Merge branch 'master' of github.com:clementfarabet/lua---nnx

2 files changed, 69 insertions, 1 deletions

diff --git a/SpatialFovea.lua b/SpatialFovea.lua
index b9e3fe4..c4ebcb8 100644
--- a/SpatialFovea.lua
+++ b/SpatialFovea.lua
@@ -116,6 +116,11 @@ function SpatialFovea:configure(width,height)
       else
          self.upsamplers[idx] = nn.SpatialUpSampling(r, r)
       end
+
+      -- set correct types
+      self.downsamplers[idx]:type(self.output:type())
+      self.padders[idx]:type(self.output:type())
+      self.upsamplers[idx]:type(self.output:type())
    end
 end
 
@@ -265,7 +270,7 @@ function SpatialFovea:backward(input, gradOutput)
    -- (4) is fovea focused ?
    if self.focused then
       for idx = 1,nscales do
-         self.gradPadded[idx] = self.gradPadded[idx] or torch.Tensor()
+         self.gradPadded[idx] = self.gradPadded[idx] or torch.Tensor():typeAs(self.output)
          self.gradPadded[idx]:resizeAs(self.padded[idx]):zero()
          local fov = self.fov
          local ox = math.floor(math.floor((self.x-1) / self.ratios[idx]) / self.sub) * self.sub + 1
@@ -327,6 +332,20 @@ function SpatialFovea:updateParameters(learningRate)
    end
 end
 
+function SpatialFovea:type(type)
+   parent.type(self,type)
+   for idx = 1,#self.processors do
+      self.processors[idx]:type(type)
+      self.upsamplers[idx]:type(type)
+      self.downsamplers[idx]:type(type)
+      self.padders[idx]:type(type)
+   end
+   for idx = 1,#self.preProcessors do
+      self.preProcessors[idx]:type(type)
+   end
+   return self
+end
+
 function SpatialFovea:write(file)
    parent.write(self, file)
    file:writeInt(self.nInputPlane)
diff --git a/Type.lua b/Type.lua
new file mode 100644
index 0000000..c65f735
--- /dev/null
+++ b/Type.lua
@@ -0,0 +1,49 @@
+local Type, parent = torch.class('nn.Type', 'nn.Sequential')
+
+function Type:__init(type)
+   parent.__init(self)
+   if not type:find('torch%..+Tensor') then
+      type = 'torch.' .. type .. 'Tensor'
+   end
+   self.type = type
+   self.defaulttype = torch.getdefaulttensortype()
+   self.convert_input = nn.Copy(self.defaulttype, self.type)
+   self.convert_gradOutput = nn.Copy(self.defaulttype, self.type)
+   self.convert_output = nn.Copy(self.type, self.defaulttype)
+   self.convert_gradInput = nn.Copy(self.type, self.defaulttype)
+end
+
+function Type:add(module)
+   parent.add(self, module:type(self.type))
+   return self
+end
+
+function Type:forward(input)
+   input = self.convert_input:forward(input)
+   local output = parent.forward(self, input)
+   self.output = self.convert_output:forward(output)
+   return self.output
+end
+
+function Type:backward(input, gradOutput)
+   input = self.convert_input:forward(input)
+   gradOutput = self.convert_gradOutput:forward(gradOutput)
+   local gradInput = parent.backward(self, input, gradOutput)
+   self.gradInput = self.convert_gradInput:forward(gradInput)
+   return self.gradInput
+end
+
+local Float, parent = torch.class('nn.Float', 'nn.Type')
+function Float:__init()
+   parent.__init(self, 'torch.FloatTensor')
+end
+
+local Double, parent = torch.class('nn.Double', 'nn.Type')
+function Double:__init()
+   parent.__init(self, 'torch.DoubleTensor')
+end
+
+local Cuda, parent = torch.class('nn.Cuda', 'nn.Type')
+function Cuda:__init()
+   parent.__init(self, 'torch.CudaTensor')
+end