2015-10-31 08:36:20 +13:00
|
|
|
-- ref: https://en.wikipedia.org/wiki/Huber_loss
|
2015-11-08 09:44:14 +13:00
|
|
|
local ClippedWeightedHuberCriterion, parent = torch.class('w2nn.ClippedWeightedHuberCriterion','nn.Criterion')
|
2015-10-31 08:36:20 +13:00
|
|
|
|
2015-11-08 09:44:14 +13:00
|
|
|
function ClippedWeightedHuberCriterion:__init(w, gamma, clip)
|
2015-10-31 08:36:20 +13:00
|
|
|
parent.__init(self)
|
2015-11-08 09:44:14 +13:00
|
|
|
self.clip = clip
|
2015-10-31 08:36:20 +13:00
|
|
|
self.gamma = gamma or 1.0
|
|
|
|
self.weight = w:clone()
|
|
|
|
self.diff = torch.Tensor()
|
|
|
|
self.diff_abs = torch.Tensor()
|
|
|
|
--self.outlier_rate = 0.0
|
|
|
|
self.square_loss_buff = torch.Tensor()
|
|
|
|
self.linear_loss_buff = torch.Tensor()
|
|
|
|
end
|
2015-11-08 09:44:14 +13:00
|
|
|
function ClippedWeightedHuberCriterion:updateOutput(input, target)
|
2015-10-31 08:36:20 +13:00
|
|
|
self.diff:resizeAs(input):copy(input)
|
2016-07-04 14:16:04 +12:00
|
|
|
self.diff:clamp(self.clip[1], self.clip[2])
|
2015-10-31 08:36:20 +13:00
|
|
|
for i = 1, input:size(1) do
|
|
|
|
self.diff[i]:add(-1, target[i]):cmul(self.weight)
|
|
|
|
end
|
|
|
|
self.diff_abs:resizeAs(self.diff):copy(self.diff):abs()
|
|
|
|
|
|
|
|
local square_targets = self.diff[torch.lt(self.diff_abs, self.gamma)]
|
|
|
|
local linear_targets = self.diff[torch.ge(self.diff_abs, self.gamma)]
|
|
|
|
local square_loss = self.square_loss_buff:resizeAs(square_targets):copy(square_targets):pow(2.0):mul(0.5):sum()
|
|
|
|
local linear_loss = self.linear_loss_buff:resizeAs(linear_targets):copy(linear_targets):abs():add(-0.5 * self.gamma):mul(self.gamma):sum()
|
|
|
|
|
|
|
|
--self.outlier_rate = linear_targets:nElement() / input:nElement()
|
|
|
|
self.output = (square_loss + linear_loss) / input:nElement()
|
|
|
|
return self.output
|
|
|
|
end
|
2015-11-08 09:44:14 +13:00
|
|
|
function ClippedWeightedHuberCriterion:updateGradInput(input, target)
|
2015-10-31 08:36:20 +13:00
|
|
|
local norm = 1.0 / input:nElement()
|
|
|
|
self.gradInput:resizeAs(self.diff):copy(self.diff):mul(norm)
|
|
|
|
local outlier = torch.ge(self.diff_abs, self.gamma)
|
|
|
|
self.gradInput[outlier] = torch.sign(self.diff[outlier]) * self.gamma * norm
|
|
|
|
return self.gradInput
|
|
|
|
end
|