PaddlePaddle
diff --git a/‎paddle/phi/kernels/cpu/unpool_grad_kernel.cc
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions b/‎paddle/phi/kernels/cpu/unpool_grad_kernel.cc
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/cpu/unpool_kernel.cc
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions b/‎paddle/phi/kernels/cpu/unpool_kernel.cc
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/funcs/pooling.h
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions b/‎paddle/phi/kernels/funcs/pooling.h
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions
diff --git a/‎paddle/phi/kernels/funcs/reduce_function.h
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions b/‎paddle/phi/kernels/funcs/reduce_function.h
Copy file name to clipboard
+4-4Lines changed: 4 additions & 4 deletions
diff --git a/‎paddle/phi/kernels/gpu/unpool_grad_kernel.cu
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions b/‎paddle/phi/kernels/gpu/unpool_grad_kernel.cu
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/gpu/unpool_kernel.cu
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions b/‎paddle/phi/kernels/gpu/unpool_kernel.cu
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/gpudnn/pool_grad_kernel.cu
Copy file name to clipboard
+5Lines changed: 5 additions & 0 deletions b/‎paddle/phi/kernels/gpudnn/pool_grad_kernel.cu
Copy file name to clipboard
+5Lines changed: 5 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/gpudnn/pool_kernel.cu
Copy file name to clipboard
+6-1Lines changed: 6 additions & 1 deletion b/‎paddle/phi/kernels/gpudnn/pool_kernel.cu
Copy file name to clipboard
+6-1Lines changed: 6 additions & 1 deletion
diff --git a/‎paddle/phi/kernels/impl/pool_grad_kernel_impl.h
Copy file name to clipboard
+12Lines changed: 12 additions & 0 deletions b/‎paddle/phi/kernels/impl/pool_grad_kernel_impl.h
Copy file name to clipboard
+12Lines changed: 12 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/impl/pool_kernel_impl.h
Copy file name to clipboard
+30Lines changed: 30 additions & 0 deletions b/‎paddle/phi/kernels/impl/pool_kernel_impl.h
Copy file name to clipboard
+30Lines changed: 30 additions & 0 deletions
@@ -81,6 +81,10 @@ void UnpoolGradKernel(const Context& dev_ctx,
                       const IntArray& output_size UNUSED,
                       const std::string& data_format UNUSED,
                       DenseTensor* x_grad) {
+  if (x_grad && x_grad->numel() == 0) {
+    dev_ctx.template Alloc<T>(x_grad);
+    return;
+  }
   const auto& indices_type = indices.dtype();
   if (indices_type == phi::DataType::INT32) {
     UnpoolGrad<T, int, Context>(dev_ctx, x, indices, out, out_grad, x_grad);
@@ -150,6 +154,10 @@ void Unpool3dGradKernel(const Context& dev_ctx,
                         const std::vector<int>& output_size UNUSED,
                         const std::string& data_format UNUSED,
                         DenseTensor* x_grad) {
+  if (x_grad && x_grad->numel() == 0) {
+    dev_ctx.template Alloc<T>(x_grad);
+    return;
+  }
   const auto& indices_type = indices.dtype();
   if (indices_type == phi::DataType::INT32) {
     Unpool3dGrad<T, int, Context>(dev_ctx, x, indices, out, out_grad, x_grad);
 
@@ -77,6 +77,10 @@ void UnpoolKernel(const Context& dev_ctx,
                   const IntArray& output_size UNUSED,
                   const std::string& data_format UNUSED,
                   DenseTensor* out) {
+  if (out && out->numel() == 0) {
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   const auto& indices_type = indices.dtype();
   if (indices_type == phi::DataType::INT32) {
     Unpool<T, int, Context>(dev_ctx, x, indices, out);
@@ -142,6 +146,10 @@ void Unpool3dKernel(const Context& dev_ctx,
                     const std::vector<int>& output_size UNUSED,
                     const std::string& data_format UNUSED,
                     DenseTensor* out) {
+  if (out && out->numel() == 0) {
+    dev_ctx.template Alloc<T>(out);
+    return;
+  }
   const auto& indices_type = indices.dtype();
   if (indices_type == phi::DataType::INT32) {
     Unpool3d<T, int, Context>(dev_ctx, x, indices, out);
 
