Merge branch 'develop' of github.com:baidu/Paddle into feature/c_api

cmake/external/python.cmake

@@ -26,6 +26,10 @@ IF(PYTHONLIBS_FOUND AND PYTHONINTERP_FOUND)
     find_python_module(wheel REQUIRED)
     find_python_module(google.protobuf REQUIRED)
     FIND_PACKAGE(NumPy REQUIRED)
+    IF(${PY_GOOGLE.PROTOBUF_VERSION} VERSION_LESS "3.0.0")
+        MESSAGE(FATAL_ERROR "Found Python Protobuf ${PY_GOOGLE.PROTOBUF_VERSION} < 3.0.0, "
+        "please use pip to upgrade protobuf.")
+    ENDIF(${PY_GOOGLE.PROTOBUF_VERSION} VERSION_LESS "3.0.0")
 ELSE(PYTHONLIBS_FOUND AND PYTHONINTERP_FOUND)
     ##################################### PYTHON ########################################
     SET(PYTHON_SOURCES_DIR ${THIRD_PARTY_PATH}/python)

cmake/python_module.cmake

@@ -26,5 +26,18 @@ function(find_python_module module)
     if(NOT PY_${module_upper}_FOUND AND ${module}_FIND_REQUIRED)
         message(FATAL_ERROR "python module ${module} is not found")
     endif()
+
+    execute_process(COMMAND "${PYTHON_EXECUTABLE}" "-c"
+        "import sys, ${module}; sys.stdout.write(${module}.__version__)"
+        OUTPUT_VARIABLE _${module}_version
+        RESULT_VARIABLE _${module}_status
+        ERROR_QUIET
+        OUTPUT_STRIP_TRAILING_WHITESPACE)
+    if(NOT _${module}_status)
+        set(PY_${module_upper}_VERSION ${_${module}_version} CACHE STRING
+            "Version of Python module ${module}")
+    endif(NOT _${module}_status)
+
     set(PY_${module_upper}_FOUND ${PY_${module_upper}_FOUND} PARENT_SCOPE)
+    set(PY_${module_upper}_VERSION ${PY_${module_upper}_VERSION} PARENT_SCOPE)
 endfunction(find_python_module)

doc/howto/usage/k8s/k8s_distributed_cn.md

@@ -159,6 +159,8 @@ docker build -t your_repo/paddle:mypaddle .
 docker push  your_repo/paddle:mypaddle
 ```
 
+注意上述命令中`your_repo`表示读者所使用的Docker镜像仓库地址，读者需要替换成自己使用的仓库地址。下文使用`your_repo/paddle:mypaddle`这个地址来表示此步骤所构建出的镜像。
+
 ### 上传训练文件
 
 本文使用PaddlePaddle官方的[recommendation demo](http://www.paddlepaddle.org/doc/demo/index.html#recommendation)作为这次训练的内容，我们将训练文件与数据放在一个job name命名的目录中，上传到MFS共享存储。完成后MFS上的文件内容大致如下：
@@ -244,6 +246,8 @@ spec:
 
 `CONF_PADDLE_GRADIENT_NUM`表示训练节点数量，即`--num_gradient_servers`参数
 
+这些参数的具体描述，读者可以查看[这里](http://www.paddlepaddle.org/doc/ui/cmd_argument/detail_introduction.html#parameter-server-and-distributed-communication)。
+
 编写完YAML文件后，可以使用Kubernetes的命令行工具创建job。
 
