95 lines (93 with data), 11.0 kB

3

░ĚA],Ń@sV
dZd
dl
Z
d
dlZd
dlmZmZmZmZm	Z	m
Z
mZmZ
d
dl
m
Z
mZ
d
dljjZd
dlmZ
d
dl
mZ
defdd	ä
ZGd
dädeâZdd
äZGddädejâZGddädejâZe
e	dťddäZe
eedťddäZ e
edťddäZ!d*e
edťddäZ"d+e
edťddäZ#e
e$e
d ťd!d"äZ%ej&ej&d#ťd$d%äZ'Gd&d'äd'âZ(Gd(d)äd)âZ)dS),zÉ
losses.py
=======================
Some additional loss functions that can be called using the pipeline, some of which still to be implemented.
ÚN)┌Any┌Callable┌Iterable┌List┌Set┌Tuple┌TypeVar┌Union)┌Tensor┌einsum)
┌distance_transform_edt)
┌nn┌cCs|s||
â
é
d
S)z░https://raw.githubusercontent.com/inferno-pytorch/inferno/0561e8a95cde6bfc5e10a3609841b7b0ca5b03ca/inferno/utils/exceptions.py
	Like assert, but with arbitrary exception types.Nę)┌	condition┌message┌exception_typerr˙B/Users/joshualevy/Documents/GitHub/PathFlowAI/pathflowai/losses.py┌assert_s
rc
@seZ
dZd
S)┌
ShapeErrorN)┌__name__┌
__module__┌__qualname__rrrrrs
rc
Cspt|j
âd
kdj|j
ââ
tâ
|jdâ
}
tt|j
ââ
â
}|d|d|d<|d<|j|Äjâ}|j	|
dâ}|S)aŹ

	https://raw.githubusercontent.com/inferno-pytorch/inferno/0561e8a95cde6bfc5e10a3609841b7b0ca5b03ca/inferno/utils/torch_utils.py
	Flattens a tensor or a variable such that the channel axis is first and the sample axis
	is second. The shapes are transformed as follows:
		(N, C, H, W) --> (C, N * H * W)
		(N, C, D, H, W) --> (C, N * D * H * W)
		(N, C) --> (C, N)
	The input must be atleast 2d.
	Úz9Tensor or variable must be atleast 2D. Got one of dim {}.Ú
r
Ú    )
r┌dim┌formatr┌size┌list┌range┌permute┌
contiguous┌view)┌input_┌num_channels┌permute_axes┌permuted┌	flattenedrrr┌flatten_sampless






r)cs*eZ
dZd
Zd	çf
ddä	ZddäZçZS)
┌GeneralizedDiceLossaé

	https://raw.githubusercontent.com/inferno-pytorch/inferno/0561e8a95cde6bfc5e10a3609841b7b0ca5b03ca/inferno/extensions/criteria/set_similarity_measures.py
	Computes the scalar Generalized Dice Loss defined in https://arxiv.org/abs/1707.03237

	This version works for multiple classes and expects predictions for every class (e.g. softmax output) and
	one-hot targets for every class.
	NFšŹÝÁá¸Ă░>cs0tt
|âjâ
|jd
|
â
||_||_||_dS)N┌weight)┌superr*┌__init__┌register_buffer┌channelwise┌eps┌add_softmax)┌selfr,r0r1r2)
┌	__class__rrr.:s





zGeneralizedDiceLoss.__init__c
Csé
|
jâ|jâkst
é
|jr(tj|
d
dŹ}
|js¬t|
â
}
t|â
jâ}|jdâ
}d||j	|j
dŹ
}|
|jdâ
|jâ}|
|jd
â
|jâ}dd||j	|j
dŹ
}nďddä}||
â
}
||â
}|jdâ
}d||j	|j
dŹ
}|
|jdâ
|jdâ
}|
|jdâ
|jdâ
}dd||j	|j
dŹ
}	|jdk	É
rv|	jâd	kÉ
rP|	j
d
â
}	|jjâ|	jâkÉ
slt
d
â
é
|j|	}	|	jâ}|S)a

		input: torch.FloatTensor or torch.cuda.FloatTensor
		target:     torch.FloatTensor or torch.cuda.FloatTensor

