Fix/auto prune error on leaf (#20056)

* test=develop, fix docker with paddle nccl problem

* test=develop, Add Variable api and refine dygraph related API

* test=develop, Add Variable api and refine dygraph related API

* test=develop, refine test for new api and error info

* test=develop, refine error info and test_layers

* test=develop, add API.spec

* test=devleop, fix to_string python2 and python3 compat error and refien doc

* test=devleop, add API spec

* test=devleop, update API spec

* test=devleop, update API spec

* test=develop, invoke ci

* test=develop, fix example code

* test=develop, update API spec

* test=develop, fix auto_prune_error_on_leaf

* test=develop, fix auto prune error on loss stop_gradient

* test=develop, remove useless error check

* test=develop, add more ut for sorted gradient

paddle/fluid/imperative/engine.cc

@@ -44,8 +44,9 @@ void BasicEngine::Init(VarBase* var, const detail::BackwardStrategy& strategy) {
   const std::vector<OpBase*> ops = var->GradVarBase()->GradOps();
   var->ClearGradOps();
 
-  if (ops.empty()) {
-    VLOG(3) << "Skip auto grad since there is no grad op for var: "
+  if (ops.empty() || var->OverridedStopGradient()) {
+    VLOG(3) << "Skip auto grad since there is no grad op for var or loss is "
+               "stop_gradient=True: "
             << var->Name();
     return;
   } else {

paddle/fluid/imperative/gradient_accumulator.cc

@@ -116,13 +116,23 @@ void EagerGradientAccumulator::Add(std::shared_ptr<VarBase> var,
   } else {
     if (!var_->Var().IsInitialized() ||
         !var_->Var().Get<framework::LoDTensor>().IsInitialized()) {
-      VLOG(6) << "Set StopGradient Grad: " << var->Name() << " as zero";
+      VLOG(6) << "Set StopGradient Grad: " << var_->Name() << " as zero ";
+
       auto* dev_ctx = platform::DeviceContextPool::Instance().Get(place);
+      if (!var_->Var().IsInitialized()) {
+        auto* tensor = var_->MutableVar()->GetMutable<framework::LoDTensor>();
+        VLOG(6) << "Dims of " << var_->Name() << " is set as: "
+                << var->Var().Get<framework::LoDTensor>().dims();
+        tensor->Resize(var->Var().Get<framework::LoDTensor>().dims());
+        tensor->mutable_data(place, var->DataType());
+        operators::math::set_constant(*dev_ctx, tensor, 0.0);
+      } else {
         auto* tensor = var_->MutableVar()->GetMutable<framework::LoDTensor>();
         tensor->mutable_data(place, var->DataType());
         operators::math::set_constant(*dev_ctx, tensor, 0.0);
       }
     }
+  }
   ++cur_cnt_;
 }
 
@@ -162,10 +172,19 @@ void SortedGradientAccumulator::Add(std::shared_ptr<VarBase> var,
         !var_->Var().Get<framework::LoDTensor>().IsInitialized()) {
       VLOG(6) << "Set StopGradient Grad: " << var->Name() << " as zero";
       auto* dev_ctx = platform::DeviceContextPool::Instance().Get(place);
+      if (!var_->Var().IsInitialized()) {
+        auto* tensor = var_->MutableVar()->GetMutable<framework::LoDTensor>();
+        VLOG(6) << "Dims of " << var_->Name() << " is set as: "
+                << var->Var().Get<framework::LoDTensor>().dims();
+        tensor->Resize(var->Var().Get<framework::LoDTensor>().dims());
+        tensor->mutable_data(place, var->DataType());
+        operators::math::set_constant(*dev_ctx, tensor, 0.0);
+      } else {
         auto* tensor = var_->MutableVar()->GetMutable<framework::LoDTensor>();
         tensor->mutable_data(place, var->DataType());
         operators::math::set_constant(*dev_ctx, tensor, 0.0);
       }
+    }
     // looks like tmp_grad_vars will not have any member but just in case
     tmp_grad_vars_.clear();
   }

