txt/2104.10602.txt

Visualizing Adapted Knowledge in Domain Transfer
Yunzhong Hou Liang Zheng
Australian National University
ffirstname.lastname [email protected]
Abstract
A source model trained on source data and a target
model learned through unsupervised domain adaptation
(UDA) usually encode different knowledge. To understand
the adaptation process, we portray their knowledge dif-
ference with image translation. Specifically, we feed a
translated image and its original version to the two mod-
els respectively, formulating two branches. Through up-
dating the translated image, we force similar outputs from
the two branches. When such requirements are met, dif-
ferences between the two images can compensate for and
hence represent the knowledge difference between models.
To enforce similar outputs from the two branches and de-
pict the adapted knowledge, we propose a source-free im-
age translation method that generates source-style images
using only target images and the two models. We visual-
ize the adapted knowledge on several datasets with differ-
ent UDA methods and find that generated images success-
fully capture the style difference between the two domains.
For application, we show that generated images enable fur-
ther tuning of the target model without accessing source
data. Code available at https://github.com/hou-
yz/DA_visualization .
1. Introduction
Domain transfer or domain adaptation aims to bridge
the distribution gap between source and target domains.
Many existing works study the unsupervised domain adap-
tation (UDA) problem, where the target domain is unla-
beled [27, 6, 46, 1, 11]. In this process, we are interested
in what knowledge neural networks learn and adapt.
Essentially, we should visualize the knowledge differ-
ence between models: a source model trained on the source
domain, and a target model learned through UDA for the
target domain. We aim to portray the knowledge difference
with image generation. Given a translated image and its
original version, we feed the two images to the source and
the target model, respectively. It is desired that differences
between image pairs can compensate for the knowledge dif-
(a) Target images (real-world)
(b) Generated source-style images
(c) Unseen source images (synthetic)
Figure 1: Visualization of adapted knowledge in unsuper-
vised domain adaptation (UDA) on the VisDA dataset [38].
To depict the knowledge difference, in our source-free im-
age translation (SFIT) approach, we generate source-style
images (b) from target images (a). Instead of accessing
source images (c), the training process is guided entirely
by the source and target models, so as to faithfully portray
the knowledge difference between them.
ference between models, leading to similar outputs from
the two branches (two images fed to two different models).
Achieving this, we could also say that the image pair repre-
sent the knowledge difference.
This visualization problem is very challenging and
heretofore yet to be studied in the literature. It focuses on a
relatively understudied field in transfer learning, where we
distill knowledge differences from models and embed it in
generated images . A related line of works, traditional image
translation, generates images in the desired style utilizing
content images and style images [7, 13, 48], and is applied
1arXiv:2104.10602v2  [cs.CV]  1 May 2021in pixel-level alignment methods for UDA [26, 2, 44, 11].
However, relying on images from both domains to indicate
the style difference, such works cannot faithfully portray
the knowledge difference between source and target models ,
and are unable to help us understand the adaptation process.
In this paper, we propose a source-free image translation
(SFIT) approach, where we translate target images to the
source style without using source images. The exclusion of
source images prevents the system from relying on image
pairs for style difference indication, and ensures that the
system only learns from the two models . Specifically, we
feed translated source-style images to the source model and
original target images to the target model, and force similar
outputs from these two branches by updating the generator
network. To this end, we use the traditional knowledge dis-
tillation loss and a novel relationship preserving loss, which
maintains relative channel-wise relationships between fea-
ture maps. We show that the proposed relationship preserv-
ing loss also helps to bridge the domain gap while chang-
ing the image style, further explaining the proposed method
from a domain adaptation point of view. Some results of
our method are shown in Fig. 1. We observe that even un-
der the source-free setting, knowledge from the two models
can still power the style transfer from the target style to the
source style (SFIT decreases color saturation and whitens
background to mimic the unseen source style).
On several benchmarks [19, 36, 39, 38], we show that
generated images from the proposed SFIT approach signifi-
cantly decrease the performance gap between the two mod-
els, suggesting a successful distillation of adapted knowl-
edge. Moreover, we find SFIT transfers the image style
at varying degrees, when we use different UDA methods
on the same dataset. This further verifies that the SFIT
visualizations are faithful to the models and that different
UDA methods can address varying degrees of style differ-
ences. For applications, we show that generated images can
serve as an additional cue and enable further tuning of target
models. This also falls into a demanding setting of UDA,
source-free domain adaptation (SFDA) [17, 20, 24], where
the system has no access to source images.
2. Related Work
Domain adaptation aims to reduce the domain gap be-
tween source and target domains. Feature-level distribution
alignment is a popular strategy [27, 6, 46, 40]. Long et
al. [27] use the maximum mean discrepancy (MMD) loss
for this purpose. Tzeng et al. [46] propose an adversarial
method, ADDA, with a loss function based on the gen-
erative adversarial network (GAN). Pixel-level alignment
with image translation is another popular choice in UDA
[26, 2, 44, 42, 1, 11]. Hoffman et al . propose the Cy-
CADA [11] method based on CycleGAN [48] image trans-
lation. Other options are also investigated. Saito et al. [40]align the task-specific decision boundaries of two classi-
fiers. Source-free domain adaptation (SFDA) does notuse
the source data and therefore greatly alleviates the privacy
concerns in releasing the source dataset. As an early at-
tempt, AdaBN [22] adapts the statistics of the batch normal-
ization layers in the source CNN to the target domain. Li et
al. [20] generate images with the same distribution of the
target images and use them to fine-tune the classifier. Liang
et al. [24] fine-tune a label smoothed [34] source model on
the target images. To the authors’ knowledge, there is still
yet to be any visualization that can indicate what models
learn during adaptation.
