Merge pull request #10 from PaddlePaddle/develop

merge to local

Author: junjun315

README.md

@@ -3,8 +3,8 @@
 English | [简体中文](./README_cn.md)
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
@@ -18,17 +18,17 @@ learning to many products at Baidu.
 Our vision is to enable deep learning for everyone via PaddlePaddle.
 Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
 
-### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
+### Latest PaddlePaddle Release: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3)
 ### Install Latest Stable Release:
 ```
 # Linux CPU
 pip install paddlepaddle
 # Linux GPU cuda9cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.2.0.post87
+pip install paddlepaddle-gpu==1.3.0.post87
 # Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.2.0.post85
+pip install paddlepaddle-gpu==1.3.0.post85
 
 # For installation on other platform, refer to http://paddlepaddle.org/
 ```
@@ -75,26 +75,26 @@ pip install paddlepaddle-gpu==1.2.0.post85
 
 ## Installation
 
-It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
+It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) on our website.
 
 ## Documentation
 
-We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
-[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation.
+We provide [English](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) and
+[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) documentation.
 
 - [Deep Learning 101](https://github.com/PaddlePaddle/book)
 
   You might want to start from this online interactive book that can run in a Jupyter Notebook.
 
-- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
+- [Distributed Training](http://paddlepaddle.org/documentation/docs/en/1.3/user_guides/howto/training/multi_node_en.html)
 
   You can run distributed training jobs on MPI clusters.
 
-- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
+- [Python API](http://paddlepaddle.org/documentation/docs/en/1.3/api/index_en.html)
 
    Our new API enables much shorter programs.
 
-- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
+- [How to Contribute](http://paddlepaddle.org/documentation/docs/en/1.3/advanced_usage/development/contribute_to_paddle/index_en.html)
 
    We appreciate your contributions!
 

README_cn.md

@@ -3,8 +3,8 @@
 [English](./README.md) | 简体中文
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
 
@@ -16,17 +16,17 @@ PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效
 
 跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
 
-### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
+### PaddlePaddle最新版本: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3)
 ### 安装最新稳定版本:
 ```
 # Linux CPU
 pip install paddlepaddle
 # Linux GPU cuda9cudnn7
 pip install paddlepaddle-gpu
 # Linux GPU cuda8cudnn7
-pip install paddlepaddle-gpu==1.2.0.post87
+pip install paddlepaddle-gpu==1.3.0.post87
 # Linux GPU cuda8cudnn5
-pip install paddlepaddle-gpu==1.2.0.post85
+pip install paddlepaddle-gpu==1.3.0.post85
 
 # 其他平台上的安装指引请参考 http://paddlepaddle.org/
 ```
@@ -57,26 +57,26 @@ pip install paddlepaddle-gpu==1.2.0.post85
 
 ## 安装
 
-推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html)
+推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/install/index_cn.html)
 
 ## 文档
 
-我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
-[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
+我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)和
+[中文](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) 文档
 
 - [深度学习101](https://github.com/PaddlePaddle/book)
 
   或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
 
-- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
+- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.3/user_guides/howto/training/multi_node.html)
 
   可以在MPI集群上运行分布式训练任务
 
-- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
+- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.3/api_cn/index_cn.html)
 
    新的API支持代码更少更简洁的程序
 
-- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
+- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.3/advanced_usage/development/contribute_to_paddle/index_cn.html)
 
    欢迎您的贡献!
 

cmake/external/mklml.cmake

@@ -39,10 +39,8 @@ IF(WIN32)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5md.lib)
     SET(MKLML_SHARED_LIB          ${MKLML_LIB_DIR}/mklml.dll)
     SET(MKLML_SHARED_IOMP_LIB     ${MKLML_LIB_DIR}/libiomp5md.dll)
-ELSE()
-    #TODO(intel-huying):
-    #  Now enable Erf function in mklml library temporarily, it will be updated as offical version later.
-    SET(MKLML_VER "VsErf_mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
+ELSE()  
+    SET(MKLML_VER "mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
     SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
     SET(MKLML_LIB                 ${MKLML_LIB_DIR}/libmklml_intel.so)
     SET(MKLML_IOMP_LIB            ${MKLML_LIB_DIR}/libiomp5.so)

paddle/fluid/API.spec

@@ -71,7 +71,7 @@ paddle.fluid.initializer.NumpyArrayInitializer.__init__ ArgSpec(args=['self', 'v
 paddle.fluid.layers.fc ArgSpec(args=['input', 'size', 'num_flatten_dims', 'param_attr', 'bias_attr', 'act', 'is_test', 'name'], varargs=None, keywords=None, defaults=(1, None, None, None, False, None))
 paddle.fluid.layers.embedding ArgSpec(args=['input', 'size', 'is_sparse', 'is_distributed', 'padding_idx', 'param_attr', 'dtype'], varargs=None, keywords=None, defaults=(False, False, None, None, 'float32'))
 paddle.fluid.layers.dynamic_lstm ArgSpec(args=['input', 'size', 'h_0', 'c_0', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'float32', None))
-paddle.fluid.layers.dynamic_lstmp ArgSpec(args=['input', 'size', 'proj_size', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'proj_activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'tanh', 'float32', None))
+paddle.fluid.layers.dynamic_lstmp ArgSpec(args=['input', 'size', 'proj_size', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'proj_activation', 'dtype', 'name', 'h_0', 'c_0', 'cell_clip', 'proj_clip'], varargs=None, keywords=None, defaults=(None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'tanh', 'float32', None, None, None, None, None))
 paddle.fluid.layers.dynamic_gru ArgSpec(args=['input', 'size', 'param_attr', 'bias_attr', 'is_reverse', 'gate_activation', 'candidate_activation', 'h_0', 'origin_mode'], varargs=None, keywords=None, defaults=(None, None, False, 'sigmoid', 'tanh', None, False))
 paddle.fluid.layers.gru_unit ArgSpec(args=['input', 'hidden', 'size', 'param_attr', 'bias_attr', 'activation', 'gate_activation', 'origin_mode'], varargs=None, keywords=None, defaults=(None, None, 'tanh', 'sigmoid', False))
 paddle.fluid.layers.linear_chain_crf ArgSpec(args=['input', 'label', 'param_attr'], varargs=None, keywords=None, defaults=(None,))
@@ -121,6 +121,7 @@ paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=
 paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
 paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0, False))
+paddle.fluid.layers.sampled_softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'num_samples', 'num_true', 'remove_accidental_hits', 'use_customized_samples', 'customized_samples', 'customized_probabilities', 'seed'], varargs=None, keywords=None, defaults=(1, True, False, None, None, 0))
 paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name', 'path_table', 'path_code', 'is_custom', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, False, False))
 paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'is_accumulated', 'name', 'return_parent_idx'], varargs=None, keywords=None, defaults=(0, True, None, False))
 paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))

