Diff of unnecessarily brutal changes to get DC-IGN working without GPU

gistfile1.diff

diff --git a/config.lua b/config.lua
index c1b1cb2..93c6984 100644
--- a/config.lua
+++ b/config.lua
@@ -1,6 +1,6 @@
 require 'modules/prequire'
 require 'modules/UnPooling'
-cunn_mod = prequire('cunn')
+-- cunn_mod = prequire('cunn')
 --Global variables for config
 bsize = 50
 imwidth = 150
@@ -9,9 +9,9 @@ TOTALFACES = 5230
 num_train_batches = 5000
 num_test_batches =  TOTALFACES-num_train_batches
 
-if not cudnn then
-  cudnn = nn
-end
+-- if not cudnn then
+  -- cudnn = nn
+-- end
 
 -- config = {
 --     learningRate = -0.0005,
@@ -89,17 +89,17 @@ function init_network2_150_mv(dim_hidden, feature_maps)
 
   encoder = nn.Sequential()
 
-  encoder:add(cudnn.SpatialConvolution(colorchannels,feature_maps,filter_size,filter_size))
-  encoder:add(cudnn.SpatialMaxPooling(2,2,2,2))
+  encoder:add(nn.SpatialConvolution(colorchannels,feature_maps,filter_size,filter_size))
+  encoder:add(nn.SpatialMaxPooling(2,2,2,2))
   encoder:add(nn.Threshold(0,1e-6))
 
-  encoder:add(cudnn.SpatialConvolution(feature_maps,feature_maps/2,filter_size,filter_size))
-  encoder:add(cudnn.SpatialMaxPooling(2,2,2,2))
+  encoder:add(nn.SpatialConvolution(feature_maps,feature_maps/2,filter_size,filter_size))
+  encoder:add(nn.SpatialMaxPooling(2,2,2,2))
   encoder:add(nn.Threshold(0,1e-6))
 
 
-  encoder:add(cudnn.SpatialConvolution(feature_maps/2,feature_maps/4,filter_size,filter_size))
-  encoder:add(cudnn.SpatialMaxPooling(2,2,2,2))
+  encoder:add(nn.SpatialConvolution(feature_maps/2,feature_maps/4,filter_size,filter_size))
+  encoder:add(nn.SpatialMaxPooling(2,2,2,2))
   encoder:add(nn.Threshold(0,1e-6))
 
   encoder:add(nn.Reshape((feature_maps/4)*15*15))
@@ -127,29 +127,29 @@ function init_network2_150_mv(dim_hidden, feature_maps)
   decoder:add(nn.Reshape((feature_maps/4),15,15))
 
   decoder:add(nn.SpatialUpSamplingNearest(2))
-  decoder:add(cudnn.SpatialConvolution(feature_maps/4,feature_maps/2, 7, 7))
+  decoder:add(nn.SpatialConvolution(feature_maps/4,feature_maps/2, 7, 7))
   decoder:add(nn.Threshold(0,1e-6))
 
   decoder:add(nn.SpatialUpSamplingNearest(2))
-  decoder:add(cudnn.SpatialConvolution(feature_maps/2,feature_maps,7,7))
+  decoder:add(nn.SpatialConvolution(feature_maps/2,feature_maps,7,7))
   decoder:add(nn.Threshold(0,1e-6))
 
   decoder:add(nn.SpatialUpSamplingNearest(2))
-  decoder:add(cudnn.SpatialConvolution(feature_maps,feature_maps,7,7))
+  decoder:add(nn.SpatialConvolution(feature_maps,feature_maps,7,7))
   decoder:add(nn.Threshold(0,1e-6))
 
   decoder:add(nn.SpatialUpSamplingNearest(2))
-  decoder:add(cudnn.SpatialConvolution(feature_maps,1,7,7))
-  decoder:add(cudnn.Sigmoid())
+  decoder:add(nn.SpatialConvolution(feature_maps,1,7,7))
+  decoder:add(nn.Sigmoid())
 
   model = nn.Sequential()
   model:add(encoder)
   model:add(nn.Reparametrize(dim_hidden))
   model:add(decoder)
 
