Merge changes from github.

Change: 124644444

ISSUE_TEMPLATE.md

@@ -16,7 +16,7 @@ Installed version of CUDA and cuDNN:
 If installed from binary pip package, provide:
 
 1. Which pip package you installed.
-2. The output from python -c "import tensorflow; print(tensorflow.__version__)".
+2. The output from `python -c "import tensorflow; print(tensorflow.__version__)"`.
 
 If installed from sources, provide the commit hash:
 

grpc.BUILD

@@ -145,6 +145,12 @@ cc_library(
     "include",
     ".",
   ],
+  defines = [
+    "GPR_BACKWARDS_COMPATIBILITY_MODE",
+  ],
+  copts = [
+    "-std=c99",
+  ],
   deps = [
   ],
 )

tensorflow/contrib/ios_examples/camera/camera_example.xcodeproj/project.pbxproj

@@ -285,7 +285,7 @@
 				GCC_WARN_UNUSED_VARIABLE = YES;
 				HEADER_SEARCH_PATHS = (
 					"$(SRCROOT)/../../makefile/gen/proto",
-					"$(SRCROOT)/../../makefile/downloads/eigen-eigen-f3a13643ac1f",
+					"$(SRCROOT)/../../makefile/downloads/eigen-eigen-d02e6a705c30",
 					"$(SRCROOT)/../../makefile/downloads",
 					"$(SRCROOT)/../../makefile/downloads/protobuf/src/",
 					"$(SRCROOT)/../../../..",
@@ -300,6 +300,12 @@
 				OTHER_LDFLAGS = (
 					"-force_load",
 					"$(SRCROOT)/../../makefile/gen/lib/libtensorflow-core.a",
+					"-Xlinker",
+					"-S",
+					"-Xlinker",
+					"-x",
+					"-Xlinker",
+					"-dead_strip",
 				);
 				PRODUCT_BUNDLE_IDENTIFIER = com.google.CameraExample;
 				PRODUCT_NAME = "$(TARGET_NAME)";
@@ -344,7 +350,7 @@
 				GCC_WARN_UNUSED_VARIABLE = YES;
 				HEADER_SEARCH_PATHS = (
 					"$(SRCROOT)/../../makefile/gen/proto",
-					"$(SRCROOT)/../../makefile/downloads/eigen-eigen-f3a13643ac1f",
+					"$(SRCROOT)/../../makefile/downloads/eigen-eigen-d02e6a705c30",
 					"$(SRCROOT)/../../makefile/downloads",
 					"$(SRCROOT)/../../makefile/downloads/protobuf/src/",
 					"$(SRCROOT)/../../../..",
@@ -359,6 +365,12 @@
 				OTHER_LDFLAGS = (
 					"-force_load",
 					"$(SRCROOT)/../../makefile/gen/lib/libtensorflow-core.a",
+					"-Xlinker",
+					"-S",
+					"-Xlinker",
+					"-x",
+					"-Xlinker",
+					"-dead_strip",
 				);
 				PRODUCT_BUNDLE_IDENTIFIER = com.google.CameraExample;
 				PRODUCT_NAME = "$(TARGET_NAME)";

tensorflow/contrib/ios_examples/simple/tf_ios_makefile_example.xcodeproj/project.pbxproj

