Downloads: 1
71 lines (70 with data), 9.8 kB
o
A®eÃ. ã @ sø d dl Z d dlm mZ d dl mZ d dlmZ d dlmZm Z m
Z
d ddZ G dd d ej Z
G d
d d ej ZG d d
d
ej ZG dd deZG dd deZG dd dej ZG dd dej ZG dd dej Zdd Zdd Zd d ZdS ) é N)Únn)ÚVariable)ÚMSELossÚ SmoothL1LossÚL1Lossçíµ ÷Æ°>c C s~ | ¡ | ¡ ks J dt| } t|}| ¡ }| | d¡}|dur'|| }| | d¡|| d¡ }d||j|d S )aÓ
Computes DiceCoefficient as defined in https://arxiv.org/abs/1606.04797 given a multi channel input and target.
Assumes the input is a normalized probability, e.g. a result of Sigmoid or Softmax function.
Args:
input (torch.Tensor): NxCxSpatial input tensor
target (torch.Tensor): NxCxSpatial target tensor
epsilon (float): prevents division by zero
weight (torch.Tensor): Cx1 tensor of weight per channel/class
ú-'input' and 'target' must have the same shapeéÿÿÿÿNé ©Úmin)ÚsizeÚflattenÚfloatÚsumÚclamp)ÚinputÚtargetÚepsilonÚweightÚ intersectÚ denominator© r úDc:\Users\zafry\Downloads\bone_paper_codes\bone_paper_codes\losses.pyÚcompute_per_channel_dice s
r c ó( e Zd ZdZ fddZdd Z ZS )Ú_MaskingLossWrapperzv
Loss wrapper which prevents the gradient of the loss to be computed where target is equal to `ignore_index`.
c s. t t| ¡ |d usJ d|| _|| _d S )Nzignore_index cannot be None)Úsuperr Ú__init__ÚlossÚ ignore_index)Úselfr r ©Ú __class__r r r ) s
z _MaskingLossWrapper.__init__c C s2 | ¡ | j¡}d|_|| }|| }| ||¡S )NF)ÚcloneÚne_r Ú
requires_gradr )r! r r Úmaskr r r Úforward/ s
z_MaskingLossWrapper.forward©Ú__name__Ú
__module__Ú __qualname__Ú__doc__r r( Ú
__classcell__r r r" r r $ s r c s* e Zd ZdZd fdd Zdd Z ZS )Ú SkipLastTargetChannelWrapperz>
Loss wrapper which removes additional target channel
Fc s t t| ¡ || _|| _d S ©N)r r/ r r Úsqueeze_channel)r! r r1 r" r r r A s
z%SkipLastTargetChannelWrapper.__init__c C sL | d¡dks J d|d d
d d
df }| jr tj|dd}| ||¡S )Né zFTarget tensor has a singleton channel dimension, cannot remove channelr .©Údim)r
r1 ÚtorchÚsqueezer ©r! r r r r r r( F s
z$SkipLastTargetChannelWrapper.forward)Fr) r r r" r r/ < s r/ c s2 e Zd ZdZd
fdd Zdd Zdd Z ZS ) Ú_AbstractDiceLossz@
Base class for different implementations of Dice loss.
NÚsigmoidc sd t t| ¡ | d|¡ |dv sJ |dkr t ¡ | _d S |dkr+tjdd| _d S dd | _d S ) Nr )r9 ÚsoftmaxÚnoner9 r: r2 r3 c S s | S r0 r )Úxr r r Ú<lambda>e s z,_AbstractDiceLoss.__init__.<locals>.<lambda>)r r8 r Úregister_bufferr ÚSigmoidÚ
normalizationÚSoftmax©r! r r@ r" r r r W s z_AbstractDiceLoss.__init__c C s t r0 )ÚNotImplementedError©r! r r r r r r Údiceg s z_AbstractDiceLoss.dicec C s* | |¡}| j||| jd}dt |¡ S )N©r ç ð?)r@ rE r r5 Úmean)r! r r Zper_channel_dicer r r r( k s
z_AbstractDiceLoss.forward©Nr9 )r* r+ r, r- r rE r( r. r r r" r r8 R s
r8 c ó* e Zd ZdZd fdd Zdd Z ZS ) ÚDiceLossa Computes Dice Loss according to https://arxiv.org/abs/1606.04797.
For multi-class segmentation `weight` parameter can be used to assign different weights per class.
The input to the loss function is assumed to be a logit and will be normalized by the Sigmoid function.
