Merge branch 'master' of https://git.codeplex.com/cntk

MachineLearning/CNTKComputationNetworkLib/ComputationNetwork.cpp

@@ -1102,7 +1102,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
                 pLeft->FunctionValues() = redU;
                 pRight->FunctionValues() = redVT;
 
-                shared_ptr<ComputationNode<ElemType>> pTimes = AddNodeToNetAndAttachInputs(New<TimesNode<ElemType>>(m_deviceId, name + L"-SVD", true /*createOutputMatrix*/), pLeft, pRight);
+                shared_ptr<ComputationNode<ElemType>> pTimes = AddNodeToNetAndAttachInputs(New<TimesNode<ElemType>>(m_deviceId, name + L"-SVD"), pLeft, pRight);
 
                 //========================================
                 // Step 3. remove old node

MachineLearning/CNTKComputationNetworkLib/LinearAlgebraNodes.h

@@ -392,18 +392,10 @@ namespace Microsoft { namespace MSR { namespace CNTK {
         static const std::wstring TypeName() { return L"Times"; }
     public:
 
-        // TODO: The createOutputMatrix parameter here is temporarily added to allow creating the function values
-        // matrix for the times node added during SVD decomposition. Since ValidateSubNetwork is called after addition
-        // of the times node, the validation crashes if the function values matrix has not yet been allocated
-        // This can be removed after the  Validation has been fixed to not access the function values matrix at all
         DeclareConstructorFromConfigWithNumInputs(TimesNode);
-        TimesNode(DEVICEID_TYPE deviceId, const wstring & name, bool createOutputMatrix = false) :
+        TimesNode(DEVICEID_TYPE deviceId, const wstring & name) :
             Base(deviceId, name)
         {
-            if (createOutputMatrix)
-            {
-                CreateMatrixIfNull(m_functionValues);
-            }
         }
 
         virtual void /*ComputationNode::*/ComputeInputPartial(const size_t inputIndex, const FrameRange & frameRange) override

Tests/Image/QuickE2E/baseline.windows.debug.cpu.txt

@@ -893,8 +893,8 @@ already there from last epoch
 Starting minibatch loop.
 randomordering: 21 retries for 100 elements (21.0%) to ensure window condition
 randomordering: recached sequence for seed 11: 6, 31, ...
- Epoch[12 of 12]-Minibatch[   1-  10 of 10]: SamplesSeen = 100; TrainLossPerSample =  0.37213734; EvalErr[0]PerSample = 0.00000000; TotalTime = 0.65604s; TotalTimePerSample = 6.56038ms; SamplesPerSecond = 152
+ Epoch[12 of 12]-Minibatch[   1-  10 of 10]: SamplesSeen = 100; TrainLossPerSample =  0.37077690; EvalErr[0]PerSample = 0.00000000; TotalTime = 0.65604s; TotalTimePerSample = 6.56038ms; SamplesPerSecond = 152
-Finished Epoch[12 of 12]: [Training Set] TrainLossPerSample = 0.37213734; EvalErrPerSample = 0; Ave LearnRatePerSample = 0.004999999888; EpochTime=0.656382
+Finished Epoch[12 of 12]: [Training Set] TrainLossPerSample = 0.37077689; EvalErrPerSample = 0; Ave LearnRatePerSample = 0.004999999888; EpochTime=0.656382
 CNTKCommandTrainEnd: Train
 
 
@@ -2269,8 +2269,8 @@ reading from record 0 to 100 to be positioned properly for epoch
 Starting minibatch loop.
 randomordering: 21 retries for 100 elements (21.0%) to ensure window condition
 randomordering: recached sequence for seed 11: 6, 31, ...
- Epoch[12 of 12]-Minibatch[   1-  10 of 10]: SamplesSeen = 100; TrainLossPerSample =  0.37792297; EvalErr[0]PerSample = 0.00000000; TotalTime = 1.34518s; TotalTimePerSample = 13.45185ms; SamplesPerSecond = 74
+ Epoch[12 of 12]-Minibatch[   1-  10 of 10]: SamplesSeen = 100; TrainLossPerSample =  0.37650299; EvalErr[0]PerSample = 0.00000000; TotalTime = 1.34518s; TotalTimePerSample = 13.45185ms; SamplesPerSecond = 74
-Finished Epoch[12 of 12]: [Training Set] TrainLossPerSample = 0.37792295; EvalErrPerSample = 0; Ave LearnRatePerSample = 0.004999999888; EpochTime=1.371377
+Finished Epoch[12 of 12]: [Training Set] TrainLossPerSample = 0.37650299; EvalErrPerSample = 0; Ave LearnRatePerSample = 0.004999999888; EpochTime=1.371377
 CNTKCommandTrainEnd: Train
 
