[BUG Fixs] adadelta lr support (#49732)

paddle/fluid/operators/optimizers/adadelta_op.cc

@@ -39,6 +39,7 @@ class AdadeltaOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("AvgSquaredGrad", "(Tensor) Input average of squared gradient");
     AddInput("AvgSquaredUpdate",
              "(Tensor) Input average of squared parameter updates");
+    AddInput("LearningRate", "(Tensor) Learning rate");
     AddInput("MasterParam", "FP32 master weight for AMP.").AsDispensable();
 
     AddOutput("ParamOut", "(Tensor) Output parameter");

paddle/fluid/pybind/eager_generator.h

@@ -220,7 +220,12 @@ std::map<std::string, std::set<std::string>> op_ins_map = {
     {"sgd", {"Param", "LearningRate", "Grad", "MasterParam"}},
     {"adagrad", {"Param", "Grad", "Moment", "LearningRate", "MasterParam"}},
     {"adadelta",
-     {"Param", "Grad", "AvgSquaredGrad", "AvgSquaredUpdate", "MasterParam"}},
+     {"Param",
+      "Grad",
+      "AvgSquaredGrad",
+      "AvgSquaredUpdate",
+      "LearningRate",
+      "MasterParam"}},
     {"graph_khop_sampler", {"Row", "Eids", "Col_Ptr", "X"}},
     {"nce",
      {"Input",

paddle/phi/api/yaml/legacy_ops.yaml

@@ -11,7 +11,7 @@
   backward : abs_grad
 
 - op : adadelta_
-  args : (Tensor param, Tensor grad, Tensor avg_squared_grad, Tensor avg_squared_update, Tensor master_param, float rho, float epsilon, bool multi_precision)
+  args : (Tensor param, Tensor grad, Tensor avg_squared_grad, Tensor avg_squared_update, Tensor learning_rate, Tensor master_param, float rho, float epsilon, bool multi_precision)
   output : Tensor(param_out), Tensor(moment_out), Tensor(inf_norm_out), Tensor(master_param_out)
   infer_meta :
     func : AdadeltaInferMeta

paddle/phi/infermeta/multiary.cc

@@ -40,6 +40,7 @@ void AdadeltaInferMeta(const MetaTensor& param,
                        const MetaTensor& grad,
                        const MetaTensor& avg_squared_grad,
                        const MetaTensor& avg_squared_update,
+                       const MetaTensor& learning_rate,
                        const MetaTensor& master_param,
                        float rho,
                        float epsilon,
@@ -48,6 +49,11 @@ void AdadeltaInferMeta(const MetaTensor& param,
                        MetaTensor* avg_squared_grad_out,
                        MetaTensor* avg_squared_update_out,
                        MetaTensor* master_param_out) {
+  auto lr_dims = learning_rate.dims();
+  PADDLE_ENFORCE_EQ(
+      phi::product(lr_dims),
+      1,
+      phi::errors::InvalidArgument("LearningRate should have one element"));
   auto param_dims = param.dims();
   PADDLE_ENFORCE_EQ(
       param_dims,

paddle/phi/infermeta/multiary.h

@@ -43,6 +43,7 @@ void AdadeltaInferMeta(const MetaTensor& param,
                        const MetaTensor& grad,
                        const MetaTensor& avg_squared_grad,
                        const MetaTensor& avg_squared_update,
+                       const MetaTensor& learning_rate,
                        const MetaTensor& master_param,
                        float rho,
                        float epsilon,

paddle/phi/kernels/adadelta_kernel.h

@@ -24,6 +24,7 @@ void AdadeltaKernel(const Context& dev_ctx,
                     const DenseTensor& grad,
                     const DenseTensor& avg_squared_grad,
                     const DenseTensor& avg_squared_update,
+                    const DenseTensor& learning_rate,
                     const paddle::optional<DenseTensor>& master_param,
                     float rho,
                     float epsilon,

paddle/phi/kernels/impl/adadelta_kernel_impl.h

@@ -13,11 +13,14 @@
 // limitations under the License.
 
