学习进度笔记

学习进度笔记12

TensorFlow双向循环神经网络

from __future__ import print_function  

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_rnn",one_hot=True)  

#Traning Parameters  

learning_rate=0.001  

training_step=10000  

batch_size=128  

display_step=400  

#Network Parmeters  

num_input=28  

timestep=28  

num_hidden=128  

num_classes=10  

#tf Graph input  

X=tf.placeholder("float32",[None,timestep,num_input])  

Y=tf.placeholder("float32",[None,num_classes])  

#Define weights  

weights={  

    'out':tf.Variable(tf.random_normal([2*num_hidden,num_classes]))  

}  

biases={  

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

}  

  

def BiRNN(X,weights,biases):  

    x=tf.unstack(X,timestep,1)  

    #define lstm cells with tensorflow  

    #Forward direction cell  

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

    #Backward direction cell  

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

    #Get lstm cell output  

    try:  

        outputs,_,_=rnn.static_bidirectional_rnn(lstm_fw_cell,lstm_bw_cell,x,dtype=tf.float32)  

    except Exception:  

        outputs=rnn.static_bidirectional_rnn(lstm_fw_cell,lstm_bw_cell,x,dtype=tf.float32)  

    # Linaer activation,using rnn inner loop last output  

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

  

logits=BiRNN(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  

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

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

#Initialize Variable  

init=tf.global_variables_initializer()  

#start training  

with tf.Session() as sess:  

    # Run the initializer  

    sess.run(init)  

  

    for step in range(1,training_step+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,timestep,num_input))  

        # Run optimizetion 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)+ ",Minbatch 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,timestep,num_input))  

    test_label=mnist.test.labels[:test_len]  

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