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

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -252,9 +252,9 @@ std::vector<ir::Node *> SortOpsAndDelayOptimizeOp(const ir::Graph &graph) {
   std::vector<ir::Node *> sorted_ret;
   for (size_t i = 0; i < ret.size(); ++i) {
     if (i < last_backward) {
-      if (boost::get<int>(ret[i]->Op()->GetAttr(
-              OpProtoAndCheckerMaker::OpRoleAttrName())) ==
-          static_cast<int>(OpRole::kOptimize)) {
+      if (static_cast<bool>(boost::get<int>(ret[i]->Op()->GetAttr(
+                                OpProtoAndCheckerMaker::OpRoleAttrName())) &
+                            static_cast<int>(OpRole::kOptimize))) {
         optimize_ops.push_back(ret[i]);
       } else {
         sorted_ret.push_back(ret[i]);

paddle/fluid/framework/op_proto_maker.cc

@@ -71,6 +71,8 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
            static_cast<int>(OpRole::kLoss) | static_cast<int>(OpRole::kForward),
            static_cast<int>(OpRole::kLoss) |
                static_cast<int>(OpRole::kBackward),
+           static_cast<int>(OpRole::kOptimize) |
+               static_cast<int>(OpRole::kLRSched),
            static_cast<int>(OpRole::kNotSpecified)})
       .SetDefault(static_cast<int>(OpRole::kNotSpecified));
   AddAttr<std::vector<std::string>>(OpRoleVarAttrName(),

paddle/fluid/framework/op_proto_maker.h

@@ -20,6 +20,9 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+//////////////////////////
+// Don't add more roles to make this too complicated!
+//////////////////////////
 enum class OpRole {
   kForward = 0x0000,
   kBackward = 0x0001,

paddle/fluid/framework/parallel_executor.cc

@@ -156,12 +156,6 @@ ParallelExecutor::ParallelExecutor(
                            params, member_->local_scopes_, member_->use_cuda_);
 #endif
 
-  // If the loss_var_name is given, the number of graph should be only one.
-  if (loss_var_name.size()) {
-    PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1,
-                      "The number of graph should be only one");
-  }
-
   if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
     member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
         exec_strategy, member_->local_scopes_, places, std::move(graph)));

python/paddle/fluid/clip.py

@@ -333,7 +333,8 @@ def append_gradient_clip_ops(param_grads):
     for p, g in param_grads:
         if g is None:
             continue
-        with p.block.program._optimized_guard([p, g]):
+        with p.block.program._optimized_guard(
+            [p, g]), framework.name_scope('append_clip'):
             clip_attr = getattr(p, 'gradient_clip_attr', NullGradientClipAttr())
             if clip_attr is None:
                 clip_attr = NullGradientClipAttr()
@@ -348,7 +349,8 @@ def append_gradient_clip_ops(param_grads):
     for p, g in param_grads:
         if g is None:
             continue
-        with p.block.program._optimized_guard([p, g]):
+        with p.block.program._optimized_guard(
+            [p, g]), framework.name_scope('append_graident_clip'):
             res.append(clip_attr._create_operators(param=p, grad=g))
 
     return res

python/paddle/fluid/framework.py

@@ -1496,6 +1496,9 @@ class Program(object):
             >>> with program._optimized_guard([p,g]):
             >>>     p = p - 0.001 * g
         """
+        tmp_role = self._current_role
+        tmp_var = self._op_role_var
+
         OpRole = core.op_proto_and_checker_maker.OpRole
         self._current_role = OpRole.Optimize
         self._op_role_var = [
@@ -1503,11 +1506,11 @@ class Program(object):
             for var in param_and_grads
         ]
         yield
-        self._op_role_var = []
-        self._current_role = OpRole.Forward
+        self._op_role_var = tmp_var
+        self._current_role = tmp_role
 
     @contextlib.contextmanager
-    def _lr_schedule_guard(self):
+    def _lr_schedule_guard(self, is_with_opt=False):
         """
         A with guard to set :code:`LRSched` :code:`OpRole` and
         :code:`OpRoleVar` automatically. The :code:`OpRoleVar` is
@@ -1515,6 +1518,10 @@ class Program(object):
 
         Notes: This is a very low level API. Users should not use it directly.
 
