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

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

Dockerfile

@@ -70,7 +70,7 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 # specify sphinx version as 1.5.6 and remove -U option for [pip install -U
 # sphinx-rtd-theme] since -U option will cause sphinx being updated to newest
 # version(1.7.1 for now), which causes building documentation failed.
-RUN pip install --upgrade pip==9.0.3 && \
+RUN easy_install -U pip && \
     pip install -U wheel && \
     pip install -U docopt PyYAML sphinx==1.5.6 && \
     pip install sphinx-rtd-theme==0.1.9 recommonmark

cmake/external/mkldnn.cmake

@@ -53,11 +53,9 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "v0.14"
+    GIT_TAG             "db3424ad44901513c03a1ea31ccaacdf633fbe9f"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
-    # Patch MKLDNN to compile with gcc 4.8, the related issue is in intel/mkl-dnn#237.
-    PATCH_COMMAND       ${CMAKE_COMMAND} -E copy_if_different ${CMAKE_CURRENT_SOURCE_DIR}/patches/mkldnn.hpp ${MKLDNN_SOURCES_DIR}/src/extern_mkldnn/include/mkldnn.hpp
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
     CMAKE_ARGS          -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} 
     CMAKE_ARGS          -DMKLROOT=${MKLML_ROOT}

cmake/inference_lib.cmake

@@ -98,6 +98,14 @@ elseif (WITH_MKLML)
     )
 endif()
 
+if(WITH_MKLDNN)
+  set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mkldnn")
+  copy(mkldnn_lib
+    SRCS ${MKLDNN_INC_DIR} ${MKLDNN_SHARED_LIB}
+    DSTS ${dst_dir} ${dst_dir}/lib
+  )
+endif()
+
 if(NOT MOBILE_INFERENCE AND NOT RPI)
   set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/snappy")
   copy(snappy_lib

paddle/fluid/framework/data_device_transform.cc

@@ -36,9 +36,11 @@ void TransDataDevice(const Tensor& in, const platform::Place& dst_place,
   VLOG(3) << "DeviceTransform in, src_place " << in.place()
           << " dst_place: " << dst_place;
   auto* dev_ctx = GetDeviceContext(in.place(), dst_place);
-  dev_ctx->Wait();
+
   TensorCopy(in, dst_place, *dev_ctx, out);
-  dev_ctx->Wait();
+  if (platform::is_gpu_place(in.place()) && platform::is_cpu_place(dst_place)) {
+    dev_ctx->Wait();
+  }
 }
 
 }  // namespace framework

paddle/fluid/framework/executor.cc

@@ -228,7 +228,8 @@ static bool has_fetch_operators(
 void Executor::Run(const ProgramDesc& program, Scope* scope,
                    std::map<std::string, const LoDTensor*>* feed_targets,
                    std::map<std::string, LoDTensor*>* fetch_targets,
-                   bool create_vars, const std::string& feed_holder_name,
+                   bool create_local_scope, bool create_vars,
+                   const std::string& feed_holder_name,
                    const std::string& fetch_holder_name) {
   platform::RecordBlock b(kProgramId);
   bool has_feed_ops =
@@ -290,8 +291,9 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
   }
 
   auto ctx = Prepare(*copy_program, 0);
-  RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets, create_vars,
-                     feed_holder_name, fetch_holder_name);
+  RunPreparedContext(ctx.get(), scope, feed_targets, fetch_targets,
+                     create_local_scope, create_vars, feed_holder_name,
+                     fetch_holder_name);
 }
 
 std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
@@ -366,8 +368,9 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 void Executor::RunPreparedContext(
     ExecutorPrepareContext* ctx, Scope* scope,
     std::map<std::string, const LoDTensor*>* feed_targets,
-    std::map<std::string, LoDTensor*>* fetch_targets, bool create_vars,
-    const std::string& feed_holder_name, const std::string& fetch_holder_name) {
+    std::map<std::string, LoDTensor*>* fetch_targets, bool create_local_scope,
+    bool create_vars, const std::string& feed_holder_name,
+    const std::string& fetch_holder_name) {
   auto& global_block = ctx->prog_.Block(ctx->block_id_);
 
