Sync Clarke fork

.github/workflows/ci.yml

@@ -22,7 +22,7 @@ jobs:
   linux-x86_64:
     if: github.event_name == 'push' || contains(github.event.pull_request.labels.*.name, 'CI build')
     runs-on: ubuntu-latest
-    container: nvidia/cuda:10.1-cudnn7-devel-centos7
+    container: centos:7
     strategy:
       matrix:
         ext: ["", -mkl, -gpu, -mkl-gpu]
@@ -39,6 +39,17 @@ jobs:
           echo Downloading Bazel
           curl -L https://github.com/bazelbuild/bazel/releases/download/2.0.0/bazel-2.0.0-installer-linux-x86_64.sh -o bazel.sh --retry 10
           bash bazel.sh
+          echo Installing CUDA
+          curl -L https://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda-repo-rhel7-10-1-local-10.1.243-418.87.00-1.0-1.x86_64.rpm -o $HOME/cuda.rpm
+          curl -L https://developer.download.nvidia.com/compute/redist/cudnn/v7.6.5/cudnn-10.1-linux-x64-v7.6.5.32.tgz -o $HOME/cudnn.tgz
+          curl -L https://developer.download.nvidia.com/compute/redist/nccl/v2.4/nccl_2.4.8-1+cuda10.1_x86_64.txz -o $HOME/nccl.txz
+          rpm -i $HOME/cuda.rpm
+          cd /var/cuda-repo-10-1-local-10.1.243-418.87.00/; rpm -i --nodeps cuda*.rpm libc*.rpm
+          ln -sf /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/libcuda.so
+          ln -sf /usr/local/cuda/lib64/stubs/libnvidia-ml.so /usr/local/cuda/lib64/libnvidia-ml.so
+          tar hxvf $HOME/cudnn.tgz -C /usr/local/
+          tar hxvf $HOME/nccl.txz --strip-components=1 -C /usr/local/cuda/
+          mv /usr/local/cuda/lib/* /usr/local/cuda/lib64/
       - name: Checkout repository
         uses: actions/checkout@v1
       - name: Build project
@@ -110,7 +121,7 @@ jobs:
           curl.exe -L https://github.com/bazelbuild/bazel/releases/download/2.0.0/bazel-2.0.0-windows-x86_64.exe -o C:/bazel/bazel.exe --retry 10
           echo Installing CUDA
           curl.exe -L http://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.243_426.00_windows.exe -o cuda.exe
-          curl.exe -L https://developer.download.nvidia.com/compute/redist/cudnn/v7.6.4/cudnn-10.1-windows7-x64-v7.6.4.38.zip -o cudnn.zip
+          curl.exe -L https://developer.download.nvidia.com/compute/redist/cudnn/v7.6.5/cudnn-10.1-windows7-x64-v7.6.5.32.zip -o cudnn.zip
           cuda.exe -s
           mkdir cuda
           unzip.exe cudnn.zip

README.md

@@ -25,16 +25,29 @@ migrated from Bazel to Maven, which is more familiar for most Java developers.
 The following describes the layout of the repository and its different artifacts:
 
 * `tensorflow-core`
-  * All artifacts that build up the core language bindings of TensorFlow for Java. 
-  * Those artifacts provide the minimal support required to use the TensorFlow runtime on a JVM.
+  * All artifacts that build up the core language bindings of TensorFlow for Java
+  * Intended audience: projects that provide their own APIs or frameworks on top of 
+    TensorFlow and just want a thin layer to access the TensorFlow runtime from the JVM 
 
 * `tensorflow-framework`
-  * High-level APIs built on top of the core libraries to simplify neural network training and inference 
-    using TensorFlow.
+  * Complete but fairly primitive API for building and training neural networks with TensorFlow
+  * Intended audience: expert neural network developers who prefer to make explicit, detailed decisions 
+    about their models and training algorithms
+
+* `tensorflow-keras` (early WIP; only defined in `dev` profile)
+  * Partially covers the framework API to allow simpler definition of models and training algorithms
+  * Intended to be familiar if you know the Python Keras API, but prioritizes clean, idiomatic Java 
+    over fidelity to Python
+  * Provides defaults based on common best practices
+  * Allows developers to selectively be more explicit by overriding defaults or dipping into the framework API
+  * Intended audience: neural network developers across the spectrum from beginner to expert who prefer to
+    rely mostly on best-practice defaults and then selectively fine-tune
 
 * `ndarray`
-  * Generic utility library for n-dimensional data I/O operations. It is used by TensorFlow without depending
-    on it, making its usage eligible to any type of projects, using TensorFlow or not.
+  * Generic utility library for n-dimensional data I/O operations
+  * Used by TensorFlow but does not depend on TensorFlow
+  * Intended audience: any developer who needs a Java n-dimensional array implementation, whether or not they
+    use it with TensorFlow
 
