本篇文章主要介紹通過RNN實現MNIST手寫資料集分類。

示例程式碼：

import numpy as np
from keras.datasets import mnist
from keras.utils import np_utils
from keras.models import Sequential
from keras.layers import Dense, Activation, SimpleRNN
from keras.optimizers import Adam

# 使多次生成的隨機數相同
np.random.seed(1337)

# 超引數
TIME_STEP = 28  # 和圖片的高度相同
INPUT_SIZE = 28  # 和圖片的寬度相同
BATCH_SIZE = 50
BATCH_INDEX = 0
OUTPUT_SIZE = 10  # [0 0 0 0 1 0 0 0 0 0]->4
CELL_SIZE = 50  # RNN裡面hidden unit的數量
LR = 0.001

# 下載資料集
# X_shape(60000 28x28),y shape(10000)
(X_train, y_train), (X_test, y_test) = mnist.load_data()

# 預處理資料
'''
X_train.reshape(X_train.shape[0], -1) 將60000個28x28的資料變為60000x784
/255：把資料標準化到[0,1]
'''
# 除以255為進行標準化
X_train = X_train.reshape(-1, 28, 28) / 255   # -1:sample個數， 1：channel, 28x28：長寬
X_test = X_test.reshape(-1, 28, 28) / 255
# 將標籤變為one-hot形式
y_train = np_utils.to_categorical(y_train, num_classes=10)
y_test = np_utils.to_categorical(y_test, num_classes=10)

# 建立模型
model = Sequential()

# RNN cell
model.add(SimpleRNN(
    batch_input_shape=(None, TIME_STEP, INPUT_SIZE),
    output_dim=CELL_SIZE,
))
# output layer
model.add(Dense(OUTPUT_SIZE))
model.add(Activation('softmax'))  # softmax進行分類

# 優化器
adam = Adam(LR)
model.compile(optimizer=adam,
              loss='categorical_crossentropy',
              metrics=['accuracy'])

# 訓練
for step in range(4001):
    # data shape = (batch_num, steps, input/output)
    x_batch = X_train[BATCH_INDEX: BATCH_SIZE + BATCH_INDEX, :, :]
    y_batch = y_train[BATCH_INDEX: BATCH_SIZE + BATCH_INDEX, :]
    cost = model.train_on_batch(x_batch, y_batch)

    BATCH_INDEX += BATCH_SIZE
    BATCH_INDEX = 0 if BATCH_INDEX >= X_train.shape[0] else BATCH_INDEX

    if step % 500 == 0:
        cost, accuracy = model.evaluate(X_test, y_test, batch_size=y_test.shape[0], verbose=False)
        print('test cost:', cost, 'test accuracy:', accuracy)
 

分類結果：

test cost: 2.4057323932647705 test accuracy: 0.03909999877214432
test cost: 0.6151769757270813 test accuracy: 0.8148999810218811
test cost: 0.43123072385787964 test accuracy: 0.8719000220298767
test cost: 0.3669806122779846 test accuracy: 0.890999972820282
test cost: 0.34460538625717163 test accuracy: 0.8988999724388123
test cost: 0.27395951747894287 test accuracy: 0.9190000295639038
test cost: 0.2991882860660553 test accuracy: 0.9136999845504761
test cost: 0.23682279884815216 test accuracy: 0.9291999936103821
test cost: 0.2040248066186905 test accuracy: 0.9419000148773193