   PADDLE_ENFORCE(
@@ -387,7 +390,7 @@ void Executor::RunPreparedContext(
     }
   }
 
-  RunPreparedContext(ctx, scope, create_vars, create_vars);
+  RunPreparedContext(ctx, scope, create_local_scope, create_vars);
 
   // obtain the data of fetch_targets from fetch_holder
   for (auto* op : global_block.AllOps()) {

paddle/fluid/framework/executor.h

@@ -57,7 +57,7 @@ class Executor {
   void Run(const ProgramDesc& program, Scope* scope,
            std::map<std::string, const LoDTensor*>* feed_targets,
            std::map<std::string, LoDTensor*>* fetch_targets,
-           bool create_vars = true,
+           bool create_local_scope = true, bool create_vars = true,
            const std::string& feed_holder_name = "feed",
            const std::string& fetch_holder_name = "fetch");
 
@@ -76,6 +76,7 @@ class Executor {
   void RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
                           std::map<std::string, const LoDTensor*>* feed_targets,
                           std::map<std::string, LoDTensor*>* fetch_targets,
+                          bool create_local_scope = true,
                           bool create_vars = true,
                           const std::string& feed_holder_name = "feed",
                           const std::string& fetch_holder_name = "fetch");

paddle/fluid/inference/tests/test_helper.h

@@ -149,7 +149,7 @@ void TestInference(const std::string& dirname,
     state = paddle::platform::ProfilerState::kCPU;
   } else {
 #ifdef PADDLE_WITH_CUDA
-    state = paddle::platform::ProfilerState::kCUDA;
+    state = paddle::platform::ProfilerState::kAll;
     // The default device_id of paddle::platform::CUDAPlace is 0.
     // Users can get the device_id using:
     //   int device_id = place.GetDeviceId();
@@ -172,7 +172,7 @@ void TestInference(const std::string& dirname,
   }
   // Disable the profiler and print the timing information
   paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
-                                    "load_program_profiler.txt");
+                                    "load_program_profiler");
   paddle::platform::ResetProfiler();
 
   // 3. Get the feed_target_names and fetch_target_names
@@ -208,10 +208,10 @@ void TestInference(const std::string& dirname,
     if (PrepareContext) {
       ctx = executor.Prepare(*inference_program, 0);
       executor.RunPreparedContext(ctx.get(), scope, &feed_targets,
-                                  &fetch_targets, CreateVars);
+                                  &fetch_targets, true, CreateVars);
     } else {
       executor.Run(*inference_program, scope, &feed_targets, &fetch_targets,
-                   CreateVars);
+                   true, CreateVars);
     }
 
     // Enable the profiler
@@ -236,8 +236,7 @@ void TestInference(const std::string& dirname,
 
     // Disable the profiler and print the timing information
     paddle::platform::DisableProfiler(
-        paddle::platform::EventSortingKey::kDefault,
-        "run_inference_profiler.txt");
+        paddle::platform::EventSortingKey::kDefault, "run_inference_profiler");
     paddle::platform::ResetProfiler();
   }
 

paddle/fluid/operators/CMakeLists.txt

@@ -186,11 +186,7 @@ endif()
 
 add_subdirectory(detail)
 if(WITH_DISTRIBUTE)
-    if(WITH_GPU)
-        op_library(gen_nccl_id_op DEPS nccl_common)
-    else()
-        set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
-    endif()
+    
     set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf)
     set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
     op_library(send_op DEPS ${DISTRIBUTE_DEPS})
@@ -207,7 +203,13 @@ if(WITH_DISTRIBUTE)
     set_source_files_properties(send_barrier_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
     cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op listen_and_serv_op sum_op executor)
-    cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op listen_and_serv_op executor)
+    if(WITH_GPU)
+        cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS send_op listen_and_serv_op executor)
+        op_library(gen_nccl_id_op DEPS nccl_common sendrecvop_grpc)
+        set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+    else()
+        set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op)
+    endif()
 else()
     set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op gen_nccl_id_op)
 endif()

paddle/fluid/operators/detail/grpc_server.cc

@@ -184,7 +184,7 @@ class RequestPrefetch final : public RequestBase {
     framework::Scope* local_scope = &scope_->NewScope();
     auto* var = local_scope->FindVar(var_name);
     InitializeVariable(var, var_desc->GetType());
-    executor_->RunPreparedContext(prefetch_ctx_, scope_, false, false);
+    executor_->RunPreparedContext(prefetch_ctx_, scope_);
 
     SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply);
 

paddle/fluid/operators/listen_and_serv_op.cc

@@ -57,8 +57,7 @@ static void ParallelExecuteBlocks(
         framework::Async([&executor, &prepared, &program, &scope, idx]() {
           int run_block = idx;  // thread local
           try {
-            executor->RunPreparedContext(prepared[run_block].get(), scope,
-                                         false, false);
+            executor->RunPreparedContext(prepared[run_block].get(), scope);
           } catch (std::exception &e) {
             LOG(ERROR) << "run sub program error " << e.what();
           }
@@ -211,8 +210,8 @@ static void AsyncUpdateThread(
     }
     auto fs = framework::Async([var_name, &executor, &v, prepared] {
       try {
-        executor->RunPreparedContext(prepared, v.second->GetMutableLocalScope(),
-                                     false, false);
+        executor->RunPreparedContext(prepared,
+                                     v.second->GetMutableLocalScope());
       } catch (std::exception &e) {
         LOG(ERROR) << "run sub program error " << e.what();
       }

paddle/fluid/operators/roi_pool_op.cu

@@ -38,10 +38,10 @@ __global__ void GPUROIPoolForward(
   int index = blockIdx.x * blockDim.x + threadIdx.x;
   int offset = blockDim.x * gridDim.x;
   for (size_t i = index; i < nthreads; i += offset) {
-    int pw = index % pooled_width;
-    int ph = (index / pooled_width) % pooled_height;
-    int c = (index / pooled_width / pooled_height) % channels;
-    int n = index / pooled_width / pooled_height / channels;
+    int pw = i % pooled_width;
+    int ph = (i / pooled_width) % pooled_height;
+    int c = (i / pooled_width / pooled_height) % channels;
+    int n = i / pooled_width / pooled_height / channels;
 
     const int64_t* offset_input_rois = input_rois + n * kROISize;
     int roi_batch_ind = roi_batch_id_data[n];
@@ -52,14 +52,19 @@ __global__ void GPUROIPoolForward(
 
     int roi_width = max(roi_end_w - roi_start_w + 1, 1);
     int roi_height = max(roi_end_h - roi_start_h + 1, 1);
-    T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
-    T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
-
-    int hstart = static_cast<int>(floor(static_cast<T>(ph) * bin_size_h));
-    int wstart = static_cast<int>(floor(static_cast<T>(pw) * bin_size_w));
-    int hend = static_cast<int>(ceil(static_cast<T>(ph + 1) * bin_size_h));
-    int wend = static_cast<int>(ceil(static_cast<T>(pw + 1) * bin_size_w));
 