 ```bash

paddle/math/Matrix.cpp

@@ -1311,7 +1311,9 @@ void GpuMatrix::paramReluForward(Matrix& data, Matrix& W) {
   real* w = W.getData();
   size_t numElements = data.getWidth();
   size_t numSamples = data.getHeight();
-  size_t partial_sum = numElements / (W.getHeight() * W.getWidth());
+  size_t paraSize = W.getHeight() * W.getWidth();
+  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
+  size_t partial_sum = numElements / paraSize;
   real* output = getData();
   hl_param_relu_forward(output, input, w, numElements, numSamples, partial_sum);
 }
@@ -1324,7 +1326,9 @@ void GpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) {
   real* wgrad = data_;
   size_t numElements = data.getWidth();
   size_t numSamples = data.getHeight();
-  size_t partial_sum = numElements / (this->getHeight() * this->getWidth());
+  size_t paraSize = this->getHeight() * this->getWidth();
+  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
+  size_t partial_sum = numElements / paraSize;
   hl_param_relu_backward_w(
       wgrad, ograd, input, numElements, numSamples, partial_sum);
 }
@@ -1336,7 +1340,9 @@ void GpuMatrix::paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) {
   real* w = W.getData();
   size_t numElements = data.getWidth();
   size_t numSamples = data.getHeight();
-  size_t partial_sum = numElements / (W.getHeight() * W.getWidth());
+  size_t paraSize = W.getHeight() * W.getWidth();
+  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
+  size_t partial_sum = numElements / paraSize;
   hl_param_relu_backward_diff(
       ograd, input, w, diff, numElements, numSamples, partial_sum);
 }
@@ -3764,7 +3770,9 @@ void CpuMatrix::paramReluForward(Matrix& data, Matrix& W) {
   real* w = W.getData();
   size_t numElements = data.getWidth();
   size_t numSamples = data.getHeight();
-  size_t partial_sum = numElements / (W.getHeight() * W.getWidth());
+  size_t paraSize = W.getHeight() * W.getWidth();
+  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
+  size_t partial_sum = numElements / paraSize;
   for (size_t n = 0, k = 0; n < numSamples; ++n) {
     for (size_t i = 0; i < numElements; ++i, ++k) {
       data_[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
@@ -3778,7 +3786,9 @@ void CpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) {
   real* wgrad = data_;
   size_t numElements = data.getWidth();
   size_t numSamples = data.getHeight();
-  size_t partial_sum = numElements / (this->getHeight() * this->getWidth());
+  size_t paraSize = this->getHeight() * this->getWidth();
+  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
+  size_t partial_sum = numElements / paraSize;
   for (size_t n = 0, k = 0; n < numSamples; ++n) {
     for (size_t i = 0; i < numElements; ++i, ++k) {
       wgrad[i / partial_sum] += ograd[k] * (input[k] > 0 ? 0 : input[k]);
@@ -3793,7 +3803,9 @@ void CpuMatrix::paramReluBackwardDiff(Matrix& oGrad, Matrix& data, Matrix& W) {
   real* w = W.getData();
   size_t numElements = data.getWidth();
   size_t numSamples = data.getHeight();
-  size_t partial_sum = numElements / (W.getHeight() * W.getWidth());
+  size_t paraSize = W.getHeight() * W.getWidth();
+  CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
+  size_t partial_sum = numElements / paraSize;
   for (size_t n = 0, k = 0; n < numSamples; ++n) {
     for (size_t i = 0; i < numElements; ++i, ++k) {
       diff[k] += ograd[k] * (input[k] > 0 ? 1 : w[i / partial_sum]);

paddle/math/tests/test_Matrix.cpp

@@ -224,10 +224,11 @@ void testParamReluBackwardW(int height, int width, int w_height, int w_width) {
 }
 
 TEST(Matrix, paramRelu) {
-  for (auto height : {10, 100}) {
-    for (auto width : {10, 100}) {
+  for (auto height : {10, 40, 100}) {
+    for (auto width : {10, 40, 100}) {
       for (auto w_height : {1, 2}) {
         for (auto w_width : {1, 2}) {
+          if (width % (w_height * w_width)) continue;
           testParamReluForward(height, width, w_height, w_width);
           testParamReluBackwardW(height, width, w_height, w_width);
         }

paddle/math/tests/test_matrixCompare.cpp

@@ -773,10 +773,11 @@ void testParamReluBackwardDiff(int height,
 }
 
 TEST(Matrix, paramReluBackwardDiff) {
-  for (auto height : {10, 100}) {
-    for (auto width : {10, 100}) {
+  for (auto height : {10, 40, 100}) {
+    for (auto width : {10, 40, 100}) {
       for (auto w_height : {1, 2}) {
         for (auto w_width : {1, 2}) {
+          if (width % (w_height * w_width)) continue;
           testParamReluBackwardDiff(height, width, w_height, w_width);
         }
       }

paddle/setup.py.in

@@ -70,6 +70,6 @@ setup(name="py_paddle",
   include_dirs = include_dirs,
   install_requires = [
     'numpy>=1.8.0',      # The numpy is required.
-    'protobuf>=2.4.1' # The paddle protobuf version
+    'protobuf>=3.0.0'    # The paddle protobuf version
   ],
 )