add clear_grad for amp sample code

paddle/fluid/imperative/layer.cc

@@ -187,6 +187,7 @@ size_t VarBase::GradOpNum() const {
 }
 
 void VarBase::ClearGradient() {
+  VLOG(4) << "ClearGradient " << Name();
   if (grad_var_) {
     if (grad_var_->Var().IsType<framework::SelectedRows>()) {
       auto* grad_t =

python/paddle/amp/grad_scaler.py

@@ -62,6 +62,7 @@ class GradScaler(AmpScaler):
             scaled = scaler.scale(loss)  # scale the loss 
             scaled.backward()            # do backward
             scaler.minimize(optimizer, scaled)  # update parameters     
+            optimizer.clear_grad()
     """
 
     def __init__(self,
@@ -105,6 +106,7 @@ def scale(self, var):
                 scaled = scaler.scale(loss)  # scale the loss 
                 scaled.backward()            # do backward
                 scaler.minimize(optimizer, scaled)  # update parameters  
+                optimizer.clear_grad()
         """
         return super(GradScaler, self).scale(var)
 
@@ -140,5 +142,6 @@ def minimize(self, optimizer, *args, **kwargs):
                 scaled = scaler.scale(loss)  # scale the loss 
                 scaled.backward()            # do backward
                 scaler.minimize(optimizer, scaled)  # update parameters  
+                optimizer.clear_grad()
         """
         return super(GradScaler, self).minimize(optimizer, *args, **kwargs)