Merge branch 'develop' into feature/op_creation_methods

Dockerfile

@@ -25,7 +25,7 @@ COPY ./paddle/scripts/docker/root/ /root/
 RUN apt-get update && \
     apt-get install -y \
     git python-pip python-dev openssh-server bison  \
-    wget unzip tar xz-utils bzip2 gzip coreutils ntp \
+    wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
     python-numpy python-matplotlib gcc g++ \
     automake locales clang-format-3.8 swig doxygen cmake  \

go/cmd/master/master.go

@@ -11,6 +11,7 @@ import (
 
 	"github.com/namsral/flag"
 	log "github.com/sirupsen/logrus"
+	"github.com/topicai/candy"
 
 	"github.com/PaddlePaddle/Paddle/go/master"
 	"github.com/PaddlePaddle/Paddle/go/utils/networkhelper"
@@ -20,11 +21,18 @@ func main() {
 	port := flag.Int("port", 8080, "port of the master server.")
 	ttlSec := flag.Int("ttl", 60, "etcd lease TTL in seconds.")
 	endpoints := flag.String("endpoints", "http://127.0.0.1:2379", "comma separated etcd endpoints. If empty, fault tolerance will not be enabled.")
-	taskTimeoutDur := flag.Duration("task_timout_dur", 20*time.Minute, "task timout duration.")
-	taskTimeoutMax := flag.Int("task_timeout_max", 3, "max timtout count for each task before it being declared failed task.")
-	chunkPerTask := flag.Int("chunk_per_task", 10, "chunk per task.")
+	taskTimeoutDur := flag.Duration("task-timout-dur", 20*time.Minute, "task timout duration.")
+	taskTimeoutMax := flag.Int("task-timeout-max", 3, "max timtout count for each task before it being declared failed task.")
+	chunkPerTask := flag.Int("chunk-per-task", 10, "chunk per task.")
+	logLevel := flag.String("log-level", "info",
+		"log level, possible values: debug, info, warning, error, fatal, panic")
 	flag.Parse()
 
+	level, e := log.ParseLevel(*logLevel)
+	candy.Must(e)
+
+	log.SetLevel(level)
+
 	if *endpoints == "" {
 		log.Warningln("-endpoints not set, fault tolerance not be enabled.")
 	}

go/cmd/pserver/pserver.go

@@ -40,7 +40,7 @@ func main() {
 		idx = *index
 	} else {
 		e = pserver.NewEtcdClient(*etcdEndpoint, *numPservers, *etcdTimeout)
-		idx, err = e.Register()
+		idx, err = e.Register(*port)
 		candy.Must(err)
 
 		cp, err = pserver.NewCheckpointFromFile(*checkpointPath, idx, e)

go/master/client.go

@@ -2,6 +2,7 @@ package master
 
 import (
 	"os"
+	"time"
 
 	"github.com/PaddlePaddle/Paddle/go/connection"
 	"github.com/PaddlePaddle/recordio"
@@ -36,9 +37,9 @@ func (c *Client) getRecords() {
 	for {
 		t, err := c.getTask()
 		if err != nil {
-			// TODO(helin): wait before move on with next
 			// getTask call.
-			log.Errorln(err)
+			log.Errorf("Get task failed, sleep 3 seconds and continue, %s", err)
+			time.Sleep(3 * time.Second)
 			continue
 		}
 

go/master/service.go

@@ -215,6 +215,7 @@ func readChunks(globPaths []string) ([]Chunk, error) {
 		}
 
 		count := index.NumChunks()
+		log.Infof("readChunks: file %s has %d chunks", path, count)
 		for i := 0; i < count; i++ {
 			chunk := Chunk{
 				Path:  path,

go/pserver/client/c/test/test_train.py

@@ -1,5 +1,23 @@
 import paddle.v2 as paddle
 import paddle.v2.dataset.uci_housing as uci_housing
+import paddle.v2.master as master
+import os
+import cPickle as pickle
+
+etcd_ip = os.getenv("MASTER_IP", "127.0.0.1")
+etcd_endpoint = "http://" + etcd_ip + ":2379"
+
+
+def cloud_reader():
+    print "connecting to master, etcd endpoints: ", etcd_endpoint
+    master_client = master.client(etcd_endpoint, 5, 64)
+    master_client.set_dataset(
+        ["/pfs/dlnel/public/dataset/uci_housing/uci_housing-*-of-*"])
+    while 1:
+        r, e = master_client.next_record()
+        if not r:
+            break
+        yield pickle.loads(r)
 
 
 def main():
@@ -22,13 +40,13 @@ def main():
     # create optimizer of new remote updater to pserver
     optimizer = paddle.optimizer.Momentum(momentum=0)
 