@@ -461,11 +461,11 @@ inline T PoolOutputSize(T input_size,
             stride +
         1;
   }
-  PADDLE_ENFORCE_GT(
+  PADDLE_ENFORCE_GE(
       output_size,
       0,
       errors::InvalidArgument(
-          "the output size must be greater than 0. But received: "
+          "the output size must be greater than or equal to 0. But received: "
           "output_size = %d due to the settings of input_size(%d), "
           "padding(%d,%d), "
           "k_size(%d) and stride(%d). Please check again!",
 
@@ -1089,10 +1089,10 @@ void ReduceKernel(const KPDevice& dev_ctx,
                   phi::DenseTensor* y,
                   const TransformOp& transform,
                   const std::vector<int>& origin_reduce_dims) {
-  PADDLE_ENFORCE_GT(x.numel(),
-                    0,
-                    common::errors::InvalidArgument(
-                        "Tensor need be reduced must not empty."));
+  if (x.numel() == 0) {
+    dev_ctx.Alloc<Ty>(y);
+    return;
+  }
 #ifdef PADDLE_WITH_XPU_KP
   auto stream = dev_ctx.x_context()->xpu_stream;
 #else
 
@@ -159,6 +159,9 @@ void UnpoolGradKernel(const Context& dev_ctx,
                       const std::string& data_format,
                       DenseTensor* x_grad) {
   T* input_grad_data = dev_ctx.template Alloc<T>(x_grad);
+  if (x_grad && x_grad->numel() == 0) {
+    return;
+  }
   const T* output_grad_data = out_grad.data<T>();
   phi::funcs::SetConstant<Context, T> zero;
   zero(dev_ctx, x_grad, static_cast<T>(0));
@@ -185,6 +188,9 @@ void Unpool3dGradKernel(const Context& dev_ctx,
                         const std::string& data_format,
                         DenseTensor* x_grad) {
   T* input_grad_data = dev_ctx.template Alloc<T>(x_grad);
+  if (x_grad && x_grad->numel() == 0) {
+    return;
+  }
   const T* output_grad_data = out_grad.data<T>();
   phi::funcs::SetConstant<Context, T> zero;
   zero(dev_ctx, x_grad, static_cast<T>(0));
 
@@ -141,6 +141,9 @@ void UnpoolKernel(const Context& dev_ctx,
                   const std::string& data_format,
                   DenseTensor* out) {
   T* output_data = dev_ctx.template Alloc<T>(out);
+  if (out->numel() == 0) {
+    return;
+  }
   if (output_data) {
     phi::funcs::SetConstant<Context, T> set_zero;
     set_zero(dev_ctx, out, static_cast<T>(0));
@@ -167,6 +170,9 @@ void Unpool3dKernel(const Context& dev_ctx,
                     const std::string& data_format,
                     DenseTensor* out) {
   T* output_data = dev_ctx.template Alloc<T>(out);
+  if (out->numel() == 0) {
+    return;
+  }
   if (output_data) {
     phi::funcs::SetConstant<Context, T> set_zero;
     set_zero(dev_ctx, out, static_cast<T>(0));
 
@@ -46,6 +46,11 @@ void PoolGradRawGPUDNNKernel(const Context& ctx,
       true,
       errors::InvalidArgument("Pool operator CUDA kernel must use CUDAPlace "
                               "rather than CPUPlace."));
+
+  if (dx && dx->numel() == 0) {
+    ctx.template Alloc<T>(dx);
+    return;
+  }
   auto run_cuda_kernel = [&]() {
     PoolGradRawKernel<T, GPUContext>(ctx,
                                      x,
 
