Merged with master

zpbappi · May 24, 2016 · 2a333f3 · 2a333f3
2 parents 7376f25 + 6cef7b0
commit 2a333f3
Show file tree

Hide file tree

Showing 251 changed files with 68,656 additions and 30,949 deletions.
diff --git a/.gitattributes b/.gitattributes
@@ -40,6 +40,10 @@ run-test-common text eol=lf
 run-timit-test-common text eol=lf
 make_binary_drop_linux text eol=lf
 
+# Used from Unix / Cygwin 'md5sum -c', needs to have LF line endings:
+Tests/EndToEndTests/Examples/Speech/TIMIT/WriteBottleneck/expected_output_md5sum.*.txt eol=lf
+Tests/EndToEndTests/Examples/Speech/TIMIT/WriteScaledLogLike/expected_output_md5sum.*.txt eol=lf
+
 Makefile text
 *.sln text
 *.vcxproj text

diff --git a/.gitignore b/.gitignore
@@ -152,7 +152,9 @@ ModelManifest.xml
 
 # Python
 *.pyc
-__pychache__/
+__pycache__/
+contrib/Python/doc/_build/*
+contrib/Python/_cntk_default/*
 
 # =========================
 # Windows detritus

diff --git a/CNTK.sln b/CNTK.sln
diff --git a/Examples/Image/MNIST/Config/02_Convolution.cntk b/Examples/Image/MNIST/Config/02_Convolution.cntk
@@ -26,7 +26,8 @@ ndlMacros = "$ConfigDir$/Macros.ndl"
 traceLevel=1
 numMBsToShowResult=500
 
-prefetch=true
+# Note: turn off prefetching; known to crash UCIFastReader occasionally.
+prefetch=false
 
 # If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
 initOnCPUOnly=true

diff --git a/Examples/Image/Miscellaneous/CIFAR-10/01_Conv.cntk b/Examples/Image/Miscellaneous/CIFAR-10/01_Conv.cntk
@@ -16,7 +16,8 @@ imageLayout = "cudnn"
 # If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
 initOnCPUOnly=true
 
-prefetch = "true"
+# Note: turn off prefetching; known to crash UCIFastReader occasionally.
+prefetch = "false"
 
 command = Train:Test
 

diff --git a/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv.cntk b/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv.cntk
@@ -16,7 +16,8 @@ imageLayout = "cudnn"
 # If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
 initOnCPUOnly=true
 
-prefetch = "true"
+# Note: turn off prefetching; known to crash UCIFastReader occasionally.
+prefetch = "false"
 
 command = Train:Test
 

diff --git a/Examples/Image/Miscellaneous/CIFAR-10/05_ConvLocal.cntk b/Examples/Image/Miscellaneous/CIFAR-10/05_ConvLocal.cntk
@@ -13,7 +13,8 @@ imageLayout = "cudnn"
 # override the above as follows when running on CPU:
 # deviceId = -1
 
-prefetch = "true"
+# Note: turn off prefetching; known to crash UCIFastReader occasionally.
+prefetch = "false"
 
 command = Train:Test
 

diff --git a/Examples/Image/Miscellaneous/ImageNet/ResNet/README.md b/Examples/Image/Miscellaneous/ImageNet/ResNet/README.md
@@ -1,12 +1,22 @@
 # CNTK example: ImageNet ResNet 
 
-**Disclaimer: network configurations and experiment settings in this this folder try to follow those published in the [ResNet paper](http://arxiv.org/abs/1512.03385) as close as possible. However, these samples are NOT endorsed or verified by the researchers who published the original work. It is NOT guaranteed that you get the same (or even close) results as those in the paper.**
-
 ## Overview
 
 |Data:     |The ILSVRC2012 dataset (http://www.image-net.org/challenges/LSVRC/2012/) of images.
 |:---------|:---
 |Purpose   |This example demonstrates usage of the NDL (Network Description Language) to define networks similar to ResNet.
-|Network   |NDLNetworkBuilder, deep convolutional networks resembling ResNet networks.
+|Network   |NDLNetworkBuilder, deep convolutional residual networks (ResNet).
 |Training  |Stochastic gradient descent with momentum.
 
+## Details
+The network configurations and experiment settings in this this folder resemble the ones in the original [ResNet paper](http://arxiv.org/abs/1512.03385) with few minor changes inspired by [this work](https://github.com/facebook/fb.resnet.torch).
+The following table contains results as well as links to pre-trained models that can be used in various applications.
+
+| Network       | Top-1 error | Top-5 error | Model
+| ------------- | ----------- | ----------- | ----------
+| ResNet-18     | 29.57       | 10.41       | [Download](https://www.cntk.ai/resnet/ResNet_18.model)
+| ResNet-34     | 27.31       | 8.97        | [Download](https://www.cntk.ai/resnet/ResNet_34.model)
+| ResNet-50     | 24.74       | 7.56        | [Download](https://www.cntk.ai/resnet/ResNet_50.model)
+
+## Notes
+This work is an implementation of ResNets in CNTK. If you are interested in the original implementation of ResNet, follow [this link](https://github.com/KaimingHe/deep-residual-networks).