-    #TODO(zhihong) : replace optimizer with new OptimizerConfig
-
+    print "etcd endoint: ", etcd_endpoint
     trainer = paddle.trainer.SGD(cost=cost,
                                  parameters=parameters,
                                  update_equation=optimizer,
                                  is_local=False,
-                                 pserver_spec="localhost:3000")
+                                 pserver_spec=etcd_endpoint,
+                                 use_etcd=True)
 
     # event_handler to print training and testing info
     def event_handler(event):
@@ -47,11 +65,11 @@ def main():
                 print "Test %d, %.2f" % (event.pass_id, result.cost)
 
     # training
+    # NOTE: use uci_housing.train() as reader for non-paddlecloud training
     trainer.train(
         reader=paddle.batch(
             paddle.reader.shuffle(
-                uci_housing.train(), buf_size=500),
-            batch_size=2),
+                cloud_reader, buf_size=500), batch_size=2),
         feeding={'x': 0,
                  'y': 1},
         event_handler=event_handler,

go/pserver/client/etcd_client.go

@@ -12,6 +12,7 @@ import (
 )
 
 const (
+	// DefaultEtcdTimeout is the default etcd timeout
 	DefaultEtcdTimeout time.Duration = 5 * time.Second
 )
 
@@ -66,10 +67,10 @@ func (p *EtcdClient) List() []Server {
 	for {
 		for i := 0; i < psDesired; i++ {
 			ctx, cancel := context.WithTimeout(context.Background(), p.timeout)
-			cancel()
 			psKey := pserver.PsPath + strconv.Itoa(i)
 			log.Debugf("checking %s", psKey)
 			resp, err := p.client.Get(ctx, psKey)
+			cancel()
 			if err != nil {
 				log.Infof("Get psKey=%s error, %v", psKey, err)
 				time.Sleep(p.timeout)

go/pserver/etcd_client.go

@@ -49,7 +49,7 @@ func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *Et
 // Register registers the pserver on etcd
 //
 // Register returns the index of the current pserver.
-func (e *EtcdClient) Register() (int, error) {
+func (e *EtcdClient) Register(port int) (int, error) {
 
 	var err error
 	e.externalIP, err = networkhelper.GetExternalIP()
@@ -116,7 +116,7 @@ func (e *EtcdClient) Register() (int, error) {
 	for {
 		ctx, cancel := context.WithTimeout(context.Background(), time.Second)
 		var err error
-		pserverIdx, err = e.registerPserverEtcd(ctx)
+		pserverIdx, err = e.registerPserverEtcd(ctx, port)
 		cancel()
 		if err != nil {
 			log.Warn(err)
@@ -140,7 +140,7 @@ func (e *EtcdClient) initDesiredPservers(ctx context.Context, numPservers int) (
 }
 
 // registerPserverEtcd registers pserver node on etcd using transaction.
-func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) {
+func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, error) {
 	var idx int
 	_, err := concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
 		registered := false
@@ -156,8 +156,9 @@ func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) {
 					log.Fatal(err)
 				}
 				// find the first id and write info
-				c.Put(psKey, e.externalIP, clientv3.WithLease(resp.ID))
-				log.Debugf("set pserver node %s with value %s", psKey, e.externalIP)
+				pserverAddr := e.externalIP + ":" + strconv.Itoa(port)
+				c.Put(psKey, pserverAddr, clientv3.WithLease(resp.ID))
+				log.Debugf("set pserver node %s with value %s", psKey, pserverAddr)
 				ch, kaerr := e.etcdClient.KeepAlive(context.TODO(), resp.ID)
 				if kaerr != nil {
 					log.Errorf("keepalive etcd node error: %v", kaerr)

paddle/api/PaddleAPI.h

@@ -843,7 +843,8 @@ public:
                                                bool useSparseUpdater);
   static ParameterUpdater* createNewRemoteUpdater(
       OptimizationConfig* config,
-      const std::string pserverSpec) throw(UnsupportError);
+      const std::string pserverSpec,
+      const bool useEtcd) throw(UnsupportError);
   ~ParameterUpdater();
 
