DataAugmentation
"""
ToEltype(T)
Converts any `AbstractArrayItem` to an `AbstractArrayItem{N, T}`.
Supports `apply!`.
## Examples
{cell=ToEltype}
```julia
using DataAugmentation
tfm = ToEltype(Float32)
item = ArrayItem(rand(Int, 10))
apply(tfm, item)
```
"""
struct
ToEltype
{
T
}
<:
Transform
end
ToEltype
(
T
::
Type
)
=
ToEltype
{
T
}
(
)
apply
(
::
ToEltype
{
T
}
,
item
::
AbstractArrayItem
{
N
,
<:
T
}
;
randstate
=
nothing
)
where
{
N
,
T
}
=
item
function
apply
(
::
ToEltype
{
T1
}
,
item
::
AbstractArrayItem
{
N
,
T2
}
;
randstate
=
nothing
)
where
{
N
,
T1
,
T2
}
newdata
=
map
(
x
->
convert
(
T1
,
x
)
,
itemdata
(
item
)
)
item
=
setdata
(
item
,
newdata
)
return
item
end
function
apply!
(
buf
,
::
ToEltype
,
item
::
AbstractArrayItem
;
randstate
=
nothing
)
## copy! does type conversion under the hood
copy!
(
itemdata
(
buf
)
,
itemdata
(
item
)
)
return
buf
end
"""
Normalize(means, stds)
Normalizes the last dimension of an `AbstractArrayItem{N}`.
Supports `apply!`.
## Examples
Preprocessing a 3D image with 3 color channels.
{cell=Normalize}
```julia
using DataAugmentation, Images
image = Image(rand(RGB, 20, 20, 20))
tfms = ImageToTensor() |> Normalize((0.1, -0.2, -0.1), (1,1,1.))
apply(tfms, image)
```
"""
struct
Normalize
{
N
}
<:
Transform
means
::
SVector
{
N
}
stds
::
SVector
{
N
}
end
function
Normalize
(
means
,
stds
)
length
(
means
)
==
length
(
stds
)
||
error
(
"
`means` and `stds` must have same length
"
)
N
=
length
(
means
)
return
Normalize
{
N
}
(
SVector
{
N
}
(
means
)
,
SVector
{
N
}
(
stds
)
)
end
function
apply
(
tfm
::
Normalize
,
item
::
ArrayItem
{
N
,
T
}
;
randstate
=
nothing
)
where
{
N
,
T
}
return
ArrayItem
(
normalize
(
itemdata
(
item
)
,
tfm
.
means
,
tfm
.
stds
)
)
end
function
apply!
(
buf
,
tfm
::
Normalize
,
item
::
ArrayItem
;
randstate
=
nothing
)
copy!
(
itemdata
(
buf
)
,
itemdata
(
item
)
)
normalize!
(
itemdata
(
buf
)
,
tfm
.
means
,
tfm
.
stds
)
return
buf
end
function
normalize!
(
a
::
AbstractArray
{
T
,
N
}
,
means
::
SVector
{
M
}
,
stds
::
SVector
{
M
}
)
where
{
N
,
T
,
M
}
means
=
reshape
(
convert
.
(
T
,
means
)
,
(
1
for
_
=
2
:
N
)
...
,
M
)
stds
=
reshape
(
convert
.
(
T
,
stds
)
,
(
1
for
_
=
2
:
N
)
...
,
M
)
a
.-=
means
a
./=
stds
return
a
end
normalize
(
a
,
means
,
stds
)
=
normalize!
(
copy
(
a
)
,
means
,
stds
)
function
denormalize!
(
a
::
AbstractArray
{
T
,
N
}
,
means
::
SVector
{
M
}
,
stds
::
SVector
{
M
}
)
where
{
N
,
T
,
M
}
means
=
reshape
(
convert
.
(
T
,
means
)
,
(
1
for
_
=
2
:
N
)
...
,
M
)
stds
=
reshape
(
convert
.
(
T
,
stds
)
,
(
1
for
_
=
2
:
N
)
...
,
M
)
a
.*=
stds
a
.+=
means
return
a
end
denormalize
(
a
,
means
,
stds
)
=
denormalize!
(
copy
(
a
)
,
means
,
stds
)
NormalizeIntensity]
"""
NormalizeIntensity()
Normalizes the pixels of an array based on calculated mean and std.
"""
struct
NormalizeIntensity
<:
Transform
end
function
apply
(
tfm
::
NormalizeIntensity
,
item
::
ArrayItem
;
randstate
=
nothing
)
array
=
itemdata
(
item
)
slices
=
ones
(
Bool
,
size
(
array
)
)
means
=
mean
(
array
[
slices
]
)
stds
=
std
(
array
[
slices
]
)
array
[
slices
]
=
(
array
[
slices
]
.-
means
)
/
stds
return
ArrayItem
(
array
)
end
ImageToTensor]
"""
ImageToTensor()
Expands an `Image{N, T}` of size `sz` to an `ArrayItem{N+1}` with
size `(sz..., ch)` where `ch` is the number of color channels of `T`.
