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

Author: Yancey0623

benchmark/fluid/mnist.py

@@ -159,6 +159,7 @@ def run_benchmark(model, args):
         paddle.dataset.mnist.train(), batch_size=args.batch_size)
 
     accuracy = fluid.metrics.Accuracy()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
     iters, num_samples, start_time = 0, 0, time.time()
     for pass_id in range(args.pass_num):
         accuracy.reset()
@@ -175,17 +176,20 @@ def run_benchmark(model, args):
             y_data = np.array(map(lambda x: x[1], data)).astype("int64")
             y_data = y_data.reshape([len(y_data), 1])
 
-            outs = exe.run(
-                fluid.default_main_program(),
+            outs = train_exe.run(
                 feed={"pixel": img_data,
                       "label": y_data},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor]
+                fetch_list=[
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ]
             )  # The accuracy is the accumulation of batches, but not the current batch.
-            accuracy.update(value=outs[1], weight=outs[2])
+            accuracy.update(
+                value=np.array(np.mean(outs[1])),
+                weight=np.mean(np.array(outs[2])))
             iters += 1
             num_samples += len(y_data)
-            loss = np.array(outs[0])
-            acc = np.array(outs[1])
+            loss = np.mean(np.array(outs[0]))
+            acc = np.mean(np.array(outs[1]))
             train_losses.append(loss)
             train_accs.append(acc)
             print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %

benchmark/fluid/resnet.py

@@ -241,6 +241,7 @@ def test(exe):
     exe = fluid.Executor(place)
     exe.run(fluid.default_startup_program())
     accuracy = fluid.average.WeightedAverage()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
     if args.use_fake_data:
         data = train_reader().next()
         image = np.array(map(lambda x: x[0].reshape(dshape), data)).astype(
@@ -264,14 +265,17 @@ def test(exe):
                                      data)).astype('float32')
                 label = np.array(map(lambda x: x[1], data)).astype('int64')
                 label = label.reshape([-1, 1])
-            loss, acc, weight = exe.run(
-                fluid.default_main_program(),
+            loss, acc, weight = train_exe.run(
                 feed={'data': image,
                       'label': label},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor])
+                fetch_list=[
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ])
             iters += 1
             num_samples += len(label)
-            accuracy.add(value=acc, weight=weight)
+            accuracy.add(value=np.array(np.mean(acc)), weight=np.mean(weight))
+            loss = np.mean(np.array(loss))
+            acc = np.mean(np.array(acc))
             train_losses.append(loss)
             train_accs.append(acc)
             print("Pass: %d, Iter: %d, Loss: %f, Accuracy: %f" %

benchmark/fluid/vgg.py

@@ -169,6 +169,7 @@ def test(exe):
 
     iters, num_samples, start_time = 0, 0, time.time()
     accuracy = fluid.average.WeightedAverage()
+    train_exe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
     for pass_id in range(args.pass_num):
         accuracy.reset()
         train_accs = []
@@ -184,14 +185,17 @@ def test(exe):
             y_data = np.array(map(lambda x: x[1], data)).astype("int64")
             y_data = y_data.reshape([-1, 1])
 
-            loss, acc, weight = exe.run(
-                fluid.default_main_program(),
+            loss, acc, weight = train_exe.run(
                 feed={"pixel": img_data,
                       "label": y_data},
-                fetch_list=[avg_cost, batch_acc, batch_size_tensor])
-            accuracy.add(value=acc, weight=weight)
+                fetch_list=[
+                    avg_cost.name, batch_acc.name, batch_size_tensor.name
+                ])
+            accuracy.add(value=np.array(np.mean(acc)), weight=np.mean(weight))
             iters += 1
             num_samples += len(y_data)
+            loss = np.mean(np.array(loss))
+            acc = np.mean(np.array(acc))
             print(
                 "Pass = %d, Iter = %d, Loss = %f, Accuracy = %f" %
                 (pass_id, iters, loss, acc)

cmake/external/eigen.cmake

@@ -21,11 +21,12 @@ else()
     ExternalProject_Add(
         extern_eigen3
         ${EXTERNAL_PROJECT_LOG_ARGS}
-        GIT_REPOSITORY  "https://github.com/RLovelett/eigen.git"
+        GIT_REPOSITORY  "https://github.com/eigenteam/eigen-git-mirror"
         # eigen on cuda9.1 missing header of math_funtions.hpp
         # https://stackoverflow.com/questions/43113508/math-functions-hpp-not-found-when-using-cuda-with-eigen
         GIT_TAG         917060c364181f33a735dc023818d5a54f60e54c
         PREFIX          ${EIGEN_SOURCE_DIR}
+        DOWNLOAD_NAME   "eigen"
         UPDATE_COMMAND  ""
         CONFIGURE_COMMAND ""
         BUILD_COMMAND     ""

cmake/external/snappy.cmake

@@ -47,8 +47,6 @@ ExternalProject_Add(
                      -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPY_INSTALL_DIR}/lib
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
-    BUILD_COMMAND   make -j8
-    INSTALL_COMMAND make install
 )
 
 add_library(snappy STATIC IMPORTED GLOBAL)

cmake/external/snappystream.cmake

@@ -46,8 +46,6 @@ ExternalProject_Add(
                         -DCMAKE_INSTALL_PREFIX:PATH=${SNAPPYSTREAM_INSTALL_DIR}
                         -DCMAKE_INSTALL_LIBDIR:PATH=${SNAPPYSTREAM_INSTALL_DIR}/lib
                         -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
-        BUILD_COMMAND   make -j8
-        INSTALL_COMMAND make install
         DEPENDS snappy
 )
 

cmake/inference_lib.cmake

@@ -70,6 +70,12 @@ copy(glog_lib
   DSTS ${dst_dir} ${dst_dir}/lib
 )
 
