def CCsigmoid_focal_loss(pred,
target,
weight=None,
gamma=2.0,
alpha=0.25,
reduction='mean',
avg_factor=None):
# Function.apply does not accept keyword arguments, so the decorator
# "weighted_loss" is not applicable
loss1 = _sigmoid_focal_loss(pred, target[0], gamma, alpha)
loss2 = _sigmoid_focal_loss(pred, target[1], gamma, alpha)
weight1 = weight[0].view(-1, 1)
weight2 = weight[1].view(-1, 1)
loss1 = loss1 * weight1
loss2 = loss2 * weight2
loss = 0.5 * loss1 + 0.5 * loss2
background_label = loss1.shape[1]
pos_ind_only1 = ((target[0] < background_label) &
(target[1] == background_label)).nonzero().reshape(-1)
pos_ind_only2 = ((target[1] < background_label) &
(target[0] == background_label)).nonzero().reshape(-1)
pos_ind_both = ((target[1] < background_label) &
(target[0] < background_label)).nonzero().reshape(-1)
sig_ind_only1 = target[0][pos_ind_both]
sig_ind_only2 = target[1][pos_ind_both]
loss[pos_ind_only1] = loss1[pos_ind_only1]
loss[pos_ind_only2] = loss2[pos_ind_only2]
loss[pos_ind_both, sig_ind_only1] = loss1[pos_ind_both, sig_ind_only1]
loss[pos_ind_both, sig_ind_only2] = loss2[pos_ind_both, sig_ind_only2]
loss = weight_reduce_loss(
loss, None, reduction,
avg_factor) # set weight == None since already multiplied
return loss
def sigmoid_focal_loss(pred,
target,
weight=None,
gamma=2.0,
alpha=0.25,
reduction='mean',
avg_factor=None,
use_py_version=False):
# Function.apply does not accept keyword arguments, so the decorator
# "weighted_loss" is not applicable
if use_py_version:
loss = py_sigmoid_focal_loss(pred,
target,
weight=weight,
gamma=gamma,
alpha=alpha,
reduction=reduction,
avg_factor=avg_factor)
else:
loss = _sigmoid_focal_loss(pred, target, gamma, alpha)
# TODO: find a proper way to handle the shape of weight
if weight is not None:
weight = weight.view(-1, 1)
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
def sigmoid_focal_loss(pred,
target,
weight=None,
gamma=2.0,
alpha=0.25,
reduction='mean',
avg_factor=None):
# Function.apply does not accept keyword arguments, so the decorator
# "weighted_loss" is not applicable
loss = _sigmoid_focal_loss(pred, target, gamma, alpha)
# TODO: find a proper way to handle the shape of weight
if weight is not None:
if weight.shape != loss.shape:
if weight.size(0) == loss.size(0):
# For most cases, weight is of shape (num_priors, ),
# which means it does not have the second axis num_class
weight = weight.view(-1, 1)
else:
# Sometimes, weight per anchor per class is also needed. e.g.
# in FSAF. But it may be flattened of shape
# (num_priors x num_class, ), while loss is still of shape
# (num_priors, num_class).
assert weight.numel() == loss.numel()
weight = weight.view(loss.size(0), -1)
assert weight.ndim == loss.ndim
# if weight is not None:
# weight = weight.view(-1, 1)
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
def sigmoid_focal_loss(pred,
target,
weight=None,
gamma=2.0,
alpha=0.25,
reduction='mean',
avg_factor=None):
"""A warpper of cuda version
`Focal Loss <https://arxiv.org/abs/1708.02002>`_
Args:
pred (torch.Tensor): The prediction with shape (N, C), C is the number
of classes.
target (torch.Tensor): The learning label of the prediction.
weight (torch.Tensor, optional): Sample-wise loss weight.
gamma (float, optional): The gamma for calculating the modulating
factor. Defaults to 2.0.
alpha (float, optional): A balanced form for Focal Loss.
Defaults to 0.25.
reduction (str, optional): The method used to reduce the loss into
a scalar. Defaults to 'mean'. Options are "none", "mean" and "sum".
avg_factor (int, optional): Average factor that is used to average
the loss. Defaults to None.
"""
# Function.apply does not accept keyword arguments, so the decorator
# "weighted_loss" is not applicable
loss = _sigmoid_focal_loss(pred, target, gamma, alpha)
if weight is not None:
if weight.shape != loss.shape:
if weight.size(0) == loss.size(0):
# For most cases, weight is of shape (num_priors, ),
# which means it does not have the second axis num_class
weight = weight.view(-1, 1)
else:
# Sometimes, weight per anchor per class is also needed. e.g.
# in FSAF. But it may be flattened of shape
# (num_priors x num_class, ), while loss is still of shape
# (num_priors, num_class).
assert weight.numel() == loss.numel()
weight = weight.view(loss.size(0), -1)
assert weight.ndim == loss.ndim
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
def weighted_focal_loss(pred,
target,
weight=None,
gamma=2.0,
alpha=0.25,
reduction='mean',
avg_factor=None):
# Function.apply does not accept keyword arguments, so the decorator
# "weighted_loss" is not applicable
loss = _sigmoid_focal_loss(pred, target, gamma, alpha)
# TODO: find a proper way to handle the shape of weight
weight = loss.new(19).fill_(1)
weight[11] = 0.5
weight[17] = 0.1
weight[10] = 0.25
# large-vehicle
# small-vehicle
# ship
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
def sigmoid_focal_loss(pred,
target,
weight=None,
gamma=2.0,
alpha=0.25,
reduction='mean',
avg_factor=None,
cate_loss_weight=None):
# Function.apply does not accept keyword arguments, so the decorator
# "weighted_loss" is not applicable
loss = _sigmoid_focal_loss(pred, target, gamma, alpha)
if cate_loss_weight is None:
cate_loss_weight = [1.0] * loss.shape[1]
cate_loss_weight = torch.tensor(cate_loss_weight).unsqueeze(0).cuda()
loss = loss * cate_loss_weight
# TODO: find a proper way to handle the shape of weight
if weight is not None:
weight = weight.view(-1, 1)
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
def sigmoid_focal_loss(pred,
target,
weight=None,
gamma=2.0,
alpha=0.25,
reduction='mean',
avg_factor=None):
'''
pred: [num_points, C]
targets: [num_points] value in {-1,0,1,...,C}, -1 means ignorance, 0 means background
p = sigmoid(pred)
loss = - alpha*(label==对应label) * (1-p)**gamma * log(p) - (1-alpha)*(label!=对应label or label==negative) * p**gamma * log(1-p)
参考: https://github.com/pytorch/pytorch/blob/master/modules/detectron/sigmoid_focal_loss_op.cc
'''
# Function.apply does not accept keyword arguments, so the decorator
# "weighted_loss" is not applicable
loss = _sigmoid_focal_loss(pred, target, gamma, alpha)
# TODO: find a proper way to handle the shape of weight
if weight is not None:
weight = weight.view(-1, 1)
loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
return loss
