import tensorflow as tfimport pandas as pdimport numpy as npfrom sklearn.metrics import confusion_matrixfrom tensorflow.python.ops import rnn,rnn_cellimport timefrom datetime import timedeltaimport mathimport randomimport osimport cv2from utils import *num_channels = 3img_size = 227num_unitlstm=64img_size_flat = img_size * img_size * num_channelsimg_shape = (img_size, img_size)train_tfdata=['cifar10_train_1.tfrecords','cifar10_train_2.tfrecords']test_tfdata=['cifar10_test.tfrecords']classes=range(10)num_classes = len(classes)starter_learning_rate = 0.001#test_path = data_dirglobal_step = tf.Variable(0, trainable=False)x = tf.placeholder(tf.float32,shape=[None,img_size,img_size,num_channels])x_image = tf.reshape(x, [-1, img_size, img_size, num_channels])y= tf.placeholder(tf.float32,shape=[None])y_true= tf.one_hot(tf.cast(y,tf.int32),num_classes)y_true_cls = tf.argmax(y_true, dimension=1)phase_train=tf.placeholder(bool)keep_prob = tf.placeholder(tf.float32)print"creating network model"with tf.name_scope('conv1') as scope: kernel = tf.get_variable(name='conv1',shape=[11,11,3,96],initializer=tf.contrib.layers.xavier_initializer_conv2d()) conv = tf.nn.conv2d(x, kernel, [1, 4, 4, 1], padding='SAME') biases = tf.Variable(tf.constant(0.0, shape=[96], dtype=tf.float32),trainable=True, name='biases') bias = tf.nn.bias_add(conv, biases) bias=batch_norm(bias, phase_train) conv1 = tf.nn.relu(bias, name=scope)pool1 = tf.nn.max_pool(conv1,ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1],padding='VALID',name='pool1')# conv2with tf.name_scope('conv2') as scope: kernel = tf.get_variable('conv2',shape=[5,5,96,256],initializer=tf.contrib.layers.xavier_initializer_conv2d()) conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1], padding='SAME') biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),trainable=True, name='biases') bias = tf.nn.bias_add(conv, biases) bias=batch_norm(bias, phase_train) conv2 = tf.nn.relu(bias, name=scope)# pool2pool2 = tf.nn.max_pool(conv2,ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1],padding='VALID',name='pool2')# conv3with tf.name_scope('conv3') as scope: kernel = tf.get_variable('conv3',shape=[3,3,256,384],initializer=tf.contrib.layers.xavier_initializer_conv2d()) conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME') biases = tf.Variable(tf.constant(0.0, shape=[384], dtype=tf.float32),trainable=True, name='biases') bias = tf.nn.bias_add(conv, biases) bias=batch_norm(bias, phase_train) conv3 = tf.nn.relu(bias, name=scope)# conv4with tf.name_scope('conv4') as scope: kernel = tf.get_variable('conv4',shape=[3,3,384,384],initializer=tf.contrib.layers.xavier_initializer_conv2d()) conv = tf.nn.conv2d(conv3, kernel, [1, 1, 1, 1], padding='SAME') biases = tf.Variable(tf.constant(0.0, shape=[384], dtype=tf.float32),trainable=True, name='biases') bias = tf.nn.bias_add(conv, biases) bias=batch_norm(bias, phase_train) conv4 = tf.nn.relu(bias, name=scope)# conv5with tf.name_scope('conv5') as scope: kernel = tf.get_variable('conv5',shape=[3,3,384,256],initializer=tf.contrib.layers.xavier_initializer_conv2d()) conv = tf.nn.conv2d(conv4, kernel, [1, 1, 1, 1], padding='SAME') biases = tf.Variable(tf.constant(0.0, shape=[256], dtype=tf.float32),trainable=True, name='biases') bias = tf.nn.bias_add(conv, biases) bias=batch_norm(bias, phase_train) conv5 = tf.nn.relu(bias, name=scope) pool5 = tf.nn.max_pool(conv5,ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1],padding='VALID',name='pool5')layer_flat, num_features = flatten_layer(pool5)layer_fc1 = new_fc_layer(input=layer_flat, num_inputs=num_features,num_outputs=num_features,name='fc1',use_relu=True)layer_fc1_split=tf.split(1,256,layer_fc1)with tf.variable_scope('lstm'): lstm=rnn_cell.BasicLSTMCell(num_unitlstm,forget_bias=1.0)with tf.variable_scope('RNN'): output,state=rnn.rnn(lstm,layer_fc1_split,dtype=tf.float32)rnn_output=tf.concat(1,output)layer_fc3 = new_fc_layer(input=rnn_output,num_inputs=num_unitlstm*256,num_outputs=num_classes,name='fc3',use_relu=False)y_pred = tf.nn.softmax(layer_fc3)cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=layer_fc3,labels=y_true)cost = tf.reduce_mean(cross_entropy)tf.summary.scalar('cost', cost)#learning_rate = tf.train.exponential_decay(starter_learning_rate, global_step,# 10000, 0.1, staircase=True)#tf.summary.scalar('learning_rate', learning_rate)#optimizer =tf.train.MomentumOptimizer(learning_rate,momentum=0.9).minimize(cost, global_step=global_step)optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)y_pred_cls = tf.argmax(y_pred, dimension=1)correct_prediction = tf.equal(y_pred_cls, y_true_cls)accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))tf.summary.scalar('accuracy', accuracy)train_img,train_label=load_tfrecords(train_tfdata,img_size,True)test_img,test_label=load_tfrecords(test_tfdata,img_size,False)train_img_batch,train_label_batch=tf.train.shuffle_batch([train_img,train_label],batch_size=64,capacity=3000,min_after_dequeue=1000)test_img_batch,test_label_batch=tf.train.shuffle_batch([test_img,test_label],batch_size=32,capacity=3000,min_after_dequeue=1000)print" starting Session"sess = tf.Session()sess.run(tf.global_variables_initializer())saver = tf.train.Saver()summary_op = tf.summary.merge_all()train_summary_writer = tf.summary.FileWriter('rnnsummary' + 'train', sess.graph)validation_summary_writer = tf.summary.FileWriter('rnnsummary'+ 'validation', sess.graph)threads=tf.train.start_queue_runners(sess=sess)print "loadinf train_batch"num_iterations=25000total_iterations = 0start_time = time.time()acc_total=[]for i in range(total_iterations,total_iterations + num_iterations): train_batch = sess.run([train_img_batch, train_label_batch]) feed_dict_train={x: train_batch[0], y: train_batch[1],phase_train:True,keep_prob:0.5} summary_train,_,train_acc=sess.run([summary_op,optimizer,accuracy], feed_dict=feed_dict_train) train_summary_writer.add_summary(summary_train, i) if (i+1) % 20 == 0: test_batch = sess.run([test_img_batch,test_label_batch]) feed_dict_validate={x: test_batch[0], y: test_batch[1],phase_train:False,keep_prob:1} summary_valid,val_loss ,val_acc= sess.run([summary_op,cost,accuracy], feed_dict=feed_dict_validate) validation_summary_writer.add_summary(summary_valid, i) msg = "Epoch {0} --- Training Accuracy: {1:>6.1%}, Validation Accuracy: {2:>6.1%}, Validation Loss: {3:.3f}" print(msg.format(i + 1, train_acc, val_acc, val_loss)) acc_total.append(val_acc) if(i+1)%10000==0: save_path = saver.save(sess, "model.ckpt")np.save('acc_total',acc_total) end_time = time.time()time_dif = end_time - start_timeprint("Time elapsed: " + str(timedelta(seconds=int(round(time_dif))))) 
