andreaskoepf

require 'image'
require 'nngraph'
require 'optim'

-- mini single layer denoising auto-encoder experiment 
-- (v1: no weight tying yet)
-- Andreas Köpf 2015-09-16 

input = image.lena() -- 3x512x512
input_size = input:size()

andreaskoepf

require 'nn'

-- mini bottleneck auto-encoder weight tying demo
net = nn.Sequential()
net:add(nn.Linear(8, 8))
net:add(nn.PReLU())
net:add(nn.Linear(8, 3))
net:add(nn.PReLU())
net:add(nn.Linear(3, 8))

andreaskoepf

local RotationDistance, parent = torch.class('RotationDistance', 'nn.Criterion')

function RotationDistance:__init(weights)
   parent.__init(self)
end

function RotationDistance:updateOutput(input, target)
  -- acos(abs(<a,b> / norm(a) / norm(b))) + abs(1 - norm(a))
  local one = torch.ones(input:size(1)):cuda() 
  local a = torch.cmul(input, target):sum(2):cdiv(torch.cmul(input:norm(2, 2), target:norm(2, 2))):abs():clamp(-1, 1):acos()

andreaskoepf

using System;
using System.Collections.Concurrent;
using System.Collections.Generic;
using System.Linq;
using System.Reactive;
using System.Reactive.Concurrency;
using System.Reactive.Disposables;
using System.Reactive.Linq;
using System.Reactive.Subjects;
using System.Threading;

andreaskoepf

--
local Foo, parent = torch.class('Foo')
Foo.__version = 1

function Foo:__init()
  parent.__init(self)
end

-- serialize 'old' object
old = Foo.new()

andreaskoepf

require 'nn'

x = torch.rand(1,5,5)
a = nn.SpatialConvolution(1,1,3,3)
a.bias:zero()
ay1 =torch.xcorr2(x,a.weight,'V')
ay2 = a:forward(x)

b = nn.SpatialFullConvolution(1,1,3,3)
b.bias:zero()

andreaskoepf

local SpatialUnpooling, parent = torch.class('nn.SpatialUnpooling', 'nn.Module')

function SpatialUnpooling:__init(kW, kH, dW, dH, padW, padH)
   parent.__init(self)
   self.dW = dW or kW
   self.dH = dH or kH
   self.padW = padW or 0
   self.padH = padH or 0
   self.indices = torch.LongTensor()
   self._indexTensor = torch.LongTensor()

andreaskoepf

local SpatialUnpooling, parent = torch.class('nn.SpatialUnpooling', 'nn.Module')

function SpatialUnpooling:__init(kW, kH, dW, dH, padW, padH)
   parent.__init(self)
   self.dW = dW or kW
   self.dH = dH or kH
   self.padW = padW or 0
   self.padH = padH or 0
   self.indices = torch.LongTensor()
   self._indexTensor = torch.LongTensor()

andreaskoepf

x = torch.DoubleTensor()
y = torch.FloatTensor()

function testfunc(a,b)
return a + b
end

f = {}
f[x:type()] = {}
f[y:type()] = {}

andreaskoepf

require 'cunn'

--test cuda & non-cuda version
function testVolumetricFullConvolution()
  local input = torch.rand(1,2,10,10,10) * 2 - 1
  local a = nn.VolumetricFullConvolution(2,3, 3,3,3, 1,1,1)
  local b = nn.VolumetricFullConvolution(2,3, 3,3,3, 1,1,1)
  b:cuda()
  
  b.weight = a.weight:cuda()