Adding support for SpatialFullConvolution.

Since SpatialFullConvolution is just the transpose of the regular
convolution operator, we can use CUDNN by swapping the forward and
backward passes.  This can be a substantial speed improvement over
cunn.SpatialFullConvolution, which works by explicitly building the full
matrix for a GEMM operation.

1 files changed, 95 insertions, 0 deletions

diff --git a/test/test_spatialFullConv.lua b/test/test_spatialFullConv.lua
new file mode 100644
index 0000000..3659864
--- /dev/null
+++ b/test/test_spatialFullConv.lua
@@ -0,0 +1,95 @@
+require 'cudnn'
+require 'cunn'
+
+local cudnntest = {}
+local precision_forward = 1e-4
+local precision_backward = 1e-2
+local precision_jac = 1e-3
+local nloop = 1
+local times = {}
+local mytester
+
+
+local function testSpatialFullConv (imageWidth, imageHeight, nPlanesIn, nPlanesOut, kW, kH, dW, dH, padW, padH, adjW, adjH)
+
+    print ("Running testSpatialFullConv (" ..
+            "imageWidth = " .. imageWidth .. ", " ..
+            "imageHeight = " .. imageHeight .. ", " ..
+            "nPlanesIn = " .. nPlanesIn .. ", " ..
+            "nPlanesOut = " .. nPlanesOut .. ", " ..
+            "kW = " .. kW .. ", " ..
+            "kH = " .. kH .. ", " ..
+            "dW = " .. dW .. ", " ..
+            "dH = " .. dH .. ", " ..
+            "padW = " .. padW .. ", " ..
+            "padH = " .. padH .. ", " ..
+            "adjW = " .. adjW .. ", " ..
+            "adjH = " .. adjH)
+
+    local layerInput = torch.randn(1, nPlanesIn, imageHeight, imageWidth):cuda()
+
+    local modelGT = nn.SpatialFullConvolution (nPlanesIn, nPlanesOut, kW, kH, dW, dH, padW, padH, adjW, adjH)
+    local modelCUDNN = cudnn.SpatialFullConvolution (nPlanesIn, nPlanesOut, kW, kH, dW, dH, padW, padH, adjW, adjH)
+    modelCUDNN.weight:copy (modelGT.weight)
+    modelCUDNN.bias:copy (modelGT.bias)
+
+    modelGT:cuda()
+    modelCUDNN:cuda()
+
+    local outputGT = modelGT:forward (layerInput)
+    local outputCUDNN = modelCUDNN:forward (layerInput)
+
+    local errorOutput = outputCUDNN:float() - outputGT:float()
+    mytester:assertlt(errorOutput:abs():max(), precision_forward, 'error on state (forward) ')
+
+    -- Now check the backwards diffs
+    local crit = nn.MSECriterion()
+    crit:cuda()
+    local target = outputGT:clone()
+    target:zero()
+    target:cuda()
+
+    local f = crit:forward (outputGT, target)
+    local df_do = crit:backward (outputGT, target)
+
+    local gradCUDNN = modelCUDNN:updateGradInput (layerInput, df_do)
+    local gradGT = modelGT:updateGradInput (layerInput, df_do)
+    local errorGradInput = gradCUDNN:float() - gradGT:float()
+    mytester:assertlt(errorGradInput:abs():max(), precision_backward, 'error on grad input (backward) ')
+
+    modelCUDNN:zeroGradParameters()
+    modelCUDNN:accGradParameters (layerInput, df_do, 1.0)
+    modelGT:zeroGradParameters()
+    modelGT:accGradParameters (layerInput, df_do:cuda(), 1.0)
+
+    local errorGradBias = (modelCUDNN.gradBias - modelGT.gradBias)
+    mytester:assertlt(errorGradBias:abs():max(), precision_backward, 'error on grad bias (backward) ')
+
+    local errorGradWeight = (modelCUDNN.gradWeight - modelGT.gradWeight)
+    mytester:assertlt(errorGradWeight:abs():max(), precision_backward, 'error on grad weight (backward) ')
+end
+
+function cudnntest.SpatialConvolution_params()
+    -- Test with a wide variety of different parameter values:
+    testSpatialFullConv (5, 5, 1, 1, 3, 3, 2, 2, 0, 0, 0, 0)
+    testSpatialFullConv (5, 5, 1, 1, 3, 3, 2, 2, 1, 1, 0, 0)
+    testSpatialFullConv (5, 7, 1, 1, 3, 1, 2, 2, 1, 1, 0, 0)
+    testSpatialFullConv (7, 5, 1, 1, 3, 1, 1, 1, 1, 1, 0, 0)
+    testSpatialFullConv (8, 5, 3, 1, 3, 3, 2, 2, 1, 1, 0, 0)
+    testSpatialFullConv (5, 5, 1, 3, 3, 3, 2, 2, 1, 1, 0, 0)
+    testSpatialFullConv (5, 5, 5, 3, 3, 3, 2, 2, 1, 1, 1, 1)
+    testSpatialFullConv (9, 9, 3, 3, 3, 5, 2, 3, 0, 1, 1, 0)
+end
+
+torch.setdefaulttensortype('torch.FloatTensor')
+math.randomseed(os.time())
+mytester = torch.Tester()
+mytester:add(cudnntest)
+
+for i=1,cutorch.getDeviceCount() do
+   print('Running test on device: ' .. i)
+   cutorch.setDevice(i)
+   mytester:run(tests)
+end
+
+os.execute('rm -f modelTemp.t7')