+    int hstart = static_cast<int>(floor(static_cast<double>(ph) *
+                                        static_cast<double>(roi_height) /
+                                        static_cast<double>(pooled_height)));
+    int wstart = static_cast<int>(floor(static_cast<double>(pw) *
+                                        static_cast<double>(roi_width) /
+                                        static_cast<double>(pooled_width)));
+    int hend = static_cast<int>(ceil(static_cast<double>(ph + 1) *
+                                     static_cast<double>(roi_height) /
+                                     static_cast<double>(pooled_height)));
+    int wend = static_cast<int>(ceil(static_cast<double>(pw + 1) *
+                                     static_cast<double>(roi_width) /
+                                     static_cast<double>(pooled_width)));
     hstart = min(max(hstart + roi_start_h, 0), height);
     hend = min(max(hend + roi_start_h, 0), height);
     wstart = min(max(wstart + roi_start_w, 0), width);
@@ -79,9 +84,9 @@ __global__ void GPUROIPoolForward(
         }
       }
     }
-    output_data[index] = maxval;
+    output_data[i] = maxval;
     if (argmax_data) {
-      argmax_data[index] = maxidx;
+      argmax_data[i] = maxidx;
     }
   }
 }
@@ -96,10 +101,10 @@ __global__ void GPUROIPoolBackward(
   int index = blockIdx.x * blockDim.x + threadIdx.x;
   int offset = blockDim.x * gridDim.x;
   for (int i = index; i < nthreads; i += offset) {
-    int pw = index % pooled_width;
-    int ph = (index / pooled_width) % pooled_height;
-    int c = (index / pooled_width / pooled_height) % channels;
-    int n = index / pooled_width / pooled_height / channels;
+    int pw = i % pooled_width;
+    int ph = (i / pooled_width) % pooled_height;
+    int c = (i / pooled_width / pooled_height) % channels;
+    int n = i / pooled_width / pooled_height / channels;
 
     int roi_batch_ind = roi_batch_id_data[n];
     int input_offset = (roi_batch_ind * channels + c) * height * width;
@@ -138,6 +143,7 @@ class GPUROIPoolOpKernel : public framework::OpKernel<T> {
     int width = in_dims[3];
 
     int rois_num = rois->dims()[0];
+
     if (rois_num == 0) return;
 
     int output_size = out->numel();

paddle/fluid/operators/send_recv_op_test.cc

@@ -92,12 +92,16 @@ void InitSelectedRowsInScope(const p::CPUPlace &place, f::Scope *scope) {
 
 void AddOp(const std::string &type, const f::VariableNameMap &inputs,
            const f::VariableNameMap &outputs, f::AttributeMap attrs,
-           f::BlockDesc *block) {
+           f::BlockDesc *block, bool is_sparse) {
   // insert output
   for (auto kv : outputs) {
     for (auto v : kv.second) {
       auto var = block->Var(v);
       var->SetDataType(f::proto::VarType::FP32);
+      var->SetPersistable(true);
+      if (is_sparse) {
+        var->SetType(f::proto::VarType::SELECTED_ROWS);
+      }
     }
   }
 
@@ -128,7 +132,8 @@ void StartServerNet(bool is_sparse, std::atomic<bool> *initialized) {
   auto *optimize_block = program.AppendBlock(root_block);
   auto *prefetch_block = program.AppendBlock(root_block);
   // X for server side tensors, RX for received tensors, must be of same shape.
-  AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block);
+  AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block,
+        is_sparse);
   f::AttributeMap attrs;
   attrs.insert({"endpoint", std::string("127.0.0.1:0")});
   attrs.insert({"Fanin", 1});

paddle/fluid/platform/profiler.cc

@@ -173,8 +173,9 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
 }
 
 RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
-    : start_ns_(PosixInNsec()) {
+    : is_enabled_(false), start_ns_(PosixInNsec()) {
   if (g_state == ProfilerState::kDisabled) return;
+  is_enabled_ = true;
   dev_ctx_ = dev_ctx;
   name_ = name;
   PushEvent(name_, dev_ctx_);
@@ -183,7 +184,7 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
 }
 
 RecordEvent::~RecordEvent() {
-  if (g_state == ProfilerState::kDisabled) return;
+  if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
   DeviceTracer* tracer = GetDeviceTracer();
   if (tracer) {
     tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(),
@@ -193,14 +194,16 @@ RecordEvent::~RecordEvent() {
   PopEvent(name_, dev_ctx_);
 }
 
-RecordBlock::RecordBlock(int block_id) : start_ns_(PosixInNsec()) {
+RecordBlock::RecordBlock(int block_id)
+    : is_enabled_(false), start_ns_(PosixInNsec()) {
   if (g_state == ProfilerState::kDisabled) return;
+  is_enabled_ = true;
   SetCurBlock(block_id);
   name_ = string::Sprintf("block_%d", block_id);
 }
 