@@ -276,7 +276,7 @@
 					"$(SRCROOT)/../../../..",
 					"$(SRCROOT)/../../makefile/downloads/protobuf/src/",
 					"$(SRCROOT)/../../makefile/downloads",
-					"$(SRCROOT)/../../makefile/downloads/eigen-eigen-f3a13643ac1f",
+					"$(SRCROOT)/../../makefile/downloads/eigen-eigen-d02e6a705c30",
 					"$(SRCROOT)/../../makefile/gen/proto",
 				);
 				INFOPLIST_FILE = "$(SRCROOT)/RunModel-Info.plist";
@@ -289,6 +289,12 @@
 				OTHER_LDFLAGS = (
 					"-force_load",
 					"$(SRCROOT)/../../makefile/gen/lib/libtensorflow-core.a",
+					"-Xlinker",
+					"-S",
+					"-Xlinker",
+					"-x",
+					"-Xlinker",
+					"-dead_strip",
 				);
 				PRODUCT_BUNDLE_IDENTIFIER = "com.google.TF-Test";
 				PRODUCT_NAME = "$(TARGET_NAME)";
@@ -304,7 +310,7 @@
 					"$(SRCROOT)/../../../..",
 					"$(SRCROOT)/../../makefile/downloads/protobuf/src/",
 					"$(SRCROOT)/../../makefile/downloads",
-					"$(SRCROOT)/../../makefile/downloads/eigen-eigen-f3a13643ac1f",
+					"$(SRCROOT)/../../makefile/downloads/eigen-eigen-d02e6a705c30",
 					"$(SRCROOT)/../../makefile/gen/proto",
 				);
 				INFOPLIST_FILE = "$(SRCROOT)/RunModel-Info.plist";
@@ -314,10 +320,16 @@
 					"$(SRCROOT)/../../makefile/gen/protobuf_ios/lib",
 					"$(SRCROOT)/../../makefile/gen/lib",
 				);
-				ONLY_ACTIVE_ARCH = NO;
+				ONLY_ACTIVE_ARCH = YES;
 				OTHER_LDFLAGS = (
 					"-force_load",
 					"$(SRCROOT)/../../makefile/gen/lib/libtensorflow-core.a",
+					"-Xlinker",
+					"-S",
+					"-Xlinker",
+					"-x",
+					"-Xlinker",
+					"-dead_strip",
 				);
 				PRODUCT_BUNDLE_IDENTIFIER = "com.google.TF-Test";
 				PRODUCT_NAME = "$(TARGET_NAME)";

tensorflow/contrib/learn/python/learn/io/data_feeder.py

@@ -358,7 +358,7 @@ def _feed_dict_fn():
         else:
           if self.n_classes > 1:
             if len(self.output_shape) == 2:
-              out.itemset((i, self.y[sample]), 1.0)
+              out.itemset((i, int(self.y[sample])), 1.0)
             else:
               for idx, value in enumerate(self.y[sample]):
                 out.itemset(tuple([i, idx, value]), 1.0)

tensorflow/contrib/learn/python/learn/ops/batch_norm_ops.py

@@ -54,22 +54,31 @@ def batch_normalize(tensor_in,
         initializer=init_ops.random_normal_initializer(1., 0.02))
     beta = vs.get_variable("beta", [shape[-1]],
                            initializer=init_ops.constant_initializer(0.))
-    ema = moving_averages.ExponentialMovingAverage(decay=decay)
-    if convnet:
-      assign_mean, assign_var = nn.moments(tensor_in, [0, 1, 2])
-    else:
-      assign_mean, assign_var = nn.moments(tensor_in, [0])
-    ema_assign_op = ema.apply([assign_mean, assign_var])
-    ema_mean, ema_var = ema.average(assign_mean), ema.average(assign_var)
+    moving_mean = vs.get_variable(
+        'moving_mean',
+        shape=[shape[-1]],
+        initializer=init_ops.zeros_initializer,
+        trainable=False)
+    moving_var = vs.get_variable(
+        'moving_var',
+        shape=[shape[-1]],
+        initializer=init_ops.ones_initializer,
+        trainable=False)
 
     def _update_mean_var():
       """Internal function that updates mean and variance during training."""
-      with ops.control_dependencies([ema_assign_op]):
-        return array_ops_.identity(assign_mean), array_ops_.identity(assign_var)
+      axis = [0, 1, 2] if convnet else [0]
+      mean, var = nn.moments(tensor_in, axis)
+      update_moving_mean = moving_averages.assign_moving_average(
+          moving_mean, mean, decay)
+      update_moving_var = moving_averages.assign_moving_average(
+          moving_var, var, decay)
+      with ops.control_dependencies([update_moving_mean, update_moving_var]):
+        return array_ops_.identity(mean), array_ops_.identity(var)
 