 #pragma once
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/operators/eigen/eigen_function.h"
 
 #include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/kernels/adadelta_kernel.h"
 #include "paddle/phi/kernels/funcs/eigen/common.h"
-#include "paddle/phi/kernels/funcs/eigen/eigen_function.h"
+#include "paddle/phi/kernels/funcs/math_function.h"
 
 namespace phi {
 
@@ -27,6 +30,7 @@ void AdadeltaKernel(const Context& dev_ctx,
                     const DenseTensor& grad,
                     const DenseTensor& avg_squared_grad,
                     const DenseTensor& avg_squared_update,
+                    const DenseTensor& learning_rate,
                     const paddle::optional<DenseTensor>& master_param,
                     float rho,
                     float epsilon,
@@ -56,29 +60,36 @@ void AdadeltaKernel(const Context& dev_ctx,
   auto eigen_avg_squared_update_out =
       EigenVector<MPDType>::Flatten(*avg_squared_update_out);
   auto& place = *dev_ctx.eigen_device();
-
   auto eigen_grad_cast = eigen_grad.template cast<MPDType>();
-
   eigen_avg_squared_grad_out.device(place) =
       rho_ * eigen_avg_squared_grad + (1 - rho_) * eigen_grad_cast.square();
-  auto update = -((eigen_avg_squared_update + epsilon_) /
-                  (eigen_avg_squared_grad_out + epsilon_))
-                     .sqrt() *
-                eigen_grad_cast;
-  eigen_avg_squared_update_out.device(place) =
-      rho_ * eigen_avg_squared_update + (1 - rho_) * update.square();
-
+  auto update =
+      -(((eigen_avg_squared_update + epsilon_).sqrt()) /
+        ((eigen_avg_squared_grad_out + epsilon_).sqrt()) * eigen_grad_cast);
+  Eigen::DSizes<int, 1> m_dsize(avg_squared_update_out->numel());
+  if (paddle::platform::is_cpu_place(dev_ctx.GetPlace())) {
+    auto* lr = learning_rate.data<T>();
+    eigen_param_out.device(place) =
+        eigen_param + lr[0] * update.template cast<T>();
+  } else {
+    auto lr = EigenVector<MPDType>::Flatten(learning_rate);
     if (multi_precision) {
       auto eigen_master_param_out =
           EigenVector<MPDType>::Flatten(*master_param_outs);
       auto eigen_master_param = EigenVector<MPDType>::Flatten(*master_param);
 
-    eigen_master_param_out.device(place) = eigen_master_param + update;
-    eigen_param_out.device(place) =
-        (eigen_param.template cast<MPDType>() + update).template cast<T>();
+      eigen_master_param_out.device(place) =
+          eigen_master_param + lr.broadcast(m_dsize) * update;
+      eigen_param_out.device(place) = (eigen_param.template cast<MPDType>() +
+                                       lr.broadcast(m_dsize) * update)
+                                          .template cast<T>();
     } else {
-    eigen_param_out.device(place) = eigen_param + update.template cast<T>();
+      eigen_param_out.device(place) =
+          eigen_param + (lr.broadcast(m_dsize) * update).template cast<T>();
     }
+  }
+  eigen_avg_squared_update_out.device(place) =
+      rho_ * eigen_avg_squared_update + (1 - rho_) * update.square();
 }
 
 }  // namespace phi

paddle/phi/kernels/xpu/adadelta_kernel.cc

@@ -25,6 +25,7 @@ void AdadeltaKernel(const Context& dev_ctx,
                     const DenseTensor& grad,
                     const DenseTensor& avg_squared_grad,
                     const DenseTensor& avg_squared_update,
+                    const DenseTensor& learning_rate,
                     const paddle::optional<DenseTensor>& master_param,
                     float rho,
                     float epsilon,

