import keras
from keras.layers import LSTM
from keras.layers import Dense, Activation
from keras.datasets import mnist
from keras.models import Sequential
from keras.optimizers import Adam

learning_rate = 0.001
training_iters = 20
batch_size = 128
display_step = 10

n_input = 28
n_step = 28
n_hidden = 128
n_classes = 10

(x_train, y_train), (x_test, y_test) = mnist.load_data()

x_train = x_train.reshape(-1, n_step, n_input)
x_test = x_test.reshape(-1, n_step, n_input)
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255

y_train = keras.utils.to_categorical(y_train, n_classes)
y_test = keras.utils.to_categorical(y_test, n_classes)

model = Sequential()
model.add(LSTM(n_hidden,
               batch_input_shape=(None, n_step, n_input),
               unroll=True))

model.add(Dense(n_classes))
model.add(Activation('softmax'))

adam = Adam(lr=learning_rate)
model.summary()
model.compile(optimizer=adam,
              loss='categorical_crossentropy',
              metrics=['accuracy'])

model.fit(x_train, y_train,
          batch_size=batch_size,
          epochs=training_iters,
          verbose=1,
          validation_data=(x_test, y_test))

scores = model.evaluate(x_test, y_test, verbose=0)
print('LSTM test score:', scores[0])
print('LSTM test accuracy:', scores[1])
 

使用如下結構，10個epochs，就可以使得測試集的準確率達到98.3%左右

_________________________________________________________________

Layer (type)                 Output Shape              Param #   
=================================================================
lstm_1 (LSTM)                (None, 128)               80384     
_________________________________________________________________
dense_1 (Dense)              (None, 10)                1290      
_________________________________________________________________
activation_1 (Activation)    (None, 10)                0         
=================================================================
Total params: 81,674
Trainable params: 81,674
Non-trainable params: 0
_________________________________________________________________

總共有81674個引數需要訓練。