!14065 New add grad accumulation for network.

From: @linqingke
Reviewed-by: @guoqi1024,@xu-yfei
Signed-off-by: @guoqi1024

mindspore/ccsrc/backend/optimizer/pass/optimize_dependence.cc

@@ -27,23 +27,86 @@
 namespace mindspore {
 namespace opt {
 constexpr auto kSingleInputIndex = 1;
+constexpr auto kIsolatedDependRealInputIndex = 0;
+constexpr auto kIsolatedDependVirtualInputIndex = 1;
 namespace {
+CNodePtr CreateNewDependNode(const FuncGraphPtr &func_graph, const CNodePtr &cnode,
+                             const std::vector<AnfNodePtr> &new_depend_inputs) {
+  MS_EXCEPTION_IF_NULL(func_graph);
+  MS_EXCEPTION_IF_NULL(cnode);
+  auto kernel_graph = func_graph->cast<std::shared_ptr<session::KernelGraph>>();
+  CNodePtr new_depend = nullptr;
+  if (kernel_graph == nullptr) {
+    new_depend = func_graph->NewCNode(new_depend_inputs);
+    MS_EXCEPTION_IF_NULL(new_depend);
+    new_depend->set_abstract(cnode->abstract());
+    new_depend->set_scope(cnode->scope());
+  } else {
+    new_depend = kernel_graph->NewCNode(cnode);
+    MS_EXCEPTION_IF_NULL(new_depend);
+    new_depend->set_inputs(new_depend_inputs);
+  }
+  func_graph->manager()->Replace(cnode, new_depend);
+  return new_depend;
+}
+
+CNodePtr CheckIsolatedVirtualNode(const CNodePtr &cnode) {
+  MS_EXCEPTION_IF_NULL(cnode);
+  if (AnfAlgo::GetCNodeName(cnode) != prim::kPrimDepend->name()) {
+    return nullptr;
+  }
+  auto virtual_input_op = AnfAlgo::GetInputNode(cnode, kIsolatedDependVirtualInputIndex);
+  if (!AnfAlgo::CheckPrimitiveType(virtual_input_op, prim::kPrimUpdateState)) {
+    return nullptr;
+  }
+  auto real_input_op = AnfAlgo::GetInputNode(cnode, kIsolatedDependRealInputIndex);
+  if (!real_input_op->isa<CNode>()) {
+    return nullptr;
+  }
+  auto real_input_cnode = real_input_op->cast<CNodePtr>();
+  return real_input_cnode;
+}
+
+AnfNodePtr EliminateIsolatedVirtualNodeInput(const FuncGraphPtr &func_graph, const CNodePtr &cnode,
+                                             const CNodePtr &eliminate_node) {
+  MS_EXCEPTION_IF_NULL(func_graph);
+  MS_EXCEPTION_IF_NULL(cnode);
+  MS_EXCEPTION_IF_NULL(eliminate_node);
+  auto replace_node = eliminate_node->input(kSingleInputIndex);
+  std::vector<AnfNodePtr> new_depend_inputs = cnode->inputs();
+  new_depend_inputs[kIsolatedDependRealInputIndex + 1] = replace_node;
+  auto new_cnode = CreateNewDependNode(func_graph, cnode, new_depend_inputs);
+  auto new_node = new_cnode->cast<AnfNodePtr>();
+  return new_node;
+}
+
 AnfNodePtr GetReplaceNode(const FuncGraphPtr &func_graph, const AnfNodePtr &node) {
+  MS_EXCEPTION_IF_NULL(func_graph);
   MS_EXCEPTION_IF_NULL(node);
   if (!node->isa<CNode>()) {
     return nullptr;
   }
   auto cnode = node->cast<CNodePtr>();
   MS_EXCEPTION_IF_NULL(cnode);
-  string op_name = AnfAlgo::GetCNodeName(cnode);
+  auto replace_cnode = cnode;
+  // Process updatestate and depend as isolated node env.
+  auto isolated_cnode = CheckIsolatedVirtualNode(replace_cnode);
+  if (isolated_cnode != nullptr) {
+    replace_cnode = isolated_cnode;
+  }
+  string op_name = AnfAlgo::GetCNodeName(replace_cnode);
   // Currently we only eliminate transdata or cast nodes.
   if (op_name != kTransDataOpName && op_name != prim::kPrimCast->name()) {
     return nullptr;
   }
-  if (!IsNotRealUsedByOthers(func_graph, cnode)) {
+  if (!IsNotRealUsedByOthers(func_graph, replace_cnode)) {
     return nullptr;
   }
-  CheckCNodeInputSize(cnode, kSingleInputIndex);
+  CheckCNodeInputSize(replace_cnode, kSingleInputIndex);
+  if (isolated_cnode != nullptr) {
+    auto new_depend_node = EliminateIsolatedVirtualNodeInput(func_graph, cnode, replace_cnode);
+    return new_depend_node;
+  }
   return cnode->input(kSingleInputIndex);
 }
 
