test_data = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_data = test_data.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(tf.keras.Model):

    def __init__(self, units):
        super(MyRNN, self).__init__()

        self.state0 = [tf.zeros([batchsz, units]), tf.zeros([batchsz, units])]
        self.state1 = [tf.zeros([batchsz, units]), tf.zeros([batchsz, units])]
        # transform text to embedding representation
        # [b, 80] => [b, 80, 100]
        self.embedding = tf.keras.layers.Embedding(total_words, embedding_len,
                                                   input_length=max_review_len)
        # [b, 80, 100], h_dim: 64
        # RNN: cell1, cell2, cell3
        # SimpleRNN
        self.rnn_cell0 = tf.keras.layers.LSTMCell(units, dropout=0.5)
        self.rnn_cell1 = tf.keras.layers.LSTMCell(units, dropout=0.5)
        # fc, [b, 80, 100] => [b, 64] => [b, 1]
        self.out_layer = tf.keras.layers.Dense(1)

    def call(self, inputs, training=None, mask=None):
        """
        net(x) net(x, training=True): train mode
        net(x, training=False): test
        :param inputs: [b, 80]
        :param training:
        :param mask:
        :return:
        """
        # [b, 80}
        x = inputs
        # embedding: [b, 80] => [b, 80, 100]
        x = self.embedding(x)
        # run 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.out_layer(out1)
        # p(y is pos|x)
        prob = tf.sigmoid(x)

        return prob


def main():
    units = 64
    epochs = 4

    t0 = time.time()
    model = MyRNN(units)
    model.compile(optimizer=tf.keras.optimizers.Adam(0.001),
                  loss=tf.losses.BinaryCrossentropy(),
                  experimental_run_tf_function=False,
                  metrics=['accuracy'])
    model.fit(train_data, epochs=epochs, validation_data=test_data, validation_freq=1)

    model.evaluate(test_data)

    t1 = time.time()
    print('total time cost:', t1 - t0)


if __name__ == '__main__':
    main()

• LSTM layer

import os
import time
import tensorflow as tf
import numpy as np

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
assert tf.__version__.startswith('2.')
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0
tf.config.experimental.set_memory_growth(physical_devices[0], True)

tf.random.set_seed(22)
np.random.seed(22)

batchsz = 128
# the most frequest words
total_words = 10000
max_review_len = 80
embedding_len = 100
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.imdb.load_data(num_words=total_words)
# x_train: [b, 80]
# x_test:  [b, 80]
x_train = tf.keras.preprocessing.sequence.pad_sequences(x_train, maxlen=max_review_len)
x_test = tf.keras.preprocessing.sequence.pad_sequences(x_test, maxlen=max_review_len)

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(1000).batch(batchsz, drop_remainder=True)
test_data = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_data = test_data.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(tf.keras.Model):

    def __init__(self, units):
        super(MyRNN, self).__init__()

        # transform text to embedding representation
        # [b, 80] => [b, 80, 100]
        self.embedding = tf.keras.layers.Embedding(total_words, embedding_len,
                                                   input_length=max_review_len)
        # [b, 80, 100], h_dim: 64
        # RNN: cell1, cell2, cell3
        # SimpleRNN
        self.rnn = tf.keras.Sequential([
            tf.keras.layers.LSTM(units, dropout=0.5, return_sequences=True),
            tf.keras.layers.LSTM(units, dropout=0.5)
        ])
        # fc, [b, 80, 100] => [b, 64] => [b, 1]
        self.out_layer = tf.keras.layers.Dense(1)

    def call(self, inputs, training=None, mask=None):
        """
        net(x) net(x, training=True): train mode
        net(x, training=False): test
        :param inputs: [b, 80]
        :param training:
        :param mask:
        :return:
        """
        # [b, 80}
        x = inputs
        # embedding: [b, 80] => [b, 80, 100]
        x = self.embedding(x)
        # run cell compute
        # [b, 80, 100] => [b, 64]
        x = self.rnn(x)
        # out: [b, 64] => [b, 1]
        x = self.out_layer(x)
        # p(y is pos|x)
        prob = tf.sigmoid(x)

        return prob


def main():
    units = 64
    epochs = 4
    t0 = time.time()
    model = MyRNN(units)
    model.compile(optimizer=tf.keras.optimizers.Adam(0.001),
                  loss=tf.losses.BinaryCrossentropy(),
                  metrics=['accuracy'])
    model.fit(train_data, epochs=epochs, validation_data=test_data, validation_freq=1)

    model.evaluate(test_data)

    # accuracy: 0.8124 total time cost: 25.03
    t1 = time.time()
    print('total time cost:', t1 - t0)


if __name__ == '__main__':
    main()

GRU Cell

import os
import time
import tensorflow as tf
import numpy as np

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
assert tf.__version__.startswith('2.')
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0
tf.config.experimental.set_memory_growth(physical_devices[0], True)

tf.random.set_seed(22)
np.random.seed(22)

batchsz = 128
# the most frequest words
total_words = 10000
max_review_len = 80
embedding_len = 100
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.imdb.load_data(num_words=total_words)
# x_train: [b, 80]
# x_test:  [b, 80]
x_train = tf.keras.preprocessing.sequence.pad_sequences(x_train, maxlen=max_review_len)
x_test = tf.keras.preprocessing.sequence.pad_sequences(x_test, maxlen=max_review_len)