paddle/phi/ops/compat/adadelta_sig.cc

@@ -18,9 +18,13 @@ namespace phi {
 
 KernelSignature AdadeltaOpArgumentMapping(const ArgumentMappingContext& ctx) {
   if (ctx.IsDenseTensorInput("Grad")) {
-    return KernelSignature(
-        "adadelta",
-        {"Param", "Grad", "AvgSquaredGrad", "AvgSquaredUpdate", "MasterParam"},
+    return KernelSignature("adadelta",
+                           {"Param",
+                            "Grad",
+                            "AvgSquaredGrad",
+                            "AvgSquaredUpdate",
+                            "LearningRate",
+                            "MasterParam"},
                            {"rho", "epsilon", "multi_precision"},
                            {"ParamOut",
                             "AvgSquaredGradOut",

python/paddle/fluid/optimizer.py

@@ -3215,6 +3215,7 @@ class AdadeltaOptimizer(Optimizer):
                 param_and_grad[1],
                 avg_squared_grad_acc,
                 avg_squared_update_acc,
+                self._create_param_lr(param_and_grad),
                 master_weight,
                 self._rho,
                 self._epsilon,
@@ -3227,6 +3228,7 @@ class AdadeltaOptimizer(Optimizer):
                 "Grad": param_and_grad[1],
                 "AvgSquaredGrad": avg_squared_grad_acc,
                 "AvgSquaredUpdate": avg_squared_update_acc,
+                "LearningRate": self._create_param_lr(param_and_grad),
             }
             outputs = {
                 "ParamOut": param_and_grad[0],

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

@@ -26,6 +26,7 @@ def adadelta_wrapper(
     Grad,
     AvgSquaredGrad,
     AvgSquaredUpdate,
+    LearningRate,
     master_weight=None,
     rho=0.95,
     epsilon=1e-6,
@@ -35,12 +36,13 @@ def adadelta_wrapper(
         Grad,
         AvgSquaredGrad,
         AvgSquaredUpdate,
+        LearningRate,
         None,
         rho,
         epsilon,
         False,
     )
-    return Param, AvgSquaredGrad, AvgSquaredUpdate
+    return Param, AvgSquaredGrad, AvgSquaredUpdate, LearningRate
 
 
 class TestAdadeltaOp1(OpTest):
@@ -58,11 +60,13 @@ class TestAdadeltaOp1(OpTest):
         rho = 0.95
         epsilon = 1e-6
 
+        learning_rate = 1.0
         self.inputs = {
             'Param': param,
             'Grad': grad,
             'AvgSquaredGrad': avg_squared_grad,
             'AvgSquaredUpdate': avg_squared_update,
+            'LearningRate': np.array([learning_rate]).astype("float32"),
         }
 
         self.attrs = {'rho': rho, 'epsilon': epsilon}
@@ -113,12 +117,13 @@ class TestAdadeltaOp2(OpTest):
         epsilon = 1e-6
 
         self.attrs = {'rho': rho, 'epsilon': epsilon}
-
+        learning_rate = 1.0
         self.inputs = {
             'Param': param,
             'Grad': grad,
             'AvgSquaredGrad': avg_squared_grad,
             'AvgSquaredUpdate': avg_squared_update,
+            'LearningRate': np.array([learning_rate]).astype("float32"),
         }
 
         avg_squared_grad_out = rho * avg_squared_grad + (1 - rho) * np.square(

python/paddle/optimizer/adadelta.py

@@ -197,6 +197,7 @@ class Adadelta(Optimizer):
                     param_and_grad[1],
                     avg_squared_grad_acc,
                     avg_squared_update_acc,
+                    self._create_param_lr(param_and_grad),
                     master_weight,
                     self._rho,
                     self._epsilon,
@@ -213,6 +214,7 @@ class Adadelta(Optimizer):
                 "Grad": param_and_grad[1],
                 "AvgSquaredGrad": avg_squared_grad_acc,
                 "AvgSquaredUpdate": avg_squared_update_acc,
+                "LearningRate": self._create_param_lr(param_and_grad),
             }
             outputs = {
                 "ParamOut": param_and_grad[0],