@@ -137,20 +200,10 @@ const AnfNodePtr OptimizeDependence::Process(const FuncGraphPtr &func_graph, con
     return nullptr;
   }
   new_depend_inputs[replace_index] = replace_node;
-  // Because depend's input has been changed, so a new depend(UpdateState) node will be created to replaced the old one.
+  auto new_depend = CreateNewDependNode(func_graph, depend_cnode, new_depend_inputs);
-  auto kernel_graph = func_graph->cast<std::shared_ptr<session::KernelGraph>>();
+  if (new_depend == nullptr) {
-  CNodePtr new_depend = nullptr;
+    return nullptr;
-  if (kernel_graph == nullptr) {
-    new_depend = func_graph->NewCNode(new_depend_inputs);
-    MS_EXCEPTION_IF_NULL(new_depend);
-    new_depend->set_abstract(depend_cnode->abstract());
-    new_depend->set_scope(depend_cnode->scope());
-  } else {
-    new_depend = kernel_graph->NewCNode(depend_cnode);
-    MS_EXCEPTION_IF_NULL(new_depend);
-    new_depend->set_inputs(new_depend_inputs);
   }
-  func_graph->manager()->Replace(depend_cnode, new_depend);
   return nullptr;
 }
 

mindspore/ccsrc/frontend/optimizer/auto_monad_eliminate.cc

@@ -107,6 +107,20 @@ std::vector<std::vector<size_t>> SplitGroup(const std::vector<AnfNodePtr> &topos
 // a = Load(para1, u1)
 // ...
 // b = Load(para1, u2)
+// u3 = UpdateState(u2, b)
+//==>
+// delete the UpdateState
+void DeleteLoadUserUpdateState(const FuncGraphManagerPtr &manager, const AnfNodePtr &load_user,
+                               const AnfNodePtr &load) {
+  const auto &load_cnode = load->cast<CNodePtr>();
+  const auto &u = load_cnode->input(2);
+  manager->Replace(load_user, u);
+}
+
+// Pattern2======================================
+// a = Load(para1, u1)
+// ...
+// b = Load(para1, u2)
 // t = make_tuple(x, b)
 // u3 = UpdateState(u2, t)
 //==>
@@ -127,7 +141,7 @@ void DeleteLoadUserMakeTuple(const FuncGraphManagerPtr &manager, const CNodePtr
   manager->Replace(make_tuple, other_input);
 }
 
-// Pattern2======================================
+// Pattern3======================================
 // a = Load(para1, u1)
 // ...
 // b = Load(para1, u2)
@@ -153,6 +167,11 @@ void ReplaceLoadUserMakeTuple(const FuncGraphManagerPtr &manager, const FuncGrap
 void ReplaceLoadUser(const FuncGraphManagerPtr &manager, const FuncGraphPtr &fg, const AnfNodePtr &load) {
   auto load_users = manager->node_users()[load];
   for (const auto &load_user : load_users) {
+    // Pattern1
+    if (IsPrimitiveCNode(load_user.first, prim::kPrimUpdateState)) {
+      DeleteLoadUserUpdateState(manager, load_user.first, load);
+      continue;
+    }
     if (IsPrimitiveCNode(load_user.first, prim::kPrimMakeTuple)) {
       const auto &make_tuple = load_user.first->cast<CNodePtr>();
       auto &maketuple_users = manager->node_users()[make_tuple];
@@ -161,12 +180,12 @@ void ReplaceLoadUser(const FuncGraphManagerPtr &manager, const FuncGraphPtr &fg,
       if (!maketuple_as_input_of_update) {
         continue;
       }
-      // Pattern1
+      // Pattern2
       if (make_tuple->size() == 3) {
         DeleteLoadUserMakeTuple(manager, make_tuple, load);
         continue;
       }
-      // Pattern2
+      // Pattern3
       if (make_tuple->size() > 3) {
         ReplaceLoadUserMakeTuple(manager, fg, make_tuple, load);
       }

mindspore/nn/wrap/cell_wrapper.py

@@ -17,10 +17,11 @@ from types import FunctionType, MethodType
 
 from mindspore.parallel._utils import (_get_device_num, _get_gradients_mean,
                                        _get_parallel_mode)
-from mindspore.context import ParallelMode
+from mindspore.context import ParallelMode, get_auto_parallel_context
 from mindspore._checkparam import Validator as validator
 from ...common import dtype as mstype
 from ...common.parameter import Parameter, ParameterTuple
+from ...common.tensor import Tensor
 from ...ops import composite as C
 from ...ops import functional as F
 from ...ops import operations as P
@@ -62,6 +63,19 @@ def _tensors_cast_datatype(datatype, param):
     """
     return F.cast(param, datatype)
 