 RecordBlock::~RecordBlock() {
-  if (g_state == ProfilerState::kDisabled) return;
+  if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
   DeviceTracer* tracer = GetDeviceTracer();
   if (tracer) {
     // We try to put all blocks at the same nested depth in the

paddle/fluid/platform/profiler.h

@@ -74,6 +74,7 @@ struct RecordEvent {
 
   ~RecordEvent();
 
+  bool is_enabled_;
   uint64_t start_ns_;
   // The device context is used by Event to get the current cuda stream.
   const DeviceContext* dev_ctx_;
@@ -89,6 +90,7 @@ struct RecordBlock {
   ~RecordBlock();
 
  private:
+  bool is_enabled_;
   std::string name_;
   uint64_t start_ns_;
 };

paddle/scripts/docker/build.sh

@@ -198,7 +198,7 @@ EOF
     # run paddle version to install python packages first
     RUN apt-get update &&\
         ${NCCL_DEPS}\
-        apt-get install -y wget python-pip dmidecode python-tk && pip install -U pip==9.0.3 && \
+        apt-get install -y wget python-pip dmidecode python-tk && easy_install -U pip && \
         pip install /*.whl; apt-get install -f -y && \
         apt-get clean -y && \
         rm -f /*.whl && \

paddle/scripts/paddle_build.sh

@@ -405,17 +405,19 @@ EOF
 
 function gen_dockerfile() {
     # Set BASE_IMAGE according to env variables
+    CUDA_MAJOR="$(echo $CUDA_VERSION | cut -d '.' -f 1).$(echo $CUDA_VERSION | cut -d '.' -f 2)"
+    CUDNN_MAJOR=$(echo $CUDNN_VERSION | cut -d '.' -f 1)
     if [[ ${WITH_GPU} == "ON" ]]; then
-    BASE_IMAGE="nvidia/cuda:8.0-cudnn5-runtime-ubuntu16.04"
+        BASE_IMAGE="nvidia/cuda:${CUDA_MAJOR}-cudnn${CUDNN_MAJOR}-runtime-ubuntu16.04"
     else
-    BASE_IMAGE="ubuntu:16.04"
+        BASE_IMAGE="ubuntu:16.04"
     fi
 
     DOCKERFILE_GPU_ENV=""
     DOCKERFILE_CUDNN_DSO=""
     if [[ ${WITH_GPU:-OFF} == 'ON' ]]; then
         DOCKERFILE_GPU_ENV="ENV LD_LIBRARY_PATH /usr/lib/x86_64-linux-gnu:\${LD_LIBRARY_PATH}"
-        DOCKERFILE_CUDNN_DSO="RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.5 /usr/lib/x86_64-linux-gnu/libcudnn.so"
+        DOCKERFILE_CUDNN_DSO="RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.${CUDNN_MAJOR} /usr/lib/x86_64-linux-gnu/libcudnn.so"
     fi
 
     cat <<EOF
@@ -449,7 +451,7 @@ EOF
     # run paddle version to install python packages first
     RUN apt-get update &&\
         ${NCCL_DEPS}\
-        apt-get install -y wget python-pip dmidecode python-tk && pip install -U pip==9.0.3 && \
+        apt-get install -y wget python-pip dmidecode python-tk && easy_install -U pip && \
         pip install /*.whl; apt-get install -f -y && \
         apt-get clean -y && \
         rm -f /*.whl && \
@@ -490,7 +492,7 @@ function gen_fluid_inference_lib() {
     Deploying fluid inference library ...
     ========================================
 EOF
-        make inference_lib_dist
+        make -j `nproc` inference_lib_dist
     fi
 }
 

python/paddle/fluid/inferencer.py

@@ -12,11 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import contextlib
+
 import core
 
 import executor
 import framework
 import io
+import parallel_executor
 import unique_name
 from trainer import check_and_get_place
 