Supports `apply!`.
## Examples
{cell=ImageToTensor}
```julia
using DataAugmentation, Images
image = Image(rand(RGB, 50, 50))
tfm = ImageToTensor()
apply(tfm, image)
```
"""
struct
ImageToTensor
{
T
}
<:
Transform
end
ImageToTensor
(
T
::
Type
{
<:
Number
}
=
Float32
)
=
ImageToTensor
{
T
}
(
)
function
apply
(
::
ImageToTensor
{
T
}
,
image
::
Image
;
randstate
=
nothing
)
where
T
return
ArrayItem
(
imagetotensor
(
itemdata
(
image
)
,
T
)
)
end
function
apply!
(
buf
,
::
ImageToTensor
,
image
::
Image
;
randstate
=
nothing
)
imagetotensor!
(
buf
.
data
,
image
.
data
)
return
buf
end
function
imagetotensor
(
image
::
AbstractArray
{
C
,
N
}
,
T
=
Float32
)
where
{
C
<:
Color
,
N
}
T
.
(
PermutedDimsArray
(
_channelview
(
image
)
,
(
(
i
for
i
in
2
:
N
+
1
)
...
,
1
)
)
)
end
function imagetotensor(image::AbstractArray, T = Float32) where {TC, C<:Color{TC, 1}, N} return T.(_channelview(image)) end TODO: relax color type constraint, implement for other colors single-channel colors need a
channelview that also expands the array
function
imagetotensor!
(
buf
,
image
::
AbstractArray
{
<:
Color
,
N
}
)
where
N
permutedims!
(
buf
,
_channelview
(
image
)
,
(
2
:
N
+
1
...
,
1
)
)
end
function
tensortoimage
(
a
::
AbstractArray
)
nchannels
=
size
(
a
)
[
end
]
if
nchannels
==
3
return
tensortoimage
(
RGB
,
a
)
elseif
nchannels
==
1
return
tensortoimage
(
Gray
,
a
)
else
error
(
"
Found image tensor with
$
nchannels
color channels. Pass in color type
explicitly.
"
)
end
end
function
tensortoimage
(
C
::
Type
{
<:
Color
}
,
a
::
AbstractArray
{
T
,
N
}
)
where
{
T
,
N
}
perm
=
(
N
,
1
:
N
-
1
...
)
return
_colorview
(
C
,
PermutedDimsArray
(
a
,
perm
)
)
end
function
_channelview
(
img
)
chview
=
channelview
(
img
)
# for single-channel colors, expand the color dimension anyway
if
size
(
img
)
==
size
(
chview
)
chview
=
reshape
(
chview
,
1
,
size
(
chview
)
...
)
end
return
chview
end
function
_colorview
(
C
::
Type
{
<:
Color
}
,
img
)
where
T
if
size
(
img
,
1
)
==
1
img
=
reshape
(
img
,
size
(
img
)
[
2
:
end
]
)
end
return
colorview
(
C
,
img
)
endOneHot encoding
"""
OneHot([T = Float32])
One-hot encodes a `MaskMulti` with `n` classes and size `sz` into
an array item of size `(sz..., n)` with element type `T`. Supports [`apply!`](#).
```julia
item = MaskMulti(rand(1:4, 100, 100), 1:4)
apply(OneHot(), item)
```
"""
struct
OneHot
{
T
}
<:
Transform
end
OneHot
(
)
=
OneHot
{
Float32
}
(
)
function
apply
(
tfm
::
OneHot
{
T
}
,
item
::
MaskMulti
;
randstate
=
nothing
)
where
T
mask
=
parent
(
itemdata
(
item
)
)
a
=
zeros
(
T
,
size
(
mask
)
...
,
length
(
item
.
classes
)
)
for
I
in
CartesianIndices
(
mask
)
a
[
I
,
mask
[
I
]
]
=
one
(
T
)
end
return
ArrayItem
(
a
)
end
function
apply!
(
buf
,
tfm
::
OneHot
{
T
}
,
item
::
MaskMulti
;
randstate
=
nothing
)
where
T
mask
=
parent
(
itemdata
(
item
)
)
a
=
itemdata
(
buf
)
fill!
(
a
,
zero
(
T
)
)
for
I
in
CartesianIndices
(
mask
)
a
[
I
,
mask
[
I
]
]
=
one
(
T
)
end
return
buf
end
function
onehot
(
T
,
x
::
Int
,
n
::
Int
)
v
=
fill
(
zero
(
T
)
,
n
)
v
[
x
]
=
one
(
T
)
return
v
end
onehot
(
x
,
n
)
=
onehot
(
Float32
,
x
,
n
)