     is_training = array_ops_.squeeze(ops.get_collection("IS_TRAINING"))
     mean, variance = control_flow_ops.cond(is_training, _update_mean_var,
-                                           lambda: (ema_mean, ema_var))
+                                           lambda: (moving_mean, moving_var))
     return nn.batch_norm_with_global_normalization(
         tensor_in,
         mean,

tensorflow/contrib/makefile/Makefile

@@ -35,8 +35,8 @@ HOST_OBJDIR := $(MAKEFILE_DIR)/gen/host_obj/
 HOST_BINDIR := $(MAKEFILE_DIR)/gen/host_bin/
 HOST_GENDIR := $(MAKEFILE_DIR)/gen/host_obj/
 
-# Which Eigen version we're using.
-EIGEN_HASH := d02e6a705c30
+# Find the current Eigen version name from the Bazel build file
+EIGEN_HASH := $(shell cat eigen.BUILD | grep archive_dir | head -1 | cut -f3 -d- | cut -f1 -d\")
 
 # Settings for the host compiler.
 HOST_CXX := gcc
@@ -56,9 +56,15 @@ HOST_LIBS := \
 
 # If we're on Linux, also link in the dl library.
 ifeq ($(HOST_OS),LINUX)
-	HOST_LIBS += -ldl
+	HOST_LIBS += -ldl -lpthread
 endif
 
+# If we're on a Pi, link in pthreads and dl
+ifeq ($(HOST_OS),PI)
+	HOST_LIBS += -ldl -lpthread
+endif
+
+
 # proto_text is a tool that converts protobufs into a form we can use more
 # compactly within TensorFlow. It's a bit like protoc, but is designed to
 # produce a much more minimal result so we can save binary space.
@@ -125,13 +131,13 @@ ifeq ($(TARGET),LINUX)
 endif
 # If we're on Linux, also link in the dl library.
 ifeq ($(TARGET),LINUX)
-	LIBS += -ldl
+	LIBS += -ldl -lpthread
 endif
 # If we're cross-compiling for the Raspberry Pi, use the right gcc.
 ifeq ($(TARGET),PI)
-	CXX := arm-linux-gnueabihf-g++
-	LDFLAGS := -L$(GENDIR)protobuf_pi/lib/ -Wl,--no-whole-archive
-	LIBS += -ldl
+	CXXFLAGS += -D__ANDROID_TYPES_SLIM__
+	LDFLAGS := -Wl,--no-whole-archive
+	LIBS += -ldl -lpthread
 	LIBFLAGS += -Wl,--allow-multiple-definition -Wl,--whole-archive
 endif
 
@@ -169,12 +175,16 @@ ifeq ($(TARGET),IOS)
 		-Wno-c++11-narrowing \
 		-mno-thumb \
 		-DTF_LEAN_BINARY \
+		-D__ANDROID_TYPES_SLIM__ \
 		-DMIN_LOG_LEVEL=0 \
 		-fno-exceptions \
 		-isysroot \
 		${IPHONEOS_SYSROOT}
 		LDFLAGS := -arch armv7 \
 		-miphoneos-version-min=${MIN_SDK_VERSION} \
+		-Xlinker -S \
+		-Xlinker -x \
+		-Xlinker -dead_strip \
 		-all_load \
 		-L$(GENDIR)protobuf_ios/lib \
 		-lz
@@ -186,6 +196,7 @@ ifeq ($(TARGET),IOS)
 		-Wno-c++11-narrowing \
 		-mno-thumb \
 		-DTF_LEAN_BINARY \
+		-D__ANDROID_TYPES_SLIM__ \
 		-DMIN_LOG_LEVEL=0 \
 		-fno-exceptions \
 		-isysroot \
@@ -205,6 +216,7 @@ ifeq ($(TARGET),IOS)
 		-D__thread= \
 		-Wno-c++11-narrowing \
 		-DTF_LEAN_BINARY \
+		-D__ANDROID_TYPES_SLIM__ \
 		-DMIN_LOG_LEVEL=0 \
 		-fno-exceptions \
 		-isysroot \
@@ -224,6 +236,7 @@ ifeq ($(TARGET),IOS)
 		-D__thread= \
 		-Wno-c++11-narrowing \
 		-DTF_LEAN_BINARY \
+		-D__ANDROID_TYPES_SLIM__ \
 		-DMIN_LOG_LEVEL=0 \
 		-fno-exceptions \
 		-isysroot \
@@ -243,6 +256,7 @@ ifeq ($(TARGET),IOS)
 		-D__thread= \
 		-Wno-c++11-narrowing \
 		-DTF_LEAN_BINARY \
+		-D__ANDROID_TYPES_SLIM__ \
 		-DMIN_LOG_LEVEL=0 \
 		-fno-exceptions \
 		-isysroot \

