Method of unsupervised domain adaptation in ordinal regression
A method of jointly training of a transferable feature extractor network, an ordinal regressor network, and an order classifier network in an ordinal regression unsupervised domain adaption network by providing a source of labeled source images and unlabeled target images; outputting image representations from a transferable feature extractor network by performing a minimax optimization procedure on the source of labeled source images and unlabeled target images; training a domain discriminator network, using the image representations from the transferable feature extractor network, to distinguish between source images and target images; training an ordinal regressor network using a full set of source images from the transferable feature extractor network; and training an order classifier network using a full set of source images from said transferable feature extractor network.
1. An ordinal regression unsupervised domain adaption network for jointly training of a transferable feature extractor network, an ordinal regressor network, and an order classifier network, comprising:
a source of labeled source images and unlabeled target images;
a transferable feature extractor network, operatively connected to said source of labeled source images and unlabeled target images, to output image representations, said image representations being realized by a minimax optimization procedure;
a domain discriminator network operatively connected to said transferable feature extractor network;
an ordinal regressor network operatively connected to said transferable feature extractor network; and
an order classifier network operatively connected to said transferable feature extractor network and said domain discriminator network;
said domain discriminator network being trained, using said image representations from said transferable feature extractor network, to distinguish between source images and target images;
said ordinal regressor network being trained, using a full set of source images from said transferable feature extractor network;
said order classifier network being trained, using a pair of source images from said transferable feature extractor network.
2. The ordinal regression unsupervised domain adaption network as claimed in claim 1 , wherein said transferable feature extractor network is trained by maximizing a loss of said domain discriminator.
3. The ordinal regression unsupervised domain adaption network as claimed in claim 1 , wherein a total loss for training the ordinal regression universal domain adaptation network is given as:
ℒ
(
F
,
G
r
,
G
o
,
G
d
)
=
ℒ
or
(
F
,
G
r
)
+
γ
1
ℒ
ord
(
F
,
G
o
)
+
γ
2
ℒ
dom
(
F
,
G
d
)
,
γ 1 , γ 2 are hyper-parameters controlling an importance of order and domain discrimination adversarial losses;
wherein
F
⋆
,
G
r
*
,
G
o
*
=
arg
min
F
,
G
r
,
G
o
max
G
d
ℒ
(
F
,
G
r
,
G
o
,
G
d
)
is solved by alternating between optimizing F, G r , G o , and G d until the total loss converges.
4. The ordinal regression unsupervised domain adaption network as claimed in claim 3 , wherein the loss for said ordinal regressor network is defined on labeled source images,
ℒ
or
(
F
,
G
o
)
=
𝔼
x
i
∈
D
s
L
coral
(
G
r
(
F
(
x
i
)
)
,
y
i
)
,
where L coral is Coral loss.
5. The ordinal regression unsupervised domain adaption network as claimed in claim 3 , wherein the loss for said order classifier network is defined on pairs of source images and order relationships thereof,
ℒ
ord
(
F
,
G
o
)
=
𝔼
x
i
,
x
j
~
D
s
L
ord
(
G
o
,
F
′
,
x
i
≺
x
j
)
,
where L ord is the cross entropy loss, and F′ compares the two image feature vectors.
6. The ordinal regression unsupervised domain adaption network as claimed in claim 3 , wherein the loss for said domain discriminator network is
ℒ
d
=
𝔼
x
s
∈
D
s
w
(
x
s
)
log
G
d
(
F
(
x
s
)
)
+
𝔼
x
t
∈
D
t
w
(
x
t
)
log
(
1
-
G
d
(
F
(
x
t
)
)
,
where weights w(x t ) and w(x s ) are obtained by target image weighting and source class weighting, w(x t )=p(x t ∈Y), w(x s )=p(x s ∈Y)=p(y s ∈Y), y s =y(x s ).
7. A method of jointly training of a transferable feature extractor network, an ordinal regressor network, and an order classifier network in an ordinal regression unsupervised domain adaption network, comprising:
(a) providing a source of labeled source images and unlabeled target images;
(b) outputting image representations from a transferable feature extractor network by performing a minimax optimization procedure on the source of labeled source images and unlabeled target images;
(c) training a domain discriminator network, using the image representations from the transferable feature extractor network, to distinguish between source images and target images;
(d) training an ordinal regressor network using a full set of source images from the transferable feature extractor network; and
(e) training an order classifier network using a full set of source images from said transferable feature extractor network.
8. The method as claimed in claim 7 , further comprising:
(f) training the transferable feature extractor network by maximizing a loss of the domain discriminator.
9. The method as claimed in claim 7 , wherein a total loss for training the ordinal regression universal domain adaptation network is given as:
ℒ
(
F
,
G
r
,
G
o
,
G
d
)
=
ℒ
or
(
F
,
G
r
)
+
γ
1
ℒ
ord
(
F
,
G
o
)
+
γ
2
ℒ
dom
(
F
,
G
d
)
,
γ 1 , γ 2 are hyper-parameters controlling an importance of order and domain discrimination adversarial losses;
wherein
F
⋆
,
G
r
*
,
G
o
*
=
arg
min
F
,
G
r
,
G
o
max
G
d
ℒ
(
F
,
G
r
,
G
o
,
G
d
)
is solved by alternating between optimizing F, G r , G o , and G d until the total loss converges.
10. The method as claimed in claim 9 , wherein the loss for the ordinal regressor network is defined on labeled source images,
ℒ
or
(
F
,
G
o
)
=
𝔼
x
i
∈
D
s
L
coral
(
G
r
(
F
(
x
i
)
)
,
y
i
)
,
where L coral is Coral loss.
11. The method as claimed in claim 9 , wherein the loss for the order classifier network is defined on pairs of source images and order relationships thereof,
ℒ
ord
(
F
,
G
o
)
=
𝔼
x
i
,
x
j
~
D
s
L
ord
(
G
o
,
F
′
,
x
i
≺
x
j
)
,
where L ord is the cross entropy loss, and F compares the two image feature vectors.
12. The ordinal regression unsupervised domain adaption network as claimed in claim 9 , wherein the loss for the domain discriminator network is
ℒ
d
=
𝔼
x
s
∈
D
s
w
(
x
s
)
log
G
d
(
F
(
x
s
)
)
+
𝔼
x
t
∈
D
t
w
(
x
t
)
log
(
1
-
G
d
(
F
(
x
t
)
)
,
where weights w(x t ) and w(x s ) are obtained by target image weighting and source class weighting, w(x t )=p(x t ∈Y), w(x s )=p(x s ∈Y)=p(y s ∈Y), y s =y(x s ).