The analytic gradient checks out.

[dagnn.git] / test-dagnn.lua

diff --git a/test-dagnn.lua b/test-dagnn.lua
index d7179cc..cac5a94 100755 (executable)
--- a/test-dagnn.lua
+++ b/test-dagnn.lua
@@ -5,6 +5,44 @@ require 'nn'
 
 require 'dagnn'
 
+function checkGrad(model, criterion, input, target)
+   local params, gradParams = model:getParameters()
+
+   local epsilon = 1e-5
+
+   local output = model:forward(input)
+   local loss = criterion:forward(output, target)
+   local gradOutput = criterion:backward(output, target)
+   gradParams:zero()
+   model:backward(input, gradOutput)
+   local analyticalGradParam = gradParams:clone()
+
+   for i = 1, params:size(1) do
+      local x = params[i]
+
+      params[i] = x - epsilon
+      local output0 = model:forward(input)
+      local loss0 = criterion:forward(output0, target)
+
+      params[i] = x + epsilon
+      local output1 = model:forward(input)
+      local loss1 = criterion:forward(output1, target)
+
+      params[i] = x
+
+      local ana = analyticalGradParam[i]
+      local num = (loss1 - loss0) / (2 * epsilon)
+      local err = torch.abs(num - ana) / torch.abs(num)
+
+      print(
+         err .. ' checkGrad ' .. i
+            .. ' analytical ' .. ana
+            .. ' numerical ' .. num
+      )
+   end
+
+end
+
 function printTensorTable(t)
    if torch.type(t) == 'table' then
       for i, t in pairs(t) do
@@ -20,44 +58,60 @@ end
 torch.setdefaulttensortype('torch.DoubleTensor')
 torch.manualSeed(2)
 
+--                     +--> c ----> e --+
+--                    /            /     \
+--                   /            /       \
+--  input --> a --> b ---> d ----+         g --> output
+--                          \             /
+--                           \           /
+--                            +--> f ---+
+
 a = nn.Linear(10, 10)
 b = nn.ReLU()
 c = nn.Linear(10, 3)
 d = nn.Linear(10, 3)
 e = nn.CMulTable()
-f = nn.Linear(3, 2)
+f = nn.Linear(3, 3)
+g = nn.CAddTable()
 
---   a -----> b ---> c ----> e ---
---             \           /
---              \--> d ---/
---                    \
---                     \---> f ---
-
-g = nn.DAG()
+----------------------------------------------------------------------
 
-g:addEdge(c, e)
-g:addEdge(a, b)
-g:addEdge(d, e)
-g:addEdge(b, c)
-g:addEdge(b, d)
-g:addEdge(d, f)
+model = nn.DAG()
 
-g:setInput({{a}})
-g:setOutput({ e, f })
+model:addEdge(a, b)
+model:addEdge(b, c)
+model:addEdge(b, d)
+model:addEdge(c, e)
+model:addEdge(d, e)
+model:addEdge(d, f)
+model:addEdge(e, g)
+model:addEdge(f, g)
 
-g:print()
+model:setInput(a)
+model:setOutput(g)
 
 input = torch.Tensor(3, 10):uniform()
 
-output = g:updateOutput({{ input }})
+print('******************************************************************')
+print('** updateOutput **************************************************')
+print('******************************************************************')
 
-printTensorTable(output)
+output = model:updateOutput(input):clone()
 
-----------------------------------------------------------------------
+printTensorTable(output)
 
 print('******************************************************************')
 print('** updateGradInput ***********************************************')
 print('******************************************************************')
-gradInput = g:updateGradInput({{input}}, output)
+
+gradInput = model:updateGradInput(input, output)
 
 printTensorTable(gradInput)
+
+print('******************************************************************')
+print('** checkGrad *****************************************************')
+print('******************************************************************')
+
+output:uniform()
+
+checkGrad(model, nn.MSECriterion(), input, output)