tensorflow/contrib/makefile/README.md

@@ -42,7 +42,7 @@ at `tensorflow/contrib/makefile/gen/bin/benchmark`. To run the executable, use:
 tensorflow/contrib/makefile/gen/bin/benchmark --graph=tensorflow_inception_graph.pb
 ```
 
-You should download the example graph from [http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz](http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz).
+You should download the example graph from [https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip](https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip).
 
 ## Supported Systems
 
@@ -132,24 +132,31 @@ static library in a simple app.
 
 ## Raspberry Pi
 
-The easiest way to build for the Raspberry Pi is to cross-compile from Linux.
-To use this makefile to do that, you first need to install the right version of
-the compiler to target the Pi, using a command like this on your Linux machine:
+Building on the Raspberry Pi is similar to a normal Linux system, though we
+recommend starting by compiling and installing protobuf:
 
 ```bash
-sudo apt-get install g++-arm-linux-gnueabihf
+cd tensorflow/contrib/makefile/downloads/protobuf/
+./autogen.sh 
+./configure
+make
+sudo make install
+cd ../../../../..
 ```
 
-After that, run `tensorflow/contrib/makefile/compile_pi_protobuf.sh` to build a
-version of the protobuf library aimed at the Pi. Then you should be able to run:
+Once that's done, you can use make to build the library and example:
 
 ```bash
-make -f tensorflow/contrib/makefile/Makefile TARGET=PI
+make -f tensorflow/contrib/makefile/Makefile HOST_OS=PI TARGET=PI OPTFLAGS="-Os"
 ```
 
-This will build the static library, and the example benchmark executable. You
-can then copy the `tensorflow/contrib/makefile/gen/bin/benchmark` program over
-to your Raspberry Pi, and run it there.
+If you're only interested in building for Raspberry Pi's 2 and 3, you can supply
+some extra optimization flags to give you code that will run faster:
+
+```bash
+make -f tensorflow/contrib/makefile/Makefile HOST_OS=PI TARGET=PI \
+OPTFLAGS="-Os -mfpu=neon-vfpv4 -funsafe-math-optimizations -ftree-vectorize"
+```
 
 ## Dependencies
 

tensorflow/contrib/makefile/download_dependencies.sh

@@ -18,7 +18,12 @@ DOWNLOADS_DIR=tensorflow/contrib/makefile/downloads
 
 mkdir ${DOWNLOADS_DIR}
 
-EIGEN_HASH=d02e6a705c30
+EIGEN_HASH=62a2305d5734
+if [ -f eigen.BUILD ]; then
+	# Grab the current Eigen version name from the Bazel build file
+	EIGEN_HASH=$(cat eigen.BUILD | grep archive_dir | head -1 | cut -f3 -d- | cut -f1 -d\")
+fi
+
 curl "https://bitbucket.org/eigen/eigen/get/${EIGEN_HASH}.tar.gz" \
 -o /tmp/eigen-${EIGEN_HASH}.tar.gz
 tar xzf /tmp/eigen-${EIGEN_HASH}.tar.gz -C ${DOWNLOADS_DIR}