   /**

paddle/api/ParameterUpdater.cpp

@@ -33,11 +33,12 @@ ParameterUpdater *ParameterUpdater::createLocalUpdater(
 
 ParameterUpdater *ParameterUpdater::createNewRemoteUpdater(
     OptimizationConfig *config,
-    const std::string pserverSpec) throw(UnsupportError) {
+    const std::string pserverSpec,
+    const bool useEtcd) throw(UnsupportError) {
 #ifndef PADDLE_WITHOUT_GOLANG
   auto updater = new ParameterUpdater();
   updater->m->updater.reset(new paddle::NewRemoteParameterUpdater(
-      config->m->getConfig(), pserverSpec));
+      config->m->getConfig(), pserverSpec, useEtcd));
   return updater;
 #else
   throw UnsupportError();

paddle/framework/CMakeLists.txt

@@ -11,8 +11,10 @@ proto_library(op_proto SRCS op_proto.proto DEPS attr_type)
 cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
 proto_library(op_desc SRCS op_desc.proto DEPS attr_type)
 cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
+
 cc_library(operator SRCS operator.cc DEPS op_desc device_context)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
+
 cc_library(op_registry SRCS op_registry.cc DEPS op_proto op_desc)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry operator)
 py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.proto)
@@ -21,4 +23,5 @@ add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch
 add_dependencies(framework_py_proto framework_py_proto_init)
 
 proto_library(net_proto SRCS net_proto.proto DEPS op_proto)
-cc_library(net SRCS net.cc DEPS net_proto)
+cc_library(net SRCS net.cc DEPS operator net_proto op_registry)
+cc_test(net_op_test SRCS net_op_test.cc DEPS net)

paddle/framework/net.cc

@@ -19,18 +19,41 @@
 namespace paddle {
 namespace framework {
 
-PlainNet::PlainNet(const NetDesc& def) {}
-
-void PlainNet::InferShape(const ScopePtr& scope) const {
+void PlainNet::CompleteAddOp() {
+  std::unordered_set<std::string> input_set;
+  std::unordered_set<std::string> output_set;
+  std::unordered_set<std::string> temp_output;
   for (auto& op : ops_) {
-    op.InferShape();
+    for (auto& ipt : op->inputs_) {
+      if (!Contains(output_set, ipt)) {  // Not other op's output
+        input_set.insert(ipt);
+      } else {
+        temp_output.insert(ipt);
       }
     }
 
-void PlainNet::Run(const ScopePtr& scope, const DeviceContext& ctx) const {
-  for (auto& op : ops_) {
-    op.Run(ctx);
+    for (auto& opt : op->outputs_) {
+      output_set.insert(opt);
+    }
+  }
+  inputs_.reserve(input_set.size());
+  std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs_));
+
+  outputs_.reserve(output_set.size());
+  std::vector<int> tmp_index;
+  tmp_index.reserve(temp_output.size());
+  int idx = 0;
+  for (auto& opt : output_set) {
+    if (Contains(temp_output, opt)) {
+      tmp_index.push_back(idx);
     }
+    outputs_.push_back(opt);
+    ++idx;
   }
+
+  attrs_["temporary_index"] = tmp_index;
+  add_op_done_ = true;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/net.h

@@ -14,86 +14,38 @@
 
 #pragma once
 
+#include <paddle/framework/op_desc.pb.h>
+#include <paddle/framework/operator.h>
 #include "paddle/framework/net_proto.pb.h"
 #include "paddle/framework/op_proto.pb.h"
+#include "paddle/framework/op_registry.h"
 #include "paddle/framework/scope.h"
 #include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace framework {
-using namespace paddle::platform;
-
-// operator's index stored in a network.
-typedef int OpIndex;
-/**
- * NOTE following codes are some definitions of unimplemented concepts.
- * We write some basic implementation to make Net compilable. These APIs will
- * keep updating if the concepts related are implemented.
- */
-
-struct OpDesc;
-struct OpAttrs {};
-
-class Operator {
- public:
-  Operator(const OpDesc &def) {}
-  void InferShape() const {}
-  void Run(const DeviceContext &ctx) const {}
-};
-
 /**
- * @brief Network that manage the operators it has.
+ * @brief Network is also a type of Operator
+ *
+ * It will manage the operators it has.
  *
- * Network is the container and controller of a set of operators, user can build
- * a real network from a NetDesc which is a protobuf message and use
- * Network.Run() * to run all the operators in the network.
+ * Network is the container and controller of a set of operators.
 
  * A network object knows all Operators belonging to this network. Variables,
  * which are inputs and outputs of these operators, are created and managed by a
  * hierarchy of Scope objects.
  *
- * This is the base class of network, all the networks should implement the apis
+ * This is the base class of network, all the networks should implement the APIs
  * it defines.
  */
-class Net {
+class Net : public OperatorBase {
  public:
-  /**
-   * @brief Infer shapes of all inputs and outputs of operators.
-   */
-  virtual void InferShape(const ScopePtr &scope) const = 0;
-  /**
-   * @brief Run the network.
-   *
-   * Run all the operators and return success(true) or not, with all the
-   * variables are located in `scope`. `context` describes the detail execution
-   * environment for ops. `begin` and `end` specify the scope of `ops_` to run,
-   * If no positive indexes are provided, all operators in `ops_` will run.
-   */
-  virtual void Run(const ScopePtr &scope, const DeviceContext &ctx) const = 0;
-
-  /**
-   * @brief Add an Operator according to `def`.
-   */
-  virtual OpIndex AddOp(const OpProto &def) = 0;
-
-  /**
-   * @brief Add optimizer operators acctording to `attrs`.
-   */
-  virtual void AddOptimizerOps(const OpAttrs &attrs) = 0;
-
-  /**
-   * @brief Add backward operators.
-   */
-  virtual void AddBackwardOps() = 0;
-
-  /**
-   * @brief Create a network.
-   */
-  static std::unique_ptr<Net> Create(const NetDesc &def = NetDesc());
-
-  virtual ~Net() {}
+  virtual void AddOp(const OperatorPtr& op) = 0;
+  virtual void CompleteAddOp() = 0;
 };
 
