def focal_loss_alt(self, x, y, iw_local):
'''Focal loss alternative.
Args:
x: (tensor) sized [N,D].
y: (tensor) sized [N,].
Return:
(tensor) focal loss.
'''
alpha = 0.25
#n=iw.size()[0]
t = one_hot_embedding(y.data.cpu(), 1 + self.num_classes)
t = t[:, 1:]
t = Variable(t).cuda()
xt = x * (2 * t - 1) # xt = x if t > 0 else -x
pt = (2 * xt + 1).sigmoid()
pt = pt.clamp(min=1e-15)
w = alpha * t + (1 - alpha) * (1 - t)
loss = -w * pt.log() / 2
w_loss = iw_local.unsqueeze(1) * loss
'''
m=loss.size()[0]
i=0
w_loss=0
while i<n:
w_loss+=iw[i]*loss[i*m/n:(i+1)*m/n].sum()
i+=1
'''
return w_loss.sum()
def focal_loss(self, x, y):
'''Focal loss.
Args:
x: (tensor) sized [N,D].
y: (tensor) sized [N,].
Return:
(tensor) focal loss.
'''
alpha = 0.25
gamma = 2
t = one_hot_embedding(y.data.cpu(), 1+self.num_classes) # [N,21]
t = t[:,1:] # exclude background
t = Variable(t).cuda() # [N,20]
p = x.sigmoid()
pt = p*t + (1-p)*(1-t) # pt = p if t > 0 else 1-p
w = alpha*t + (1-alpha)*(1-t) # w = alpha if t > 0 else 1-alpha
w = w * (1-pt).pow(gamma)
return F.binary_cross_entropy_with_logits(x, t, w, size_average=False)