【cherry-pick】fix decay param and overflow in match_matrix (#22107)

* fix decay param in DecayAdagrad test=develop (#22026)

* fix integer overflow in match_matrix (#22036)

* fix integer overflow in match_matrix test=develop

* fix integer overflow in match_matrix test=develop

* fix typo test=develop

paddle/fluid/operators/match_matrix_tensor_op.cc

@@ -56,8 +56,8 @@ void MatchMatrixTensorOP::InferShape(framework::InferShapeContext* ctx) const {
   PADDLE_ENFORCE_EQ(w_dims[2], y_dims[1],
                     "W 's shape must satisfy: W[2] = Y[1]");
 
-  int out_dim_0 = -1;
-  int tmp_dim_0 = -1;
+  int64_t out_dim_0 = -1;
+  int64_t tmp_dim_0 = -1;
   if (ctx->IsRuntime()) {
     framework::Variable* x_var =
         boost::get<framework::Variable*>(ctx->GetInputVarPtrs("X")[0]);
@@ -86,8 +86,8 @@ void MatchMatrixTensorOP::InferShape(framework::InferShapeContext* ctx) const {
 
     out_dim_0 = 0;
     for (size_t i = 1; i < x_lod_0.size(); i++) {
-      int x_len = x_lod_0[i] - x_lod_0[i - 1];
-      int y_len = y_lod_0[i] - y_lod_0[i - 1];
+      int64_t x_len = x_lod_0[i] - x_lod_0[i - 1];
+      int64_t y_len = y_lod_0[i] - y_lod_0[i - 1];
       out_dim_0 += (x_len * y_len);
     }
     out_dim_0 *= dim_t;
@@ -173,17 +173,17 @@ class CPUMatchMatrixTensorOPKernel : public framework::OpKernel<T> {
     auto* tmp = ctx.Output<LoDTensor>("Tmp");
 
     int dim_t = ctx.Attr<int>("dim_t");
-    int dim_in = x->dims()[1];
+    int64_t dim_in = x->dims()[1];
 
     const auto& offset_l = x->lod()[0];
     const auto& offset_r = y->lod()[0];
 
     std::vector<size_t> top_offset;
-    int top_size = 0;
+    size_t top_size = 0;
     top_offset.push_back(top_size);
     for (size_t b = 0; b < x->lod()[0].size() - 1; b++) {
-      int len_l = offset_l[b + 1] - offset_l[b];
-      int len_r = offset_r[b + 1] - offset_r[b];
+      size_t len_l = offset_l[b + 1] - offset_l[b];
+      size_t len_r = offset_r[b + 1] - offset_r[b];
       top_size += dim_t * len_l * len_r;
       top_offset.push_back(top_size);
     }
@@ -204,8 +204,8 @@ class CPUMatchMatrixTensorOPKernel : public framework::OpKernel<T> {
 
     for (size_t b = 0; b < x->lod()[0].size() - 1; b++) {
       for (int t = 0; t < dim_t; t++) {
-        int len_l = offset_l[b + 1] - offset_l[b];
-        int len_r = offset_r[b + 1] - offset_r[b];
+        size_t len_l = offset_l[b + 1] - offset_l[b];
+        size_t len_r = offset_r[b + 1] - offset_r[b];
         auto* top_data = out_data + top_offset[b] + t * len_l * len_r;
         const auto* l_t_data =
             bottom_l_trans_data + offset_l[b] * dim_t * dim_in + t * dim_in;
@@ -234,16 +234,16 @@ class CPUMatchMatrixTensorOPGradKernel : public framework::OpKernel<T> {
     auto* tmp = ctx.Input<LoDTensor>("Tmp");
 
     int dim_t = ctx.Attr<int>("dim_t");
-    int dim_in = x->dims()[1];
+    int64_t dim_in = x->dims()[1];
 
     const auto& offset_l = x->lod()[0];
     const auto& offset_r = y->lod()[0];
-    std::vector<int> top_offset;
-    int top_size = 0;
+    std::vector<size_t> top_offset;
+    size_t top_size = 0;
     top_offset.push_back(top_size);
     for (size_t b = 0; b < x->lod()[0].size() - 1; b++) {
-      int len_l = offset_l[b + 1] - offset_l[b];
-      int len_r = offset_r[b + 1] - offset_r[b];
+      size_t len_l = offset_l[b + 1] - offset_l[b];
+      size_t len_r = offset_r[b + 1] - offset_r[b];
       top_size += dim_t * len_l * len_r;
       top_offset.push_back(top_size);
     }
@@ -270,11 +270,11 @@ class CPUMatchMatrixTensorOPGradKernel : public framework::OpKernel<T> {
 
     for (size_t b = 0; b < x->lod()[0].size() - 1; b++) {
       for (int t = 0; t < dim_t; t++) {
-        int len_l = offset_l[b + 1] - offset_l[b];
-        int len_r = offset_r[b + 1] - offset_r[b];
+        size_t len_l = offset_l[b + 1] - offset_l[b];
+        size_t len_r = offset_r[b + 1] - offset_r[b];
 
-        for (int i = 0; i < len_l; i++) {
-          for (int j = 0; j < len_r; j++) {
+        for (size_t i = 0; i < len_l; i++) {
+          for (size_t j = 0; j < len_r; j++) {
             auto diff =
                 top_diff[top_offset[b] + t * len_l * len_r + i * len_r + j];
             auto* l_trans_data = bottom_l_trans_data +
@@ -324,11 +324,7 @@ REGISTER_OPERATOR(match_matrix_tensor_grad, ops::MatchMatrixTensorOpGrad);
 REGISTER_OP_CPU_KERNEL(match_matrix_tensor,
                        ops::CPUMatchMatrixTensorOPKernel<
                            paddle::platform::CPUDeviceContext, float>);
-//     ops::CPUMatchMatrixTensorOPKernel<paddle::platform::CPUDeviceContext,
-//                                       double>
 
 REGISTER_OP_CPU_KERNEL(match_matrix_tensor_grad,
                        ops::CPUMatchMatrixTensorOPGradKernel<
                            paddle::platform::CPUDeviceContext, float>);
-//     ops::CPUMatchMatrixTensorOPGradKernel<paddle::platform::CPUDeviceContext,
-//                                           double>

python/paddle/fluid/optimizer.py

@@ -2060,7 +2060,8 @@ class DecayedAdagradOptimizer(Optimizer):
             },
             outputs={"ParamOut": param_and_grad[0],
                      "MomentOut": moment_acc},
-            attrs={"epsilon": self._epsilon},
+            attrs={"epsilon": self._epsilon,
+                   "decay": self._decay},
             stop_gradient=True)
 
         return decayed_adagrad_op