+using NetPtr = std::shared_ptr<Net>;
+
 /**
  * @brief a basic implementation of Net.
  *
@@ -103,18 +55,14 @@ class Net {
 class PlainNet : public Net {
  public:
   /**
-   * @brief Initialize a PlainNet.
-   *
-   * Initialize from  a network describe by `def`. NetDesc is the definition of
-   * a network.
-   */
-  PlainNet(const NetDesc &def);
-
-  /**
-   * Infer all the operators' input and output varialbes' shapes, will be called
+   * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch
    */
-  virtual void InferShape(const ScopePtr &scope) const override;
+  void InferShape(const ScopePtr& scope) const override {
+    for (auto& op : ops_) {
+      op->InferShape(scope);
+    }
+  }
 
   /**
    * @brief Run the network.
@@ -123,49 +71,32 @@ class PlainNet : public Net {
    * scope will be used instead. If no OpContext is provicded, default context
    * will be used.
    */
-  virtual void Run(const ScopePtr &scope,
-                   const DeviceContext &ctx) const override;
+  void Run(const ScopePtr& scope,
+           const platform::DeviceContext& dev_ctx) const override {
+    for (auto& op : ops_) {
+      op->Run(scope, dev_ctx);
+    }
+  }
 
   /**
-   * @brief Add an operator to this network.
+   * @brief Add an operator by ptr
    */
-  virtual OpIndex AddOp(const OpProto &def) override;
+  void AddOp(const OperatorPtr& op) override {
+    PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
+    ops_.push_back(op);
+  }
 
-  /**
-   * @brief Add all optimizer operators related into the network.
-   */
-  virtual void AddOptimizerOps(const OpAttrs &attrs) override;
+  void CompleteAddOp() override;
 
-  /**
-   * @brief Add all backward operators related into the network.
-   */
-  virtual void AddBackwardOps() override;
-
-  virtual ~PlainNet() override {}
-
- protected:
-  /**
-   * @brief Build the network.
-   *
-   * Create operators accordding to `def`, will be called by the constructor.
-   */
-  void BuildNet(const NetDesc &def);
-
-  /**
-   * @brief Add an operator into this network.
-   *
-   * Add a operator which is identified as `type` and has attributes described
-   * in `attrs`, the `inputs` are the keys of readonly input variables,
-   * `outputs` are keys of mutable output variables. An `OpIndex` will be
-   * returned to indicate the offset of the new operator in `ops_`.
-   */
-  OpIndex AddOp(const std::string &type, const std::vector<std::string> &inputs,
-                const std::vector<std::string> &outputs,
-                const OpAttrs &attrs = OpAttrs());
+  std::vector<OperatorPtr> ops_;
 
  private:
-  // the operators owned by `Network`.
-  std::vector<Operator> ops_;
+  bool add_op_done_{false};
+
+  template <typename T, typename KeyType>
+  static bool Contains(T container, KeyType key) {
+    return container.find(key) != container.end();
+  }
 };
 
