学习进度笔记

学习进度笔记11

TensorFlow循环神经网络

import numpy as np  

import tensorflow as tf  

from tensorflow.contrib import rnn  

from tensorflow.examples.tutorials.mnist import input_data  

import os  

os.environ["CUDA_VISIBLE_DEVICES"]="0"  

  

mnist=input_data.read_data_sets("/home/yxcx/tf_data",one_hot=True)  

 

#Training Parameters  

learning_rate=0.001  

training_steps=10000  

batch_size=128  

display_step=200  

 

#Network Parameters  

num_input=28  

timesteps=28  

num_hidden=128  

num_classes=10  

 

#tf Graph input  

X=tf.placeholder("float",[None,timesteps,num_input])  

Y=tf.placeholder("float",[None,num_classes])  

 

# Define weights  

weights={  

    'out':tf.Variable(tf.random_normal([num_hidden,num_classes]))  

}  

biases={  

    'out':tf.Variable(tf.random_normal([num_classes]))  

}  

  

def RNN(x,weights,biases):  

    x=tf.unstack(x,timesteps,1)  

    #define a lstm cell with tensorflow  

    lstm_cell=rnn.BasicLSTMCell(num_hidden,forget_bias=1.0)  

 

    #Get lstm cell ouput  

    outputs,states=rnn.static_rnn(lstm_cell,x,dtype=tf.float32)  

 

    #Linear activation ,using rnn inner loop last output  

    return tf.matmul(outputs[-1],weights['out'])+biases['out']  

  

logits=RNN(X,weights,biases)  

prediction=tf.nn.softmax(logits)  

 

#Define loss and  optimizer  

loss_op=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(  

    logits=logits,labels=Y  

))  

optimizer=tf.train.GradientDescentOptimizer(learning_rate=learning_rate)  

train_op=optimizer.minimize(loss_op)  

 

#Evaluate model(with test logits,for dropout to be disabled)  

corrent_pred=tf.equal(tf.argmax(prediction,1),tf.argmax(Y,1))  

accuracy=tf.reduce_mean(tf.cast(corrent_pred,tf.float32))  

 

#Initialize the variables  

init=tf.global_variables_initializer()  

 

#Start Training  

with tf.Session() as sess:  

    # Run the initializer  

    sess.run(init)  

    for step in range(1,training_steps+1):  

        batch_x,batch_y=mnist.train.next_batch(batch_size)  

        # Reshape data to get 28 seq of 28 elements  

        batch_x=batch_x.reshape((batch_size,timesteps,num_input))  

        #Run optimization op  

        sess.run(train_op,feed_dict={X:batch_x,Y:batch_y})  

        if step % display_step ==0 or step==1:  

            #Calculate batch loss and accuracy  

            loss,acc=sess.run([loss_op,accuracy],feed_dict={X:batch_x,Y:batch_y})  

            print('Step'+str(step)+" ,Minibatch Loss"+"{:.4f}".format(loss)+  

                  ",Training Accuracy="+"{:.3f}".format(acc))  

    print("Optimization Finished!")  

 

    #Calculate accuracy for 128 mnist test images  

    test_len=128  

    test_data=mnist.test.images[:test_len].reshape((-1,timesteps,num_input))  

    test_label=mnist.test.labels[:test_len]  

    print("Testing Accuracy:",sess.run(accuracy,feed_dict={X:test_data,Y:test_label}))