用tensorflow实现usps和mnist数据集的迁移学习

本程序环境：tensorflow+python，用到的库：numpy，os，Image，random。

基于论文：《Deep Transfer Network: Unsupervised Domain Adaptation》

前面我已经对这篇文章做过简单的导读：【深度学习】论文导读：无监督域适应（Deep Transfer Network: Unsupervised Domain Adaptation）

首先我们要用到两个数据集usps和mnist，它们都是用来完成手写数字识别任务的，以下提供两个数据集的下载链接

mnist数据集下载

usps数据集下载

上面两个链接下载下来，会发现mnist是以图片存储，而usps是以数值方式存储。这也是为了程序方便。

一.数据准备

首先需要读入usps数据集，因为usps是简单的以字符形式存储，所以读入也比较方便，程序如下，最后返回traindata矩阵以及trainlabel矩阵

def read_usps_dataset():
    filename= '数据集文件路径'
    fr = open(filename)
    numberOfLines = len(fr.readlines())
    traindataMat = zeros((numberOfLines,256))        #prepare matrix to return
    trainlabelMat = zeros((numberOfLines),dtype=int32)
    fr = open(filename)
    index = 0
    for line in fr.readlines():
        line = line.strip()                     #delete the /r/n
        listFromLine = line.split(' ')
        trainlabelMat[index] = listFromLine[0]
        traindataMat[index, :] = float32(listFromLine[1:])
        index += 1
    print "the size of source dataset:",numberOfLines
    trainlabelMat=array(trainlabelMat)
    traindataMat=array(traindataMat)/float32(2)
    trainlabelMat.astype(int)
    return traindataMat,trainlabelMat

然后处理mnist，这部分稍微麻烦一点，我们之所以使用mnist图片，是为了方便缩小，因为我们需要mnist数据和usps一起输入到神经网络，它们的维度应该是一样的。usps是16×16的大小，而mnist是28×28的大小，这里需要将mnist图片缩小到16×16

def preprocess_mnist():
    image=[]
    label=[]
    i=0
    for labels in range(10):
        pathDir =os.listdir('MNIST/trainimage/pic2/'+str(labels)+'/')
        for allDir in pathDir:
            child = os.path.join('%s%s' % ('MNIST/trainimage/pic2/'+str(labels)+'/', allDir))
            img = Image.open(child)
            img=img.resize((16, 16))
            img_array=array(img)
            img_array=img_array[:,:,0]
            img_array=reshape(img_array,-1)
            image.append(img_array)
            label.append(labels)
            i=i+1
    image = array(image)/float32(256)
    label = array(label)
    print "the size of target dataset:",i
    return image,label

数据预处理部分还没有结束，别忘了把数值型的标签转化成01型的

def dense_to_one_hot(labels_dense, num_classes):
  num_labels = labels_dense.shape[0]
  index_offset = arange(num_labels) * num_classes
  labels_one_hot = zeros((num_labels, num_classes))
  labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1
  return labels_one_hot

二.创建M矩阵

M矩阵用于计算empirical Maximum Mean Discrepancy（MMD），我们一开始就要将它初始化

def createMmetrix():
    mat1=tf.constant(float32(1)/(square(BATCHSIZE/2)),shape=[BATCHSIZE/2,BATCHSIZE/2],dtype=tf.float32)
    mat2=-mat1
    mat3=tf.concat(1,[mat1,mat2])
    mat4=tf.concat(1,[mat2,mat1])
    mat5=tf.concat(0,[mat3,mat4])
    return mat5

三.建立模型

采用两层卷积+全连接的结构，首先定义权值/偏置初始化函数，卷积/池化函数

def weight_variable(shape):
	initial = tf.truncated_normal(shape, stddev=0.1)
	return tf.Variable(initial)

def bias_variable(shape):
	initial = tf.constant(0.1, shape = shape)
	return tf.Variable(initial)

# convolution
def conv2d(x, W):
	return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
# pooling
def max_pool_2x2(x):
	return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

建立模型，注意obj_func的形式，我们已经将它替换成论文当中的公式了

#first convolutinal layer
w_conv1 = weight_variable([3, 3, 1, 32])
b_conv1 = bias_variable([32])
x_image = tf.reshape(X, [-1, 16, 16, 1])
h_conv1 = tf.nn.relu(conv2d(x_image, w_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

# second convolutional layer
w_conv2 = weight_variable([3, 3, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, w_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

# densely connected layer
w_fc1 = weight_variable([4*4*64, 256])
b_fc1 = bias_variable([256])
h_pool2_flat = tf.reshape(h_pool2, [-1, 4*4*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1) + b_fc1)

# softmax layer
w_fc2 = weight_variable([256, 10])
b_fc2 = bias_variable([10])
y_conv = tf.nn.softmax(tf.matmul(h_fc1, w_fc2) + b_fc2)

obj_func= -tf.reduce_sum(Y * tf.log(y_conv))+tf.constant(lamda,dtype=tf.float32)*tf.trace(tf.matmul(tf.matmul(h_fc1,M,transpose_a=True),h_fc1))+tf.constant(miu,dtype=tf.float32)*tf.trace(tf.matmul(tf.matmul(y_conv,M,transpose_a=True),y_conv))
optimizer = tf.train.GradientDescentOptimizer(learningrate).minimize(obj_func)

correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))