@@ -24,40 +27,53 @@ __all__ = ['Inferencer', ]
 
 
 class Inferencer(object):
-    def __init__(self, infer_func, param_path, place=None):
+    def __init__(self, infer_func, param_path, place=None, parallel=False):
         """
         :param infer_func: a function that will return predict Variable
         :param param_path: the path where the inference model is saved by fluid.io.save_params
         :param place: place to do the inference
+        :param parallel: use parallel_executor to run the inference, it will use multi CPU/GPU.
         """
         self.param_path = param_path
         self.scope = core.Scope()
+        self.parallel = parallel
+        self.place = check_and_get_place(place)
 
         self.inference_program = framework.Program()
         with framework.program_guard(self.inference_program):
             with unique_name.guard():
                 self.predict_var = infer_func()
 
-        self.exe = executor.Executor(check_and_get_place(place))
-        with executor.scope_guard(self.scope):
+        with self._prog_and_scope_guard():
             # load params from param_path into scope
-            io.load_params(self.exe, param_path, self.inference_program)
+            io.load_params(executor.Executor(self.place), param_path)
+
+        if parallel:
+            with self._prog_and_scope_guard():
+                self.exe = parallel_executor.ParallelExecutor(
+                    use_cuda=isinstance(self.place, core.CUDAPlace),
+                    loss_name=self.predict_var.name)
+        else:
+            self.exe = executor.Executor(self.place)
 
-    def infer(self, inputs, return_numpy=True):
+    def infer(self, inputs):
         """
         :param inputs: a map of {"input_name": input_var} that will be feed into the inference program
         to get the predict value
-        :param return_numpy: if return numpy value for row tensor
         :return: the predict value of the inference model
         """
         if not isinstance(inputs, dict):
             raise ValueError(
                 "inputs should be a map of {'input_name': input_var}")
 
-        with executor.scope_guard(self.scope):
-            results = self.exe.run(self.inference_program,
-                                   feed=inputs,
-                                   fetch_list=[self.predict_var],
-                                   return_numpy=return_numpy)
+        with self._prog_and_scope_guard():
+            results = self.exe.run(feed=inputs,
+                                   fetch_list=[self.predict_var.name])
 
         return results
+
+    @contextlib.contextmanager
+    def _prog_and_scope_guard(self):
+        with framework.program_guard(main_program=self.inference_program):
+            with executor.scope_guard(self.scope):
+                yield

python/paddle/fluid/tests/book/high-level-api/CMakeLists.txt

@@ -8,3 +8,4 @@ endforeach()
 
 add_subdirectory(fit_a_line)
 add_subdirectory(recognize_digits)
+add_subdirectory(image_classification)

python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py

@@ -57,22 +57,20 @@ def train(use_cuda, train_program, save_dirname):
         optimizer=fluid.optimizer.SGD(learning_rate=0.001))
 
     def event_handler(event):
-        if isinstance(event, fluid.EndEpochEvent):
-            test_metrics = trainer.test(
-                reader=test_reader, feed_order=['x', 'y'])
-            print test_metrics
-            '''
-            
-            ...
-            ['25.768919467926025']
-            ['15.343549569447836']
-            ...
-            
-            '''
-            if float(test_metrics[0]) < 20.0:
+        if isinstance(event, fluid.EndStepEvent):
+            if event.step == 10:
+                test_metrics = trainer.test(
+                    reader=test_reader, feed_order=['x', 'y'])
+                print test_metrics
+                '''
+                ...
+                ['25.768919467926025']
+                ['15.343549569447836']
+                ...
+                '''
                 if save_dirname is not None:
                     trainer.save_params(save_dirname)
-            return
+                trainer.stop()
 
     trainer.train(
         reader=train_reader,
@@ -94,7 +92,7 @@ def infer(use_cuda, inference_program, save_dirname=None):
     tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32")
 
     results = inferencer.infer({'x': tensor_x})
-    print("infer results: ", results[0])
+    print("infer results: ", numpy.array(results[0]))
 
 
 def main(use_cuda):

python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt

+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+# default test
+foreach(src ${TEST_OPS})
+    py_test(${src} SRCS ${src}.py)
+endforeach()