Nr9 c s t ||¡ d S r0 )r r rB r" r r r | s zDiceLoss.__init__c C s t ||| jdS )NrF )r r rD r r r rE s z
DiceLoss.dicerI ©r* r+ r, r- r rE r. r r r" r rK v s rK c rJ ) ÚGeneralizedDiceLossz_Computes Generalized Dice Loss (GDL) as described in https://arxiv.org/pdf/1707.03237.pdf.
r9 r c s t jd |d || _d S )N©r r@ )r r r )r! r@ r r" r r r s
z GeneralizedDiceLoss.__init__c C sÚ | ¡ | ¡ ks J dt|}t|}| ¡ }| d¡dkr5tj|d| fdd}tj|d| fdd}| d¡}d|| j| jd }d|_|| d¡}|| }|| d¡}|| j| jd}d| ¡ | ¡ S ) Nr r r2 r3 r r Fr
) r
r r r5 Úcatr r r r& )r! r r r Zw_lr r r r r rE s
zGeneralizedDiceLoss.dice)r9 r rL r r r" r rM s rM c r )Ú BCEDiceLossz)Linear combination of BCE and Dice lossesc s0 t t| ¡ || _t ¡ | _|| _t | _ d S r0 )
r rP r Úalphar ÚBCEWithLogitsLossÚbceÚbetarK rE )r! rQ rT r" r r r © s
zBCEDiceLoss.__init__c C s$ | j | ||¡ | j| ||¡ S r0 )rQ rS rT rE r7 r r r r( ° s $zBCEDiceLoss.forwardr) r r r" r rP ¦ s rP c s6 e Zd ZdZd fdd Zdd Zedd Z ZS )
ÚWeightedCrossEntropyLosszXWeightedCrossEntropyLoss (WCE) as described in https://arxiv.org/pdf/1707.03237.pdf
r c s t t| ¡ || _d S r0 )r rU r r )r! r r" r r r ¸ s
z!WeightedCrossEntropyLoss.__init__c C s | |¡}tj|||| jdS )N©r r )Ú_class_weightsÚFÚ
cross_entropyr rD r r r r( ¼ s
z WeightedCrossEntropyLoss.forwardc C sB t j| dd} t| }d| d¡}| d¡}t|| dd}|S )Nr2 r3 rG r F)r& )rX r: r r r )r Ú flattenedÚ nominatorr Ú
class_weightsr r r rW À s
z'WeightedCrossEntropyLoss._class_weights)r ) r* r+ r, r- r r( Ú staticmethodrW r. r r r" r rU ´ s rU c s* e Zd Zd fdd Z fddZ ZS )ÚWeightedSmoothL1LossTc s$ t jdd || _|| _|| _d S )Nr; )Ú reduction)r r Ú thresholdÚapply_below_thresholdr )r! r` Úinitial_weightra r" r r r Ì s
z WeightedSmoothL1Loss.__init__c sD t ||¡}| jr|| jk }n|| jk}|| | j ||< | ¡ S r0 )r r( ra r` r rH )r! r r Úl1r' r" r r r( Ò s
z WeightedSmoothL1Loss.forward)T)r* r+ r, r r( r. r r r" r r^ Ë s r^ c C s: | d¡}dttd| ¡ }| |¡}| ¡ |d¡S )zFlattens a given tensor such that the channel axis is first.
The shapes are transformed as follows:
(N, C, D, H, W) -> (C, N * D * H * W)
r2 )r2 r r
r )r
ÚtupleÚranger4 ÚpermuteÚ
contiguousÚview)ÚtensorÚCZ
axis_orderÚ
transposedr r r r ß s
r c C sÎ d| v sJ d| d }| d¡}| dd¡}| dd¡}| dd¡}|dur1t |¡ | d ¡}| d
d¡}|durEt |¡ | d ¡}t|||||}|du sZ|d v sZt||}|ret|| d d¡}|S )
zÆ
Returns the loss function based on provided configuration
:param config: (dict) a top level configuration object containing the 'loss' key
:return: an instance of the loss function
r z*Could not find loss function configurationÚnamer NÚskip_last_targetFr ÚdeviceÚ
pos_weight)ÚCrossEntropyLossrU r1 )Úpopr5 ri ÚtoÚ _create_lossr r/ Úget)ÚconfigÚ loss_configrl r rm r ro r r r r Úget_loss_criterionî s"
rw c C s0 | dkr
t j|dS | dkr | dd¡}| dd¡}t||S | dkr0|d u r)d}t j||d S | d
kr?|d u r:d}t|d S | d krIt||d
S | dkrX| dd¡}t|dS | dkrh| dd¡}t||dS | dkrot S | dkrvt
S | dkr}t S | dkrt |d |d | dd¡d S t
d | d ) NrR )ro rP ZalphsrG rT rp iÿÿÿrV rU )r ÚPixelWiseCrossEntropyLoss)r\ r rM r@ r9 )r@ rK rN r r r r^ r` rb ra T)r` rb ra z Unsupported loss function: 'ú')r rR rt rP rp rU rx rM rK r r r r^ Ú RuntimeError)rl rv r r ro rQ rT r@ r r r rs sD
þrs )r N)r5 Útorch.nn.functionalr Ú
functionalrX Útorch.autogradr Útorch.nnr r r r ÚModuler r/ r8 rK rM rP rU r^ r rw rs r r r r Ú<module> s"
$
# %