Merge pull request #10546 from chengduoZH/feature/change_pe_strategy

Balance parameter_opt between cards

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -37,20 +37,26 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
     const std::vector<Scope *> &local_scopes,
-    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale)
+    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale,
+    bool balance_parameter_opt_between_cards)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes),
-      nccl_ctxs_(nccl_ctxs) {
+      nccl_ctxs_(nccl_ctxs),
+      balance_parameter_opt_between_cards_(
+          balance_parameter_opt_between_cards) {
 #else
 MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale)
+    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale,
+    bool balance_parameter_opt_between_cards)
     : loss_var_name_(loss_var_name),
       places_(places),
-      local_scopes_(local_scopes) {
+      local_scopes_(local_scopes),
+      balance_parameter_opt_between_cards_(
+          balance_parameter_opt_between_cards) {
 #endif
   for (auto &p : params) {
     grad_names_.insert(GradVarName(p));
@@ -124,6 +130,12 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
   // Find "send" op first for split is in front of send.
   OpDesc *send_op = GetSendOpDesc(program);
 
+  size_t cur_device_id = 0;
+  std::vector<std::unordered_set<std::string>> var_name_on_devices;
+  std::vector<std::unordered_set<std::string>> bcast_var_name_set;
+  var_name_on_devices.resize(places_.size());
+  bcast_var_name_set.resize(places_.size());
+
   bool is_forwarding = true;
   for (auto *op : program.Block(0).AllOps()) {
     if (op->Type() == "send") {
@@ -139,12 +151,27 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       }
       is_forwarding = false;
     } else {
+      int op_dev_id = GetOpDeviceID(var_name_on_devices, *op);
+      if (op_dev_id == -1) {  // var on all device
         CreateComputationalOps(&result, *op, places_.size());
+      } else {
+        CreateComputationalOp(&result, *op, op_dev_id);
+        for (auto &var_name : op->OutputArgumentNames()) {
+          var_name_on_devices[op_dev_id].emplace(var_name);
+        }
+      }
       if (!is_forwarding && places_.size() > 1) {
         // Currently, we assume that once gradient is generated, it can be
         // broadcast, and each gradient is only broadcast once.
         for (auto &og : op->OutputArgumentNames()) {
           if (IsParameterGradientOnce(og, &og_has_been_broadcast)) {
+            if (balance_parameter_opt_between_cards_) {
+              CreateReduceOp(&result, og, cur_device_id);
+              var_name_on_devices[cur_device_id].emplace(og);
+              bcast_var_name_set[cur_device_id].emplace(
+                  og.substr(0, og.size() - strlen(kGradVarSuffix)));
+              cur_device_id = (cur_device_id + 1) % places_.size();
+            } else {
               if (IsSparseGradient(var_types, og)) {
                 CreateReduceOp(&result, og, 0);
                 CreateBroadcastOp(&result, og, 0);
@@ -156,7 +183,15 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
         }
       }
     }
+  }
 
+  // Insert BCast Ops
+  for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
+    auto &to_bcast_set = bcast_var_name_set[dev_id];
+    for (auto &bcast_name : to_bcast_set) {
+      CreateBroadcastOp(&result, bcast_name, dev_id);
+    }
+  }
   /*
     Dependency graph has been constructed. However, there are still data
     harzaeds need to be handled.
@@ -265,6 +300,26 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
   return is_pg_once;
 }
 
+int MultiDevSSAGraphBuilder::GetOpDeviceID(
+    const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
+    const OpDesc &op) const {
+  if (!balance_parameter_opt_between_cards_) {
+    return -1;
+  }
+
+  int var_dev_id = -1;
+  for (auto &var_name : op.InputArgumentNames()) {
+    if (var_dev_id != -1) break;
+    for (size_t i = 0; i < var_name_on_devices.size(); ++i) {
+      if (var_name_on_devices[i].count(var_name)) {
+        var_dev_id = static_cast<int>(i);
+        break;
+      }
+    }
+  }
+  return var_dev_id;
+}
+
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const {
   for (size_t i = 0; i < places_.size(); ++i) {
 // Insert ScaleCost OpHandle