python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+CIFAR dataset.
+
+This module will download dataset from
+https://www.cs.toronto.edu/~kriz/cifar.html and parse train/test set into
+paddle reader creators.
+
+The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes,
+with 6000 images per class. There are 50000 training images and 10000 test
+images.
+
+The CIFAR-100 dataset is just like the CIFAR-10, except it has 100 classes
+containing 600 images each. There are 500 training images and 100 testing
+images per class.
+
+"""
+
+import cPickle
+import itertools
+import numpy
+import paddle.v2.dataset.common
+import tarfile
+
+__all__ = ['train10']
+
+URL_PREFIX = 'https://www.cs.toronto.edu/~kriz/'
+CIFAR10_URL = URL_PREFIX + 'cifar-10-python.tar.gz'
+CIFAR10_MD5 = 'c58f30108f718f92721af3b95e74349a'
+
+
+def reader_creator(filename, sub_name, batch_size=None):
+    def read_batch(batch):
+        data = batch['data']
+        labels = batch.get('labels', batch.get('fine_labels', None))
+        assert labels is not None
+        for sample, label in itertools.izip(data, labels):
+            yield (sample / 255.0).astype(numpy.float32), int(label)
+
+    def reader():
+        with tarfile.open(filename, mode='r') as f:
+            names = (each_item.name for each_item in f
+                     if sub_name in each_item.name)
+
+            batch_count = 0
+            for name in names:
+                batch = cPickle.load(f.extractfile(name))
+                for item in read_batch(batch):
+                    if isinstance(batch_size, int) and batch_count > batch_size:
+                        break
+                    batch_count += 1
+                    yield item
+
+    return reader
+
+
+def train10(batch_size=None):
+    """
+    CIFAR-10 training set creator.
+
+    It returns a reader creator, each sample in the reader is image pixels in
+    [0, 1] and label in [0, 9].
+
+    :return: Training reader creator
+    :rtype: callable
+    """
+    return reader_creator(
+        paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5),
+        'data_batch',
+        batch_size=batch_size)

python/paddle/fluid/tests/book/high-level-api/image_classification/notest_image_classification_resnet.py → python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py

@@ -17,6 +17,7 @@ from __future__ import print_function
 import paddle
 import paddle.fluid as fluid
 import numpy
+import cifar10_small_test_set
 
 
 def resnet_cifar10(input, depth=32):
@@ -81,46 +82,50 @@ def train_network():
     cost = fluid.layers.cross_entropy(input=predict, label=label)
     avg_cost = fluid.layers.mean(cost)
     accuracy = fluid.layers.accuracy(input=predict, label=label)
-    return avg_cost, accuracy
+    return [avg_cost, accuracy]
 
 
-def train(use_cuda, save_path):
+def train(use_cuda, train_program, save_dirname):
     BATCH_SIZE = 128
     EPOCH_NUM = 1
 
     train_reader = paddle.batch(
         paddle.reader.shuffle(
-            paddle.dataset.cifar.train10(), buf_size=128 * 10),
+            cifar10_small_test_set.train10(batch_size=10), buf_size=128 * 10),
         batch_size=BATCH_SIZE)
 
     test_reader = paddle.batch(
         paddle.dataset.cifar.test10(), batch_size=BATCH_SIZE)
 
     def event_handler(event):
-        if isinstance(event, fluid.EndIteration):
-            if (event.batch_id % 10) == 0:
-                avg_cost, accuracy = trainer.test(reader=test_reader)
+        if isinstance(event, fluid.EndStepEvent):
+            avg_cost, accuracy = trainer.test(
+                reader=test_reader, feed_order=['pixel', 'label'])
 
-                print('BatchID {1:04}, Loss {2:2.2}, Acc {3:2.2}'.format(
-                    event.batch_id + 1, avg_cost, accuracy))
+            print('Loss {0:2.2}, Acc {1:2.2}'.format(avg_cost, accuracy))
 
