我读过一些关于使用“自助交叉熵损失”来训练分割网络的论文。这个想法是只关注最难的k%(比如说15%)的像素,以提高学习性能,特别是当易于处理的像素占主导地位时。
目前,我正在使用标准的交叉熵:
loss = F.binary_cross_entropy(mask, gt)
如何在PyTorch中高效地将此转换为引导版本?
我读过一些关于使用“自助交叉熵损失”来训练分割网络的论文。这个想法是只关注最难的k%(比如说15%)的像素,以提高学习性能,特别是当易于处理的像素占主导地位时。
目前,我正在使用标准的交叉熵:
loss = F.binary_cross_entropy(mask, gt)
如何在PyTorch中高效地将此转换为引导版本?
通常我们还会给损失函数添加“热身”时间,使网络首先适应简单区域,然后过渡到更难的区域。
此实现从k=100
开始,并持续20000次迭代,然后线性衰减到k=15
,再进行50000次迭代。
class BootstrappedCE(nn.Module):
def __init__(self, start_warm=20000, end_warm=70000, top_p=0.15):
super().__init__()
self.start_warm = start_warm
self.end_warm = end_warm
self.top_p = top_p
def forward(self, input, target, it):
if it < self.start_warm:
return F.cross_entropy(input, target), 1.0
raw_loss = F.cross_entropy(input, target, reduction='none').view(-1)
num_pixels = raw_loss.numel()
if it > self.end_warm:
this_p = self.top_p
else:
this_p = self.top_p + (1-self.top_p)*((self.end_warm-it)/(self.end_warm-self.start_warm))
loss, _ = torch.topk(raw_loss, int(num_pixels * this_p), sorted=False)
return loss.mean(), this_p
除了@hkchengrex提供的自我回答(为了未来自我和与PyTorch的API相似性),
可以先实现functional
版本(在原始torch.nn.functional.cross_entropy
中提供一些额外的参数),像这样(我更喜欢将reduction
作为callable
而不是预定义字符串):
import typing
import torch
def bootstrapped_cross_entropy(
inputs,
targets,
iteration,
p: float,
warmup: typing.Union[typing.Callable[[float, int], float], int] = -1,
weight=None,
ignore_index=-100,
reduction: typing.Callable[[torch.Tensor], torch.Tensor] = torch.mean,
):
if not 0 < p < 1:
raise ValueError("p should be in [0, 1] range, got: {}".format(p))
if isinstance(warmup, int):
this_p = 1.0 if iteration < warmup else p
elif callable(warmup):
this_p = warmup(p, iteration)
else:
raise ValueError(
"warmup should be int or callable, got {}".format(type(warmup))
)
# Shortcut
if this_p == 1.0:
return torch.nn.functional.cross_entropy(
inputs, targets, weight, ignore_index=ignore_index, reduction=reduction
)
raw_loss = torch.nn.functional.cross_entropy(
inputs, targets, weight=weight, ignore_index=ignore_index, reduction="none"
).view(-1)
num_pixels = raw_loss.numel()
loss, _ = torch.topk(raw_loss, int(num_pixels * this_p), sorted=False)
return reduction(loss)
同时,可以将warmup
指定为callable
(接受p
和当前iteration
)或int
,从而实现灵活或轻松的调度。
并且可以基于_WeightedLoss
和自动递增的iteration
创建一个类(因此只需要传递inputs
和targets
):
class BoostrappedCrossEntropy(torch.nn.modules.loss._WeightedLoss):
def __init__(
self,
p: float,
warmup: typing.Union[typing.Callable[[float, int], float], int] = -1,
weight=None,
ignore_index=-100,
reduction: typing.Callable[[torch.Tensor], torch.Tensor] = torch.mean,
):
self.p = p
self.warmup = warmup
self.ignore_index = ignore_index
self._current_iteration = -1
super().__init__(weight, size_average=None, reduce=None, reduction=reduction)
def forward(self, inputs, targets):
self._current_iteration += 1
return bootstrapped_cross_entropy(
inputs,
targets,
self._current_iteration,
self.p,
self.warmup,
self.weight,
self.ignore_index,
self.reduction,
)