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(1000).batch(batchsz, drop_remainder=True)
test_data = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_data = test_data.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(tf.keras.Model):

    def __init__(self, units):
        super(MyRNN, self).__init__()

        self.state0 = [tf.zeros([batchsz, units])]
        self.state1 = [tf.zeros([batchsz, units])]
        # transform text to embedding representation
        # [b, 80] => [b, 80, 100]
        self.embedding = tf.keras.layers.Embedding(total_words, embedding_len,
                                                   input_length=max_review_len)
        # [b, 80, 100], h_dim: 64
        # RNN: cell1, cell2, cell3
        # SimpleRNN
        self.rnn_cell0 = tf.keras.layers.GRUCell(units, dropout=0.5)
        self.rnn_cell1 = tf.keras.layers.GRUCell(units, dropout=0.5)
        # fc, [b, 80, 100] => [b, 64] => [b, 1]
        self.out_layer = tf.keras.layers.Dense(1)

    def call(self, inputs, training=None, mask=None):
        """
        net(x) net(x, training=True): train mode
        net(x, training=False): test
        :param inputs: [b, 80]
        :param training:
        :param mask:
        :return:
        """
        # [b, 80}
        x = inputs
        # embedding: [b, 80] => [b, 80, 100]
        x = self.embedding(x)
        # run 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.out_layer(out1)
        # p(y is pos|x)
        prob = tf.sigmoid(x)

        return prob


def main():
    units = 64
    epochs = 4

    t0 = time.time()
    model = MyRNN(units)
    model.compile(optimizer=tf.keras.optimizers.Adam(0.001),
                  loss=tf.losses.BinaryCrossentropy(),
                  experimental_run_tf_function=False,
                  metrics=['accuracy'])
    model.fit(train_data, epochs=epochs, validation_data=test_data, validation_freq=1)

    model.evaluate(test_data)

    t1 = time.time()
    print('total time cost:', t1 - t0)


if __name__ == '__main__':
    main()

GRU Layer

import os
import time
import tensorflow as tf
import numpy as np

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
assert tf.__version__.startswith('2.')
physical_devices = tf.config.experimental.list_physical_devices('GPU')
assert len(physical_devices) > 0
tf.config.experimental.set_memory_growth(physical_devices[0], True)

tf.random.set_seed(22)
np.random.seed(22)

batchsz = 128
# the most frequest words
total_words = 10000
max_review_len = 80
embedding_len = 100
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.imdb.load_data(num_words=total_words)
# x_train: [b, 80]
# x_test:  [b, 80]
x_train = tf.keras.preprocessing.sequence.pad_sequences(x_train, maxlen=max_review_len)
x_test = tf.keras.preprocessing.sequence.pad_sequences(x_test, maxlen=max_review_len)

train_data = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_data = train_data.shuffle(1000).batch(batchsz, drop_remainder=True)
test_data = tf.data.Dataset.from_tensor_slices((x_test, y_test))
test_data = test_data.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(tf.keras.Model):

    def __init__(self, units):
        super(MyRNN, self).__init__()

        # transform text to embedding representation
        # [b, 80] => [b, 80, 100]
        self.embedding = tf.keras.layers.Embedding(total_words, embedding_len,
                                                   input_length=max_review_len)
        # [b, 80, 100], h_dim: 64
        # RNN: cell1, cell2, cell3
        # SimpleRNN
        self.rnn = tf.keras.Sequential([
            tf.keras.layers.GRU(units, dropout=0.5, return_sequences=True),
            tf.keras.layers.GRU(units, dropout=0.5)
        ])
        # fc, [b, 80, 100] => [b, 64] => [b, 1]
        self.out_layer = tf.keras.layers.Dense(1)

    def call(self, inputs, training=None, mask=None):
        """
        net(x) net(x, training=True): train mode
        net(x, training=False): test
        :param inputs: [b, 80]
        :param training:
        :param mask:
        :return:
        """
        # [b, 80}
        x = inputs
        # embedding: [b, 80] => [b, 80, 100]
        x = self.embedding(x)
        # run cell compute
        # [b, 80, 100] => [b, 64]
        x = self.rnn(x)
        # out: [b, 64] => [b, 1]
        x = self.out_layer(x)
        # p(y is pos|x)
        prob = tf.sigmoid(x)

        return prob


def main():
    units = 64
    epochs = 4
    t0 = time.time()
    model = MyRNN(units)
    model.compile(optimizer=tf.keras.optimizers.Adam(0.001),
                  loss=tf.losses.BinaryCrossentropy(),
                  metrics=['accuracy'])
    model.fit(train_data, epochs=epochs, validation_data=test_data, validation_freq=1)

    model.evaluate(test_data)

    # accuracy: 0.8311 total time cost: 26.32
    t1 = time.time()
    print('total time cost:', t1 - t0)


if __name__ == '__main__':
    main()