-                if accuracy > 0.01:  # Low threshold for speeding up CI
-                    trainer.params.save(save_path)
-                    return
+            if accuracy > 0.01:  # Low threshold for speeding up CI
+                if save_dirname is not None:
+                    trainer.save_params(save_dirname)
+                return
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     trainer = fluid.Trainer(
-        train_network,
+        train_func=train_program,
         optimizer=fluid.optimizer.Adam(learning_rate=0.001),
-        place=place,
-        event_handler=event_handler)
-    trainer.train(train_reader, EPOCH_NUM, event_handler=event_handler)
+        place=place)
 
+    trainer.train(
+        reader=train_reader,
+        num_epochs=EPOCH_NUM,
+        event_handler=event_handler,
+        feed_order=['pixel', 'label'])
 
-def infer(use_cuda, save_path):
-    params = fluid.Params(save_path)
+
+def infer(use_cuda, inference_program, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    inferencer = fluid.Inferencer(inference_network, params, place=place)
+    inferencer = fluid.Inferencer(
+        infer_func=inference_program, param_path=save_dirname, place=place)
 
     # The input's dimension of conv should be 4-D or 5-D.
     # Use normilized image pixels as input data, which should be in the range
@@ -135,8 +140,14 @@ def main(use_cuda):
     if use_cuda and not fluid.core.is_compiled_with_cuda():
         return
     save_path = "image_classification_resnet.inference.model"
-    train(use_cuda, save_path)
-    infer(use_cuda, save_path)
+
+    train(
+        use_cuda=use_cuda, train_program=train_network, save_dirname=save_path)
+
+    infer(
+        use_cuda=use_cuda,
+        inference_program=inference_network,
+        save_dirname=save_path)
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py

@@ -71,24 +71,18 @@ def train(use_cuda, train_program, save_dirname):
         if isinstance(event, fluid.EndEpochEvent):
             test_reader = paddle.batch(
                 paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
-            test_metrics = trainer.test(
+            avg_cost, acc = trainer.test(
                 reader=test_reader, feed_order=['img', 'label'])
-            avg_cost_set = test_metrics[0]
-            acc_set = test_metrics[1]
-
-            # get test acc and loss
-            acc = numpy.array(acc_set).mean()
-            avg_cost = numpy.array(avg_cost_set).mean()
 
             print("avg_cost: %s" % avg_cost)
             print("acc     : %s" % acc)
 
-            if float(acc) > 0.2:  # Smaller value to increase CI speed
+            if acc > 0.2:  # Smaller value to increase CI speed
                 trainer.save_params(save_dirname)
             else:
                 print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
-                    event.epoch + 1, float(avg_cost), float(acc)))
-                if math.isnan(float(avg_cost)):
+                    event.epoch + 1, avg_cost, acc))
+                if math.isnan(avg_cost):
                     sys.exit("got NaN loss, training failed.")
         elif isinstance(event, fluid.EndStepEvent):
             print("Step {0}, Epoch {1} Metrics {2}".format(
@@ -118,7 +112,7 @@ def infer(use_cuda, inference_program, save_dirname=None):
 
     results = inferencer.infer({'img': tensor_img})
 
-    print("infer results: ", results[0])
+    print("infer results: ", numpy.array(results[0]))
 
 
 def main(use_cuda):

python/paddle/fluid/tests/unittests/test_dist_train.py

@@ -55,15 +55,18 @@ class TestSendOp(unittest.TestCase):
             serv = layers.ListenAndServ(
                 "127.0.0.1:0", ["X"], optimizer_mode=False)
             with serv.do():
+                out_var = main.global_block().create_var(
+                    name="scale_0.tmp_0",
+                    psersistable=True,
+                    dtype="float32",
+                    shape=[32, 32])
                 x = layers.data(
                     shape=[32, 32],
                     dtype='float32',
                     name="X",
                     append_batch_size=False)
                 fluid.initializer.Constant(value=1.0)(x, main.global_block())
-                o = layers.scale(x=x, scale=10.0)
-            main.global_block().create_var(
-                name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape)
+                layers.scale(x=x, scale=10.0, out=out_var)
 
         self.server_exe = fluid.Executor(place)
         self.server_exe.run(main)