 ## Building Sources
 

tensorflow-core/pom.xml

@@ -59,7 +59,7 @@
     <javacpp.platform.macosx-x86_64.extension>macosx-x86_64${javacpp.platform.extension}</javacpp.platform.macosx-x86_64.extension>
     <javacpp.platform.windows-x86.extension>windows-x86${javacpp.platform.extension}</javacpp.platform.windows-x86.extension>
     <javacpp.platform.windows-x86_64.extension>windows-x86_64${javacpp.platform.extension}</javacpp.platform.windows-x86_64.extension>
-    <javacpp.version>1.5.4-SNAPSHOT</javacpp.version>
+    <javacpp.version>1.5.4</javacpp.version>
     <mkl-dnn.version>0.21.5-${javacpp.version}</mkl-dnn.version>
   </properties>
 

tensorflow-core/tensorflow-core-api/build.sh

@@ -24,6 +24,7 @@ fi
 
 if [[ "${EXTENSION:-}" == *gpu* ]]; then
     export BUILD_FLAGS="$BUILD_FLAGS --config=cuda"
+    export TF_CUDA_COMPUTE_CAPABILITIES="3.5,7.0"
     if [[ -z ${TF_CUDA_PATHS:-} ]] && [[ -d ${CUDA_PATH:-} ]]; then
         # Work around some issue with Bazel preventing it from detecting CUDA on Windows
         export TF_CUDA_PATHS="$CUDA_PATH"
@@ -35,9 +36,9 @@ BUILD_FLAGS="$BUILD_FLAGS --experimental_repo_remote_exec --python_path="$PYTHON
 # Always allow distinct host configuration since we rely on the host JVM for a few things (this was disabled by default on windows)
 BUILD_FLAGS="$BUILD_FLAGS --distinct_host_configuration=true"
 
-# Build C API of TensorFlow itself including a target to generate ops for Java
+# Build C/C++ API of TensorFlow itself including a target to generate ops for Java
 bazel build $BUILD_FLAGS \
-    @org_tensorflow//tensorflow:tensorflow \
+    @org_tensorflow//tensorflow:tensorflow_cc \
     @org_tensorflow//tensorflow/tools/lib_package:jnilicenses_generate \
     :java_proto_gen_sources \
     :java_op_generator \
@@ -49,23 +50,23 @@ export BAZEL_BIN=$(pwd -P)/bazel-bin
 export TENSORFLOW_BIN=$BAZEL_BIN/external/org_tensorflow/tensorflow
 
 # Normalize some paths with symbolic links
-TENSORFLOW_SO=($TENSORFLOW_BIN/libtensorflow.so.?.?.?)
+TENSORFLOW_SO=($TENSORFLOW_BIN/libtensorflow_cc.so.?.?.?)
 if [[ -f $TENSORFLOW_SO ]]; then
     export TENSORFLOW_LIB=$TENSORFLOW_SO
-    ln -sf $(basename $TENSORFLOW_SO) $TENSORFLOW_BIN/libtensorflow.so
-    ln -sf $(basename $TENSORFLOW_SO) $TENSORFLOW_BIN/libtensorflow.so.2
+    ln -sf $(basename $TENSORFLOW_SO) $TENSORFLOW_BIN/libtensorflow_cc.so
+    ln -sf $(basename $TENSORFLOW_SO) $TENSORFLOW_BIN/libtensorflow_cc.so.2
 fi
-TENSORFLOW_DYLIB=($TENSORFLOW_BIN/libtensorflow.?.?.?.dylib)
+TENSORFLOW_DYLIB=($TENSORFLOW_BIN/libtensorflow_cc.?.?.?.dylib)
 if [[ -f $TENSORFLOW_DYLIB ]]; then
     export TENSORFLOW_LIB=$TENSORFLOW_DYLIB
-    ln -sf $(basename $TENSORFLOW_DYLIB) $TENSORFLOW_BIN/libtensorflow.dylib
-    ln -sf $(basename $TENSORFLOW_DYLIB) $TENSORFLOW_BIN/libtensorflow.2.dylib
+    ln -sf $(basename $TENSORFLOW_DYLIB) $TENSORFLOW_BIN/libtensorflow_cc.dylib
+    ln -sf $(basename $TENSORFLOW_DYLIB) $TENSORFLOW_BIN/libtensorflow_cc.2.dylib
 fi
-TENSORFLOW_DLLS=($TENSORFLOW_BIN/tensorflow.dll.if.lib $TENSORFLOW_BIN/libtensorflow.dll.ifso)
+TENSORFLOW_DLLS=($TENSORFLOW_BIN/tensorflow_cc.dll.if.lib $TENSORFLOW_BIN/libtensorflow_cc.dll.ifso)
 for TENSORFLOW_DLL in ${TENSORFLOW_DLLS[@]}; do
     if [[ -f $TENSORFLOW_DLL ]]; then
-        export TENSORFLOW_LIB=$TENSORFLOW_BIN/tensorflow.dll
-        ln -sf $(basename $TENSORFLOW_DLL) $TENSORFLOW_BIN/tensorflow.lib
+        export TENSORFLOW_LIB=$TENSORFLOW_BIN/tensorflow_cc.dll
+        ln -sf $(basename $TENSORFLOW_DLL) $TENSORFLOW_BIN/tensorflow_cc.lib
     fi
 done
 echo "Listing $TENSORFLOW_BIN:" && ls -l $TENSORFLOW_BIN

