[Accuracy diff No.146] Fix accuracy diff for paddle.Tensor.mean API (#73320)

NKNaN · web-flow · commit 7f6ca18047f3 · 2025-06-19T20:33:45.000+08:00
diff --git a/paddle/phi/kernels/reduce_mean_kernel.cc b/paddle/phi/kernels/reduce_mean_kernel.cc
@@ -16,6 +16,7 @@
 
 #include "paddle/phi/backends/all_context.h"
 #include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/cast_kernel.h"
 #include "paddle/phi/kernels/reduce_kernel_impl.h"
 
 namespace phi {
@@ -27,7 +28,24 @@ void MeanKernel(const Context& dev_ctx,
                 bool keep_dim,
                 DenseTensor* out) {
   bool reduce_all = recompute_reduce_all(x, dims);
-  MeanRawKernel<T>(dev_ctx, x, dims, keep_dim, reduce_all, out);
+  if (std::is_same<T, int>::value || std::is_same<T, int64_t>::value ||
+      std::is_same<T, bool>::value) {
+    using Type =
+        typename std::conditional<std::is_same<T, int>::value ||
+                                      std::is_same<T, int64_t>::value ||
+                                      std::is_same<T, bool>::value,
+                                  float,
+                                  T>::type;
+    DenseTensor x_float =
+        phi::Cast<T, Context>(dev_ctx, x, phi::DataType::FLOAT32);
+    DenseTensor* out_float = new DenseTensor();
+    out_float->Resize(out->dims());
+    MeanRawKernel<Type>(
+        dev_ctx, x_float, dims, keep_dim, reduce_all, out_float);
+    phi::CastKernel<Type, Context>(dev_ctx, *out_float, x.dtype(), out);
+  } else {
+    MeanRawKernel<T>(dev_ctx, x, dims, keep_dim, reduce_all, out);
+  }
 }
 
 }  // namespace phi
diff --git a/test/legacy_test/test_mean_op.py b/test/legacy_test/test_mean_op.py
@@ -806,6 +806,64 @@ def test_errors(self):
         )
 
 
+class TestMeanAPIInt32(unittest.TestCase):
+    def setUp(self):
+        self.x_shape = [2, 3, 4, 5]
+        self.dtype = "int32"
+        self.x_np = np.random.randint(-1, 10000, self.x_shape).astype(
+            self.dtype
+        )
+        self.places = [paddle.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            self.places.append(paddle.CUDAPlace(0))
+
+    def test_dygraph(self):
+        for place in self.places:
+            with base.dygraph.guard(place):
+                x = paddle.to_tensor(self.x_np)
+                out = paddle.mean(x=x)
+            np.testing.assert_equal(
+                out.numpy(),
+                np.mean(self.x_np.astype("float32")).astype(self.dtype),
+            )
+
+    def test_static(self):
+        paddle.enable_static()
+        for place in self.places:
+            with base.program_guard(base.Program(), base.Program()):
+                x = paddle.static.data(
+                    "x", shape=self.x_shape, dtype=self.dtype
+                )
+                out = paddle.mean(x=x)
+                exe = base.Executor(place)
+                res = exe.run(feed={"x": self.x_np}, fetch_list=[out])
+            np.testing.assert_equal(
+                res[0], np.mean(self.x_np.astype("float32")).astype(self.dtype)
+            )
+
+
+class TestMeanAPIInt64(TestMeanAPIInt32):
+    def setUp(self):
+        self.x_shape = [2, 3, 4, 5]
+        self.dtype = "int64"
+        self.x_np = np.random.randint(-1, 10000, self.x_shape).astype(
+            self.dtype
+        )
+        self.places = [paddle.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            self.places.append(paddle.CUDAPlace(0))
+
+
+class TestMeanAPIBool(TestMeanAPIInt32):
+    def setUp(self):
+        self.x_shape = [2, 3, 4, 5]
+        self.dtype = "bool"
+        self.x_np = np.random.uniform(-1, 1, self.x_shape).astype(self.dtype)
+        self.places = [paddle.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            self.places.append(paddle.CUDAPlace(0))
+
+
 class TestMeanWithTensorAxis1(TestReduceOPTensorAxisBase):
     def init_data(self):
         self.pd_api = paddle.mean