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -36,13 +36,15 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
                           platform::NCCLContextMap *nccl_ctxs,
-                          bool use_default_grad_scale);
+                          bool use_default_grad_scale,
+                          bool balance_parameter_opt_between_cards);
 #else
   MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
-                          bool use_default_grad_scale);
+                          bool use_default_grad_scale,
+                          bool balance_parameter_opt_between_cards);
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
@@ -60,6 +62,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #ifdef PADDLE_WITH_CUDA
   platform::NCCLContextMap *nccl_ctxs_;
 #endif
+  bool balance_parameter_opt_between_cards_;
   bool use_default_grad_scale_;
 
   bool IsScaleLossOp(const OpDesc &op) const;
@@ -84,6 +87,10 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
       const std::string &og,
       std::unordered_set<std::string> *og_has_been_broadcast) const;
 
+  int GetOpDeviceID(
+      const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
+      const OpDesc &op) const;
+
   void InsertNCCLAllReduceOp(SSAGraph *result, const std::string &og) const;
 
   void CreateBroadcastOp(SSAGraph *result, const std::string &p_name,

paddle/fluid/framework/parallel_executor.cc

@@ -58,7 +58,7 @@ ParallelExecutor::ParallelExecutor(
     const std::unordered_set<std::string> &bcast_vars,
     const ProgramDesc &main_program, const std::string &loss_var_name,
     Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay,
-    bool use_default_grad_scale)
+    bool use_default_grad_scale, bool balance_parameter_opt_between_cards)
     : member_(new ParallelExecutorPrivate(places)) {
   member_->global_scope_ = scope;
 
@@ -93,11 +93,12 @@ ParallelExecutor::ParallelExecutor(
 #ifdef PADDLE_WITH_CUDA
   details::MultiDevSSAGraphBuilder builder(
       member_->places_, loss_var_name, params, member_->local_scopes_,
-      member_->nccl_ctxs_.get(), use_default_grad_scale);
+      member_->nccl_ctxs_.get(), use_default_grad_scale,
+      balance_parameter_opt_between_cards);
 #else
-  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
-                                           params, member_->local_scopes_,
-                                           use_default_grad_scale);
+  details::MultiDevSSAGraphBuilder builder(
+      member_->places_, loss_var_name, params, member_->local_scopes_,
+      use_default_grad_scale, balance_parameter_opt_between_cards);
 #endif
   auto graph = builder.Build(main_program);
 

paddle/fluid/framework/parallel_executor.h

@@ -40,7 +40,8 @@ class ParallelExecutor {
                             const ProgramDesc& main_program,
                             const std::string& loss_var_name, Scope* scope,
                             const std::vector<Scope*>& local_scopes,
-                            bool allow_op_delay, bool use_default_grad_scale);
+                            bool allow_op_delay, bool use_default_grad_scale,
+                            bool balance_parameter_opt_between_cards);
 
   ~ParallelExecutor();
 

paddle/fluid/pybind/pybind.cc

@@ -502,11 +502,13 @@ All parameter, weight, gradient are variables in Paddle.
               const std::unordered_set<std::string> &bcast_vars,
               const ProgramDesc &main_program, const std::string &loss_var_name,
               Scope *scope, std::vector<Scope *> &local_scopes,
-              bool allow_op_delay, bool use_default_grad_scale) {
+              bool allow_op_delay, bool use_default_grad_scale,
+              bool balance_parameter_opt_between_cards) {
              new (&self) ParallelExecutor(
                  num_threads, use_event, places, params, bcast_vars,
                  main_program, loss_var_name, scope, local_scopes,
-                 allow_op_delay, use_default_grad_scale);
+                 allow_op_delay, use_default_grad_scale,
+                 balance_parameter_opt_between_cards);
            })
       .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
       // NOTE: even we return a vec<Scope*>* to Python use reference policy.

python/paddle/fluid/parallel_executor.py

@@ -30,7 +30,8 @@ class ParallelExecutor(object):
                  num_threads=None,
                  allow_op_delay=False,
                  share_vars_from=None,
-                 use_default_grad_scale=True):
+                 use_default_grad_scale=True,
+                 balance_parameter_opt_between_cards=False):
         """
         ParallelExecutor can run program in parallel.
 
@@ -51,6 +52,9 @@ class ParallelExecutor(object):
                 gradients of each device and scaled gradients would be
                 aggregated. Otherwise, a customized scale value should be fed
                 to the network.
+            balance_parameter_opt_between_cards(bool, default True): Whether
+                updating different gradients on different cards. Currently, it
+                is not recommended.
 