-  if cunn_mod then
-    model:cuda()
-  end
+  -- if cunn_mod then
+  --   model:cuda()
+  -- end
   collectgarbage()
   return model
 end
diff --git a/modules/Reparametrize.lua b/modules/Reparametrize.lua
index 43bf6a7..a531080 100644
--- a/modules/Reparametrize.lua
+++ b/modules/Reparametrize.lua
@@ -8,17 +8,17 @@ function Reparametrize:__init(dimension)
     self.size = torch.LongStorage()
     self.dimension = dimension
     self.gradInput = {}
-end 
+end
 
 function Reparametrize:updateOutput(input)
     self.eps = torch.randn(input[2]:size(1),self.dimension)
 
-    if torch.typename(input[1]) == 'torch.CudaTensor' then
-        self.eps = self.eps:cuda()
-        self.output = torch.CudaTensor():resizeAs(input[2]):fill(0.5)
-    else
+    -- if torch.typename(input[1]) == 'torch.CudaTensor' then
+    --     self.eps = self.eps:cuda()
+    --     self.output = torch.CudaTensor():resizeAs(input[2]):fill(0.5)
+    -- else
         self.output = torch.Tensor():resizeAs(input[2]):fill(0.5)
-    end
+    -- end
 
     self.output:cmul(input[2]):exp():cmul(self.eps)
 
@@ -31,13 +31,13 @@ end
 function Reparametrize:updateGradInput(input, gradOutput)
     -- Derivative with respect to mean is 1
     self.gradInput[1] = gradOutput:clone()
-    
+
     --test gradient with Jacobian
-    if torch.typename(input[1]) == 'torch.CudaTensor' then
-        self.gradInput[2] = torch.CudaTensor():resizeAs(input[2]):fill(0.5)
-    else
+    -- if torch.typename(input[1]) == 'torch.CudaTensor' then
+    --     self.gradInput[2] = torch.CudaTensor():resizeAs(input[2]):fill(0.5)
+    -- else
         self.gradInput[2] = torch.Tensor():resizeAs(input[2]):fill(0.5)
-    end
+    -- end
 
     self.gradInput[2]:cmul(input[2]):exp():mul(0.5):cmul(self.eps)
     self.gradInput[2]:cmul(gradOutput)
diff --git a/modules/UnPooling.lua b/modules/UnPooling.lua
index 5359713..d8f80df 100644
--- a/modules/UnPooling.lua
+++ b/modules/UnPooling.lua
@@ -17,9 +17,9 @@ function UnPooling:updateOutput(input)
    local bsize = input:size()[1]
    local dim  = input:size()[3]
    local sdim = dim*self.scale
-   
+
    self.output = torch.zeros(bsize, input:size()[2] , sdim, sdim )
-   
+
    local ii,jj,kk; ii=1;jj=1;kk=1;
 
    self.mapping = {} -- store non-zero mappings for gradient calc
@@ -34,7 +34,7 @@ function UnPooling:updateOutput(input)
       ii = ii + 1;
    end
 
-   self.output = self.output:cuda()
+   -- self.output = self.output:cuda()
    return self.output
 end
 
@@ -44,18 +44,18 @@ function UnPooling:updateGradInput(input, gradOutput)
    input = input:float()
 
    local dim  = input:size()[3]
-   
+
    self.gradInput = torch.zeros(bsize, input:size()[2], dim, dim)
 
    for ii=1,dim do
       for jj=1,dim do
          local t = self.mapping[ii .. jj]
          i = t[1]; j = t[2];
-         self.gradInput[{{},{},ii,jj}] = gradOutput[{{},{}, i,j}]   
+         self.gradInput[{{},{},ii,jj}] = gradOutput[{{},{}, i,j}]
       end
    end
 