		Expected shape of the inputs:
			- if not channelwise: (batch_size, nb_classes, ...)
			- if channelwise:     (batch_size, nb_channels, nb_classes, ...)
		r)
rg­?)
┌ming@c
Ss||jâ}
|
d
kst
é
|jdâ
}|jdâ
}tt|
â
â
}|d|d|d|d<|d<|d<|j|Äjâ}|j||dâ}|S)NÚrrr
r)r┌AssertionErrorrrr r!r"r#)┌tensorZ
tensor_dimr%┌num_classesr&r'r(rrr┌flatten_and_preserve_channels]s






(


zBGeneralizedDiceLoss.forward.<locals>.flatten_and_preserve_channelsNrzy`weight` should have shape (nb_channels, ),
					   `target` should have shape (batch_size, nb_channels, nb_classes, ...)rrrrrrrr)rr7r2┌
F┌softmaxr0r)┌float┌sum┌clampr1r,r┌squeeze)
r3┌input┌targetZsum_targetsZ
class_weigths┌numer┌denom┌lossr:Zchannelwise_lossrrr┌forwardAs6	
















zGeneralizedDiceLoss.forward)NFr+F)rrr┌__doc__r.rF┌
__classcell__rr)
r4rr*2s
r*cs*eZ
dZd
Zdçf
ddä	Zdd	äZçZS)┌	FocalLossa8
	# https://raw.githubusercontent.com/Hsuxu/Loss_ToolBox-PyTorch/master/FocalLoss/FocalLoss.py
	This is a implementation of Focal Loss with smooth label cross entropy supported which is proposed in
	'Focal Loss for Dense Object Detection. (https://arxiv.org/abs/1708.02002)'
		Focal_Loss= -1*alpha*(1-pt)*log(pt)
	:param num_class:
	:param alpha: (tensor) 3D or 4D the scalar factor for this criterion
	:param gamma: (float,double) gamma > 0 reduces the relative loss for well-classified examples (p>0.5) putting more
					focus on hard misclassified example
	:param smooth: (float,double) smooth value when cross entropy
	:param balance_index: (int) balance class index, should be specific when alpha is float
	:param size_average: (bool, optional) By default, the losses are averaged over each loss element in the batch.
	NrrTcs
tt
|âjâ
|
|_||_||_||_||_|jdkrHtj	|jd
â|_nĺt
|jttj
fârśt|jâ
|jksnté
tj|â
j|jd
â|_|j|jjâ|_nBt
|jtârĎtj	|jd
â}|d
|j}|j||<||_ntdâ
é
|jdk	É
r|jdksŘ|jdkÉ
rtdâ
é
dS)NrzNot support alpha typer
g­?zsmooth value should be in [0,1])r-rIr.┌	num_class┌alpha┌gamma┌smooth┌size_average┌torch┌ones┌
isinstancer┌np┌ndarray┌lenr7┌FloatTensorr#r>r=┌	TypeError┌
ValueError)r3rJrKrLZ
balance_indexrMrN)
r4rrr.ôs*



















zFocalLoss.__init__cCsJ
|
jâd
krJ|
j
|
jdâ
|
jdâ
dâ}
|
jdd
dâjâ}
|
j
d|
jdâ
â}
|j
d	dâ}d}|j}|jtjkrx|jtjâ
}|j	âj
â}tj|jdâ
|j
âjâ}|jd|dâ}|j|
jkr┬|j|
jâ
}|jrŔtj||j|j
dd|jâ}||
jdâ
|}|jâ}|j}	||}d
|tjd||	â|}
|jÉ
r>|
jâ}
n|
jâ}
|
S)Nrr
rg╗ŻÎ┘▀|█=g­?rrrrr)rr#rr!r"rK┌devicerA┌to┌cpu┌longrOrUrJ┌zero_┌scatter_rMr?r>┌logrL┌powrN┌mean)r3┌logitrB┌epsilonrK┌idxZone_hot_key┌ptZlogptrLrErrrrFşs4
















