用RNN做MNIST分类

1.前言

RNN常用作NLP中，像图片生成文字、自动生成古诗词等。这篇文章用RNN做MNIST手写数字识别，分类效果虽然没有CNN效果好，但准确率也能够达到96%。

2.环境

Mac os系统，python：3.5，Keras

3.代码实现

import numpy as np
np.random.seed(1337)  
 
from keras.datasets import mnist
from keras.utils import np_utils
from keras.models import Sequential
from keras.layers import SimpleRNN, Activation, Dense
from keras.optimizers import Adam
 
TIME_STEPS = 28     
INPUT_SIZE = 28    
BATCH_SIZE = 50
BATCH_INDEX = 0
OUTPUT_SIZE = 10
CELL_SIZE = 50
LR = 0.001
 
(X_train, y_train), (X_test, y_test) = mnist.load_data()
 
# data pre-processing
X_train = X_train.reshape(-1, 28, 28) / 255.      # normalize
X_test = X_test.reshape(-1, 28, 28) / 255.        # normalize
y_train = np_utils.to_categorical(y_train, num_classes=10)
y_test = np_utils.to_categorical(y_test, num_classes=10)
 
# build RNN model
model = Sequential()
 
# RNN cell
model.add(SimpleRNN(
    batch_input_shape=(None, TIME_STEPS, INPUT_SIZE),       
    output_dim=CELL_SIZE,
    unroll=True,
))
 
# output layer
model.add(Dense(OUTPUT_SIZE))
model.add(Activation('softmax'))
 
# optimizer
adam = Adam(LR)
model.compile(optimizer=adam,
              loss='categorical_crossentropy',
              metrics=['accuracy'])
 
# training
for step in range(40001):
    X_batch = X_train[BATCH_INDEX: BATCH_INDEX+BATCH_SIZE, :, :]
    Y_batch = y_train[BATCH_INDEX: BATCH_INDEX+BATCH_SIZE, :]
    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)

4.结果