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

@@ -241,6 +241,72 @@ class TestImperativeAutoPrune(unittest.TestCase):
             self.assertTrue((fc._w.gradient() == 0).all())
             self.assertTrue((out1.gradient() == 0).all())
 
+    def test_auto_prune8(self):
+        with fluid.dygraph.guard():
+            value0 = np.arange(26).reshape(2, 13).astype("float32")
+            value1 = np.arange(6).reshape(2, 3).astype("float32")
+            value2 = np.arange(10).reshape(2, 5).astype("float32")
+            fc = fluid.FC("fc1", size=5, dtype="float32")
+            fc2 = fluid.FC("fc2", size=3, dtype="float32")
+            a = fluid.dygraph.to_variable(value0)
+            b = fluid.dygraph.to_variable(value1)
+            c = fluid.dygraph.to_variable(value2)
+            out1 = fc(a)
+            fc_origin = fc._w.numpy()
+            out2 = fc2(out1)
+            fc2_origin = fc2._w.numpy()
+            fc2._w.stop_gradient = True
+            out2.backward()
+            optimizer = fluid.optimizer.SGD(learning_rate=0.003)
+            optimizer.minimize(out2)
+            self.assertTrue(np.array_equal(fc2_origin, fc2._w.numpy()))
+            self.assertFalse(np.array_equal(fc_origin, fc._w.numpy()))
+
+    def test_auto_prune9(self):
+        with fluid.dygraph.guard():
+            value0 = np.arange(26).reshape(2, 13).astype("float32")
+            value1 = np.arange(6).reshape(2, 3).astype("float32")
+            value2 = np.arange(10).reshape(2, 5).astype("float32")
+            fc = fluid.FC("fc1", size=5, dtype="float32")
+            fc2 = fluid.FC("fc2", size=3, dtype="float32")
+            a = fluid.dygraph.to_variable(value0)
+            b = fluid.dygraph.to_variable(value1)
+            c = fluid.dygraph.to_variable(value2)
+            out1 = fc(a)
+            fc_origin = fc._w.numpy()
+            out2 = fc2(out1)
+            fc2_origin = fc2._w.numpy()
+            out2.stop_gradient = True
+            out2.backward()
+            optimizer = fluid.optimizer.SGD(learning_rate=0.003)
+            optimizer.minimize(out2)
+            self.assertTrue(np.array_equal(fc2_origin, fc2._w.numpy()))
+            self.assertTrue(np.array_equal(fc_origin, fc._w.numpy()))
+            try:
+                fc2._w.gradient()
+            except ValueError as e:
+                assert type(e) == ValueError
+
+    def test_auto_prune10(self):
+        with fluid.dygraph.guard():
+            value0 = np.arange(26).reshape(2, 13).astype("float32")
+            value1 = np.arange(6).reshape(2, 3).astype("float32")
+            value2 = np.arange(10).reshape(2, 5).astype("float32")
+            fc = fluid.FC("fc1", size=5, dtype="float32")
+            fc2 = fluid.FC("fc2", size=3, dtype="float32")
+            a = fluid.dygraph.to_variable(value0)
+            b = fluid.dygraph.to_variable(value1)
+            c = fluid.dygraph.to_variable(value2)
+            out1 = fc(a)
+            out2 = fc2(b)
+            out1.stop_gradient = True
+            out = fluid.layers.concat(input=[out1, out2, c], axis=1)
+            backward_strategy = fluid.dygraph.BackwardStrategy()
+            backward_strategy.sort_sum_gradient = True
+            out.backward(backward_strategy)
+            self.assertTrue((fc._w.gradient() == 0).all())
+            self.assertTrue((out1.gradient() == 0).all())
+
     def test_auto_prune_with_optimizer(self):
         vocab_size = 100
         size = 20