@@ -16,10 +16,10 @@ limitations under the License. */
 
 #include "paddle/phi/backends/gpu/gpu_dnn.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/math_function.h"
 #include "paddle/phi/kernels/funcs/pooling.h"
 #include "paddle/phi/kernels/gpudnn/pool_gpudnn.h"
-
 #include "paddle/phi/kernels/impl/pool_kernel_impl.h"
 
 namespace phi {
@@ -37,6 +37,11 @@ void PoolRawGPUDNNKernel(const Context& ctx,
                          bool adaptive,
                          const std::string& padding_algorithm,
                          DenseTensor* out) {
+  if (x.numel() == 0) {
+    phi::Full<T, Context>(
+        ctx, phi::IntArray(common::vectorize(out->dims())), NAN, out);
+    return;
+  }
   PADDLE_ENFORCE_EQ(
       ctx.GetPlace().GetType() == phi::AllocationType::GPU,
       true,
 
@@ -39,6 +39,10 @@ void PoolGradRawKernel(const Context& ctx,
                        const std::string& padding_algorithm,
                        const float norm_type,
                        DenseTensor* dx) {
+  if (dx && dx->numel() == 0) {
+    ctx.template Alloc<T>(dx);
+    return;
+  }
   const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");
   std::vector<int64_t> paddings_ = paddings;
   std::vector<int64_t> kernel_size_ = kernel_size;
@@ -176,6 +180,10 @@ void MaxPoolWithIndexGradRawKernel(const Context& ctx,
                                    bool global_pooling,
                                    bool adaptive,
                                    DenseTensor* dx) {
+  if (dx && dx->numel() == 0) {
+    ctx.template Alloc<T1>(dx);
+    return;
+  }
   std::vector<int> paddings_ = paddings;
   std::vector<int> kernel_size_ = kernel_size;
 
@@ -401,6 +409,10 @@ void FractionalMaxPoolGradRawKernel(const Context& ctx,
                                     float random_u,
                                     bool return_mask,
                                     DenseTensor* dx) {
+  if (dx && dx->numel() == 0) {
+    ctx.template Alloc<T1>(dx);
+    return;
+  }
   std::vector<int> output_size_ = output_size;
 
   if (dx) {
 
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <algorithm>
 
 #include "paddle/common/ddim.h"
+#include "paddle/phi/kernels/full_kernel.h"
 #include "paddle/phi/kernels/funcs/pooling.h"
 #include "paddle/phi/kernels/pool_kernel.h"
 
@@ -63,6 +64,11 @@ void PoolRawKernel(const Context& ctx,
                    const std::string& padding_algorithm,
                    const float norm_type,
                    DenseTensor* out) {
+  if (x.numel() == 0) {
+    phi::Full<T, Context>(
+        ctx, phi::IntArray(common::vectorize(out->dims())), NAN, out);
+    return;
+  }
   const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");
   std::vector<int64_t> paddings_ = paddings;
   std::vector<int64_t> kernel_size_ = kernel_size;
@@ -221,6 +227,18 @@ void MaxPoolWithIndexRawKernel(const Context& ctx,
                                bool adaptive,
                                DenseTensor* out,
                                DenseTensor* mask) {
+  if (x.numel() == 0) {
+    if (out) {
+      phi::Full<T1, Context>(
+          ctx, phi::IntArray(common::vectorize(out->dims())), NAN, out);
+    }
+    if (mask) {
+      phi::Full<T2, Context>(
+          ctx, phi::IntArray(common::vectorize(mask->dims())), 0, mask);
+    }
+    return;
+  }
+
   std::vector<int> paddings_ = paddings;
   std::vector<int> kernel_size_ = kernel_size;
 
@@ -390,6 +408,18 @@ void FractionalMaxPoolRawKernel(const Context& ctx,
                                 bool return_mask,
                                 DenseTensor* out,
                                 DenseTensor* mask) {
+  if (x.numel() == 0) {
+    if (out) {
+      phi::Full<T1, Context>(
+          ctx, phi::IntArray(common::vectorize(out->dims())), NAN, out);
+    }
+    if (mask) {
+      phi::Full<T2, Context>(
+          ctx, phi::IntArray(common::vectorize(mask->dims())), 0, mask);
+    }
+    return;
+  }
+
   std::vector<int> output_size_ = output_size;
 
   switch (output_size_.size()) {