NarrowLookupTable:accGradParameters works (unit tested)

1 files changed, 33 insertions, 9 deletions

diff --git a/test/test-all.lua b/test/test-all.lua
index a2832d4..28254a7 100644
--- a/test/test-all.lua
+++ b/test/test-all.lua
@@ -516,23 +516,34 @@ function nnxtest.NarrowLookupTable()
    local embedSize = 32
    local deltaSize = 4
    
+   -- 1D input
    local input = torch.randperm(dictSize):narrow(1,1,nIndex)
    local nlt = nn.NarrowLookupTable(deltaSize, dictSize, embedSize)
    local output = nlt:forward(input)
    
-   local narrowSize = embedSize
    local output2 = torch.Tensor(120):zero()
+   local narrowSize = embedSize
    local idx = 1
    for i=1,nIndex do
       output2:narrow(1, idx, narrowSize):copy(nlt.weight[input[i]]:narrow(1,1,narrowSize))
-      if i == nIndex then
-         break
-      end
       idx = idx + narrowSize
       narrowSize = narrowSize - deltaSize
    end
    mytester:assertTensorEq(output, output2, 0.000001, "1D forward error")
    
+   nlt:zeroGradParameters()
+   local gradWeight2 = nlt.gradWeight:clone()
+   nlt:backward(input, output)
+   local idx = 1
+   local narrowSize = embedSize
+   for i=1,nIndex do
+      gradWeight2[input[i]]:narrow(1, 1, narrowSize):copy(output:narrow(1,idx,narrowSize))
+      idx = idx + narrowSize
+      narrowSize = narrowSize - deltaSize
+   end
+   mytester:assertTensorEq(nlt.gradWeight, gradWeight2, 0.000001, "1D backward error")
+   
+   -- 2D input
    local input = torch.randperm(dictSize):narrow(1,1,nIndex*batchSize):view(8,-1)
    local output = nlt:forward(input)
    local output2 = torch.Tensor(batchSize, 120):zero()
@@ -542,15 +553,28 @@ function nnxtest.NarrowLookupTable()
       local narrowSize = embedSize
       local idx = 1
       for i=1,nIndex do
-         output2:narrow(1, idx, narrowSize):copy(nlt.weight[input[i]]:narrow(1,1,narrowSize))
-         if i == nIndex then
-            break
-         end
+         output2:narrow(1, idx, narrowSize):add(nlt.weight[input[i]]:narrow(1,1,narrowSize))
          idx = idx + narrowSize
          narrowSize = narrowSize - deltaSize
       end
    end
-   mytester:assertTensorEq(output, output2, 0.000001, "1D forward error")
+   mytester:assertTensorEq(output, output2, 0.000001, "2D forward error")
+   
+   nlt:zeroGradParameters()
+   local gradWeight2 = nlt.gradWeight:clone()
+   nlt:backward(input, output)
+   for k=1,batchSize do
+      local input = input[k]
+      local output = output[k]
+      local idx = 1
+      local narrowSize = embedSize
+      for i=1,nIndex do
+         gradWeight2[input[i]]:narrow(1,1,narrowSize):add(output:narrow(1,idx,narrowSize))
+         idx = idx + narrowSize
+         narrowSize = narrowSize - deltaSize
+      end
+   end
+   mytester:assertTensorEq(nlt.gradWeight, gradWeight2, 0.000001, "2D backward error")
 end
 
 function nnx.test(tests)