Knowledge distillation transfers knowledge from a pre-
trained teacher model to a student model [10], by maxi-
mizing the mutual information between teacher outputs and
student outputs. Some existing works consider the relation-
ship between instance or pixels for better distillation per-
formance [45, 23, 37]. Instead of distilling teacher knowl-
edge on a given training dataset, data-free knowledge dis-
tillation (DFKD) [30, 35, 3, 33, 8, 47] first generates train-
ing data and then learns a student network on this gener-
ated dataset. Training data can be generated by aligning
feature statistics [30, 8, 47], enforcing high teacher confi-
dence [30, 35, 3, 8, 47], and adversarial generation of hard
examples for the student [33, 47]. In [8, 47], batch normal-
ization statistics are matched as regularization. Our work,
while also assuming no access to source images, differs sig-
nificantly from these works in that our image translation
has to portray the transferred knowledge, whereas data-free
knowledge distillation just generates whatever images that
satisfy the teacher networks.
Image translation renders the same content in a differ-
ent artistic style. Some existing works adopt a GAN-based
system for this task [26, 44, 14, 48, 11], while others use a
pre-trained feature extractor for style transfer [7, 15, 32, 13].
Zhuet al. adopt a cycle consistency loss in the image trans-
lation loop to train the CycleGAN system [48]. Gatys
et al . consider a content loss on high-level feature maps,
and a style loss on feature map statistics for style transfer
[7]. Huang and Belongie [13] propose a real-time AdaIN
style transfer method by changing the statistics in instance
normalization layers. Based on AdaIN, Karras et al. pro-
pose StyleGAN for state-of-the-art image generation [16].
Our work differs from traditional image translations in that
rather than images from the two domains, only models from
two domains are used to guide the image update.
3. Problem Formulation
To achieve our goal, i.e.,visualizing adapted knowledge
in UDA, we translate a image xfrom a certain domain to
a new imageex. It is hoped that feeding the original image
to its corresponding model (trained for that certain domain)
and the generated image to the other model can minimize
2target image 𝒙𝒙 generated image �𝒙𝒙generatorsource
CNN
target
CNNrelationship 
preserving loss
classifier classifierknowledge
distillation lossFigure 2: The proposed source-free image translation (SFIT) method for visualizing the adapted knowledge in UDA. The
system includes two branches: original target images are fed to the target CNN, whereas generated source-style images are
fed to the source CNN. We minimize the knowledge distillation loss and the relationship preserving loss, and update the
generator network accordingly. If the two branches get similar results while adopting different models, then the difference
between the original target image xand the generated source-style image exshould be able to mitigate and therefore exhibit
the knowledge difference between models. Dashed lines indicate fixed network parameters.
the output difference between these two branches. The up-
date process is directed only by the source model fS(
)and
the target model fT(
), and we prevent access to the images
from the other domain to avoid distractions. We formulate
the task of visualizing adapted knowledge as a function of
the source model, the target model, and the image from a
certain domain,
G(fS; fT;x)!ex: (1)
In contrast, traditional image translation needs access to im-
ages from both domains for content and style specification.
In addition to the source image xSand the target image xT,
traditional image translation also relies on certain neural
network d(
)as the criterion. Instead of the source and tar-
get models, ImageNet [4] pre-trained VGG [43] and adver-
sarially trained discriminator networks are used for this task
in style transfer [7, 13] and GAN-based methods [48, 11],
respectively. Traditional image translation task can thus be
formulated as,
G(d;xS;xT)!ex: (2)
Comparing our goal in Eq. 1 and traditional image transla-
tion in Eq. 2, we can see a clear gap between them. Tradi-
tional image translation learns the style difference indicated
byimages from both domains, whereas our goal is to learn
to visualize the knowledge difference between the source
and target models fS(
); fT(
).
4. Method
To investigate what neural networks learn in do-
main adaptation, we propose source-free image translation
(SFIT), a novel method that generates source-style images
from original target images, so as to mitigate and represent
the knowledge difference between models.4.1. Overview
Following many previous UDA works [6, 27, 46, 24], we
assume that only the feature extractor CNN in the source
model is adapted to the target domain. Given a source CNN
fS(
)and a target CNN fT(
)sharing the same classifier
p(
), we train a generator g(
)for the SFIT task. We discuss
why we choose this translation direction in Section 4.3. As
the training process is source-free, for simplicity, we refer
to the target image as xinstead of xTin what follows.
As shown in Fig. 2, given a generated image ex=g(x),
the source model outputs a feature map fS(ex)and a prob-
ability distribution p(fS(ex))over all Cclasses. To depict
the adapted knowledge in the generated image, in addition
to the traditional knowledge distillation loss, we introduce a
novel relationship preserving loss, which maintains relative
channel-wise relationships between the target-image-target-
model feature map fT(x)and the generated-image-source-
model feature map fS(ex).
