23 lines (23 with data), 2.5 kB

3

i`âŃ@s4dd
lZ
dd
ljjZGddädâZGddädâZd
S)ÚNc@seZ
dZd
ZddäZdS)┌Dicez#
    N-D dice for segmentation
    c	Cszt|j
âjââ
d
}ttd|dââ
}d
tj|
||â}tj|
||â}ttjdâr^tjj	ntj
}tj|||ââ
}|S)NÚÚ
┌
divide_no_nan)┌len┌	get_shape┌as_list┌list┌range┌tf┌
reduce_sum┌hasattr┌mathr┌
div_no_nan┌reduce_mean)	┌self┌y_true┌y_pred┌ndimsZvol_axes┌top┌bottomrZdiceęr˙../src/utils/losses.py┌loss
s




z	Dice.lossN)┌__name__┌
__module__┌__qualname__┌__doc__rrrrrrsrc@s*eZ
dZd
Zd
ddä
ZddäZdd	äZdS)┌Gradz█
    N-D gradient loss.
    loss_mult can be used to scale the loss value - this is recommended if
    the gradient is computed on a downsampled vector field (where loss_mult
    is equal to the downsample factor).
    ┌l1NcCs|
|_||_
dS)
N)┌penalty┌	loss_mult)rr r!rrr┌__init__s

z
Grad.__init__c	CsÔ|
jâj
âd
dů}t|â
}dg
|}xÂt|â
D]¬}|d
}|f
t|â
t|d
|dâĽ}tj|
|â}
|
d
důdf|
ddůdf}td
|d
âdt|d
|dâĽ}|f
td
|âdt|d
|dâĽ}tj||â||<q0W|S)	Nrr.r
Ú    r#)
r
)
r
)rrrr
┌
K┌permute_dimensions)	r┌
yZ	vol_shaper┌df┌
i┌
d┌
rZdfirrr┌_diffs#s






 "
"
zGrad._diffscCsä|jd
kr ddä|j
|â
Dâ
}n,|jdks8td|jâ
é
ddä|j
|â
Dâ
}ddä|Dâ
}tj|â
t|â
}|jdk	rÇ||j9}|S)Nrc
Ssg|]}
tj
|
â
ĹqSr)r┌abs)┌.0┌
frrr˙
<listcomp>;szGrad.loss.<locals>.<listcomp>┌l2z%penalty can only be l1 or l2. Got: %sc
Ssg|]}
|
|
ĹqSrr)r-r.rrrr/>sc
Ss g|]}
tj
tj|
â
dd
ŹĹqS)r)
┌axisr#)rrr$┌
batch_flatten)r-r.rrrr/@s)r r+┌AssertionErrorr┌add_nrr!)r┌
_r┌difr'┌gradrrrr8s






z	Grad.loss)rN)rrrrr"r+rrrrrrs
r)┌
tensorflowrZtensorflow.keras.backend┌keras┌backendr$rrrrrr┌<module>
s