RNN GRU LSTM 的学习来自:
https://zybuluo.com/hanbingtao/note/541458
https://zybuluo.com/hanbingtao/note/581764
RNN的实例化
import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.datasets import mnist
# import keras
physical_devices = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0],True)
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train = x_train.astype('float32') / 255.0
x_test = x_test.astype('float32') / 255.0
model = keras.Sequential()
model.add(keras.Input(shape=(None, 28)))
model.add(
layers.SimpleRNN(512, return_sequences=True, activation='relu')
)
model.add(layers.SimpleRNN(512, activation='relu'))
model.add(layers.Dense(10))
print(model.summary())
model.compile(
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
optimizer=keras.optimizers.Adam(lr=0.001),
metrics=['accuracy']
)
model.fit(x_train, y_train, batch_size=64, epochs=10, verbose=2)
model.evaluate(x_test, y_test, batch_size=64, verbose=2)
Epoch 10/10
938/938 - 174s - loss: 0.0755 - accuracy: 0.9798
157/157 - 5s - loss: 0.0602 - accuracy: 0.9834
GRU
model = keras.Sequential()
model.add(keras.Input(shape=(None, 28)))
model.add(
layers.GRU(512, return_sequences=True, activation='relu')
)
model.add(layers.GRU(512, activation='relu'))
model.add(layers.Dense(10))
Epoch 10/10
938/938 - 523s - loss: 0.0161 - accuracy: 0.9950
157/157 - 29s - loss: 0.0411 - accuracy: 0.9879
LSTM and 双向
model = keras.Sequential()
model.add(keras.Input(shape=(None, 28)))
model.add(
layers.Bidirectional(
layers.GRU(512, return_sequences=True, activation='relu')
)
)
model.add(layers.Bidirectional(
layers.GRU(512, activation='relu'))
)
model.add(layers.Dense(10))