+        Args:
+            is_with_opt: Only set to true if these ops a in the middle
+                 of a bunch of optimize ops so that it can be treated
+                 correctly. For example, sgd->lr_op->sgd->lr_op->sgd.
 
         Examples:
 
@@ -1528,6 +1535,8 @@ class Program(object):
 
         OpRole = core.op_proto_and_checker_maker.OpRole
         self._current_role = OpRole.LRSched
+        if is_with_opt:
+            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
         # TODO(typhoonzero): how to set target learning rate var
         self._op_role_var = []
         yield

python/paddle/fluid/optimizer.py

@@ -111,7 +111,9 @@ class Optimizer(object):
             if param_lr == 1.0:
                 return self._global_learning_rate()
             else:
-                with default_main_program()._lr_schedule_guard():
+                with default_main_program()._lr_schedule_guard(
+                        is_with_opt=True), framework.name_scope(
+                            'scale_with_param_lr'):
                     return self._global_learning_rate() * param_lr
 
     def _create_accumulators(self, block, parameters):
@@ -602,7 +604,8 @@ class AdamOptimizer(Optimizer):
         for param, grad in param_and_grads:
             if grad is None:
                 continue
-            with param.block.program._optimized_guard([param, grad]):
+            with param.block.program._optimized_guard(
+                [param, grad]), name_scope("optimizer"):
                 beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                                       param)
                 beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str,
@@ -740,7 +743,8 @@ class AdamaxOptimizer(Optimizer):
         for param, grad in parameters_and_grads:
             if grad is None:
                 continue
-            with param.block.program._optimized_guard([param, grad]):
+            with param.block.program._optimized_guard(
+                [param, grad]), name_scope('adamx'):
                 beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str,
                                                       param)
                 main_block.append_op(
@@ -1279,7 +1283,8 @@ class ModelAverage(Optimizer):
         for param, grad in self.params_grads:
             if grad is None:
                 continue
-            with param.block.program._optimized_guard([param, grad]):
+            with param.block.program._optimized_guard(
+                [param, grad]), name_scope('move_average'):
                 self._append_average_accumulate_op(param)
 
         self.apply_program = Program()

python/paddle/fluid/regularizer.py

@@ -47,7 +47,8 @@ def append_regularization_ops(parameters_and_grads, regularization=None):
         if grad is None:
             params_and_grads.append((param, grad))
             continue
-        with param.block.program._optimized_guard([param, grad]):
+        with param.block.program._optimized_guard(
+            [param, grad]), framework.name_scope('regularization'):
             regularization_term = None
             if param.regularizer is not None:
                 # Add variable for regularization term in grad block

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -78,9 +78,9 @@ if(WITH_DISTRIBUTE)
         set_tests_properties(test_dist_word2vec PROPERTIES TIMEOUT 200)
         py_test_modules(test_dist_se_resnext MODULES test_dist_se_resnext)
         set_tests_properties(test_dist_se_resnext PROPERTIES TIMEOUT 1000)
-
-        py_test_modules(test_dist_transformer MODULES test_dist_transformer)
-        set_tests_properties(test_dist_transformer PROPERTIES TIMEOUT 1000)
+        # FIXME(typhoonzero): add this back
+	#py_test_modules(test_dist_transformer MODULES test_dist_transformer)
+	#set_tests_properties(test_dist_transformer PROPERTIES TIMEOUT 1000)
     endif(NOT APPLE)
     py_test_modules(test_dist_transpiler MODULES test_dist_transpiler)
 endif()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -49,6 +49,7 @@ LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
 OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
 RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName(
 )
+OPT_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Optimize
 RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC
 DIST_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Dist
 LR_SCHED_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.LRSched
@@ -1717,8 +1718,10 @@ to transpile() call.")
         lr_ops = []
         block = self.origin_program.global_block()
         for op in block.ops:
-            if int(op.attr(RPC_OP_ROLE_ATTR_NAME)) == int(
-                    LR_SCHED_OP_ROLE_ATTR_VALUE):
+            role_id = int(op.attr(RPC_OP_ROLE_ATTR_NAME))
+            if role_id == int(LR_SCHED_OP_ROLE_ATTR_VALUE) or \
+                role_id == int(LR_SCHED_OP_ROLE_ATTR_VALUE) | \
+                    int(OPT_OP_ROLE_ATTR_VALUE):
                 lr_ops.append(op)
                 log("append lr op: ", op.type)
         return lr_ops