The code of ‘Sampler.lua’ in torch7 is there:
require ‘nn’
local Sampler, parent = torch.class(‘nn.Sampler’, ‘nn.Module’)
function Sampler:__init()
parent.__init(self)
self.gradInput = {}
end
function Sampler:updateOutput(input)
self.eps = self.eps or input[1].new()
self.eps:resizeAs(input[1]):copy(torch.randn(input[1]:size()))
self.ouput = self.output or self.output.new() self.output:resizeAs(input[2]):copy(input[2]) self.output:mul(0.5):exp():cmul(self.eps)
self.output:add(input[1])
return self.output
end
function Sampler:updateGradInput(input, gradOutput)
self.gradInput[1] = self.gradInput[1] or input[1].new()
self.gradInput[1]:resizeAs(gradOutput):copy(gradOutput)self.gradInput[2] = self.gradInput[2] or input[2].new() self.gradInput[2]:resizeAs(gradOutput):copy(input[2]) self.gradInput[2]:mul(0.5):exp():mul(0.5):cmul(self.eps) self.gradInput[2]:cmul(gradOutput)
return self.gradInput
end
the input has 2 elements:(In lua, elements start with 1 instead of 2)
input[1] = (batch_size, size) stand for mean
input[2] = (batch_size, size) strand for log_var
for example:
input[1] = (100, 20)
input[2] = (100, 20)
use the mean and log_var sampling. After that get output(100, 20)
It looks like there is no ‘nn.Sampler’ in pytorch.
So how to do it? Thanks~