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

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into Adapting_to_the_BatchNorm_structure_to_support_3D_data

cmake/external/openblas.cmake

@@ -25,7 +25,12 @@ IF(NOT ${CBLAS_FOUND})
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         CACHE FILEPATH "openblas library." FORCE)
 
+    IF(APPLE)
+        SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
+        SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
+    ELSE()
         SET(COMMON_ARGS CC=${CMAKE_C_COMPILER} NO_SHARED=1 NO_LAPACK=1 libs)
+    ENDIF()
 
     IF(CMAKE_CROSSCOMPILING)
         IF(ANDROID)
@@ -40,11 +45,11 @@ IF(NOT ${CBLAS_FOUND})
             SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=${TARGET} ARM_SOFTFP_ABI=1 USE_THREAD=0)
         ELSEIF(RPI)
             # use hardfp
-            SET(OPENBLAS_COMMIT "v0.2.19")
+            SET(OPENBLAS_COMMIT "v0.2.20")
             SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 USE_THREAD=0)
         ENDIF()
     ELSE()
-        SET(OPENBLAS_COMMIT "v0.2.19")
+        SET(OPENBLAS_COMMIT "v0.2.20")
         SET(OPTIONAL_ARGS "")
         IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
             SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)

paddle/framework/tensor_impl.h

@@ -117,6 +117,8 @@ inline void Tensor::CopyFrom(const Tensor& src,
     memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
   }
+  PADDLE_ENFORCE(cudaStreamSynchronize(0),
+                 "cudaStreamSynchronize failed in Tensor CopyFrom");
 
 #endif
 }

paddle/gserver/layers/CrossEntropyOverBeam.cpp

@@ -223,7 +223,7 @@ void CrossEntropyOverBeam::checkInputs() {
                                 << inputLayers_[i * 3]->getName()
                                 << " should be a nested sequence";
       CHECK_EQ(getInputValue(i * 3 + 1)->getWidth(), beamSize_);
-      CHECK_EQ(scores.getNumSequences(), batchSize_);
+      CHECK_EQ(batchSize_, static_cast<size_t>(scores.getNumSequences()));
       CHECK_EQ(scores.getNumSubSequences(), selCandidates.getBatchSize());
     } else {
       CHECK(scores.hasSeq()) << "input " << i << " "
@@ -231,10 +231,10 @@ void CrossEntropyOverBeam::checkInputs() {
                              << " should be a sequence";
       batchSize_ = scores.getNumSequences();
       beamSize_ = getInputValue(i * 3 + 1)->getWidth();
-      CHECK_EQ(batchSize_, selCandidates.getBatchSize());
+      CHECK_EQ(batchSize_, static_cast<size_t>(selCandidates.getBatchSize()));
     }
     CHECK_EQ(1U, scores.value->getWidth());
-    CHECK_EQ(batchSize_, goldSeq.getBatchSize());
+    CHECK_EQ(batchSize_, static_cast<size_t>(goldSeq.getBatchSize()));
   }
 }
 
@@ -377,8 +377,8 @@ void CrossEntropyOverBeam::forward(PassType passType) {
 
   MatrixPtr outputValue = getOutputValue();
   for (size_t i = 0; i < batchSize_; ++i) {
-    beamCosts_[i].setData(
-        std::move(std::make_shared<BeamExpansion>(beamPerSeq_[i])), beamSize_);
+    BeamExpansionPtr ptr = std::make_shared<BeamExpansion>(beamPerSeq_[i]);
+    beamCosts_[i].setData(std::move(ptr), beamSize_);
     outputValue->getData()[i] = beamCosts_[i].forward();
   }
 }

paddle/operators/CMakeLists.txt

+file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
+string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
 function(op_library TARGET)
     # op_library is a function to create op library. The interface is same as
     # cc_library. But it handle split GPU/CPU code and link some common library
     # for ops.
+    set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
     set(cc_srcs)
     set(cu_srcs)
     set(op_common_deps operator op_registry)
@@ -43,33 +46,26 @@ endfunction()
 
 add_subdirectory(math)
 
