Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into pad_op

paddle/gserver/layers/DetectionOutputLayer.cpp

@@ -139,7 +139,13 @@ void DetectionOutputLayer::forward(PassType passType) {
                                        allDecodedBBoxes,
                                        &allIndices);
 
+  if (numKept > 0) {
     resetOutput(numKept, 7);
+  } else {
+    MatrixPtr outV = getOutputValue();
+    outV = NULL;
+    return;
+  }
   MatrixPtr outV = getOutputValue();
   getDetectionOutput(confBuffer_->getData(),
                      numKept,

paddle/gserver/layers/DetectionUtil.cpp

@@ -469,7 +469,7 @@ size_t getDetectionIndices(
     const size_t numClasses,
     const size_t backgroundId,
     const size_t batchSize,
-    const size_t confThreshold,
+    const real confThreshold,
     const size_t nmsTopK,
     const real nmsThreshold,
     const size_t keepTopK,

paddle/gserver/layers/DetectionUtil.h

@@ -275,7 +275,7 @@ size_t getDetectionIndices(
     const size_t numClasses,
     const size_t backgroundId,
     const size_t batchSize,
-    const size_t confThreshold,
+    const real confThreshold,
     const size_t nmsTopK,
     const real nmsThreshold,
     const size_t keepTopK,

paddle/platform/enforce.h

@@ -25,10 +25,6 @@ limitations under the License. */
 #include "paddle/string/printf.h"
 #include "paddle/string/to_string.h"
 
-#ifdef __GNUC__
-#include <cxxabi.h>  // for __cxa_demangle
-#endif
-
 #ifndef PADDLE_ONLY_CPU
 
 #include "paddle/platform/dynload/cublas.h"
@@ -46,19 +42,6 @@ limitations under the License. */
 namespace paddle {
 namespace platform {
 
-namespace {
-#ifdef __GNUC__
-inline std::string demangle(std::string name) {
-  int status = -4;  // some arbitrary value to eliminate the compiler warning
-  std::unique_ptr<char, void (*)(void*)> res{
-      abi::__cxa_demangle(name.c_str(), NULL, NULL, &status), std::free};
-  return (status == 0) ? res.get() : name;
-}
-#else
-inline std::string demangle(std::string name) { return name; }
-#endif
-}
-
 struct EnforceNotMet : public std::exception {
   std::exception_ptr exp_;
   std::string err_str_;
@@ -79,7 +62,7 @@ struct EnforceNotMet : public std::exception {
       Dl_info info;
       for (int i = 0; i < size; ++i) {
         if (dladdr(call_stack[i], &info)) {
-          auto demangled = demangle(info.dli_sname);
+          auto demangled = info.dli_sname;
           auto addr_offset = static_cast<char*>(call_stack[i]) -
                              static_cast<char*>(info.dli_saddr);
           sout << string::Sprintf("%-3d %*0p %s + %zd\n", i,

paddle/scripts/docker/build.sh

@@ -30,6 +30,8 @@ Configuring cmake in /paddle/build ...
       -DCMAKE_BUILD_TYPE=Release
       -DWITH_DOC=OFF
       -DWITH_GPU=${WITH_GPU:-OFF}
+      -DWITH_MKLDNN=${WITH_MKLDNN:-ON}
+      -DWITH_MKLML=${WITH_MKLML:-ON}
       -DWITH_AVX=${WITH_AVX:-OFF}
       -DWITH_GOLANG=${WITH_GOLANG:-ON}
       -DWITH_SWIG_PY=ON
@@ -50,6 +52,8 @@ cmake .. \
       -DCMAKE_BUILD_TYPE=Release \
       -DWITH_DOC=OFF \
       -DWITH_GPU=${WITH_GPU:-OFF} \
+      -DWITH_MKLDNN=${WITH_MKLDNN:-ON} \
+      -DWITH_MKLML=${WITH_MKLML:-ON} \
       -DWITH_AVX=${WITH_AVX:-OFF} \
       -DWITH_GOLANG=${WITH_GOLANG:-ON} \
       -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \

python/paddle/trainer/config_parser.py

@@ -3748,8 +3748,8 @@ class SwitchOrderLayer(LayerBase):
     def __init__(self, name, inputs, reshape, **xargs):
         super(SwitchOrderLayer, self).__init__(
             name, 'switch_order', 0, inputs=inputs, **xargs)
-        self.config.reshape_conf.heightAxis.extend(reshape['height'])
-        self.config.reshape_conf.widthAxis.extend(reshape['width'])
+        self.config.reshape_conf.height_axis.extend(reshape['height'])
+        self.config.reshape_conf.width_axis.extend(reshape['width'])
 
 
 # Deprecated, use a new layer specific class instead

python/paddle/trainer_config_helpers/layers.py

@@ -1223,7 +1223,8 @@ def detection_output_layer(input_loc,
                            name=None):
     """
     Apply the NMS to the output of network and compute the predict bounding
-    box location.
+    box location. The output of this layer could be None if there is no valid
+    bounding box.
 
     :param name: The Layer Name.
     :type name: basestring
@@ -6460,6 +6461,7 @@ def switch_order_layer(input,
     return LayerOutput(
         name=name,
         layer_type=LayerType.SWITCH_ORDER_LAYER,
+        activation=act,
         parents=input,
         size=l.config.size)
 

python/paddle/v2/event.py

@@ -53,10 +53,13 @@ class BeginPass(object):
 class EndPass(WithMetric):
     """
     Event On One Pass Training Complete.
+    To get the output of a specific layer, add "event.gm.getLayerOutputs('predict_layer')"
+    in your event_handler call back
     """
 
-    def __init__(self, pass_id, evaluator):
+    def __init__(self, pass_id, evaluator, gm):
         self.pass_id = pass_id
+        self.gm = gm
         WithMetric.__init__(self, evaluator)
 
 
@@ -73,10 +76,13 @@ class BeginIteration(object):
 class EndIteration(WithMetric):
     """
     Event On One Batch Training Complete.
+    To get the output of a specific layer, add "event.gm.getLayerOutputs('predict_layer')"
+    in your event_handler call back
     """
 
-    def __init__(self, pass_id, batch_id, cost, evaluator):
+    def __init__(self, pass_id, batch_id, cost, evaluator, gm):
         self.pass_id = pass_id
         self.batch_id = batch_id
         self.cost = cost
+        self.gm = gm
         WithMetric.__init__(self, evaluator)

python/paddle/v2/trainer.py

@@ -174,13 +174,18 @@ class SGD(object):
                         pass_id=pass_id,
                         batch_id=batch_id,
                         cost=cost,
-                        evaluator=batch_evaluator))
+                        evaluator=batch_evaluator,
+                        gm=self.__gradient_machine__))
                 self.__parameter_updater__.finishBatch(cost)
                 batch_evaluator.finish()
 
             self.__parameter_updater__.finishPass()
             pass_evaluator.finish()
-            event_handler(v2_event.EndPass(pass_id, evaluator=pass_evaluator))
+            event_handler(
+                v2_event.EndPass(
+                    pass_id,
+                    evaluator=pass_evaluator,
+                    gm=self.__gradient_machine__))
         self.__gradient_machine__.finish()
 
     def test(self, reader, feeding=None):