4.2. Loss Functions
With a knowledge distillation loss LKDand a relationship
preserving lossLRP, we have the overall loss function,
L=LKD+LRP: (3)
In the following sections, we detail the loss terms.
Knowledge distillation loss. In the proposed source-
free image translation method, portraying the adapted
knowledge in the target model fT(
)with source model
and generator combined fS(g(
))can be regarded as a spe-
cial case of knowledge distillation, where we aim to distill
the adapted knowledge to the generator. In this case, we
include a knowledge distillation loss between generated-
image-source-model output p(fS(ex))and target-image-
3target-model output p(fT(x)),
LKD=DKL(p(fT(x)); p(fS(ex))); (4)
whereDKL(
;
)denotes the Kullback-Leibler divergence.
Relationship preserving loss. Similar classification
outputs indicate a successful depiction of the target model
knowledge on the generated images. As we assume a fixed
classifier for UDA, the global feature vectors from the tar-
get image target CNN and the generated image source CNN
should be similar after a successful knowledge distillation.
Promoting similar channel-wise relationships between fea-
ture maps fT(x)andfS(ex)helps to achieve this goal.
Previous knowledge distillation works preserve relative
batch-wise or pixel-wise relationships [45, 23]. However,
they are not suitable here for the following reasons. Relative
batch-wise relationships can not effectively supervise the
per-image generation task. Besides, the efficacy of pixel-
wise relationship preservation can be overshadowed by the
global pooling before the classifier. By contrast, channel-
wise relationships are computed on a per-image basis, and
are effective even after global pooling. As such, we choose
the channel-wise relationship preserving loss that is com-
puted in the following manner.
Given feature maps fT(x); fS(ex), we first reshape them
into feature vectors FSandFT,
fS(ex)2RDHW!F S2RDHW;
fT(x)2RDHW!F T2RDHW;(5)
where D; H , and Ware the feature map depth (number of
channels), height, and width, respectively. Next, we calcu-
late their channel-wise self correlations, or Gram matrices,
GS=FS
FT
S; G T=FT
FT
T; (6)
where GS; GT2RDD. Like other similarity preserving
losses for knowledge distillation [45, 23], we then apply the
row-wiseL2normalization,
eGS[i;:]=GS[i;:]

GS[i;:]
2;eGT[i;:]=GT[i;:]

GT[i;:]
2; (7)
where [i;:]indicates the i-th row in a matrix. At last, we
define the relationship preserving loss as mean square error
(MSE) between the normalized Gram matrices,
LRP=1
D


eGSeGT


2
F; (8)
wherek
kFdenotes the Frobenius norm (entry-wise L2
norm for matrix). In Section 4.3, we further discuss the rela-
tionship preserving loss from the viewpoint of style transfer
and domain adaptation, and show it can align feature map
distributions in a similar way as style loss [7] for style trans-
fer and MMD loss [27] for UDA, forcing the generator to
portray the knowledge difference between the two models.
(a) Relationship preserving loss
 (b) Traditional style loss
Figure 3: Comparison between the proposed relationship
preserving loss and the traditional style loss. In (a) and (b),
given 256-dimensional feature maps, we show differences
of row-wise normalized Gram matrix (Eq. 8) and original
Gram matrix (Eq. 9). Deeper colors indicate larger dif-
ferences and therefore stronger supervision. The proposed
relationship preserving loss provides evenly distributed su-
pervision for all channels, whereas the traditional style loss
focuses primarily on several channels.
4.3. Discussions
Why transfer target images to the source style. Ac-
cording to the problem formulation in Eq. 1, we should be
able to visualize the adapted knowledge by generating ei-
ther source-style images from target images, or target-style
images from source images. In this paper, we select the for-
mer direction as it might be further applied in fine-tuning
the target model (see Section 5.4 for application).
Style transfer with the relationship preserving loss.
The proposed relationship preserving loss can be regarded
as a normalized version of the traditional style loss intro-
duced by Gatys et al. [7],
Lstyle=1
D2kGSGTk2
F; (9)
which computes MSE between Gram matrices.
In the proposed relationship preserving loss (Eq. 8), in-
stead of original Gram matrices, we use a row-wise normal-
ized version. It focuses on relative relationships between
channels, rather than absolute values of self correlations as
in the traditional style loss. Preserving relative relation-
ships provides more evenly-distributed supervision for all
channels, instead of prioritizing several channels as in the
traditional style loss (Fig. 3). Experiments find this evenly-
distributed supervision better preserves the foreground ob-
ject and allows for easier training and higher performance,
while also changing the image style (see Section 5.5).
Distribution alignment with the relationship preserv-
ing loss. As proved by Li et al. [21], the traditional style
4lossLstyleis equivalent to the MMD loss [27] for UDA. We
can also see the relationship preserving loss as a modified
version of the MMD loss, which aligns the distribution of
the generated image source CNN feature map fS(ex)to the
target image target CNN feature map fT(x).
5. Experiments
5.1. Datasets
We visualize the knowledge difference between source
and target models on the following datasets.
Digits is a standard UDA benchmark that focuses on
10-class digit recognition. Specifically, we experiment on
MNIST [19], USPS, and SVHN [36] datasets.
