Skip to content

Latest commit

 

History

History
121 lines (102 loc) · 5.39 KB

README.md

File metadata and controls

121 lines (102 loc) · 5.39 KB
Raw

HoMM: Higher-order Moment Matching for Unsupervised Domain Adaptation (AAAI-2020)

HoMM-Master

  • This repository contains code for our paper HoMM: Higher-order Moment Matching for Unsupervised Domain Adaptation Download paper here
  • If you have any question about our paper or code, please don't hesitate to contact with me [email protected], we will update our repository accordingly

Setup

Training

  • MNIST You can run TrainLenet.py in HoMM-mnist.
  • Office&Office-Home You can run finetune.py in HoMM_office/resnet/.
  • We have provide four functions HoMM3, HoMM4, HoMM and KHoMM conresponding to the third-order HoMM, fourth-order HoMM, Arbitrary-order moment matching, and Kernel HoMM.

Reimplement HoMM in your work

  • Readers can reimplement the HoMM in their work very easily by using the following function.
  • In our code, xs and xt denotes source and target deep features in the adapted layer. the dimension of xs and xt is b*L where b is the batchsize and L is the number of neurons in the adapted layer. num denotes the N in our paper, which indicates the number of sampled values in the high-level tensor.
  • It is worth noting that the relu activation function can not be applied to the adapted layer, as relu activation function will make most of the values in the high-level tensor to be zero, which will make our HoMM fail. Therefore, we adopt tanh activation function in the adapted layer

HoMM3

def HoMM3_loss(self, xs, xt):
        xs = xs - tf.reduce_mean(xs, axis=0)
        xt = xt - tf.reduce_mean(xt, axis=0)
        xs=tf.expand_dims(xs,axis=-1)
        xs = tf.expand_dims(xs, axis=-1)
        xt = tf.expand_dims(xt, axis=-1)
        xt = tf.expand_dims(xt, axis=-1)
        xs_1=tf.transpose(xs,[0,2,1,3])
        xs_2 = tf.transpose(xs, [0, 2, 3, 1])
        xt_1 = tf.transpose(xt, [0, 2, 1, 3])
        xt_2 = tf.transpose(xt, [0, 2, 3, 1])
        HR_Xs=xs*xs_1*xs_2   # dim: b*L*L*L
        HR_Xs=tf.reduce_mean(HR_Xs,axis=0)   #dim: L*L*L
        HR_Xt = xt * xt_1 * xt_2
        HR_Xt = tf.reduce_mean(HR_Xt, axis=0)
        return tf.reduce_mean(tf.square(tf.subtract(HR_Xs, HR_Xt)))
  • HoMM4
  • The adapted layer has 90 neurons, we divided them into 3 group with each group 30 neurons.
def HoMM4(self,xs,xt):
	ind=tf.range(tf.cast(xs.shape[1],tf.int32))
	ind=tf.random_shuffle(ind)
	xs=tf.transpose(xs,[1,0])
	xs=tf.gather(xs,ind)
	xs = tf.transpose(xs, [1, 0])
	xt = tf.transpose(xt, [1, 0])
	xt = tf.gather(xt, ind)
	xt = tf.transpose(xt, [1, 0])
	return self.HoMM4_loss(xs[:,:30],xt[:,:30])+self.HoMM4_loss(xs[:,30:60],xt[:,30:60])+self.HoMM4_loss(xs[:,60:90],xt[:,60:90])



def HoMM4_loss(self, xs, xt):
	xs = xs - tf.reduce_mean(xs, axis=0)
	xt = xt - tf.reduce_mean(xt, axis=0)
	xs = tf.expand_dims(xs,axis=-1)
	xs = tf.expand_dims(xs, axis=-1)
	xs = tf.expand_dims(xs, axis=-1)
	xt = tf.expand_dims(xt, axis=-1)
	xt = tf.expand_dims(xt, axis=-1)
	xt = tf.expand_dims(xt, axis=-1)
	xs_1 = tf.transpose(xs,[0,2,1,3,4])
	xs_2 = tf.transpose(xs, [0, 2, 3, 1,4])
	xs_3 = tf.transpose(xs, [0, 2, 3, 4, 1])
	xt_1 = tf.transpose(xt, [0, 2, 1, 3,4])
	xt_2 = tf.transpose(xt, [0, 2, 3, 1,4])
	xt_3 = tf.transpose(xt, [0, 2, 3, 4, 1])
	HR_Xs=xs*xs_1*xs_2*xs_3    # dim: b*L*L*L*L
	HR_Xs=tf.reduce_mean(HR_Xs,axis=0)  # dim: L*L*L*L
	HR_Xt = xt * xt_1 * xt_2*xt_3
	HR_Xt = tf.reduce_mean(HR_Xt, axis=0)
	return tf.reduce_mean(tf.square(tf.subtract(HR_Xs, HR_Xt)))
  • Arbitrary-order Moment Matching
def HoMM(self,xs, xt, order=3, num=300000):
	xs = xs - tf.reduce_mean(xs, axis=0)
	xt = xt - tf.reduce_mean(xt, axis=0)
	dim = tf.cast(xs.shape[1], tf.int32)
	index = tf.random_uniform(shape=(num, dim), minval=0, maxval=dim - 1, dtype=tf.int32)
	index = index[:, :order]
	xs = tf.transpose(xs)
	xs = tf.gather(xs, index)  ##dim=[num,order,batchsize]
	xt = tf.transpose(xt)
	xt = tf.gather(xt, index)
	HO_Xs = tf.reduce_prod(xs, axis=1)
	HO_Xs = tf.reduce_mean(HO_Xs, axis=1)
	HO_Xt = tf.reduce_prod(xt, axis=1)
	HO_Xt = tf.reduce_mean(HO_Xt, axis=1)
	return tf.reduce_mean(tf.square(tf.subtract(HO_Xs, HO_Xt)))

Results

Citation

  • If you find it helpful for you, please cite our paper
@inproceedings{chen2020HoMM,
  title={HoMM: Higher-order Moment Matching for Unsupervised Domain Adaptation},
  author={Chao Chen, Zhihang Fu, Zhihong Chen, Sheng Jin, Zhaowei Cheng, Xinyu Jin, Xian-Sheng Hua},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={34},
  year={2020}
}