 

Tests/Image/QuickE2E/baseline.windows.release.cpu.txt

@@ -864,8 +864,8 @@ already there from last epoch
 Starting minibatch loop.
 randomordering: 21 retries for 100 elements (21.0%) to ensure window condition
 randomordering: recached sequence for seed 11: 6, 31, ...
- Epoch[12 of 12]-Minibatch[   1-  10 of 10]: SamplesSeen = 100; TrainLossPerSample =  0.37213734; EvalErr[0]PerSample = 0.00000000; TotalTime = 0.08724s; TotalTimePerSample = 0.87241ms; SamplesPerSecond = 1146
+ Epoch[12 of 12]-Minibatch[   1-  10 of 10]: SamplesSeen = 100; TrainLossPerSample =  0.37077690; EvalErr[0]PerSample = 0.00000000; TotalTime = 0.08724s; TotalTimePerSample = 0.87241ms; SamplesPerSecond = 1146
-Finished Epoch[12 of 12]: [Training Set] TrainLossPerSample = 0.37213734; EvalErrPerSample = 0; Ave LearnRatePerSample = 0.004999999888; EpochTime=0.087336
+Finished Epoch[12 of 12]: [Training Set] TrainLossPerSample = 0.37077689; EvalErrPerSample = 0; Ave LearnRatePerSample = 0.004999999888; EpochTime=0.087336
 CNTKCommandTrainEnd: Train
 
 
@@ -2182,8 +2182,8 @@ reading from record 0 to 100 to be positioned properly for epoch
 Starting minibatch loop.
 randomordering: 21 retries for 100 elements (21.0%) to ensure window condition
 randomordering: recached sequence for seed 11: 6, 31, ...
- Epoch[12 of 12]-Minibatch[   1-  10 of 10]: SamplesSeen = 100; TrainLossPerSample =  0.37792297; EvalErr[0]PerSample = 0.00000000; TotalTime = 0.89367s; TotalTimePerSample = 8.93670ms; SamplesPerSecond = 111
+ Epoch[12 of 12]-Minibatch[   1-  10 of 10]: SamplesSeen = 100; TrainLossPerSample =  0.37650299; EvalErr[0]PerSample = 0.00000000; TotalTime = 0.89367s; TotalTimePerSample = 8.93670ms; SamplesPerSecond = 111
-Finished Epoch[12 of 12]: [Training Set] TrainLossPerSample = 0.37792295; EvalErrPerSample = 0; Ave LearnRatePerSample = 0.004999999888; EpochTime=0.908817
+Finished Epoch[12 of 12]: [Training Set] TrainLossPerSample = 0.37650299; EvalErrPerSample = 0; Ave LearnRatePerSample = 0.004999999888; EpochTime=0.908817
 CNTKCommandTrainEnd: Train
 
 

Tests/Image/QuickE2E/cntk.config

@@ -5,6 +5,7 @@ deviceId=$DeviceId$
 ndlMacros=$ConfigDir$/Macros.ndl
 
 parallelTrain=false
+NumCPUThreads=8
 
 Train=[
     action=train

Tests/Image/QuickE2E/testcases.yml

@@ -1,9 +1,9 @@
 dataDir: ../Data
 tags:
      # running on every BVT job in 'I' (Image) leg:
-     - bvt-i  os=='windows' or device=='gpu'
+     - bvt-i  device=='gpu'
      # running every Nightly job in 'I' leg
-     - nightly-i os=='windows' or device=='gpu'
+     - nightly-i device=='gpu'
 
 testCases:
   CNTK Run must be completed: