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

JiayiFeng · JiayiFeng · commit 7834b4a47069 · 2018-08-07T14:10:29.000+08:00
diff --git a/.travis.yml b/.travis.yml
@@ -27,15 +27,6 @@ script:
     # 43min timeout
     paddle/scripts/paddle_docker_build.sh ${JOB}
     if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi;
-  - |
-    if [[ "$JOB" != "doc" ]]; then exit 0; fi;
-    # For document only
-    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
-    if [[ "$TRAVIS_BRANCH" != "develop"  && ! "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
-    export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh
-    export DOCS_DIR=`pwd`
-    cd ..
-    curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/
 notifications:
   email:
     on_success: change
diff --git a/cmake/generic.cmake b/cmake/generic.cmake
@@ -265,6 +265,7 @@ function(cc_test TARGET_NAME)
     if (${cc_test_SERIAL})
         set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
     endif()
   endif()
 endfunction(cc_test)
@@ -330,6 +331,7 @@ function(nv_test TARGET_NAME)
     if (nv_test_SERIAL)
         set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
     endif()
   endif()
 endfunction(nv_test)
@@ -577,7 +579,8 @@ function(py_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS ARGS ENVS)
     cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_test(NAME ${TARGET_NAME}
-             COMMAND env FLAGS_init_allocated_mem=true PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
+             COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
+             PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
              ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
   endif()
diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
@@ -136,6 +136,8 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
     platform::SetDeviceId(dev_id);
 #endif
   }
+  platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
+  platform::RecordEvent record_event(Type(), pool.Get(place));
   RunImpl(scope, place);
   VLOG(10) << "+ " << DebugStringEx(&scope);
 }
@@ -639,9 +641,6 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto* dev_ctx = pool.Get(place);
 
-  // For profiling, don't move out of this function because that will result
-  // in the failure of multi-GPU profiling.
-  platform::RecordEvent record_event(Type(), dev_ctx);
   // check if op[type] has kernel registered.
   auto& all_op_kernels = AllOpKernels();
   auto kernels_iter = all_op_kernels.find(type_);
diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt
@@ -74,9 +74,10 @@ if (WITH_ANAKIN) # only needed in CI
     target_link_libraries(inference_anakin_api anakin anakin_saber_common)
     target_link_libraries(inference_anakin_api_shared anakin anakin_saber_common)
     if (WITH_TESTING)
-        cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc
-                                  ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
-                                  DEPS inference_anakin_api_shared)
-        target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
+        # this test is unstable, disable it first.
+        #cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc
+                                  #ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
+                                  #DEPS inference_anakin_api_shared)
+        #target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
      endif(WITH_TESTING)
 endif()
diff --git a/paddle/fluid/operators/feed_op.cc b/paddle/fluid/operators/feed_op.cc
@@ -31,7 +31,6 @@ class FeedOp : public framework::OperatorBase {
                const platform::Place &place) const override {
     // get device context from pool
     auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-    platform::RecordEvent record_event(Type(), dev_ctx);
 
     auto feed_var_name = Input("X");
     auto *feed_var = scope.FindVar(feed_var_name);
diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc
@@ -36,12 +36,6 @@ class FetchBarrierOp : public framework::OperatorBase {
   void RunImpl(const framework::Scope& scope,
                const platform::Place& place) const override {
     std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
-
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
diff --git a/paddle/fluid/operators/fetch_op.cc b/paddle/fluid/operators/fetch_op.cc
@@ -30,9 +30,6 @@ class FetchOp : public framework::OperatorBase {
  private:
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
-    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
-    platform::RecordEvent record_event(Type(), pool.Get(place));
-
     auto fetch_var_name = Input("X");
     auto *fetch_var = scope.FindVar(fetch_var_name);
     PADDLE_ENFORCE(fetch_var != nullptr,
diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc
@@ -31,9 +31,6 @@ class LoadOp : public framework::OperatorBase {
  private:
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
-    auto *dev_ctx = platform::DeviceContextPool::Instance().Get(place);
-    platform::RecordEvent record_event(Type(), dev_ctx);
-
     // FIXME(yuyang18): We save variable to local file now, but we should change
     // it to save an output stream.
     auto filename = Attr<std::string>("file_path");
diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc
@@ -40,8 +40,6 @@ class RecvOp : public framework::OperatorBase {
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
 
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc
@@ -39,11 +39,6 @@ class SendBarrierOp : public framework::OperatorBase {
     std::vector<std::string> eps = Attr<std::vector<std::string>>("endpoints");
     bool sync_mode = Attr<bool>("sync_mode");
 
-    platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
-    auto& ctx = *pool.Get(place);
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc
@@ -42,9 +42,6 @@ class SendOp : public framework::OperatorBase {
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
 
-    // For profiling
-    platform::RecordEvent record_event(Type(), &ctx);
-
     distributed::RPCClient* rpc_client =
         distributed::RPCClient::GetInstance<RPCCLIENT_T>();
 
diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc
@@ -110,6 +110,8 @@ Event::Event(EventType type, std::string name, uint32_t thread_id,
   has_cuda_ = dev_ctx ? platform::is_gpu_place(dev_ctx->GetPlace()) : false;
   if (has_cuda_) {
     auto* cuda_dev_ctx = static_cast<const CUDADeviceContext*>(dev_ctx);
+    PADDLE_ENFORCE(cudaSetDevice(
+        boost::get<platform::CUDAPlace>(cuda_dev_ctx->GetPlace()).device));
     PADDLE_ENFORCE(cudaGetDevice(&device_));
     PADDLE_ENFORCE(cudaEventCreate(&event_));
     auto stream = cuda_dev_ctx->stream();
@@ -176,6 +178,7 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
 
 RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
     : is_enabled_(false), start_ns_(PosixInNsec()) {
+  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled) return;
   is_enabled_ = true;
   dev_ctx_ = dev_ctx;
@@ -186,6 +189,7 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
 }
 
 RecordEvent::~RecordEvent() {
+  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
   DeviceTracer* tracer = GetDeviceTracer();
   if (tracer) {
@@ -198,13 +202,15 @@ RecordEvent::~RecordEvent() {
 
 RecordBlock::RecordBlock(int block_id)
     : is_enabled_(false), start_ns_(PosixInNsec()) {
+  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled) return;
   is_enabled_ = true;
   SetCurBlock(block_id);
   name_ = string::Sprintf("block_%d", block_id);
 }
 
 RecordBlock::~RecordBlock() {
+  std::lock_guard<std::mutex> l(profiler_mu);
   if (g_state == ProfilerState::kDisabled || !is_enabled_) return;
   DeviceTracer* tracer = GetDeviceTracer();
   if (tracer) {
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
@@ -419,6 +419,25 @@ EOF
     linkchecker doc/v2/en/html/index.html
     linkchecker doc/v2/cn/html/index.html
     linkchecker doc/v2/api/en/html/index.html
+
+    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
+
+    # Deploy to the the content server if its a "develop" or "release/version" branch
+    # The "develop_doc" branch is reserved to test full deploy process without impacting the real content.
+    if [ "$TRAVIS_BRANCH" == "develop_doc" ]; then
+        PPO_SCRIPT_BRANCH=develop
+    elif [[ "$TRAVIS_BRANCH" == "develop"  ||  "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then
+        PPO_SCRIPT_BRANCH=master
+    else
+        # Early exit, this branch doesn't require documentation build
+        return 0;
+    fi
+     # Fetch the paddlepaddle.org deploy_docs.sh from the appopriate branch
+    export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/$PPO_SCRIPT_BRANCH/scripts/deploy/deploy_docs.sh
+    export PYTHONPATH=$PYTHONPATH:${PADDLE_ROOT}/build/python:/paddle/build/python
+    cd ..
+    curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH ${PADDLE_ROOT} ${PADDLE_ROOT}/build/doc/ ${PPO_SCRIPT_BRANCH}
+    cd -
 }
 
 function gen_html() {
diff --git a/paddle/scripts/paddle_docker_build.sh b/paddle/scripts/paddle_docker_build.sh
@@ -52,6 +52,9 @@ EOL
     ${DOCKER_CMD} run -it \
         ${DOCKER_ENV} \
         -e SCRIPT_NAME=$0 \
+        -e CONTENT_DEC_PASSWD=$CONTENT_DEC_PASSWD \
+        -e TRAVIS_BRANCH=$TRAVIS_BRANCH \
+        -e TRAVIS_PULL_REQUEST=$TRAVIS_PULL_REQUEST \
         -v $PADDLE_ROOT:/paddle \
         -v ${HOME}/.ccache:/root/.ccache \
         -w /paddle \
diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py
@@ -572,8 +572,6 @@ def append_backward(loss, parameter_list=None, no_grad_set=None,
 
     program.current_block_idx = current_block_idx
     program._sync_with_cpp()
-    # FIXME(zcd): prevent loss.grad optimized by mem_opt.
-    loss.block.var(_append_grad_suffix_(loss.name)).persistable = True
 
     if parameter_list is not None:
         parameters = parameter_list
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
@@ -1038,7 +1038,26 @@ def create_parameter(self, *args, **kwargs):
         global_block = self.program.global_block()
         param = Parameter(global_block, *args, **kwargs)
         if 'initializer' in kwargs:
-            kwargs['initializer'](param, self)
+
+            def _is_inited_by(block, var):
+                init_ops = []
+                for op in block.ops:
+                    if var.name in op.output_arg_names:
+                        init_ops.append(op)
+                return init_ops
+
+            initializer = kwargs['initializer']
+            init_ops = _is_inited_by(global_block, param)
+            init_ops_len = len(init_ops)
+            if init_ops_len > 1:
+                raise RuntimeError("param " + param.name +
+                                   " is inited by multiple init ops " + str(
+                                       init_ops))
+            elif init_ops_len == 1:
+                #TODO already inited, do nothing, should log a warning
+                pass
+            else:
+                initializer(param, self)
         return param
 
     def append_op(self, *args, **kwargs):
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
@@ -949,6 +949,10 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
     helper = LayerHelper('dropout', **locals())
     out = helper.create_tmp_variable(dtype=x.dtype)
     mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True)
+
+    if (seed is None or seed == 0) and helper.main_program.random_seed != 0:
+        seed = helper.main_program.random_seed
+
     helper.append_op(
         type='dropout',
         inputs={'X': [x]},
diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py
@@ -73,9 +73,18 @@ def _transpiler_instance(self, config=None):
 
         return self.transpiler
 
+    def transpiler_test_impl(self):
+        pass
 
-class TestBasicModel(TranspilerTest):
     def test_transpiler(self):
+        main = fluid.Program()
+        startup = fluid.Program()
+        with fluid.program_guard(main, startup):
+            self.transpiler_test_impl()
+
+
+class TestBasicModel(TranspilerTest):
+    def transpiler_test_impl(self):
         pserver, startup = self.get_pserver(self.pserver1_ep)
         pserver2, startup2 = self.get_pserver(self.pserver2_ep)
 
@@ -123,7 +132,7 @@ def test_transpiler(self):
 
 
 class TestBasicModelWithLargeBlockSize(TranspilerTest):
-    def test_transpiler(self):
+    def transpiler_test_impl(self):
         config = fluid.DistributeTranspilerConfig()
         config.min_block_size = 1048576
 
@@ -148,7 +157,7 @@ def test_transpiler(self):
                          ["sum", "scale", "sgd"])
         # confirm startup program
         self.assertEqual([op.type for op in startup.global_block().ops],
-                         ["fill_constant", "fill_constant", "fill_constant"])
+                         ["fill_constant", "fill_constant"])
         # the variable #fc_w will be split into two blocks
         fc_w_var = startup2.global_block().var("fc_w")
         self.assertEqual(fc_w_var.shape, (1000L, 1000L))
@@ -177,7 +186,7 @@ class TestNoSliceVar(TranspilerTest):
     def setUp(self):
         super(TestNoSliceVar, self).setUp()
 
-    def test_transpiler(self):
+    def transpiler_test_impl(self):
         config = fluid.DistributeTranspilerConfig()
         config.slice_var_up = False
 
@@ -212,7 +221,7 @@ def net_conf(self):
         sgd_optimizer.minimize(avg_cost)
         return
 
-    def test_transpiler(self):
+    def transpiler_test_impl(self):
         pserver, startup = self.get_pserver(self.pserver1_ep)
         trainer = self.get_trainer()
 
@@ -242,7 +251,7 @@ def net_conf(self):
         sgd_optimizer.minimize(avg_cost)
         return
 
-    def test_transpiler(self):
+    def transpiler_test_impl(self):
         pserver, startup = self.get_pserver(self.pserver1_ep)
         trainer = self.get_trainer()
 
@@ -291,7 +300,7 @@ def net_conf(self):
         sgd_optimizer.minimize(avg_cost)
         return
 
-    def test_transpiler(self):
+    def transpiler_test_impl(self):
         pserver, startup = self.get_pserver(self.pserver1_ep)
         trainer = self.get_trainer()
 
@@ -326,7 +335,7 @@ def net_conf(self):
         sgd_optimizer.minimize(avg_cost)
         return
 
-    def test_transpiler(self):
+    def transpiler_test_impl(self):
         pserver, startup = self.get_pserver(self.pserver1_ep)
         trainer = self.get_trainer()
 
diff --git a/python/paddle/fluid/tests/unittests/test_initializer.py b/python/paddle/fluid/tests/unittests/test_initializer.py
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py
diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py