 }  // namespace framework

paddle/framework/net_op_test.cc

@@ -0,0 +1,67 @@
+#include <gtest/gtest.h>
+#include <paddle/framework/net.h>
+#include <paddle/framework/op_registry.h>
+#include <paddle/framework/operator.h>
+
+namespace pd = paddle::framework;
+
+static int infer_shape_cnt = 0;
+static int run_cnt = 0;
+
+class TestOp : public pd::OperatorBase {
+ public:
+  void InferShape(const paddle::framework::ScopePtr& scope) const override {
+    ++infer_shape_cnt;
+  }
+  void Run(const paddle::framework::ScopePtr& scope,
+           const paddle::platform::DeviceContext& dev_ctx) const override {
+    ++run_cnt;
+  }
+};
+
+template <typename T>
+void AssertSameVectorWithoutOrder(const std::vector<T>& expected,
+                                  const std::vector<T>& actual) {
+  ASSERT_EQ(expected.size(), actual.size());
+  std::unordered_set<T> expected_set;
+  for (auto& tmp : expected) {
+    expected_set.insert(tmp);
+  }
+  for (auto& act : actual) {
+    ASSERT_NE(expected_set.end(), expected_set.find(act));
+  }
+}
+
+TEST(OpKernel, all) {
+  auto net = std::make_shared<paddle::framework::PlainNet>();
+  ASSERT_NE(net, nullptr);
+
+  auto op1 = std::make_shared<TestOp>();
+  op1->inputs_ = {"x", "w1", "b1"};
+  op1->outputs_ = {"y"};
+  net->AddOp(op1);
+
+  auto op2 = std::make_shared<TestOp>();
+  op2->inputs_ = {"y", "w2", "b2"};
+  op2->outputs_ = {"z"};
+  net->AddOp(op2);
+
+  net->CompleteAddOp();
+  AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, net->inputs_);
+  AssertSameVectorWithoutOrder({"y", "z"}, net->outputs_);
+  auto tmp_idx_iter = net->attrs_.find("temporary_index");
+  ASSERT_NE(net->attrs_.end(), tmp_idx_iter);
+  auto& tmp_idx = boost::get<std::vector<int>>(tmp_idx_iter->second);
+  ASSERT_EQ(1UL, tmp_idx.size());
+  ASSERT_EQ("y", net->outputs_[tmp_idx[0]]);
+
+  auto scope = std::make_shared<pd::Scope>();
+  paddle::platform::CPUDeviceContext dev_ctx;
+
+  net->InferShape(scope);
+  net->Run(scope, dev_ctx);
+  ASSERT_EQ(2, infer_shape_cnt);
+  ASSERT_EQ(2, run_cnt);
+
+  ASSERT_THROW(net->AddOp(op2), paddle::framework::EnforceNotMet);
+}

paddle/framework/op_registry.h

@@ -203,10 +203,9 @@ class OpRegistry {
     //! Create a OpPtr by type.
     std::string op_type = op_desc.type();
     OperatorPtr op(creators().at(op_type)());
-
     //! Fill op's data member. Not use constructor because it will be noising
     //! for Op developer.
-    op->desc_ = op_desc;
+    op->type_ = op_desc.type();
     op->inputs_.reserve((size_t)op_desc.inputs_size());
     std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
               std::back_inserter(op->inputs_));
@@ -239,7 +238,7 @@ class OpRegistry {
     static std::atomic<size_t> gUniqId(0UL);
     for (auto& outname : op->outputs_) {
       if (outname == OperatorBase::TMP_VAR_NAME()) {
-        outname += op->Type();
+        outname += op->type_;
         outname += "@";
         outname += std::to_string(gUniqId.fetch_add(1));
       }
@@ -265,12 +264,18 @@ class OpRegisterHelper {
   }
 };
 
+/**
+ * check if MACRO is used in GLOBAL NAMESPACE.
+ */
 #define STATIC_ASSERT_GLOBAL_NAMESPACE(uniq_name, msg)                        \
   struct __test_global_namespace_##uniq_name##__ {};                          \
   static_assert(std::is_same<::__test_global_namespace_##uniq_name##__,       \
                              __test_global_namespace_##uniq_name##__>::value, \
                 msg)
 
+/**
+ * Macro to Register Operator.
+ */
 #define REGISTER_OP(__op_type, __op_class, __op_maker_class)                 \
   STATIC_ASSERT_GLOBAL_NAMESPACE(__reg_op__##__op_type,                      \
                                  "REGISTER_OP must be in global namespace"); \
@@ -278,9 +283,12 @@ class OpRegisterHelper {
       __op_register_##__op_type##__(#__op_type);                             \
   int __op_register_##__op_type##_handle__() { return 0; }
 
-#define REGISTER_OP_KERNEL(type, GPU_OR_CPU, PlaceType, KernelType)       \
+/**
+ * Macro to Register OperatorKernel.
+ */
+#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, KernelType)      \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
-      __reg_op_kernel_##type##_##GPU_OR_CPU##__,                          \
+      __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
       "REGISTER_OP_KERNEL must be in global namespace");                  \
   struct __op_kernel_register__##type##__ {                               \
     __op_kernel_register__##type##__() {                                  \
@@ -291,7 +299,7 @@ class OpRegisterHelper {
     }                                                                     \
   };                                                                      \
   static __op_kernel_register__##type##__ __reg_kernel_##type##__;        \
-  int __op_kernel_register_##type##_handle_##GPU_OR_CPU##__() { return 0; }
+  int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
 
 #define REGISTER_OP_GPU_KERNEL(type, KernelType) \
   REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, KernelType)
@@ -299,6 +307,10 @@ class OpRegisterHelper {
 #define REGISTER_OP_CPU_KERNEL(type, KernelType) \
   REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, KernelType)
 
