Merge pull request #1615 from luotao1/ml

remove with_metric_learning compile option

cmake/util.cmake

@@ -71,21 +71,10 @@ function(link_paddle_exe TARGET_NAME)
         generate_rdma_links()
     endif()
 
-    if(WITH_METRIC)
-        if(WITH_GPU)
-            set(METRIC_LIBS paddle_metric_learning paddle_dserver_lib metric metric_cpu)
-        else()
-            set(METRIC_LIBS paddle_metric_learning paddle_dserver_lib metric_cpu)
-        endif()
-    else()
-        set(METRIC_LIBS "")
-    endif()
-
     target_circle_link_libraries(${TARGET_NAME}
         ARCHIVE_START
         paddle_gserver
         paddle_function
-        ${METRIC_LIBS}
         ARCHIVE_END
         paddle_pserver
         paddle_trainer_lib
@@ -95,7 +84,6 @@ function(link_paddle_exe TARGET_NAME)
         paddle_parameter
         paddle_proto
         paddle_cuda
-        ${METRIC_LIBS}
         ${EXTERNAL_LIBS}
         ${CMAKE_THREAD_LIBS_INIT}
         ${CMAKE_DL_LIBS}

doc/howto/usage/cmd_parameter/arguments_cn.md

@@ -228,16 +228,6 @@
 <td class="left"></td><td class="left"></td><td class="left">√</td><td class="left">√</td>
 </tr>
 
-<tr>
-<td class="left" rowspan = "2">度量学习(metric learning)</td><td class="left">external</td>
-<td class="left">√</td><td class="left">√</td><td class="left">√</td><td class="left">√</td>
-</tr>
-
-<tr>
-<td class="left">data_server_port</td>
-<td class="left"></td><td class="left">√</td><td class="left"></td><td class="left">√</td>
-</tr>
-
 <tr>
 <td class="left" rowspan = "16">参数服务器(PServer)</td><td class="left">start_pserver</td>
 <td class="left"></td><td class="left">√</td><td class="left"></td><td class="left">√</td>

doc/howto/usage/cmd_parameter/arguments_en.md

@@ -228,16 +228,6 @@ It looks like there are a lot of arguments. However, most of them are for develo
 <td class="left"></td><td class="left"></td><td class="left">√</td><td class="left">√</td>
 </tr>
 
-<tr>
-<td class="left" rowspan = "2">metric learning</td><td class="left">external</td>
-<td class="left">√</td><td class="left">√</td><td class="left">√</td><td class="left">√</td>
-</tr>
-
-<tr>
-<td class="left">data_server_port</td>
-<td class="left"></td><td class="left">√</td><td class="left"></td><td class="left">√</td>
-</tr>
-
 <tr>
 <td class="left" rowspan = "16">PServer</td><td class="left">start_pserver</td>
 <td class="left"></td><td class="left">√</td><td class="left"></td><td class="left">√</td>

doc/howto/usage/cmd_parameter/detail_introduction_cn.md

@@ -180,15 +180,6 @@
   - 用户可以自定义beam search的方法，编译成动态库，供PaddlePaddle加载。 该参数用于指定动态库路径.
   - 类型: string (默认: "", null).
 
-## 度量学习(Metric Learning)
-* `--external`
-   - 指示是否使用外部机器进行度量学习.
-   - 类型: bool (默认: 0).
-
-* `--data_server_port`
-  - 数据服务器(data server)的监听端口，主要用在度量学习中.
-  - 类型: int32 (默认: 21134).
-
 ## 数据支持(DataProvider)
 
 * `--memory_threshold_on_load_data`

doc/howto/usage/cmd_parameter/detail_introduction_en.md

@@ -184,15 +184,6 @@
   - Specify shared dynamic library. It can be defined out of paddle by user.
   - type: string (default: "", null).
 
-## Metric Learning
-* `--external`
-   - Whether to use external machine for metric learning.
-   - type: bool (default: 0).
-
-* `--data_server_port`
-  - Listening port for dserver (data server), dserver is mainly used in metric learning.
-  - type: int32 (default: 21134).
-
 ## DataProvider
 
 * `--memory_threshold_on_load_data`

paddle/gserver/gradientmachines/MultiGradientMachine.cpp

@@ -24,9 +24,6 @@ limitations under the License. */
 DEFINE_bool(allow_only_one_model_on_one_gpu,
             true,
             "If true, do not allow multiple models on one GPU device");
-#ifdef PADDLE_METRIC_LEARNING
-DECLARE_bool(external);
-#endif
 
 namespace paddle {
 
@@ -45,11 +42,7 @@ MultiGradientMachine::MultiGradientMachine(const ModelConfig& config,
       trainerBarrier_(FLAGS_trainer_count),
       allBarrier_(FLAGS_trainer_count + 1),
       inArgsCopied_(false) {
-#ifdef PADDLE_METRIC_LEARNING
-  isPassGrad_ = FLAGS_external;
-#else
   isPassGrad_ = false;
-#endif
   numThreads_ = FLAGS_trainer_count;
   if (useGpu) {
     //! TODO(yuyang18): When useGpu=false && paddle is not compiled with gpu,

paddle/gserver/layers/Layer.cpp

@@ -381,8 +381,7 @@ void Layer::backwardActivation() {
 void Layer::forwardDropOut() {
   auto& outV = getOutputValue();
 