Office-31 [39] is a standard benchmark for UDA that
contains 31 classes from three distinct domains: Amazon
(A), Webcam (W), and DSLR (D).
VisDA [38] is a challenging large-scale UDA benchmark
for domain adaptation from 12 classes of synthetic CAD
model images to real-world images in COCO [25].
5.2. Implementation Details
Source and target models. We adopt source and tar-
get models from a recent SFDA work SHOT-IM [24] if not
specified. SFDA is a special case of UDA, and it is even
more interesting to see what machines learn in the absence
of source data. We also include UDA methods DAN [27]
and ADDA [46] for SFIT result comparisons. For network
architectures, on digits dataset, following Long et al. [28],
we choose a LeNet [18] classifier. On Office-31 and VisDA,
we choose ResNet-50 and ResNet-101 [9], respectively.
Generator for SFIT. We use a modified CycleGAN [48]
architecture with 3 residue blocks due to memory concerns.
Training schemes. During training, we first initialize
the generator as a transparent filter, which generates im-
ages same as the original input. To this end, we use the
ID lossLID=kexxk1and the content loss Lcontent =
kfS(ex)fS(x)k2to train the generator for initialization.
The initialization performance is shown in Table 4, where
we can see a mild 1.9% accuracy drop from original tar-
get images. Then, we train the generator with the overall
loss function in Eq. 3 for visualizing the adapted knowl-
edge. Specifically, we use an Adam optimizer with a cosine
decaying [31] learning rate starting from 3104and a
batch size of 16. All experiments are finished using one
RTX-2080Ti GPU.
5.3. Evaluation
Recognition accuracy on generated images. To ex-
amine whether the proposed SFIT method can depict the
knowledge difference, in Table 1-3, we report recogni-
tion results using the generated-image-source-model branch
(referred as “generated images”). On the digits dataset,Method SVHN!MNIST USPS!MNIST MNIST!USPS
Source only [11] 67.10.6 69.63.8 82.20.8
DAN [27] 71.1 - 81.1
DANN [6] 73.8 73 85.1
CDAN+E [28] 89.2 98.0 95.6
CyCADA [11] 90.40.4 96.50.1 95.60.4
MCD [40] 96.20.4 94.10.3 94.20.7
GTA [41] 92.40.9 90.81.3 95.30.7
3C-GAN [20] 99.40.1 99.30.1 97.30.2
Source model [24] 72.30.5 90.51.6 72.72.3
Target model [24] 98.80.1 98.10.5 97.90.2
Generated images 98.60.1 97.40.3 97.60.3
Table 1: Classification accuracy (%) on digits datasets. In
Table 1-3, “Generated images” refers to feeding images
generated by SFIT to the source model.
Method A!W D!W W!D A!D D!A W!A Avg.
ResNet-50 [9] 68.4 96.7 99.3 68.9 62.5 60.7 76.1
DAN [27] 80.5 97.1 99.6 78.6 63.6 62.8 80.4
DANN [6] 82.6 96.9 99.3 81.5 68.4 67.5 82.7
ADDA [46] 86.2 96.2 98.4 77.8 69.5 68.9 82.9
JAN [29] 86.0 96.7 99.7 85.1 69.2 70.7 84.6
CDAN+E [28] 94.1 98.6 100.0 92.9 71.0 69.3 87.7
GTA [41] 89.5 97.9 99.8 87.7 72.8 71.4 86.5
3C-GAN [20] 93.7 98.5 99.8 92.7 75.3 77.8 89.6
Source model [24] 76.9 95.6 98.5 80.3 60.6 63.4 79.2
Target model [24] 90.8 98.4 99.9 88.8 73.6 71.7 87.2
Generated images 89.1 98.1 99.9 87.3 69.8 68.7 85.5
Fine-tuning 91.8 98.7 99.9 89.9 73.9 72.0 87.7
Table 2: Classification accuracy (%) on the Office-31
dataset. In Table 2 and Table 3, “Fine-tuning” refers to tar-
get model fine-tuning result with both generated images and
target images (see Section 5.4 for more details).
Method plane bcycl bus car horse knife mcycl person plant sktbrd train truck per-class
ResNet-101 [9] 55.1 53.3 61.9 59.1 80.6 17.9 79.7 31.2 81.0 26.5 73.5 8.5 52.4
DAN [27] 87.1 63.0 76.5 42.0 90.3 42.9 85.9 53.1 49.7 36.3 85.8 20.7 61.1
DANN [6] 81.9 77.7 82.8 44.3 81.2 29.5 65.1 28.6 51.9 54.6 82.8 7.8 57.4
JAN [29] 75.7 18.7 82.3 86.3 70.2 56.9 80.5 53.8 92.5 32.2 84.5 54.5 65.7
ADDA [46] 88.8 65.7 85.6 53.1 74.9 96.2 83.3 70.7 75.9 26.4 83.9 32.4 69.7
MCD [40] 87.0 60.9 83.7 64.0 88.9 79.6 84.7 76.9 88.6 40.3 83.0 25.8 71.9
CDAN+E [28] 85.2 66.9 83.0 50.8 84.2 74.9 88.1 74.5 83.4 76.0 81.9 38.0 73.9
SE [5] 95.9 87.4 85.2 58.6 96.2 95.7 90.6 80.0 94.8 90.8 88.4 47.9 84.3
3C-GAN [20] 94.8 73.4 68.8 74.8 93.1 95.4 88.6 84.7 89.1 84.7 83.5 48.1 81.6
Source model [24] 58.3 17.6 54.2 69.9 64.4 5.5 82.2 30.7 62.2 24.6 86.2 6.0 46.8
Target model [24] 92.5 84.7 81.3 54.6 90.5 94.7 80.9 79.1 90.8 81.5 87.9 50.1 80.7
Generated images 88.9 65.8 83.0 61.7 88.5 76.8 89.5 69.6 91.4 51.9 84.3 34.3 73.8
Fine-tuning 94.3 79.0 84.9 63.6 92.6 92.0 88.4 79.1 92.2 79.8 87.6 43.0 81.4
Table 3: Classification accuracy (%) on the VisDA dataset.
