-
Notifications
You must be signed in to change notification settings - Fork 17
/
BinaryClassReward.lua
82 lines (71 loc) · 2.83 KB
/
BinaryClassReward.lua
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
------------------------------------------------------------------------
--[[ BinaryClassReward ]]--
-- Variance reduced binary classification reinforcement criterion.
-- The binary class version of VRClassReward.
-- input : {class prediction, baseline reward}
-- Reward is 1 for success, Reward is 0 otherwise.
-- reward = scale*(Reward - baseline) where baseline is 2nd input element
-- Note : for RNNs with R = 1 for last step in sequence, encapsulate it
-- in nn.ModuleCriterion(BinaryClassReward, nn.SelectTable(-1))
------------------------------------------------------------------------
local BinaryClassReward, parent = torch.class("nn.BinaryClassReward", "nn.Criterion")
function BinaryClassReward:__init(module, scale, criterion)
parent.__init(self)
self.module = module -- so it can call module:reinforce(reward)
self.scale = scale or 1 -- scale of reward
self.criterion = criterion or nn.MSECriterion() -- baseline criterion
self.sizeAverage = true
self.gradInput = {torch.Tensor()}
end
function BinaryClassReward:updateOutput(input, target)
assert(torch.type(input) == 'table')
local input = input[1]
assert(input:dim() == 1)
assert(target:dim() == 1)
self._binary = self._binary or input.new()
self._binary:gt(input, 0.5)
-- max class value is class prediction
if torch.type(self._binary) ~= torch.type(target) then
self._target = self._target or self._binary.new()
self._target:resize(target:size()):copy(target)
target = self._target
end
-- reward = scale when correctly classified
self._reward = self._reward or input.new()
self._reward:eq(self._binary, target)
self.reward = self.reward or input.new()
self.reward:resize(self._reward:size(1)):copy(self._reward)
self.reward:mul(self.scale)
-- loss = -sum(reward)
self.output = -self.reward:sum()
if self.sizeAverage then
self.output = self.output/input:size(1)
end
return self.output
end
function BinaryClassReward:updateGradInput(inputTable, target)
local input, baseline = unpack(inputTable)
-- reduce variance of reward using baseline
self.vrReward = self.vrReward or self.reward.new()
self.vrReward:resizeAs(self.reward):copy(self.reward)
self.vrReward:add(-1, baseline)
if self.sizeAverage then
self.vrReward:div(input:size(1))
end
-- broadcast reward to modules
self.module:reinforce(self.vrReward)
-- zero gradInput (this criterion has no gradInput for class pred)
self.gradInput[1]:resizeAs(input):zero()
-- learn the baseline reward
self.gradInput[2] = self.criterion:backward(baseline, self.reward)
return self.gradInput
end
function BinaryClassReward:type(type)
self._binary = nil
self._target = nil
local module = self.module
self.module = nil
local ret = parent.type(self, type)
self.module = module
return ret
end