-cc_test(gather_test SRCS gather_test.cc DEPS tensor)
-op_library(gather_op SRCS gather_op.cc gather_op.cu)
-
-cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
-op_library(scatter_op SRCS scatter_op.cc scatter_op.cu)
-
-cc_library(net_op SRCS net_op.cc DEPS op_registry)
-cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
-
-op_library(add_op SRCS add_op.cc add_op.cu)
-
-op_library(mean_op SRCS mean_op.cc mean_op.cu)
+list(REMOVE_ITEM GENERAL_OPS
+     net_op
+     minus_op
+     mul_op
+     recurrent_op
+     scale_op)
 
+op_library(net_op SRCS net_op.cc)
+op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)
 op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
-op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
+op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc 
+  DEPS framework_proto tensor operator net_op)
+op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
 
-op_library(sigmoid_op SRCS sigmoid_op.cc sigmoid_op.cu)
-op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
-op_library(gaussian_random_op SRCS gaussian_random_op.cc gaussian_random_op.cu)
-op_library(cross_entropy_op SRCS cross_entropy_op.cc cross_entropy_op.cu)
-op_library(fill_zeros_like_op SRCS fill_zeros_like_op.cc fill_zeros_like_op.cu)
+foreach(src ${GENERAL_OPS})
+    op_library(${src} SRCS ${src}.cc ${src}.cu)
+endforeach()
 
-op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)
+set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
 
-op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
-    DEPS framework_proto tensor op_registry operator net_op)
-op_library(uniform_random_op SRCS uniform_random_op.cc uniform_random_op.cu)
-op_library(lookup_table_op SRCS lookup_table_op.cc lookup_table_op.cu)
-op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
-op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)
+cc_test(gather_test SRCS gather_test.cc DEPS tensor)
+cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
+cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)

paddle/pybind/CMakeLists.txt

@@ -2,21 +2,5 @@ if(WITH_PYTHON)
 cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
-    sgd_op
-    gather_op
-    scatter_op
-    add_op
-    mul_op
-    rowwise_add_op
-    sigmoid_op
-    softmax_op
-    mean_op
-    cross_entropy_op
-    recurrent_op
-    uniform_random_op
-    gaussian_random_op
-    fill_zeros_like_op
-    lookup_table_op
-    scale_op
-    minus_op)
+    ${GLOB_OP_LIB})
 endif(WITH_PYTHON)

python/paddle/trainer_config_helpers/layers.py

@@ -137,7 +137,7 @@ __all__ = [
     'clip_layer',
     'slice_projection',
     'seq_slice_layer',
-    'kmax_sequence_score_layer',
+    'kmax_seq_score_layer',
     'img_pool3d_layer',
     'scale_shift_layer',
     'img_conv3d_layer',
@@ -5996,7 +5996,7 @@ def cross_entropy_over_beam(input, name=None):
     Note that, if gold falls off the beam at search step t, then the cost is
     calculated over the beam at step t.
 
-    This cost layer always works together with kmax_sequence_score_layer,
+    This cost layer always works together with kmax_seq_score_layer,
     sub_nested_seq_layer, and sequence_slice_layer to trim the input to form a
     sub-search space.
 
@@ -6599,14 +6599,14 @@ def seq_slice_layer(input, starts, ends, name=None):
 
 @wrap_name_default()
 @layer_support()
-def kmax_sequence_score_layer(input, name=None, beam_size=1):
+def kmax_seq_score_layer(input, name=None, beam_size=1):
     """
     This layer accepts one input which are scores over a sequence or a nested
     sequence, and returns indices of beam_size sequences with highest scores.
 
