Minor changes to configs in Demo/Speech/ based on Dong's comments

glebkuznetsov · Dec 10, 2015 · 048c91b · 048c91b
1 parent 780b9ee
commit 048c91b
Show file tree

Hide file tree

Showing 6 changed files with 10 additions and 15 deletions.
diff --git a/Demos/Simple2d/README.md b/Demos/Simple2d/README.md
@@ -2,12 +2,13 @@
 
 ## Overview
 
+| | |
 |:--------|:---|
-Data:     |Two dimensional synthetic data
-Purpose:  |Showcase how to train a simple CNTK network (CPU and GPU) and how to use it for scoring (decoding)
-Network:  |SimpleNetworkBuilder, 2 hidden layers with 50 sigmoid nodes each, cross entropy with softmax
-Training: |Stochastic gradient descent with momentum
-Comments: |There are two config files: Simple.config uses a single CPU or GPU, Multigpu.config uses data-parallel SGD for training on multiple GPUs
+Data     |Two dimensional synthetic data
+Purpose  |Showcase how to train a simple CNTK network (CPU and GPU) and how to use it for scoring (decoding)
+Network  |SimpleNetworkBuilder, 2 hidden layers with 50 sigmoid nodes each, cross entropy with softmax
+Training |Stochastic gradient descent with momentum
+Comments |There are two config files: Simple.config uses a single CPU or GPU, Multigpu.config uses data-parallel SGD for training on multiple GPUs
 
 ## Running the example
 

diff --git a/Demos/Speech/Config/FeedForward.config b/Demos/Speech/Config/FeedForward.config
@@ -31,9 +31,7 @@ speechTrain = [
         trainingCriterion = "CrossEntropyWithSoftmax"
         evalCriterion = "ErrorPrediction"
         layerTypes = "Sigmoid"
-        initValueScale = 1.0
         applyMeanVarNorm = true
-        uniformInit = true
         needPrior = true
     ]
 
@@ -43,7 +41,6 @@ speechTrain = [
         learningRatesPerMB = 1.0:0.5:0.1
         numMBsToShowResult = 10
         momentumPerMB = 0.9:0.656119
-        dropoutRate = 0.0
         maxEpochs = 3
         keepCheckPointFiles = true
 
@@ -62,13 +59,7 @@ speechTrain = [
             autoAdjustMinibatch = true
             minibatchSizeTuningFrequency = 1
             minibatchSearchCriterionErrorMargin = 2
-            autoAdjustLR = "none"
-            numMiniBatch4LRSearch = 1000
-            numPrevLearnRate = 5
-            numBestSearchEpoch = 1
         ]
-
-        clippingThresholdPerSample = 1#INF
     ]
 
     reader = [

diff --git a/Demos/Speech/Config/LSTM-NDL.config b/Demos/Speech/Config/LSTM-NDL.config
@@ -28,6 +28,7 @@ truncated = true
 
 speechTrain = [
     action = "train"
+    nbrUttsIneachRecurrentIter = 16
 
     NDLNetworkBuilder = [
         networkDescription = "$ConfigDir$/lstmp-3layer-opt.ndl"
@@ -47,7 +48,6 @@ speechTrain = [
         readerType = "HTKMLFReader"
         readMethod = "blockRandomize"
         miniBatchMode = "partial"
-        nbruttsineachrecurrentiter = 16
         randomize = "auto"
         verbosity = 0
 

diff --git a/Demos/Speech/README.md b/Demos/Speech/README.md
@@ -9,6 +9,7 @@ This modified version of dataset is distributed under the terms of a AN4 license
 
 ## Overview
 
+| | |
 |:--------|:---|
 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

diff --git a/Demos/Text/README.md b/Demos/Text/README.md
@@ -6,6 +6,7 @@ Note: The data is not checked into the repository currently since a license is r
 
 ## Overview
 
+| | |
 |:--------|:---|
 Data:     |The Penn Treebank Project (https://www.cis.upenn.edu/~treebank/) annotates naturally-occuring text for linguistic structure .
 Purpose:  |Showcase how to train a recurrent network for text data.

diff --git a/ExampleSetups/Image/MNIST/README.md b/ExampleSetups/Image/MNIST/README.md
@@ -2,6 +2,7 @@
 
 ## Overview
 
+| | |
 |:--------|:---|
 Data:     |The MNIST database (http://yann.lecun.com/exdb/mnist/) of handwritten digits.
 Purpose:  |This example demonstrates usage of NDL to train neural networks on MNIST dataset.