RNN探索(2)之手写数字识别

这篇博文不介绍基础的RNN理论知识,只是初步探索如何使用Tensorflow,之后会用笔推导RNN的公式和理论,现在时间紧迫所以先使用为主~~

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
import tensorflow.examples.tutorials.mnist.input_data as input_data
from   tensorflow.contrib import rnn

mnist = input_data.read_data_sets("MNIST_data/",one_hot=True)
trainimgs, trainlabels, testimgs, testlabels 
 = mnist.train.images, mnist.train.labels, mnist.test.images, mnist.test.labels
ntrain, ntest, dim, nclasses 
 = trainimgs.shape[0], testimgs.shape[0], trainimgs.shape[1], trainlabels.shape[1]
print ("MNIST loaded")
dim_input   = 28       #28*1
dim_hidden  = 128      #28*128
dim_output  = 10       #
nsteps      = 28
weight = {
    "hidden":tf.Variable(tf.random_normal([dim_input,dim_hidden])),
    "out"   :tf.Variable(tf.random_normal([dim_hidden,dim_output]))
}
biases = {
    "hidden":tf.Variable(tf.random_normal([dim_hidden])),
    "out"   :tf.Variable(tf.random_normal([dim_output]))
}

def RNN(_X,_W,_b,_nsteps,_name):
    #[batchsize,nsteps*dim_input]-->>[batchsize,nsteps,dim_input]=[num,28,28]
    _X = tf.reshape(_X,[-1,28,28])
    #-->>[nsteps,batchsize,dim_input]==[28,num,28]
    _X = tf.transpose(_X,[1,0,2])
    #-->>[nsteps*batchsize,input]==[28*num,28]
    _X = tf.reshape(_X,[-1,28])
    #这里是共享权重,nsteps个weights全部一样的.
    _H = tf.matmul(_X,_W['hidden']) + _b["hidden"]
    _Hsplit = tf.split(_H,num_or_size_splits=nsteps,axis=0)
    with tf.variable_scope(_name,reuse=tf.AUTO_REUSE):#重复使用参数节约空间,防止报错
        #版本更新弃用
        #scop.reuse_variables()
        #设计一个计算单元
        lstm_cell = rnn.BasicLSTMCell(128,forget_bias=1.0)
        #版本更新已经弃用
        #lstm_cell = rnn_cell.BasicLSTMCell(dim_hidden,forget_bias=1.0)
        #利用RNN单元搭建网络,这里用的最简单的,其它以后在说
        _LSTM_0,_LSTM_S = rnn.static_rnn(lstm_cell,_Hsplit,dtype=tf.float32)
        #版本更新已经弃用
        #_LSTM_O, _LSTM_S = tf.nn.rnn(lstm_cell, _Hsplit,dtype=tf.float32)
    return  tf.matmul(_LSTM_0[-1],_W["out"])+_b["out"]
#使用GPU按需增长模式
config = tf.ConfigProto(allow_soft_placement=True)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.3)
config.gpu_options.allow_growth = True
if __name__== "__main__":
    learning_rate = 0.001
    x     = tf.placeholder(dtype=tf.float32,shape=[None,28*28],name="input_x")
    y     = tf.placeholder(dtype=tf.float32,shape=[None,10],name="output_y")
    pred  = RNN(x,weight,biases,nsteps,"basic")
    cost  = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
    optm  = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
    accr  = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(pred,1),tf.argmax(y,1)),dtype=tf.float32))
    init  = tf.global_variables_initializer()
    print("RNN Already")

    training_epochs = 50
    batch_size = 16
    display_step = 1
    sess = tf.Session(config=config)
    sess.run(init)
    print("Start optimization")
    for epoch in range(training_epochs):
        avg_cost = 0.
        total_batch = int(mnist.train.num_examples/batch_size)
        #total_batch = 100
        # Loop over all batches
        for i in range(total_batch):
            batch_xs, batch_ys = mnist.train.next_batch(batch_size)
            # Fit training using batch data
            feeds = {x: batch_xs, y: batch_ys}
            sess.run(optm, feed_dict=feeds)
            # Compute average loss
            avg_cost += sess.run(cost, feed_dict=feeds) / total_batch
        # Display logs per epoch step
        if epoch % display_step == 0:
            print("Epoch: %03d/%03d cost: %.9f" % (epoch, training_epochs, avg_cost))
            feeds = {x: batch_xs, y: batch_ys}
            train_acc = sess.run(accr, feed_dict=feeds)
            print(" Training accuracy: %.3f" % (train_acc))
            feeds = {x: testimgs, y: testlabels}
            test_acc = sess.run(accr, feed_dict=feeds)
            print(" Test accuracy: %.3f" % (test_acc))
    print("Optimization Finished.")

• 没有训练结束,使用的GTX1060训练了大概8分钟,如果训练结束感觉应该可以达到97%左右
• 因为是单层网络,深度不够,也没处理数据~~
• 这只是简单了解RNN工作流程,和如何用TF操作RNN
• 以后会慢慢补上~~

参考:

• 唐迪宇课程,因为版本问题会出现很多代码更新
• 其它中间忘记记录了,如有侵权请联系博主,抱歉~