         Returns:
             A ParallelExecutor object.
@@ -129,7 +133,8 @@ class ParallelExecutor(object):
             scope,
             local_scopes,
             allow_op_delay,
-            use_default_grad_scale)
+            use_default_grad_scale,
+            balance_parameter_opt_between_cards)
 
         self.scope = scope
 

python/paddle/fluid/tests/unittests/test_parallel_executor.py

@@ -205,7 +205,8 @@ class TestParallelExecutorBase(unittest.TestCase):
                                   allow_op_delay=False,
                                   feed_dict=None,
                                   seed=None,
-                                  use_parallel_executor=True):
+                                  use_parallel_executor=True,
+                                  balance_parameter_opt_between_cards=False):
         def run_executor(exe, feed, fetch_list, program=None):
             if isinstance(exe, fluid.ParallelExecutor):
                 res = exe.run(fetch_list=fetch_list, feed=feed)
@@ -234,7 +235,11 @@ class TestParallelExecutorBase(unittest.TestCase):
 
             if use_parallel_executor:
                 exe = fluid.ParallelExecutor(
-                    True, loss_name=loss.name, allow_op_delay=allow_op_delay)
+                    True,
+                    loss_name=loss.name,
+                    allow_op_delay=allow_op_delay,
+                    balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+                )
             else:
                 exe = fluid.Executor(place=place)
 
@@ -280,20 +285,27 @@ class TestMNIST(TestParallelExecutorBase):
             fluid.recordio_writer.convert_reader_to_recordio_file(
                 './mnist.recordio', reader, feeder)
 
-    def check_simple_fc_convergence(self):
+    def check_simple_fc_convergence(self, balance_parameter_opt_between_cards):
         self.check_network_convergence(simple_fc_net)
         self.check_network_convergence(simple_fc_net, allow_op_delay=True)
 
         img = np.zeros(shape=[32, 784], dtype='float32')
         label = np.ones(shape=[32, 1], dtype='int64')
         self.check_network_convergence(
-            simple_fc_net, feed_dict={"image": img,
-                                      "label": label})
+            simple_fc_net,
+            feed_dict={"image": img,
+                       "label": label},
+            balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+        )
 
     def test_simple_fc(self):
-        self.check_simple_fc_convergence()
+        self.check_simple_fc_convergence(False)
+
+    def test_simple_fc_with_new_strategy(self):
+        self.check_simple_fc_convergence(True)
 
-    def check_simple_fc_parallel_accuracy(self):
+    def check_simple_fc_parallel_accuracy(self,
+                                          balance_parameter_opt_between_cards):
         img = np.zeros(shape=[32, 784], dtype='float32')
         label = np.ones(shape=[32, 1], dtype='int64')
         single_first_loss, single_last_loss = self.check_network_convergence(
@@ -307,7 +319,9 @@ class TestMNIST(TestParallelExecutorBase):
             seed=1000,
             feed_dict={"image": img,
                        "label": label},
-            use_parallel_executor=True)
+            use_parallel_executor=True,
+            balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+        )
 
         for p_f in parallel_first_loss:
             self.assertAlmostEquals(p_f, single_first_loss[0], delta=1e-6)
@@ -315,18 +329,28 @@ class TestMNIST(TestParallelExecutorBase):
             self.assertAlmostEquals(p_l, single_last_loss[0], delta=1e-6)
 
     def test_simple_fc_parallel_accuracy(self):
-        self.check_simple_fc_parallel_accuracy()
+        self.check_simple_fc_parallel_accuracy(False)
 
-    def check_batchnorm_fc_convergence(self):
+    def test_simple_fc_parallel_accuracy_with_new_strategy(self):
+        self.check_simple_fc_parallel_accuracy(True)
+
+    def check_batchnorm_fc_convergence(self,
+                                       balance_parameter_opt_between_cards):
         self.check_network_convergence(fc_with_batchnorm)
         img = np.zeros(shape=[32, 784], dtype='float32')
         label = np.ones(shape=[32, 1], dtype='int64')
         self.check_network_convergence(
-            fc_with_batchnorm, feed_dict={"image": img,
-                                          "label": label})
+            fc_with_batchnorm,
+            feed_dict={"image": img,
+                       "label": label},
+            balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+        )
 
     def test_batchnorm_fc(self):
-        self.check_batchnorm_fc_convergence()
+        self.check_batchnorm_fc_convergence(False)
+
+    def test_batchnorm_fc_with_new_strategy(self):
+        self.check_batchnorm_fc_convergence(True)
 
 
 class TestResnet(TestParallelExecutorBase):
@@ -348,17 +372,22 @@ class TestResnet(TestParallelExecutorBase):
     #         fluid.recordio_writer.convert_reader_to_recordio_file(
     #             "./flowers.recordio", reader, feeder, compressor=fluid.core.RecordIOWriter.Compressor.NoCompress)
 
