import osimport kerasimport timeimport numpy as npimport tensorflow as tffrom random import shufflefrom keras.utils import np_utilsfrom skimage import color, data, transform, io

trainDataDirList = os.listdir("F:\\MachineLearn\\ML-xiaoxueqi\\fruits\\trainGrayImage")trainDataList = []for i in range(len(trainDataDirList)): image = io.imread("F:\\MachineLearn\\ML-xiaoxueqi\\fruits\\trainGrayImage\\"+trainDataDirList[i]) trainDataList.append(image)trainLabelNum = np.load("F:\\MachineLearn\\ML-xiaoxueqi\\fruits\\trainLabelNum.npy")

testDataDirList = os.listdir("F:\\MachineLearn\\ML-xiaoxueqi\\fruits\\testGrayImage")testDataList = []for i in range(len(testDataDirList)): image = io.imread("F:\\MachineLearn\\ML-xiaoxueqi\\fruits\\testGrayImage\\"+testDataDirList[i]) testDataList.append(image)testLabelNum = np.load("F:\\MachineLearn\\ML-xiaoxueqi\\fruits\\testLabelNum.npy")

#乱序train_images = []train_labels = []index = [i for i in range(len(trainDataList))]shuffle(index)for i in range(len(index)): train_images.append(trainDataList[index[i]]) train_labels.append(trainLabelNum[index[i]])#将标签转码train_labels=keras.utils.to_categorical(train_labels,77)#保存处理后的数据np.save("E:\\tmp\\train_images",train_images)np.save("E:\\tmp\\train_labels",train_labels)

#加载上面保存的数据train77_images = np.load("E:\\train_images.npy")train77_labeles = np.load("E:\\train_labels.npy")

#变成四维训练数据，两维标签dataset = train77_images.reshape((-1, 64, 64, 1)).astype(np.float32)labels = train77_labeles

## 配置神经网络的参数n_classes = 77batch_size = 64kernel_h = kernel_w = 5#dropout = 0.8depth_in = 1depth_out1 = 64depth_out2 = 128image_size = 64 ##图片尺寸n_sample = len(dataset) ##样本个数

#每张图片的像素大小为64*64，训练样本x = tf.placeholder(tf.float32, [None, 64, 64, 1]) #训练样本对应的真实labely=tf.placeholder(tf.float32,[None,n_classes])

# y_ = tf.placeholder(tf.float32, [None, n_classes])

#设置dropout的placeholderdropout = tf.placeholder(tf.float32)

# 扁平化fla = int((image_size * image_size / 16) * depth_out2)

#卷积函数def inference(x, dropout): #第一层卷积 with tf.name_scope('convLayer1'): Weights = tf.Variable(tf.random_normal([kernel_h, kernel_w, depth_in, depth_out1])) bias = tf.Variable(tf.random_normal([depth_out1])) x = tf.nn.conv2d(x, Weights, strides=[1, 1, 1, 1], padding="SAME") x = tf.nn.bias_add(x, bias) conv1 = tf.nn.relu(x) #可视化权值 tf.summary.histogram('convLayer1/weights1', Weights) #可视化偏置 tf.summary.histogram('convLayer1/bias1', bias) #可视化卷积结果 tf.summary.histogram('convLayer1/conv1', conv1) #对卷积的结果进行池化 pool1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME") #可视化池化结果 tf.summary.histogram('ConvLayer1/pool1', pool1)

#第二层卷积 with tf.name_scope('convLayer2'): Weights = tf.Variable(tf.random_normal([kernel_h, kernel_w, depth_out1, depth_out2])) bias = tf.Variable(tf.random_normal([depth_out2])) x = tf.nn.conv2d(pool1, Weights, strides=[1, 1, 1, 1], padding="SAME") x = tf.nn.bias_add(x, bias) conv2 = tf.nn.relu(x) #可视化权值 tf.summary.histogram('convLayer2/weights2', Weights) #可视化偏置 tf.summary.histogram('convLayer2/bias2', bias) #可视化卷积结果 tf.summary.histogram('convLayer2/conv2', conv2) #对卷积的结果进行池化 pool2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME") #可视化池化结果 tf.summary.histogram('ConvLayer2/pool2', pool2)

#扁平化处理 flatten = tf.reshape(pool2, [-1, fla])

#第一层全连接 Weights = tf.Variable(tf.random_normal([int((image_size * image_size / 16) * depth_out2), 512])) bias = tf.Variable(tf.random_normal([512])) fc1 = tf.add(tf.matmul(flatten, Weights), bias) #使用relu激活函数处理全连接层结果 fc1r = tf.nn.relu(fc1)

#第二层全连接 Weights = tf.Variable(tf.random_normal([512, 128])) bias = tf.Variable(tf.random_normal([128])) fc2 = tf.add(tf.matmul(fc1r, Weights), bias) #使用relu激活函数处理全连接层结果 fc2 = tf.nn.relu(fc2) #使用Dropout（Dropout层防止预测数据过拟合） fc2 = tf.nn.dropout(fc2, dropout)

#输出预测的结果 Weights = tf.Variable(tf.random_normal([128, n_classes])) bias = tf.Variable(tf.random_normal([n_classes])) prediction = tf.add(tf.matmul(fc2, Weights), bias) return prediction

#使用上面定义好的神经网络进行训练,得到预测的labelprediction = inference(x, dropout)#定义损失函数，使用上面的预测label与真实的label作运算cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y)) #选定一个优化器和学习率（步长）optimizer = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)merged = tf.summary.merge_all()

#评估模型（准确率）correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

#初始会话并开始训练过程with tf.Session() as sess: tf.global_variables_initializer().run() for i in range(20): for j in range(int(n_sample / batch_size) + 1): start = (j * batch_size) end = start + batch_size x_ = dataset[start:end] y_ = labels[start:end] #准备验证数据 sess.run(optimizer, feed_dict={x: x_, y: y_, dropout: 0.5}) #计算当前块训练数据的损失和准确率 loss, acc = sess.run([cross_entropy, accuracy], feed_dict={x: x_, y: y_, dropout: 0.5}) print("Epoch:", '%04d' % (i + 1), "cost=", "{:.9f}".format(loss), "Training accuracy", "{:.5f}".format(acc*100)) print('Optimization Completed')