     .. code-block:: python
 
-        kmax_indices = kmax_sequence_score_layer(input=input_layer, beam_size)
+        kmax_indices = kmax_seq_score_layer(input=input_layer, beam_size)
 
 
     :param name: The Layer Name.
@@ -6619,10 +6619,10 @@ def kmax_sequence_score_layer(input, name=None, beam_size=1):
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
-    assert isinstance(input, LayerOutput), ("kmax_sequence_score_layer "
+    assert isinstance(input, LayerOutput), ("kmax_seq_score_layer "
                                             "accepts only one input.")
     assert input.size == 1, (
-        "input of kmax_sequence_score_layer is a score"
+        "input of kmax_seq_score_layer is a score "
         "over a sequence or a nested sequence, so its width must be 1.")
 
     Layer(

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -8,7 +8,7 @@ test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
 test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
 test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
-test_kmax_seq_socre_layer test_seq_select_layers test_scale_shift_layer
+test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer
 test_seq_slice_layer test_cross_entropy_over_beam test_pooling3D_layer
 test_conv3d_layer test_deconv3d_layer test_BatchNorm3D)
 

python/paddle/trainer_config_helpers/tests/configs/protostr/test_cross_entropy_over_beam.protostr

@@ -12,7 +12,7 @@ layers {
   active_type: ""
 }
 layers {
-  name: "__kmax_sequence_score_layer_0__"
+  name: "__kmax_seq_score_layer_0__"
   type: "kmax_seq_score"
   active_type: ""
   inputs {
@@ -29,7 +29,7 @@ layers {
     input_layer_name: "sentence_states"
   }
   inputs {
-    input_layer_name: "__kmax_sequence_score_layer_0__"
+    input_layer_name: "__kmax_seq_score_layer_0__"
   }
 }
 layers {
@@ -44,7 +44,7 @@ layers {
   bias_parameter_name: "___fc_layer_0__.wbias"
 }
 layers {
-  name: "__kmax_sequence_score_layer_1__"
+  name: "__kmax_seq_score_layer_1__"
   type: "kmax_seq_score"
   active_type: ""
   inputs {
@@ -61,7 +61,7 @@ layers {
     input_layer_name: "__sub_nested_seq_layer_0__"
   }
   inputs {
-    input_layer_name: "__kmax_sequence_score_layer_1__"
+    input_layer_name: "__kmax_seq_score_layer_1__"
   }
   select_first: true
 }
@@ -77,7 +77,7 @@ layers {
   bias_parameter_name: "___fc_layer_1__.wbias"
 }
 layers {
-  name: "__kmax_sequence_score_layer_2__"
+  name: "__kmax_seq_score_layer_2__"
   type: "kmax_seq_score"
   active_type: ""
   inputs {
@@ -111,7 +111,7 @@ layers {
     input_layer_name: "sentence_scores"
   }
   inputs {
-    input_layer_name: "__kmax_sequence_score_layer_0__"
+    input_layer_name: "__kmax_seq_score_layer_0__"
   }
   inputs {
     input_layer_name: "sentences_ids"
@@ -120,7 +120,7 @@ layers {
     input_layer_name: "__fc_layer_0__"
   }
   inputs {
-    input_layer_name: "__kmax_sequence_score_layer_1__"
+    input_layer_name: "__kmax_seq_score_layer_1__"
   }
   inputs {
     input_layer_name: "start_ids"
@@ -129,7 +129,7 @@ layers {
     input_layer_name: "__fc_layer_1__"
   }
   inputs {
-    input_layer_name: "__kmax_sequence_score_layer_2__"
+    input_layer_name: "__kmax_seq_score_layer_2__"
   }
   inputs {
     input_layer_name: "end_ids"
@@ -185,13 +185,13 @@ sub_models {
   name: "root"
   layer_names: "sentence_states"
   layer_names: "sentence_scores"
-  layer_names: "__kmax_sequence_score_layer_0__"
+  layer_names: "__kmax_seq_score_layer_0__"
   layer_names: "__sub_nested_seq_layer_0__"
   layer_names: "__fc_layer_0__"
-  layer_names: "__kmax_sequence_score_layer_1__"
+  layer_names: "__kmax_seq_score_layer_1__"
   layer_names: "__seq_slice_layer_0__"
   layer_names: "__fc_layer_1__"
-  layer_names: "__kmax_sequence_score_layer_2__"
+  layer_names: "__kmax_seq_score_layer_2__"
   layer_names: "sentences_ids"
   layer_names: "start_ids"
   layer_names: "end_ids"

python/paddle/trainer_config_helpers/tests/configs/protostr/test_kmax_seq_socre_layer.protostr