in terms of performance gaps, the knowledge differ-
ences between source and target models are 26.5% on
SVHN!MNIST, 7.6% on USPS !MNIST, and 25.2%
on MNIST!USPS. Generated images from SFIT bridges
these differences to 0.2%, 0.7%, and 0.3%, respectively.
On the Office-31 dataset, the performance gap between the
two models is 8.0% on average, and the generated images
shrink this down to 1.7%. Notably, the performance drops
from the target-image-target-model branch to the generated-
image-source-model branch are especially pronounced on
D!A and W!A, two settings that transfer Amazon im-
ages with white or no background to real-world background
5(a) Target images (MNIST)
(b) Generated source-style images
(c) Unseen source images (SVHN)
Figure 4: Results from the SFIT method on digits datasets
SVHN!MNIST. In Fig. 1 and Fig. 4-6, we show in (a):
target images, (b): generated source-style images, each of
which corresponds to the target image above it, and (c):
the unseen source images. For gray-scale target images
from MNIST, our SFIT approach adds random RGB colors
to mimic the full-color style in the unseen source (SVHN)
without changing the content.
(a) Target images (Webcam)
(b) Generated source-style images
(c) Unseen source images (Amazon)
Figure 5: Results from the SFIT method on the Office-
31 dataset Amazon !Webcam. Our translation method
whitens backgrounds while increasing contrast ratios of the
object (Webcam) for more appealing appearances as in the
online shopping images (Amazon).
in Webcam or DSLR. In fact, in experiments we find gen-
erating an overall consistent colored background is very de-
manding, and the system usually generates a colored back-
ground around the outline of the object. On the VisDAdataset, generated images bridge the performance gap from
33.9% to 6.9%, even under a more demanding setting and
a larger domain gap going from real-world images to syn-
thetic CAD model images. Overall, on all three datasets,
generated images significantly mitigate the knowledge dif-
ference in terms of performance gaps, indicating that the
proposed SFIT method can successfully distill the adapted
knowledge from the target model to the generated images.
Visualization of source-free image translation results.
For digits datasets SVHN !MNIST (Fig. 4), the generator
learns to add RGB colors to the gray-scale MNIST (target)
images, which mimics the full-color SVHN (source) im-
ages. For Office-31 dataset Amazon !Webcam (Fig. 5), the
generated images whiten the background, while having a
white or no background rather than real-world background
is one of the main characteristics of the Amazon (source)
domain when compared to Webcam (target). Moreover,
Amazon online shopping images also have higher contrast
ratios for more appealing appearances, and our translated
images also capture these characteristics, e.g., keys in the
calculator, case of the desktop computer. For VisDA dataset
SYN!REAL (Fig. 1 and Fig. 6), the generator learns to
decrease the overall saturation of the real-world (target)
objects which makes them more similar to the synthetic
(source) scenario, while at the same time whitens the back-
ground, e.g., horse, truck, and plane in Fig. 1, car and skate-
board in Fig. 6, and brings out the green color in the plants.
Overall, image generation results exhibit minimal content
changes from target images, while successfully capturing
theunseen source style.
In terms of visual quality, it is noteworthy that generation
results for digits datasets SVHN !MNIST contain colors
and patterns that are not from the source domain, whereas
our results on the Office-31 dataset and VisDA dataset are
more consistent with the unseen source. Due to the lack
of source images, rather than traditional image translation
approaches [7, 13, 11, 44], SFIT only relies on source and
target models, and portrays adapted knowledge according
to the two models. Since a weaker LeNet classifier is used
for the digits dataset, it is easier to generate images that sat-
isfy the proposed loss terms without requiring the generated
images to perfectly mimic the source style. On Office-31
and VisDA datasets, given stronger models like ResNet, it
is harder to generate images that can satisfy the loss terms.
Stricter restrictions and longer training time lead to gener-
ation results more coherent with unseen source images that
also have better visual quality.
Visualization for different UDA methods. In Fig. 7,
we show SFIT visualization results using different UDA
methods. Given source and target domain, a traditional im-
age translation method generates a certain type of images
regardless of the UDA methods, indicating its incapabil-
ity of presenting the knowledge difference between mod-
6(a) Target images (real-world)
(b) Generated source-style images
(c) Unseen source images (synthetic)
Figure 6: Results from the SFIT method on the VisDA dataset SYN !REAL. Our translation method decreases the target
(real-world) image saturation and whitens the background while keeping the semantics unchanged.