zFocalLoss.forwardr)NrrNT)rrrrGr.rFrHrr)
r4rrIäs
rI)┌
a┌returnc

Cstt
j|jââ
jââ
S)
zGhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/utils.py)┌setrO┌uniquerZ┌numpy)
rerrr┌uniq┌srj)re┌subrfcCst|â
j
|
â
S)
zGhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/utils.py)rj┌issubset)rerkrrr┌ssetŮsrmcCstj
||
âjâS)
zGhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/utils.py)rO┌eq┌all)re┌
brrrrnŃsrnr)┌
trfcCs.|j|
â
j
tjâ
}tj|tjd
Ź}tj||âS)zGhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/utils.py)
┌dtype)r>┌typerO┌float32┌	ones_like┌allclose)rq┌axis┌_sumZ_onesrrr┌simplexŔs

rycCst||
âot
|d
dgâS)zGhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/utils.pyr
r)ryrm)rqrwrrr┌one_hotţsrz)┌seg┌
CrfcsŐtłj
â
d
krłjddŹ
ëtłtt|
â
â
âs0té
łj
\}}}tjçf
ddät|
â
Dâ
ddŹj	tj
â
}|j
||
||fkszté
t|â
sćté
|S)zGhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/utils.pyrr
)
rc
sg|]}
ł|
kĹqSrr)┌.0┌
c)
r{rr˙
<listcomp>˙sz!class2one_hot.<locals>.<listcomp>r)rT┌shape┌	unsqueezermrr r7rO┌stackrs┌int32rz)r{r|rp┌
w┌
h┌resr)
r{r┌
class2one_hot˛s

(

rç)r{rfc
Cs|tt
j|â
d
dŹsté
t|â
}
tj|â
}xNt|
â
D]B}||jtj	â
}|j
âr2|}t|â
|t|â
d|||<q2W|S)zGhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/utils.pyr
)
rwr)rzrOr
r7rTrR┌
zeros_liker ┌astype┌bool┌any┌distance)r{r|rćr~ZposmaskZnegmaskrrr┌one_hot2dist
s





$
rŹc@s,eZ
dZd
ZddäZeeeedťddäZdS)┌SurfaceLosszHhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/losses.pyc
Ks&|
d
|_t
d|jjŤd|
ŤŁâ

dS)N┌idczInitialized z with )rĆ┌printr4r)r3┌kwargsrrrr.
s

zSurfaceLoss.__init__)┌probs┌	dist_maps┌
_rfcCsjt|
â
st
é
t|â
st
é
|
ddů|jd
fjtjâ
}|ddů|jd
fjtjâ
}td||â}|jâ}|S)N.zbcwh,bcwh->bcwh)	ryr7rzrĆrsrOrtrr`)r3rĺrôrö┌pc┌dcZ	multipledrErrr┌__call__
s


zSurfaceLoss.__call__N)rrrrGr.r
rŚrrrrrÄ
s

rÄc@s,eZ
dZd
ZddäZeeeedťddäZdS)┌GeneralizedDicezHhttps://raw.githubusercontent.com/LIVIAETS/surface-loss/master/losses.pyc
Ks&|
d
|_t
d|jjŤd|
ŤŁâ

dS)NrĆzInitialized z with )rĆrÉr4r)r3rĹrrrr.$
s

zGeneralizedDice.__init__)rĺrBrörfcCs┬t|
â
rt|â
st
é
|
ddů|jd
fjtjâ
}|ddů|jd
fjtjâ
}dtd|âjtjâ
dd}|td||â}|td|âtd|â}ddtd|âdtd|âd}	|	jâ}
|
S)N.rzbcwh->bcg╗ŻÎ┘▀|█=rz
bcwh,bcwh->bczbc->b)ryr7rĆrsrOrtrr`)r3rĺrBrörĽ┌tcrä┌intersection┌unionZdividedrErrrrŚ)
s



$zGeneralizedDice.__call__N)rrrrGr.r
rŚrrrrrś"
s

rś)
r)
r)*rGrOrirR┌typingrrrrrrrr	r
rZtorch.nn.functionalr
┌
functionalr;Z
scipy.ndimagerrîr7rrWrr)┌Moduler*rIrjrŐrmrnryrz┌intrçrSrŹrÄrśrrrr┌<module>s(
(



RV