-   self.gradInput = self.gradInput:cuda()
+   -- self.gradInput = self.gradInput:cuda()
    return self.gradInput
 end
 
diff --git a/monovariant_main.lua b/monovariant_main.lua
index 2c4d8c3..2ff835d 100644
--- a/monovariant_main.lua
+++ b/monovariant_main.lua
@@ -14,16 +14,16 @@ require 'modules/Reparametrize'
 require 'modules/SelectiveOutputClamp'
 require 'modules/SelectiveGradientFilter'
 
-cutorch_mod = prequire('cutorch')
-cunn_mod = prequire('cunn')
+-- cutorch_mod = prequire('cutorch')
+-- cunn_mod = prequire('cunn')
 require 'optim'
 require 'testf'
 require 'utils'
 require 'config'
 
-if cutorch_mod then
-  print(  cutorch.getDeviceProperties(cutorch.getDevice()) )
-end
+-- if cutorch_mod then
+  -- print(  cutorch.getDeviceProperties(cutorch.getDevice()) )
+-- end
 
 cmd = torch.CmdLine()
 cmd:text()
@@ -38,8 +38,8 @@ cmd:option('--networks_dir',      'networks',     'the directory to save the res
 cmd:option('--name',              'default',      'the name for this network. used for saving the network and results')
 cmd:option('--datasetdir',        'DATASET',      'dataset source directory')
 
-cmd:option('--dim_hidden',        200,            'dimension of the representation layer')
-cmd:option('--feature_maps',      96,             'number of feature maps')
+cmd:option('--dim_hidden',        40,            'dimension of the representation layer')
+cmd:option('--feature_maps',      24,             'number of feature maps')
 
 cmd:option('--force_invariance',  false,          'propagate error equal to change in outputs corresponding to fixed variables')
 cmd:option('--invariance_strength',0,             'multiplier for the invariance error signal')
@@ -55,7 +55,7 @@ cmd:text()
 cmd:text()
 
 cmd:text("Probably don't change these:")
-cmd:option('--threads',2,'how many threads to use in torch')
+cmd:option('--threads',8,'how many threads to use in torch')
 cmd:option('--num_train_batches',5000,'number of batches to train with per epoch')
 cmd:option('--num_train_batches_per_type',3000,'number of available train batches of each data type')
 cmd:option('--num_test_batches',1400,'number of batches to test with')
@@ -66,7 +66,7 @@ cmd:text()
 
 opt = cmd:parse(arg)
 opt.save = paths.concat(opt.networks_dir, opt.name)
-os.execute('mkdir ' .. opt.save)
+os.execute('mkdir -p ' .. opt.save)
 
 config = {
     learningRate = opt.learning_rate,
@@ -98,12 +98,12 @@ criterion.sizeAverage = false
 KLD = nn.KLDCriterion()
 KLD.sizeAverage = false
 
-if cutorch_mod then
-  criterion:cuda()
-  KLD:cuda()
-  model:cuda()
-  cutorch.synchronize()
-end
+-- if cutorch_mod then
+--   criterion:cuda()
+--   KLD:cuda()
+--   model:cuda()
+--   cutorch.synchronize()
+-- end
 
 parameters, gradients = model:getParameters()
 print('Num before', #parameters)
@@ -175,9 +175,9 @@ while true do
       gradFilters[clampIndex].active = true
     end
 
-    if cutorch_mod then
-      batch = batch:cuda()
-    end
+    -- if cutorch_mod then
+      -- batch = batch:cuda()
+    -- end
 
     --Optimization function
     local opfunc = function(x)
diff --git a/utils.lua b/utils.lua
index 1ea13f1..b0f6600 100644
--- a/utils.lua
+++ b/utils.lua
@@ -5,7 +5,8 @@ function load_batch(id, mode)
 end
 
 function load_mv_batch(id, dataset_name, mode)
-    return torch.load(opt.datasetdir .. '/th_' .. dataset_name .. '/' .. mode .. '/batch' .. id)
+    return torch.rand(30, 1, 150, 150)
+    -- return torch.load(opt.datasetdir .. '/th_' .. dataset_name .. '/' .. mode .. '/batch' .. id)
 end
 
 function load_random_mv_batch(mode)