+set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/boost/")
+copy(boost_lib
+  SRCS ${BOOST_INCLUDE_DIR}/boost
+  DSTS ${dst_dir}
+)
+
 if(NOT PROTOBUF_FOUND)
     set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf")
     copy(protobuf_lib
@@ -92,6 +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
@@ -142,4 +156,10 @@ copy(string_lib
   DSTS ${dst_dir}/${module} ${dst_dir}/${module}/tinyformat
 )
 
+set(module "pybind")
+copy(pybind_lib
+  SRCS ${CMAKE_CURRENT_BINARY_DIR}/paddle/fluid/${module}/pybind.h
+  DSTS ${dst_dir}/${module}
+)
+
 add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) 

doc/fluid/design/concepts/functions_operators_layers.md

@@ -40,7 +40,7 @@ template <typename T>
 class FCOp : public OperatorBase {
  public:
   void Run(...) {
-    add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b");
+    add(mul(Input<T>("X"), Input<T>("W")), Input<T>("b"));
   }
 };
 REGISTER_OP(FCOp, "fc");

doc/fluid/design/dist_train/async_update.md

@@ -4,34 +4,37 @@
 
 For the typical synchronous distributed training, some significant steps are as follows:
 
-1. A Trainer will compute the gradients and SEND them to the Parameter Server(PServer) nodes.
-1. After the PServer node received gradients came from all the Trainers, It will aggregate the
+1. A trainer process will compute the gradients and **send** them to the parameter server (PS) nodes.
+1. After the PS node received gradients came from all the Trainers, It will aggregate the
 gradient variables for the same parameter into one gradient variable and then apply the aggregated
 gradient to the respective parameter, finally using an optimize algorithms(SGD, Monument...)
 to update the parameters.
-1. The Trainer would wait for the PServers finished the optimize stage, and GET the parameters from PServer,
+1. The Trainer would wait for the PS finished the optimize stage, and GET the parameters from PS,
 so all the Trainers would get the same parameters.
 
-In the synchronously distributed training, there should be a `Barrier` to synchronise the
-parameters after the optimizing stage. The performance of a distributed training job would
-depend on the slowest node if there were hundreds or thousands of training nodes in a
-Job, the performance of synchronously distributed training might be very poor because of
-the slow node. So this design doc would introduce an approach to implement
-*asynchronously* distributed training in PaddlePaddle Fluid.
+In Synchronous Distributed Training, there is a **barrier** on each PS to wait until all trainers processes
+have completed running current mini-batch. After that, all trainers can continue to run the next
+mini-batch. So, we can find that the overall performance of Synchronous Distributed Training depends 
+on the slowest node.
+
+In Asynchronous Distributed Training, we don't need to wait for a global mini-bach, the optimizer on
+the PS will run immediately when the gradient is uploaded to the PS from one trainer. This mode would
+train such models that achieve scaling, better throughput. In this design doc, we will introduce how to 
+implement the Asynchronous Distributed Training base on PaddlePaddle Fluid.
 
 ## Design
 
 <img src="./src/async_update.png" width="600"/>
 
-As the figure above, we describe a global view of asynchronously update process and use
+As the figure above, we describe a global view of the asynchronous update process and use
 the parameter `w1` as an example to introduce the steps:
 1. For each gradient variables, they may distribute on different GPU card and aggregate
 them while they are all calculated.
-1. Split the gradient variable into multiple blocks according to the number of PServer
+1. Split the gradient variable into multiple blocks according to the number of PS
 instances and then send them.
-1. PServer would run an `Optimize Block` using a specified optimize algorithm to update
+1. PS would run an `Optimize Block` using a specified optimize algorithm to update
 the specified parameter.
-1. The trainer will fetch latest parameter from PServer before running forward Op which depends
+1. The trainer will fetch the latest parameter from PS before running forward Op which depends
 on the specified parameter.
 1. Broadcast the received variable into multiple GPU cards and continue to run the next
 mini-batch.
@@ -40,8 +43,8 @@ mini-batch.
 
 - For the multiple devices distributed training, we need to aggregate the gradient
 variables which placed on different devices firstly and then schedule a `SendVars` Operator to
-send the gradient variables to the multiple PServer instances.
-- Schedule `FetchVars` operator to fetch the latest parameter from PServer before running
+send the gradient variables to the multiple PS instances.
+- Schedule `FetchVars` operator to fetch the latest parameter from PS before running
 the forward ops.
 - There could be a large number of gradient variables to be sent, so we need to use another
 thread pool(IO Threadpool) whose a number of the schedulable threads is larger than the

paddle/CMakeLists.txt

@@ -24,6 +24,6 @@ if(NOT WITH_FLUID_ONLY)
 endif()
 
 add_subdirectory(testing)
-if(NOT MOBILE_INFERENCE AND NOT RPI)
+if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API)
   add_subdirectory(fluid)
 endif()