+/**
+ * Macro to mark what Operator and Kernel we will use and tell the compiler to
+ * link them into target.
+ */
 #define USE_OP_WITHOUT_KERNEL(op_type)                      \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                           \
       __use_op_without_kernel_##op_type,                    \
@@ -316,15 +328,16 @@ class OpRegisterHelper {
       __attribute__((unused)) =                                           \
           __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__()
 
-#ifdef PADDLE_ONLY_CPU
-#define USE_OP(op_type)           \
+// use Operator with only cpu kernel.
+#define USE_OP_CPU(op_type)       \
   USE_OP_WITHOUT_KERNEL(op_type); \
-  USE_OP_KERNEL(op_type, CPU);
+  USE_OP_KERNEL(op_type, CPU)
 
+#ifdef PADDLE_ONLY_CPU
+#define USE_OP(op_type) USE_OP_CPU(op_type)
 #else
 #define USE_OP(op_type) \
-  USE_OP_WITHOUT_KERNEL(op_type); \
-  USE_OP_KERNEL(op_type, CPU);    \
+  USE_OP_CPU(op_type);  \
   USE_OP_KERNEL(op_type, GPU)
 #endif
 

paddle/framework/operator.cc

@@ -19,7 +19,7 @@ namespace framework {
 
 std::string OperatorBase::DebugString() const {
   std::stringstream ss;
-  ss << "Op(" << Type() << "), inputs:(";
+  ss << "Op(" << type_ << "), inputs:(";
   for (size_t i = 0; i < inputs_.size(); ++i) {
     ss << inputs_[i];
     if (i != inputs_.size() - 1) {

paddle/framework/operator.h

@@ -69,10 +69,8 @@ class OperatorBase {
   virtual void Run(const ScopePtr& scope,
                    const platform::DeviceContext& dev_ctx) const = 0;
 
-  std::string Type() const { return desc_.type(); }
-
  public:
-  OpDesc desc_;
+  std::string type_;
   std::vector<std::string> inputs_;
   std::vector<std::string> outputs_;
   AttributeMap attrs_;
@@ -148,7 +146,7 @@ class OperatorWithKernel : public OperatorBase {
 
   void Run(const ScopePtr& scope,
            const platform::DeviceContext& dev_ctx) const final {
-    auto& opKernel = AllOpKernels().at(Type()).at(OpKernelKey(dev_ctx));
+    auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
     opKernel->Compute(OpKernel::KernelContext(this, scope, dev_ctx));
   }
 

paddle/framework/operator_test.cc

@@ -19,14 +19,18 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class OperatorTest : public OperatorBase {
+static int op_run_num = 0;
+
+class OpWithoutKernelTest : public OperatorBase {
  public:
   void Init() override { x = 1; }
   void InferShape(const ScopePtr& scope) const override {}
   void Run(const ScopePtr& scope,
            const platform::DeviceContext& dev_ctx) const override {
-    float scale = GetAttr<float>("scale");
-    ASSERT_NEAR(scale, 3.14, 1e-5);
+    op_run_num++;
+    ASSERT_EQ((int)inputs_.size(), 1);
+    ASSERT_EQ((int)outputs_.size(), 1);
+    ASSERT_NEAR(GetAttr<float>("scale"), 3.14, 1e-5);
     ASSERT_EQ(scope->GetVariable(inputs_[0]), nullptr);
     ASSERT_EQ(x, 1);
     ASSERT_NE(scope->GetVariable(outputs_[0]), nullptr);
@@ -36,15 +40,14 @@ class OperatorTest : public OperatorBase {
   float x = 0;
 };
 
-class OperatorTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
+class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
-  OperatorTestProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+  OpeWithoutKernelTestProtoAndCheckerMaker(OpProto* proto,
+                                           OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("input", "input of test op");
     AddOutput("output", "output of test op");
-    AddAttr<float>("scale", "scale of cosine op")
-        .SetDefault(1.0)
-        .LargerThan(0.0);
+    AddAttr<float>("scale", "scale of cosine op");
     AddComment("This is test op");
   }
 };
@@ -52,8 +55,8 @@ class OperatorTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_OP(test_operator, paddle::framework::OperatorTest,
-            paddle::framework::OperatorTestProtoAndCheckerMaker);
+REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest,
+            paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
 
