chenchao666 / JDDA-Master

Joint Domain Alignment and Discriminative Feature Learning for Unsupervised Deep Domain Adaptation（AAAI-2019）

JDDA-Master

Joint Domain Alignment and Discriminative Feature Learning for Unsupervised Deep Domain Adaptation

• This repository contains code for our paper Joint Domain Alignment and Discriminative Feature Learning for Unsupervised Deep Domain Adaptation Download Paper

• Another our domain adaptation paper with much better performance is also recommended "HoMM: Higher-order Moment Matching for Unsupervised Domain Adaptation", codes are avaliable at HoMM

• Another qualified repository completed by the co-author can be seen JDDA repository

• Our code contains not only our proposed JDDA, but also other famous distribution discrepancy measures MMD, DeepCoral, KMMD and CMD(Central Moment Discrepancy), LogCoral, ect.

• If you have any question about our paper or code, please don't hesitate to contact with me [e-mail], we will update our repository accordingly

Movition of our proposal

• Most of existing work only concentrates on learning shared feature representation by minimizing the distribution discrepancy across different domains. Due to the fact that all the domain alignment approaches can only reduce, but not remove the domain shift, target domain samples distributed near the edge of the clusters, or far from their corresponding class centers are easily to be misclassified by the hyperplane learned from the source domain. To alleviate this issue, we propose to joint domain alignment and discriminative feature learning, which could benefit both domain alignment and final classification. The necessity of joint domain alignment and discriminant features learning can be seen below

Result

• Our proposed JDDA achieves a state-of-art results among those completing Discrepancy-Based Domain Adaptation methods.

• The t-sne visualization, with the incorporation of our proposed discriminative loss, the deep features become better clusterred and more separable

Run The Code

• This code requires Python 2.7 and implemented in Tensorflow 1.9. You can download all the datasets used in our paper from Dataset_mnist Dataset_office and place them in the specified directory. Run trainLenet.py to obtain the results.

trainLenet.py

• the Core Code of our proposed Instance-Based and Center-Based discriminative feature learning can be seen in trainLenet.py

        ## Instence-Based Discriminative Feature Learning
        ## Xs is the deep features from the source domain with its row-number equals to batchsize and column-number equals to number of neurons in the adapted layer
	## self.W is the indicator matrix. self.W[i,j]=1 means i-th and j-th samples are from the same calss, self.W[i,j]=0 
	## means i-th and j-th samples are from difference calsses
    def CalDiscriminativeLoss(self,method):
        if method=="InstanceBased":
            Xs = self.source_model.fc4
            norm = lambda x: tf.reduce_sum(tf.square(x), 1)
            self.F0 = tf.transpose(norm(tf.expand_dims(Xs, 2) - tf.transpose(Xs)))  #calculate pair-wise distance of Xs
            margin0 = 0
            margin1 = 100
            F0=tf.pow(tf.maximum(0.0, self.F0-margin0),2)
            F1=tf.pow(tf.maximum(0.0, margin1-self.F0),2)
            self.intra_loss=tf.reduce_mean(tf.multiply(F0, self.W))
            self.inter_loss=tf.reduce_mean(tf.multiply(F1, 1.0-self.W))
            self.discriminative_loss = (self.intra_loss+self.inter_loss) / (self.BatchSize * self.BatchSize)


        ## Center-Based Discriminative Feature Learning, Note that the center_loss.py should be import 
	## Note that when using the Center-Based Discriminative Loss, the "global class center" should be also update in each iteration by using
	## with tf.control_dependencies([self.centers_update_op]):
        ##     self.solver = tf.train.AdamOptimizer(learning_rate=self.LearningRate).minimize(self.loss)
        elif method=="CenterBased":
            Xs=self.source_model.fc4
            labels=tf.argmax(self.source_label,1)
            self.inter_loss, self.intra_loss, self.centers_update_op = get_center_loss(Xs, labels, 0.5, 10)
            self.discriminative_loss = self.intra_loss+ self.inter_loss
            self.discriminative_loss=self.discriminative_loss/(self.ClassNum*self.BatchSize+self.ClassNum*self.ClassNum)

• If you want to use these methods, you can modify them in trainLenet.py. For the domain loss, the MMD, KMMD, mmatch LCORAL(LogCoral) are also included in our codes. For the DiscriminativeLoss, we have CenterBased and InstanceBased.

        self.CalDiscriminativeLoss(method="CenterBased")
        self.CalDomainLoss(method="CORAL")

• Note that the centers_update_op needs to be added in the control_dependencies when using the Center-Based method. When using the Instance-Based method you need to Comment with tf.control_dependencies([self.centers_update_op])

with tf.control_dependencies([self.centers_update_op]): ##comment this line when using Instance-Based method
    self.solver = tf.train.AdamOptimizer(learning_rate=self.LearningRate).minimize(self.loss)