+_gradient_accumulation_op = C.MultitypeFuncGraph("gradient_accumulation_op")
+
+@_gradient_accumulation_op.register("Int64", "Tensor", "Tensor")
+def _cumulative_grad(accumulation_step, cumulative_grad, grad):
+    """Apply gradient accumulation to cumulative grad."""
+    return P.AssignAdd()(cumulative_grad, grad / accumulation_step)
+
+_gradient_clear_op = C.MultitypeFuncGraph("gradient_clear_op")
+
+@_gradient_clear_op.register("Tensor")
+def  _clear_grad(cumulative_grad):
+    zero_grad = P.ZerosLike()(cumulative_grad)
+    return F.assign(cumulative_grad, zero_grad)
 
 class WithLossCell(Cell):
     r"""
@@ -347,15 +361,28 @@ class TrainOneStepCell(Cell):
             self.mean = _get_gradients_mean()
             self.degree = _get_device_num()
             self.grad_reducer = DistributedGradReducer(self.weights, self.mean, self.degree)
+        self.use_grad_accumulation = False
+        if self.parallel_mode in (ParallelMode.DATA_PARALLEL, ParallelMode.STAND_ALONE):
+            self.use_grad_accumulation = True
+        if self.use_grad_accumulation:
+            self.max_accumulation_step = get_auto_parallel_context("grad_accumulation_step")
+            if self.max_accumulation_step <= 1:
+                self.max_accumulation_step = 1
+                self.use_grad_accumulation = False
+        if self.use_grad_accumulation:
+            self.grad_accumulation = GradientAccumulation(self.max_accumulation_step, self.optimizer)
 
     def construct(self, *inputs):
         weights = self.weights
         loss = self.network(*inputs)
         sens = P.Fill()(P.DType()(loss), P.Shape()(loss), self.sens)
-        sens = F.depend(sens, loss)
         grads = self.grad(self.network, weights)(*inputs, sens)
         grads = self.grad_reducer(grads)
-        loss = F.depend(loss, self.optimizer(grads))
+
+        if self.use_grad_accumulation:
+            loss = self.grad_accumulation(loss, grads)
+        else:
+            loss = F.depend(loss, self.optimizer(grads))
         return loss
 
 
@@ -557,3 +584,34 @@ class _BroadCastCell(Cell):
         params = self.broadcast(params)
         new_params = self.map_(F.partial(_cast_datatype), datatypes, params)
         return new_params
+
+class GradientAccumulation(Cell):
+    """
+    After accumulating the gradients of multiple steps, call to optimize its update.
+
+    Args:
+       max_accumulation_step (int): Steps to accumulate gradients.
+       optimizer(Cell):Optimizer used.
+    """
+    def __init__(self, max_accumulation_step, optimizer):
+        super(GradientAccumulation, self).__init__()
+        self._max_accumulation_step = max_accumulation_step
+        self.optimizer = optimizer
+        self.weights = optimizer.parameters
+        self.hyper_map = C.HyperMap()
+        self._grad_accumulation = self.weights.clone(prefix="grad_accumulation", init='zeros')
+        self._accumulation_step = Parameter(Tensor(0, dtype=mstype.int32), name="accumulation_step")
+
+    def construct(self, loss, grads):
+        loss = F.depend(loss, self.hyper_map(F.partial(_gradient_accumulation_op, self._max_accumulation_step),
+                                             self._grad_accumulation, grads))
+        self._accumulation_step += 1
+
+        if self._accumulation_step >= self._max_accumulation_step:
+            loss = F.depend(loss, self.optimizer(self._grad_accumulation))
+            self._accumulation_step = 0
+
+        if self._accumulation_step == 0:
+            loss = F.depend(loss, self.hyper_map(F.partial(_gradient_clear_op), self._grad_accumulation))
+
+        return loss

mindspore/nn/wrap/loss_scale.py

@@ -319,7 +319,10 @@ class TrainOneStepWithLossScaleCell(TrainOneStepCell):
         overflow = self.process_loss_scale(cond)
         # if there is no overflow, do optimize
         if not overflow:
-            loss = F.depend(loss, self.optimizer(grads))
+            if self.use_grad_accumulation:
+                loss = self.grad_accumulation(loss, grads)
+            else:
+                loss = F.depend(loss, self.optimizer(grads))
         return loss, cond, scaling_sens
 
     def set_sense_scale(self, sens):