 TEST(OperatorBase, all) {
   paddle::framework::OpDesc op_desc;
@@ -63,18 +66,17 @@ TEST(OperatorBase, all) {
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
   attr->set_type(paddle::framework::AttrType::FLOAT);
-  float scale = 3.14;
-  attr->set_f(scale);
+  attr->set_f(3.14);
 
   paddle::platform::CPUDeviceContext device_context;
   auto scope = std::make_shared<paddle::framework::Scope>();
 
   paddle::framework::OperatorPtr op =
       paddle::framework::OpRegistry::CreateOp(op_desc);
-  ASSERT_EQ(op->GetAttr<float>("scale"), scale);
   scope->CreateVariable("OUT1");
+  ASSERT_EQ(paddle::framework::op_run_num, 0);
   op->Run(scope, device_context);
-  std::cout << op->DebugString() << std::endl;
+  ASSERT_EQ(paddle::framework::op_run_num, 1);
 }
 
 namespace paddle {
@@ -86,13 +88,13 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("input", "input of test op");
     AddOutput("output", "output of test op");
-    AddAttr<float>("scale", "scale of cosine op")
-        .SetDefault(1.0)
-        .LargerThan(0.0);
+    AddAttr<float>("scale", "scale of cosine op");
     AddComment("This is test op");
   }
 };
 
+static int cpu_kernel_run_num = 0;
+
 class OpWithKernelTest : public OperatorWithKernel {
  protected:
   void InferShape(const std::vector<const Tensor*>& inputs,
@@ -102,10 +104,10 @@ class OpWithKernelTest : public OperatorWithKernel {
 class CPUKernelTest : public OpKernel {
  public:
   void Compute(const KernelContext& context) const {
-    float scale = context.op_.GetAttr<float>("scale");
-    ASSERT_NEAR(scale, 3.14, 1e-5);
-    std::cout << "this is cpu kernel" << std::endl;
-    std::cout << context.op_.DebugString() << std::endl;
+    cpu_kernel_run_num++;
+    ASSERT_EQ((int)context.op_.inputs_.size(), 1);
+    ASSERT_EQ((int)context.op_.outputs_.size(), 1);
+    ASSERT_NEAR(context.op_.GetAttr<float>("scale"), 3.14, 1e-5);
   }
 };
 
@@ -131,5 +133,7 @@ TEST(OpKernel, all) {
 
   paddle::framework::OperatorPtr op =
       paddle::framework::OpRegistry::CreateOp(op_desc);
+  ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 0);
   op->Run(scope, cpu_device_context);
+  ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
 }

paddle/function/RowConvOpGpu.cu

@@ -144,12 +144,15 @@ __global__ void KeRowConvBwWeight(real* dw, const real* x, const real* dy,
       int yoff = start + j;
 
       // transpose
-      sh_x[tidx][tidy] = (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0;
-      sh_dy[tidx][tidy + context - 1] = (xoff < width && yoff < end) ? dy[yoff * width + xoff] : 0.0;
+      sh_x[tidx][tidy] = (xoff < width && yoff < end) ?
+      x[yoff * width + xoff] : 0.0;
+      sh_dy[tidx][tidy + context - 1] = (xoff < width && yoff < end) ?
+      dy[yoff * width + xoff] : 0.0;
       __syncthreads();
       if (tidy < (context - 1)) {
         yoff = yoff - context + 1;
-        sh_dy[tidx][tidy] = (xoff < width && yoff >= start) ? dy[yoff * width + xoff] : 0.0;
+        sh_dy[tidx][tidy] = (xoff < width && yoff >= start) ?
+        dy[yoff * width + xoff] : 0.0;
       }
       __syncthreads();
 
@@ -199,11 +202,13 @@ __global__ void KeRowConvBwWeight2(real* dw, const real* x, const real* dy,
       int yoff = start + j;
 
       // transpose
-      sh_x[tidx][tidy] = (xoff < width && yoff < end) ? x[yoff * width + xoff] : 0.0;
+      sh_x[tidx][tidy] = (xoff < width && yoff < end) ?
+      x[yoff * width + xoff] : 0.0;
       __syncthreads();
 
       for (int t = 0; t < context; t++) {
-        sh_dy[tidx][tidy] = (xoff < width && (yoff - t) >= start && yoff - t < end) ? dy[(yoff - t) * width + xoff] : 0.0;
+        sh_dy[tidx][tidy] = (xoff < width && (yoff - t) >= start &&
+        yoff - t < end) ? dy[(yoff - t) * width + xoff] : 0.0;
         __syncthreads();
 
         real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx];

paddle/scripts/docker/build.sh

@@ -155,7 +155,8 @@ RUN apt-get update &&\
     paddle version
 ${DOCKERFILE_CUDNN_DSO}
 ${DOCKERFILE_GPU_ENV}
-
+ADD go/cmd/pserver/pserver /usr/bin/
+ADD go/cmd/master/master /usr/bin/
 # default command shows the paddle version and exit
 CMD ["paddle", "version"]
 EOF

paddle/trainer/NewRemoteParameterUpdater.cpp

@@ -28,6 +28,17 @@ NewRemoteParameterUpdater::NewRemoteParameterUpdater(
       newGradients_(nullptr),
       pserverSpec_(pserverSpec) {}
 