paddle/fluid/framework/details/build_strategy.cc

@@ -135,12 +135,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
   void AppendMultiDevPass(const BuildStrategy &strategy) {
     ir::Pass *multi_devices_pass;
     if (strategy_.is_distribution_) {
+      VLOG(3) << "multi device parameter server mode";
       multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
     } else {
       if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
+        VLOG(3) << "multi devices collective mode with allreduce";
         multi_devices_pass =
             AppendPass("allreduce_mode_multi_devices_pass").get();
       } else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
+        VLOG(3) << "multi deivces collective mode with reduce";
         multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
       } else {
         PADDLE_THROW("Unknown reduce strategy.");

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -937,9 +937,21 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
 }
 
 void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
-  if (need_broadcast_var_ ||
-      (UseGPU() &&
-       strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce)) {
+  // broad cast received parameters when training in parameter server mode.
+  if (need_broadcast_var_) {
+    // There are 4 conditions:
+    // 1. GPU && Reduce: Reduce gradient then broadcast gradient to other GPUS.
+    // Need to broadcast received parameters to other GPU.
+    // 2. GPU && AllReduce: AllReduce all graident to each GPU. Need to
+    // broadcast received parameters to other GPU.
+    // 3. CPU && AllReduce: AllReduce all gradient to each thread. Need to
+    // broadcast received parameters to other scope.
+    // 4. CPU && Reduce: because all parameters share the same memory, did not
+    // broadcast received parameters.
+    if (!UseGPU() &&
+        strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
+      return;
+    }
     if (strategy_.fuse_broadcast_op_) {
       CreateFusedBroadcastOp(result, bcast_var_name_set_);
     } else {

paddle/fluid/operators/CMakeLists.txt

@@ -66,7 +66,7 @@ set(COMMON_OP_DEPS ${OP_HEADER_DEPS})
 
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows lod_tensor maxouting unpooling pooling lod_rank_table context_project sequence_pooling executor)
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc)
-set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler tree2col)
+set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler sample_prob tree2col)
 set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions beam_search)
 if (WITH_GPU)
   set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv prelu)

paddle/fluid/operators/activation_op.h

@@ -11,7 +11,6 @@ limitations under the License. */
 
 #pragma once
 #include <glog/logging.h>
-#include <algorithm>
 #include <string>
 #include <unordered_set>
 #include <utility>
@@ -25,7 +24,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/eigen.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
-#include "paddle/fluid/operators/math/blas.h"
 #include "paddle/fluid/platform/float16.h"
 
 #ifdef PADDLE_WITH_MKLDNN
@@ -303,28 +301,8 @@ template <typename T>
 struct GeluFunctor : public BaseActivationFunctor<T> {
   template <typename Device, typename X, typename Out>
   void operator()(Device d, X x, Out out) const {
-// Because the execute or device context can not be deliver here, it keep the
-// marco for NVCC.
-#if defined(PADDLE_WITH_MKLML) && !defined(_WIN32) && !defined(__APPLE__) && \
-    !defined(__OSX__) && !defined(PADDLE_WITH_CUDA)
-    auto x_data = x.data();
-    auto out_data = out.data();
-    int n = std::min(x.size(), out.size());
-
-    std::memset(out_data, 0, n * sizeof(T));
-    math::CBlas<T>::AXPY(n, static_cast<T>(M_SQRT1_2), x_data, 1, out_data, 1);
-    math::CBlas<T>::VMERF(n, out_data, out_data, VML_LA);
-    for (int i = 0; i < n; i++) {
-      out_data[i] += static_cast<T>(1);
-    }
-    math::CBlas<T>::VMUL(n, x_data, out_data, out_data);
-    for (int i = 0; i < n; i++) {
-      out_data[i] *= static_cast<T>(0.5);
-    }
-#else
     auto temp = (x * static_cast<T>(M_SQRT1_2)).erf();
     out.device(d) = x * static_cast<T>(0.5) * (static_cast<T>(1) + temp);
-#endif
   }
 };
 

paddle/fluid/operators/beam_search_decode_op.h

@@ -122,7 +122,7 @@ void BeamSearchDecoder<T>::ConvertSentenceVectorToLodTensor(
 
   auto cpu_place = std::unique_ptr<paddle::platform::CPUPlace>(
       new paddle::platform::CPUPlace());
-  paddle::platform::CPUDeviceContext cpu_ctx(*cpu_place.get());
+  paddle::platform::CPUDeviceContext cpu_ctx(*cpu_place);
 
   framework::LoD lod;
   lod.push_back(source_level_lod);