@@ -17,7 +17,7 @@ layers {
   bias_parameter_name: "___fc_layer_0__.wbias"
 }
 layers {
-  name: "__kmax_sequence_score_layer_0__"
+  name: "__kmax_seq_score_layer_0__"
   type: "kmax_seq_score"
   active_type: ""
   inputs {
@@ -46,14 +46,14 @@ parameters {
   initial_smart: false
 }
 input_layer_names: "input_seq"
-output_layer_names: "__kmax_sequence_score_layer_0__"
+output_layer_names: "__kmax_seq_score_layer_0__"
 sub_models {
   name: "root"
   layer_names: "input_seq"
   layer_names: "__fc_layer_0__"
-  layer_names: "__kmax_sequence_score_layer_0__"
+  layer_names: "__kmax_seq_score_layer_0__"
   input_layer_names: "input_seq"
-  output_layer_names: "__kmax_sequence_score_layer_0__"
+  output_layer_names: "__kmax_seq_score_layer_0__"
   is_recurrent_layer_group: false
 }
 

python/paddle/trainer_config_helpers/tests/configs/test_cross_entropy_over_beam.py

@@ -7,14 +7,14 @@ beam_size = 5
 # the first beam expansion.
 sentence_states = data_layer(name="sentence_states", size=32)
 sentence_scores = data_layer(name="sentence_scores", size=1)
-topk_sentence_ids = kmax_sequence_score_layer(
+topk_sentence_ids = kmax_seq_score_layer(
     input=sentence_scores, beam_size=beam_size)
 
 # the second beam expansion.
 topk_sen = sub_nested_seq_layer(
     input=sentence_states, selected_indices=topk_sentence_ids)
 start_pos_scores = fc_layer(input=topk_sen, size=1, act=LinearActivation())
-topk_start_pos_ids = kmax_sequence_score_layer(
+topk_start_pos_ids = kmax_seq_score_layer(
     input=sentence_scores, beam_size=beam_size)
 
 # the final beam expansion.
@@ -22,7 +22,7 @@ topk_start_spans = seq_slice_layer(
     input=topk_sen, starts=topk_start_pos_ids, ends=None)
 end_pos_scores = fc_layer(
     input=topk_start_spans, size=1, act=LinearActivation())
-topk_end_pos_ids = kmax_sequence_score_layer(
+topk_end_pos_ids = kmax_seq_score_layer(
     input=end_pos_scores, beam_size=beam_size)
 
 # define the cost

python/paddle/trainer_config_helpers/tests/configs/test_kmax_seq_socre_layer.py

@@ -4,6 +4,6 @@ from paddle.trainer_config_helpers import *
 
 data = data_layer(name="input_seq", size=128)
 scores = fc_layer(input=data, size=1, act=ExpActivation())
-kmax_seq_id = kmax_sequence_score_layer(input=scores, beam_size=5)
+kmax_seq_id = kmax_seq_score_layer(input=scores, beam_size=5)
 
 outputs(kmax_seq_id)

python/paddle/v2/__init__.py

@@ -78,6 +78,8 @@ def init(**kwargs):
 
     if 'use_gpu' in kwargs:
         cp.g_command_config_args['use_gpu'] = kwargs['use_gpu']
+    if 'use_mkldnn' in kwargs:
+        cp.g_command_config_args['use_mkldnn'] = kwargs['use_mkldnn']
     assert 'parallel_nn' not in kwargs, ("currently 'parallel_nn' is not "
                                          "supported in v2 APIs.")