+NewRemoteParameterUpdater::NewRemoteParameterUpdater(
+    const OptimizationConfig &config,
+    const std::string pserverSpec,
+    const bool useEtcd)
+    : trainerConfig_(config),
+      parameterClient_(-1),
+      newParameters_(nullptr),
+      newGradients_(nullptr),
+      pserverSpec_(pserverSpec),
+      useEtcd_(useEtcd) {}
+
 void NewRemoteParameterUpdater::init(
     const std::vector<ParameterPtr> &parameters) {
   ParameterUpdater::init(parameters);
@@ -38,8 +49,13 @@ void NewRemoteParameterUpdater::init(
   }
 
   // create parameter server client.
+  if (useEtcd_) {
+    parameterClient_ = paddle_new_etcd_pserver_client(
+        (char *)pserverSpec_.c_str(), FLAGS_trainer_id == 0);
+  } else {
     parameterClient_ = paddle_new_pserver_client((char *)pserverSpec_.c_str(),
                                                  FLAGS_trainer_id == 0);
+  }
 
   // init new parameter and gradient.
   newParameters_ = initNewParameter(PARAMETER_VALUE);

paddle/trainer/NewRemoteParameterUpdater.h

@@ -32,6 +32,9 @@ class NewRemoteParameterUpdater : public ParameterUpdater {
 public:
   NewRemoteParameterUpdater(const OptimizationConfig& config,
                             const std::string pserverSpec);
+  NewRemoteParameterUpdater(const OptimizationConfig& config,
+                            const std::string pserverSpec,
+                            const bool useEtcd);
   ~NewRemoteParameterUpdater() {
     releaseNewParameter(newParameters_);
     releaseNewParameter(newGradients_);
@@ -111,6 +114,8 @@ protected:
   paddle_parameter** newGradients_;
   /// the specification of parameter server "host1:port,host1:port"
   std::string pserverSpec_;
+  /// true if pserverSpec_ is etcd endpoint, else pserverSpec_ is pserver addr
+  bool useEtcd_;
 };
 
 }  // namespace paddle

python/paddle/v2/dataset/common.py

@@ -22,6 +22,8 @@ import importlib
 import paddle.v2.dataset
 import cPickle
 import glob
+import cPickle as pickle
+import random
 
 __all__ = [
     'DATA_HOME', 'download', 'md5file', 'split', 'cluster_files_reader',
@@ -170,8 +172,6 @@ def convert(output_path,
             name_prefix,
             max_lines_to_shuffle=1000):
     import recordio
-    import cPickle as pickle
-    import random
     """
     Convert data from reader to recordio format files.
 
@@ -201,8 +201,10 @@ def convert(output_path,
     def write_data(w, lines):
         random.shuffle(lines)
         for i, d in enumerate(lines):
-            d = pickle.dumps(d, pickle.HIGHEST_PROTOCOL)
-            w[i % num_shards].write(d)
+            # FIXME(Yancey1989):
+            # dumps with protocol: pickle.HIGHEST_PROTOCOL
+            o = pickle.dumps(d)
+            w[i % num_shards].write(o)
 
     w = open_writers()
     lines = []

python/paddle/v2/dataset/mq2007.py

@@ -212,19 +212,19 @@ def gen_pair(querylist, partial_order="full"):
         for j in range(i + 1, len(querylist)):
             query_right = querylist[j]
             if query_left.relevance_score > query_right.relevance_score:
-                labels.append(1)
+                labels.append([1])
                 docpairs.append([
                     np.array(query_left.feature_vector),
                     np.array(query_right.feature_vector)
                 ])
             elif query_left.relevance_score < query_right.relevance_score:
-                labels.append(1)
+                labels.append([1])
                 docpairs.append([
                     np.array(query_right.feature_vector),
                     np.array(query_left.feature_vector)
                 ])
     for label, pair in zip(labels, docpairs):
-        yield label, pair[0], pair[1]
+        yield np.array(label), pair[0], pair[1]
 
 
 def gen_list(querylist):

python/paddle/v2/master/client.py

@@ -10,8 +10,9 @@ class client(object):
     client is a client to the master server.
     """
 
-    def __init__(self, addr, buf_size):
-        self.c = lib.paddle_new_master_client(addr, buf_size)
+    def __init__(self, etcd_endpoints, timeout, buf_size):
+        self.c = lib.paddle_new_etcd_master_client(etcd_endpoints, timeout,
+                                                   buf_size)
 
     def close(self):
         lib.paddle_release_master_client(self.c)