-    def check_resnet_convergence(self):
+    def check_resnet_convergence(self, balance_parameter_opt_between_cards):
         import functools
         batch_size = 2
         self.check_network_convergence(
             functools.partial(
                 SE_ResNeXt50Small, batch_size=batch_size),
             iter=20,
-            batch_size=batch_size)
+            batch_size=batch_size,
+            balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+        )
 
     def test_resnet(self):
-        self.check_resnet_convergence()
+        self.check_resnet_convergence(False)
+
+    def test_resnet_with_new_strategy(self):
+        self.check_resnet_convergence(True)
 
 
 class ModelHyperParams(object):
@@ -519,7 +548,7 @@ class TestTransformer(TestParallelExecutorBase):
 
 
 class ParallelExecutorTestingDuringTraining(unittest.TestCase):
-    def check_network_convergence(self):
+    def check_network_convergence(self, balance_parameter_opt_between_cards):
         main = fluid.Program()
         startup = fluid.Program()
         with fluid.program_guard(main, startup):
@@ -539,12 +568,18 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
             feed_dict = {'image': image, 'label': label}
 
             train_exe = fluid.ParallelExecutor(
-                use_cuda=True, loss_name=loss.name, main_program=main)
+                use_cuda=True,
+                loss_name=loss.name,
+                main_program=main,
+                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+            )
 
             test_exe = fluid.ParallelExecutor(
                 use_cuda=True,
                 main_program=test_program,
-                share_vars_from=train_exe)
+                share_vars_from=train_exe,
+                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+            )
 
             for i in xrange(5):
                 test_loss, = test_exe.run([loss.name], feed=feed_dict)
@@ -558,8 +593,11 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
                     "Train loss: " + str(train_loss) + "\n Test loss:" +
                     str(test_loss))
 
-    def test_parallel(self):
-        self.check_network_convergence()
+    def test_parallel_testing(self):
+        self.check_network_convergence(False)
+
+    def test_parallel_testing_with_new_strategy(self):
+        self.check_network_convergence(True)
 
 
 import paddle.dataset.conll05 as conll05
@@ -579,7 +617,7 @@ embedding_name = 'emb'
 
 
 def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
-            is_sparse, **ignored):
+            is_sparse, balance_parameter_opt_between_cards, **ignored):
     # 8 features
     predicate_embedding = fluid.layers.embedding(
         input=predicate,
@@ -648,7 +686,9 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
 
 
 class TestCRFModel(unittest.TestCase):
-    def check_network_convergence(self, is_sparse):
+    def check_network_convergence(self,
+                                  is_sparse,
+                                  balance_parameter_opt_between_cards=False):
         main = fluid.Program()
         startup = fluid.Program()
         with fluid.program_guard(main, startup):
@@ -696,7 +736,11 @@ class TestCRFModel(unittest.TestCase):
             exe = fluid.Executor(place)
             exe.run(startup)
 
-            pe = fluid.ParallelExecutor(use_cuda=True, loss_name=avg_cost.name)
+            pe = fluid.ParallelExecutor(
+                use_cuda=True,
+                loss_name=avg_cost.name,
+                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
+            )
 
             feeder = fluid.DataFeeder(
                 feed_list=[
@@ -718,6 +762,14 @@ class TestCRFModel(unittest.TestCase):
     def test_update_dense_parameter(self):
         self.check_network_convergence(is_sparse=False)
 
+    def test_update_sparse_parameter_with_new_strategy(self):
+        self.check_network_convergence(
+            is_sparse=False, balance_parameter_opt_between_cards=True)
+
+    def test_update_dense_parameter_with_new_strategy(self):
+        self.check_network_convergence(
+            is_sparse=False, balance_parameter_opt_between_cards=True)
+
 
 # test fetch all the variables of global_block