tensorflow-core/tensorflow-core-api/src/bazel/api_def/api_def_SoftmaxCrossEntropyWithLogits.pbtxt

@@ -1,6 +1,6 @@
 op {
   graph_op_name: "SoftmaxCrossEntropyWithLogits"
   endpoint {
-    name: "nn.SoftmaxCrossEntropyWithLogits"
+    name: "nn.raw.SoftmaxCrossEntropyWithLogits"
   }
 }

tensorflow-core/tensorflow-core-api/src/bazel/api_def/api_def_SparseSoftmaxCrossEntropyWithLogits.pbtxt

@@ -1,6 +1,6 @@
 op {
   graph_op_name: "SparseSoftmaxCrossEntropyWithLogits"
   endpoint {
-    name: "nn.SparseSoftmaxCrossEntropyWithLogits"
+    name: "nn.raw.SparseSoftmaxCrossEntropyWithLogits"
   }
 }

tensorflow-core/tensorflow-core-api/src/bazel/op_generator/op_generator.cc

@@ -509,9 +509,20 @@ void GenerateOp(const OpSpec& op, const EndpointSpec& endpoint,
     RenderInterfaceImpl(op, mode, &writer);
   }
   writer.EndLine();
-  for (const ArgumentSpec& output : op.outputs()) {
-    writer.WriteField(output.var(), PRIVATE);
+
+  Variable nameVariable = Variable::Create("OP_NAME", Type::Class("String")); 
+  Javadoc name_javadoc = Javadoc::Create("The name of this op, as known by TensorFlow core engine");
+  string quoted_string = "\"" + op.graph_op_name() + "\"";
+  writer.WriteFieldWithInitializer(nameVariable, PUBLIC|STATIC|FINAL, &name_javadoc, quoted_string );
+
+
+  if(!op.outputs().empty()) {
+    writer.EndLine();
+    for (const ArgumentSpec& output : op.outputs()) {
+        writer.WriteField(output.var(), PRIVATE);
+    }
   }
+
   RenderConstructor(op, op_class, &writer);
   writer.EndType();
 }

tensorflow-core/tensorflow-core-api/src/bazel/op_generator/source_writer.cc

@@ -238,6 +238,22 @@ SourceWriter& SourceWriter::WriteField(const Variable& field, int modifiers,
   return *this;
 }
 
+SourceWriter& SourceWriter::WriteFieldWithInitializer(const Variable& field,
+	int modifiers, const Javadoc* javadoc, const string& initializer) {
+  // If present, write field javadoc only as one brief line
+  if (javadoc != nullptr && !javadoc->brief().empty()) {
+    Append("/** ").Append(javadoc->brief()).Append(" */").EndLine();
+  }
+  WriteModifiers(modifiers);
+  if (!initializer.empty())
+    AppendType(field.type()).Append(" ").Append(field.name()).
+                            Append(" = ").Append(initializer).Append(";");
+  else
+    AppendType(field.type()).Append(" ").Append(field.name()).Append(";");
+  EndLine();
+  return *this;
+}
+
 SourceWriter& SourceWriter::WriteModifiers(int modifiers) {
   if (modifiers & PUBLIC) {
     Append("public ");

tensorflow-core/tensorflow-core-api/src/bazel/op_generator/source_writer.h

@@ -153,6 +153,9 @@ class SourceWriter {
   SourceWriter& WriteField(const Variable& field, int modifiers,
                            const Javadoc* javadoc = nullptr);
 
+  SourceWriter& WriteFieldWithInitializer(const Variable& field,
+        int modifiers, const Javadoc* javadoc = nullptr, const string& initializer = nullptr);
+
  protected:
   virtual void DoAppend(const StringPiece& str) = 0;
 
@@ -161,7 +164,7 @@ class SourceWriter {
   class TypeVisitor {
    public:
     virtual ~TypeVisitor() = default;
-    void Visit(const Type& type);
+    void Visit(const Type& type);	
 
    protected:
     virtual void DoVisit(const Type& type) = 0;