-  if (passType_ == PASS_TRAIN || passType_ == PASS_METRIC_TRAIN ||
-      passType_ == PASS_METRIC_TRAIN_WITH_NOERROR) {
+  if (passType_ == PASS_TRAIN) {
     // new dropOutMask_ if dropOutMask_ is null ptr
     Matrix::resizeOrCreate(dropOutMask_,
                            outV->getHeight(),

paddle/pserver/BaseClient.h

@@ -30,9 +30,6 @@ namespace paddle {
  * the first solution arms with sendThreads_/recvThreads_ and sendJobQueue_/
  * recvJobQueue_. the second solution use some shared thread pool to manage
  * connections.
- * In addition to pserver, metric learning also uses network to exchange
- * features within multi-machines, so this class just abstracts some basic
- * threads and queue buffer creation for them
  */
 class BaseClient {
 protected:

paddle/pserver/ParameterServer2.cpp

@@ -367,11 +367,8 @@ void ParameterServer2::addGradient(const SendParameterRequest& request,
                                    std::vector<Buffer>* outputBuffers) {
   VLOG(1) << "pserver: addGradient";
 
-/// forwardbackward delta from all trainers
-/// indicate the fluctuation caused by forwardbackward.
-#ifndef PADDLE_METRIC_LEARNING
-  // @TODO(yanfei):
-  // add support tuning forwardbackward balance for metric learning
+  // forwardbackward delta from all trainers
+  // indicate the fluctuation caused by forwardbackward.
   if (!numPassFinishClients_) {
     REGISTER_BARRIER_DELTA_SERVER_SET(
         *statSet_,
@@ -381,7 +378,6 @@ void ParameterServer2::addGradient(const SendParameterRequest& request,
         request.forwardbackward_time(),
         isSparseServer_ ? "_sparseUpdater" : "_denseUpdater");
   }
-#endif
 
   {
     /// approximately pure network overhead

paddle/trainer/Trainer.h

@@ -30,10 +30,6 @@ limitations under the License. */
 #include "TrainerConfigHelper.h"
 #include "TrainerInternal.h"
 
-#ifdef PADDLE_METRIC_LEARNING
-#include "paddle/internals/metric_learning/MetricTrainer.h"
-#endif
-
 DECLARE_int32(num_passes);
 
 namespace paddle {
@@ -201,12 +197,8 @@ protected:
   // parameter util
   std::unique_ptr<ParameterUtil> paramUtil_;
 
-#ifdef PADDLE_METRIC_LEARNING
-  MetricTrainer trainerInternal_;
-#else
   // trainer Internal
   TrainerInternal trainerInternal_;
-#endif
 };
 
 }  // namespace paddle

paddle/utils/Flags.cpp

@@ -30,7 +30,6 @@ DEFINE_bool(parallel_nn,
 DEFINE_int32(trainer_count, 1, "Defined how many trainers to train");
 DEFINE_int32(gpu_id, 0, "Which gpu core to use");
 DEFINE_int32(port, 20134, "Listening port for pserver");
-DEFINE_int32(data_server_port, 21134, "Listening port for dserver");
 DEFINE_int32(ports_num,
              1,
              "Number of ports for sending dense parameter,"

paddle/utils/Flags.h

@@ -19,7 +19,6 @@ limitations under the License. */
 DECLARE_bool(parallel_nn);
 DECLARE_int32(async_count);
 DECLARE_int32(port);
-DECLARE_int32(data_server_port);
 DECLARE_bool(use_gpu);
 DECLARE_int32(gpu_id);
 DECLARE_int32(trainer_count);

paddle/utils/GlobalConstants.h

@@ -23,11 +23,6 @@ enum PassType {
   PASS_TEST,    // Test pass
   PASS_GC,      // Gradient Check pass
   PASS_METRIC,  // pass for generate template output with no drop rate.
-  // pass for metric learning training with metric learning error, only used
-  // when we are doing KNN evaluation.
-  PASS_METRIC_TRAIN,
-  PASS_METRIC_TRAIN_WITH_NOERROR,  // Pass for metric learning training
-                                   // with no evaluation.
 };
 
 enum ParameterType {