Minor fixes in demos (MB rescaling, tabs, image link, Speech/LSTM)

Demos/Simple2d/Config/Multigpu.config

@@ -46,20 +46,20 @@ Multigpu_Demo_Train=[
     ]
 
     SGD = [
-	    # epochSize = 0 means epochSize is the size of the training set 
+        # epochSize = 0 means epochSize is the size of the training set 
         epochSize = 0 
         minibatchSize = 25	
         learningRatesPerMB = 0.5:0.2*20:0.1
         momentumPerMB = 0.9
         dropoutRate = 0.0
         maxEpochs = 10
-		
-		# Additional optional parameters are: distributedMBReading
+
+        # Additional optional parameters are: distributedMBReading
         parallelTrain = [
             parallelizationMethod = "DataParallelSGD"
             parallelizationStartEpoch = 2
 
-		    # Additional optional parameters are: useZeroThresholdFor1BitQuantization
+            # Additional optional parameters are: useZeroThresholdFor1BitQuantization
             dataParallelSGD = [
                 gradientBits = 1
             ]
@@ -103,12 +103,12 @@ Multigpu_Demo_Test=[
         readerType = "UCIFastReader"
         file = "$DataDir$/SimpleDataTest.txt"
         randomize = "none"
-	  
+
         features = [
             dim = 2
             start = 0
         ]
-	  
+
         labels = [
             start = 2
             dim = 1
@@ -132,19 +132,19 @@ Multigpu_Demo_Output=[
         readerType = "UCIFastReader"
         file = "$DataDir$/SimpleDataTest.txt"
         randomize = "none"
-	  
+
         features = [
             dim = 2
             start = 0
         ]
-	  
+
         labels = [
             start = 2
             dim = 1
             labelDim = 2
             labelMappingFile = "$DataDir$/SimpleMapping.txt"
         ]
     ]
-	
+
     outputPath = "$OutputDir$/MultigpuOutput"    # Dump output as text
 ]

Demos/Simple2d/README.md

@@ -68,13 +68,12 @@ training and test data as well as the plots that are provided in the folder.
 The data is synthetic 2d data representing two classes that are separated by a sinusoidal boundary. 
 SimpleDemoDataReference.png shows a plot of the training data.
 
-![training data plot](AdditionalFiles/SimpleDemoDataReference.png)
-
 ## Using a trained model
 
 The Test (e.g. Simple_Demo_Test) and the Output (e.g. Simple_Demo_Output) commands 
 specified in the config files use the trained model to compute labels for data specified in the SimpleDataTest.txt file. 
 The Test command computes prediction error, cross entropy and perplexity for the test set and outputs them to the console. 
 The Output command writes for each test instance the likelihood per label to a file `outputPath = $OutputDir$/SimpleOutput`. 
+To use the Output command either set `command=Simple_Demo_Output` in the config file or add it to the command line.
 The model that is used to compute the labels in these commands is defined 
 in the modelPath variable at the beginning of the file `modelPath=$modelDir$/simple.dnn`.

Demos/Speech/Config/FeedForward.config

@@ -58,13 +58,14 @@ speechTrain = [
             ]
         ]
 
-        autoAdjust = [
-            reduceLearnRateIfImproveLessThan = 0
-            loadBestModel = true
-            increaseLearnRateIfImproveMoreThan = 1000000000
-            learnRateDecreaseFactor = 0.5
-            learnRateIncreaseFactor = 1.382
-            autoAdjustLR = "adjustAfterEpoch"
+        autoAdjust=[
+            autoAdjustMinibatch = true
+            minibatchSizeTuningFrequency = 1
+            minibatchSearchCriterionErrorMargin = 2
+            autoAdjustLR = "none"
+            numMiniBatch4LRSearch = 1000
+            numPrevLearnRate = 5
+            numBestSearchEpoch = 1
         ]
 
         clippingThresholdPerSample = 1#INF

Demos/Speech/Config/LSTM-NDL.config

@@ -34,8 +34,8 @@ speechTrain = [
     ]
 
     SGD = [
-        epochSize = 20480
-        minibatchSize = 2048
+        epochSize = 0
+        minibatchSize = 16
         learningRatesPerMB = 0.5
         numMBsToShowResult = 10
         momentumPerMB = 0:0.9
@@ -47,7 +47,7 @@ speechTrain = [
         readerType = "HTKMLFReader"
         readMethod = "blockRandomize"
         miniBatchMode = "partial"
-        nbruttsineachrecurrentiter = 32
+        nbruttsineachrecurrentiter = 16
         randomize = "auto"
         verbosity = 0
 

Demos/Speech/README.md

@@ -13,7 +13,7 @@ This modified version of dataset is distributed under the terms of a AN4 license
 Data:     |Speech data from the CMU Audio Database aka AN4 (http://www.speech.cs.cmu.edu/databases/an4)
 Purpose:  |Showcase how to train feed forward and LSTM networks for speech data
 Network:  |SimpleNetworkBuilder for 2-layer FF, NdlNetworkBuilder for 3-layer LSTM network
-Training: |Data-parallel 1-Bit SGD with adjusted learning rate
+Training: |Data-parallel 1-Bit SGD with automatic mini batch rescaling (FF)
 Comments: |There are two config files: FeedForward.config and LSTM-NDL.config for FF and LSTM training respectively
 
 ## Running the example