(a)
 (b)
 (c)
 (d)
Figure 7: SFIT results on VisDA dataset with different UDA
methods. (a) Target images; (b) DAN [27]; (c) ADDA [46];
(d) SHOT-IM [24].
els. In contrast, the proposed SFIT method generates differ-
ent images for different UDA methods. Specifically, when
comparing visualization results of the adapted knowledge
in DAN [27], ADDA [46], and SHOT-IM [24], we find
stronger UDA methods can better transfer the target style
to the unseen source style. As shown in Fig. 7, in terms
of whitening the background for style transfer, SFIT re-sults on ADDA are less coherent than SHOT-IM but better
than DAN. This further verifies that our SFIT method in-
deed visualizes the knowledge difference between models,
and stronger adaptation methods can better endure the style
difference (leading to larger knowledge difference and thus
stronger style transfer results).
5.4. Application
The generated images from SFIT allows for further tun-
ing of the target model in SFDA systems, where no source
image is available. We include a diversity loss on all train-
ing samples to promote even class-wise distributions,
Ldiv=H
ExPtarget(x)[p(fT(x))]

; (10)
whereH(
)denotes the information entropy function. We
also incluse a pseudo-label fine-tuning loss, if pseudo
label ^yS= arg max p(fS(ex))from the generated-
image-source-model branch equals to the pseudo label
^yT= arg max p(fT(x))from the target-image-target-
model branch. We then use this pseudo label ^y= ^yS= ^yT
to fine-tune the target model,
Lpseudo =(
H(p(fT(x));^y); if^y= ^yS= ^yT;
0; else;(11)
whereH(
;
)denotes the cross entropy function. We com-
bine these two loss terms in Eq. 10 and Eq. 11 to give an
overall fine-tuning loss LFT=Ldiv+Lpseudo .
7(a)
 (b)
 (c)
 (d)
Figure 8: Visualization results on VisDA dataset with dif-
ferent distribution alignment methods. (a) Target images;
(b) BN stats alignment [12]; (c) traditional style loss [7];
(d) relationship preserving loss.
As an additional cue, supervision from generated-image-
source-model further boosts target model SFDA perfor-
mance. On Office-31, fine-tuning brings a performance
improvement of 0.4% according to Table 2. On VisDA,
fine-tuning improves the target model accuracy by 0.7% as
shown in Table 3. These improvements are statistically very
significant ( i.e.,p-value <0.001 over 5 runs), and introduce
a real-world application for images generated by SFIT.
5.5. Comparison and Variant Study
Comparison with the BatchNorm statistics alignment
method [12]. Hou et al. propose to match the batch-wise
feature map statistics so as to directly generate images that
mimic the source style. Specifically, they explore the Batch-
Norm (BN) statistics stored in the BN layers in the source
model for style indication, and match them against that of
the generated images. Using their approach, we can mildly
change the image to the unseen source style (see Fig. 8) and
slightly reduce the performance difference between the two
branches (see Table 4). With that said, their lack of output
alignments between the two branches (only supervisions
from the source branch ) results in much lower quantita-
tive performance and under-performing style transfer qual-
ity when compared to the proposed method.
Effect of the knowledge distillation loss. The knowl-
edge distillation loss transfers the adapted knowledge to the
generated images, and the removal of it results in a 1.1%
performance drop.
Effect of the relationship preserving loss. As shown in
Fig. 8, the traditional style loss can successfully transfer the
target image to the source style on its own. However, usingVariantLKDLRP accuracy (%)
Target image - - 46.8
Initialized g(
) 44.9
BN stats alignment [12] 51.7
w/oLKD 3 72.7
w/oLRP 3 71.2
LRP!L style 3Lstyle[7] 66.4
LRP!L batch 3Lbatch[45] 71.2
LRP!L pixel 3Lpixel[23] 70.9
SFIT 3 3 73.8
Table 4: Variant study on VisDA dataset. “Initialized g(
)”
refers to our transparent filter initialization in Section 5.2.
it causes a 4.8% performance drop compared to the “w/o
LRP” variant (see Table 4), suggesting it being unsuitable
for SFIT. On the other hand, the batch-wise or pixel-wise
relationship preserving variants [45, 23] are found not use-
ful, as they fail to improve over the “w/o LRP” variant.
In contrast, the proposed channel-wise relationship pre-
serving lossLRPcan effectively improve the recognition ac-
curacy on the generated images, as the inclusion of it leads
to a 2.6% performance increase. Moreover, as shown in
Fig. 8, similar to the traditional style loss, using only the re-
lationship preserving loss can also effectively transfer the
target image to the unseen source style. Besides, focus-
ing on the relative channel-wise relationship instead of the
absolute correlation values, the proposed relationship pre-
serving loss can better maintain the foreground object (less
blurry and more prominent) while transferring the overall
image style, leading to higher recognition accuracy.
6. Conclusion
In this paper, we study the scientific problem of visu-
alizing the adapted knowledge in UDA. Specifically, we
propose a source-free image translation (SFIT) approach,
which generates source-style images from original target
images under the guidance of source and target models.
Translated images on the source model achieve similar re-
sults as target images on the target model, indicating a suc-
cessful depiction of the adapted knowledge. Such images
also exhibit the source style, and the extent of style trans-
fer follows the performance of UDA methods, which fur-
ther verifies that stronger UDA methods can better address
the distribution difference between domains. We show that
the generated images can be applied to fine-tune the target
model, and might help other tasks like incremental learning.
Acknowledgement
This work was supported by the ARC Discovery Early
Career Researcher Award (DE200101283) and the ARC
Discovery Project (DP210102801).
8References
[1] Konstantinos Bousmalis, Nathan Silberman, David Dohan,
Dumitru Erhan, and Dilip Krishnan. Unsupervised pixel-
level domain adaptation with generative adversarial net-
works. In Proceedings of the IEEE conference on computer
vision and pattern recognition , pages 3722–3731, 2017. 1, 2
[2] Konstantinos Bousmalis, George Trigeorgis, Nathan Silber-
man, Dilip Krishnan, and Dumitru Erhan. Domain separa-
tion networks. In Advances in neural information processing
systems , pages 343–351, 2016. 2
[3] Hanting Chen, Yunhe Wang, Chang Xu, Zhaohui Yang,
Chuanjian Liu, Boxin Shi, Chunjing Xu, Chao Xu, and Qi
Tian. Data-free learning of student networks. In Proceed-
ings of the IEEE International Conference on Computer Vi-
sion, pages 3514–3522, 2019. 2
[4] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei.
ImageNet: A Large-Scale Hierarchical Image Database. In
CVPR09 , 2009. 3
[5] Geoff French, Michal Mackiewicz, and Mark Fisher. Self-
ensembling for visual domain adaptation. In International
Conference on Learning Representations , 2018. 5
[6] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pas-
cal Germain, Hugo Larochelle, Franc ¸ois Laviolette, Mario
Marchand, and Victor Lempitsky. Domain-adversarial train-
ing of neural networks. The Journal of Machine Learning
Research , 17(1):2096–2030, 2016. 1, 2, 3, 5
[7] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Im-
age style transfer using convolutional neural networks. In
Proceedings of the IEEE conference on computer vision and
pattern recognition , pages 2414–2423, 2016. 1, 2, 3, 4, 6, 8
[8] Matan Haroush, Itay Hubara, Elad Hoffer, and Daniel
Soudry. The knowledge within: Methods for data-free model
compression. arXiv preprint arXiv:1912.01274 , 2019. 2
[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun.
Deep residual learning for image recognition. In Proceed-
ings of the IEEE conference on computer vision and pattern
recognition , pages 770–778, 2016. 5
[10] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distill-
ing the knowledge in a neural network. arXiv preprint
arXiv:1503.02531 , 2015. 2
[11] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu,
Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Dar-
rell. CyCADA: Cycle-consistent adversarial domain adap-
tation. In Jennifer Dy and Andreas Krause, editors, Pro-
ceedings of the 35th International Conference on Machine
Learning , volume 80 of Proceedings of Machine Learning
Research , pages 1989–1998, Stockholmsm ¨assan, Stockholm
Sweden, 10–15 Jul 2018. PMLR. 1, 2, 3, 5, 6
[12] Yunzhong Hou and Liang Zheng. Source free do-
main adaptation with image translation. arXiv preprint
arXiv:2008.07514 , 2020. 8
[13] Xun Huang and Serge Belongie. Arbitrary style transfer in
real-time with adaptive instance normalization. In Proceed-
ings of the IEEE International Conference on Computer Vi-
sion, pages 1501–1510, 2017. 1, 2, 3, 6
[14] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A
Efros. Image-to-image translation with conditional adver-sarial networks. In Proceedings of the IEEE conference on
computer vision and pattern recognition , pages 1125–1134,
2017. 2
[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual
losses for real-time style transfer and super-resolution. In
European conference on computer vision , pages 694–711.
Springer, 2016. 2
[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based
generator architecture for generative adversarial networks.
InProceedings of the IEEE Conference on Computer Vision
and Pattern Recognition , pages 4401–4410, 2019. 2
[17] Jogendra Nath Kundu, Naveen Venkat, and R Venkatesh
Babu. Universal source-free domain adaptation. arXiv
preprint arXiv:2004.04393 , 2020. 2
[18] Yann LeCun, L ´eon Bottou, Yoshua Bengio, and Patrick
Haffner. Gradient-based learning applied to document recog-
nition. Proceedings of the IEEE , 86(11):2278–2324, 1998.
5
[19] Yann LeCun, Corinna Cortes, and CJ Burges. Mnist hand-
written digit database. ATT Labs [Online]. Available:
http://yann.lecun.com/exdb/mnist , 2, 2010. 2, 5
[20] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and
Si Wu. Model adaptation: Unsupervised domain adaptation
without source data. In Proceedings of the IEEE/CVF Con-
ference on Computer Vision and Pattern Recognition , pages
9641–9650, 2020. 2, 5
[21] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. De-
mystifying neural style transfer. In Proceedings of the 26th
International Joint Conference on Artificial Intelligence , IJ-
CAI’17, page 2230–2236. AAAI Press, 2017. 4
[22] Yanghao Li, Naiyan Wang, Jianping Shi, Xiaodi Hou, and
Jiaying Liu. Adaptive batch normalization for practical do-
main adaptation. Pattern Recognition , 80:109–117, 2018. 2
[23] Zeqi Li, Ruowei Jiang, and Parham Aarabi. Semantic re-
lation preserving knowledge distillation for image-to-image
translation. In European conference on computer vision .
Springer, 2020. 2, 4, 8
[24] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need
to access the source data? source hypothesis transfer for un-
supervised domain adaptation. In International Conference
on Machine Learning (ICML) , pages xx–xx, July 2020. 2, 3,
5, 7
[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays,
Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C Lawrence
Zitnick. Microsoft coco: Common objects in context. In
European conference on computer vision , pages 740–755.
Springer, 2014. 5
[26] Ming-Yu Liu and Oncel Tuzel. Coupled generative adversar-
ial networks. In Advances in neural information processing
systems , pages 469–477, 2016. 2
[27] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I
Jordan. Learning transferable features with deep adaptation
networks. arXiv preprint arXiv:1502.02791 , 2015. 1, 2, 3,
4, 5, 7
[28] Mingsheng Long, Zhangjie Cao, Jianmin Wang, and
Michael I Jordan. Conditional adversarial domain adapta-
tion. In Advances in Neural Information Processing Systems ,
pages 1645–1655, 2018. 5
9[29] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I
Jordan. Deep transfer learning with joint adaptation net-
works. In International conference on machine learning ,
pages 2208–2217. PMLR, 2017. 5
[30] Raphael Gontijo Lopes, Stefano Fenu, and Thad Starner.
Data-free knowledge distillation for deep neural networks.
arXiv preprint arXiv:1710.07535 , 2017. 2
[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochas-
tic gradient descent with warm restarts. arXiv preprint
arXiv:1608.03983 , 2016. 5
[32] Fujun Luan, Sylvain Paris, Eli Shechtman, and Kavita Bala.
Deep photo style transfer. In Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition , pages
4990–4998, 2017. 2
[33] Paul Micaelli and Amos J Storkey. Zero-shot knowledge
transfer via adversarial belief matching. In Advances in
Neural Information Processing Systems , pages 9547–9557,
2019. 2
[34] Rafael M ¨uller, Simon Kornblith, and Geoffrey E Hinton.
When does label smoothing help? In Advances in Neural
Information Processing Systems , pages 4694–4703, 2019. 2
[35] Gaurav Kumar Nayak, Konda Reddy Mopuri, Vaisakh Shaj,
Venkatesh Babu Radhakrishnan, and Anirban Chakraborty.
Zero-shot knowledge distillation in deep networks. In Ka-
malika Chaudhuri and Ruslan Salakhutdinov, editors, Pro-
ceedings of the 36th International Conference on Machine
Learning , volume 97 of Proceedings of Machine Learning
Research , pages 4743–4751, Long Beach, California, USA,
09–15 Jun 2019. PMLR. 2
[36] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bis-
sacco, Bo Wu, and Andrew Y . Ng. Reading digits in natural
images with unsupervised feature learning. In NIPS Work-
shop on Deep Learning and Unsupervised Feature Learning
2011 , 2011. 2, 5
[37] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Rela-
tional knowledge distillation. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition ,
pages 3967–3976, 2019. 2
[38] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman,
Dequan Wang, and Kate Saenko. Visda: The visual domain
adaptation challenge, 2017. 1, 2, 5
[39] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Dar-
rell. Adapting visual category models to new domains. In
European conference on computer vision , pages 213–226.
Springer, 2010. 2, 5
[40] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tat-
suya Harada. Maximum classifier discrepancy for unsuper-
vised domain adaptation. In Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition , pages
3723–3732, 2018. 2, 5
[41] Swami Sankaranarayanan, Yogesh Balaji, Carlos D Castillo,
and Rama Chellappa. Generate to adapt: Aligning domains
using generative adversarial networks. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recogni-
tion, pages 8503–8512, 2018. 5
[42] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua
Susskind, Wenda Wang, and Russell Webb. Learningfrom simulated and unsupervised images through adversarial
training. In Proceedings of the IEEE conference on computer
vision and pattern recognition , pages 2107–2116, 2017. 2
[43] Karen Simonyan and Andrew Zisserman. Very deep convo-
lutional networks for large-scale image recognition. arXiv
preprint arXiv:1409.1556 , 2014. 3
[44] Yaniv Taigman, Adam Polyak, and Lior Wolf. Unsupervised
cross-domain image generation. In ICLR , 2017. 2, 6
[45] Frederick Tung and Greg Mori. Similarity-preserving knowl-
edge distillation. In Proceedings of the IEEE International
Conference on Computer Vision , pages 1365–1374, 2019. 2,
4, 8
[46] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Dar-
rell. Adversarial discriminative domain adaptation. In Pro-
ceedings of the IEEE Conference on Computer Vision and
Pattern Recognition , pages 7167–7176, 2017. 1, 2, 3, 5, 7
[47] Hongxu Yin, Pavlo Molchanov, Jose M. Alvarez, Zhizhong
Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz.
Dreaming to distill: Data-free knowledge transfer via deep-
inversion. In The IEEE/CVF Conf. Computer Vision and Pat-
tern Recognition (CVPR) , 2020. 2
[48] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A
Efros. Unpaired image-to-image translation using cycle-
consistent adversarial networks. In Proceedings of the IEEE
international conference on computer vision , pages 2223–
2232, 2017. 1, 2, 3, 5
10