[Zero-Dim] add FLAGS_set_to_1d, control whether to hack process to 1D…

…, add ut for xpu (#51899)

paddle/fluid/pybind/CMakeLists.txt

@@ -504,6 +504,7 @@ if(WITH_PYTHON)
     list(APPEND PYBIND_DEPS tensor_api)
     list(APPEND PYBIND_DEPS eager_tensor_operants)
     list(APPEND PYBIND_DEPS pybind_util)
+    list(APPEND PYBIND_DEPS flags)
   endif()
 
   # On Linux, cc_library(paddle SHARED ..) will generate the libpaddle.so,

paddle/fluid/pybind/eager_method.cc

@@ -62,6 +62,8 @@ typedef SSIZE_T ssize_t;
 #include "paddle/phi/core/tensor_utils.h"
 #include "paddle/phi/kernels/funcs/math_function.h"
 
+DECLARE_bool(set_to_1d);
+
 namespace paddle {
 namespace pybind {
 
@@ -124,7 +126,8 @@ static PyObject* tensor_method_numpy(TensorObject* self,
   size_t numel = 1;
   if (py_rank == 0) {
     Py_ssize_t args_num = PyTuple_Size(args);
-    bool set_to_1d = true;
+    // true by default
+    bool set_to_1d = FLAGS_set_to_1d;
     if (args_num == (Py_ssize_t)1) {
       PyObject* obj = PyTuple_GET_ITEM(args, 0);
       if (obj == Py_False) {

paddle/phi/core/flags.cc

@@ -744,6 +744,16 @@ PADDLE_DEFINE_EXPORTED_int32(
     "less FLAGS_max_inplace_grad_add, than it will be use several grad_add"
     "instead of sum. Default is 0.");
 
+/**
+ * Tensor.numpy() has a hack, and this flag can close this hack
+ * [true]: set 0D Tensor to 1D Numpy
+ * [false]: not set 0D Tensor to 1D Numpy, close the hack
+ *
+ * Now, just set true by default in 2.5 transition time
+ * which will be removed in future (2.6 or 2.7) .
+ */
+PADDLE_DEFINE_EXPORTED_bool(set_to_1d, true, "set 0D Tensor to 1D numpy");
+
 /**
  * Debug related FLAG
  * Name: tracer_mkldnn_ops_on

paddle/scripts/paddle_build.bat

@@ -685,6 +685,7 @@ for /F %%# in ('wmic os get localdatetime^|findstr 20') do set start=%%#
 set start=%start:~4,10%
 
 set FLAGS_call_stack_level=2
+set FLAGS_set_to_1d=False
 dir %THIRD_PARTY_PATH:/=\%\install\openblas\lib
 dir %THIRD_PARTY_PATH:/=\%\install\openblas\bin
 dir %THIRD_PARTY_PATH:/=\%\install\zlib\bin

paddle/scripts/paddle_build.sh

@@ -61,6 +61,7 @@ function init() {
 
     # NOTE(chenweihang): For easy debugging, CI displays the C++ error stacktrace by default
     export FLAGS_call_stack_level=2
+    export FLAGS_set_to_1d=False
 }
 
 function cmake_base() {

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

@@ -1100,12 +1100,10 @@ def test_tolist(self):
 
     def test_numpy(self):
         x = paddle.full([], 0.5)
-        # 0D Tensor hack to 1D Numpy defaut, will remove in future
         x_np = x.numpy()
-        np.testing.assert_array_equal(x_np.shape, (1,))
-        np.testing.assert_array_equal(x_np, np.array([0.5]))
+        np.testing.assert_array_equal(x_np.shape, ())
+        np.testing.assert_array_equal(x_np, np.array(0.5))
 
-        # return origin correct numpy
         x_np = x.numpy(False)
         np.testing.assert_array_equal(x_np.shape, ())
         np.testing.assert_array_equal(x_np, np.array(0.5))

python/paddle/fluid/tests/unittests/xpu/test_zero_dim_tensor_xpu.py

@@ -161,6 +161,20 @@ def test_dygraph_reduce(self):
                 np.testing.assert_allclose(x.grad.numpy(), np.array(1.0))
                 np.testing.assert_allclose(out.grad.numpy(), np.array(1.0))
 
+            out1 = api(x, 0)
+            self.assertEqual(out1.shape, [])
+            self.assertEqual(out1, out)
+            out1.backward()
+
+            out2 = api(x, -1)
+            self.assertEqual(out2.shape, [])
+            self.assertEqual(out2, out)
+            out2.backward()
+
+            if x.grad is not None:
+                self.assertEqual(x.grad.shape, [])
+                np.testing.assert_allclose(x.grad.numpy(), np.array(3.0))
+
         paddle.enable_static()
 
 
@@ -463,32 +477,72 @@ def test_top_k(self):
             tmp = paddle.topk(x1, k=1, axis=2)
 
     def test_argmin(self):
+        # 1) x is 0D
         x = paddle.rand([])
         out1 = paddle.argmin(x, 0)
         out2 = paddle.argmin(x, -1)
         out3 = paddle.argmin(x, None)
+
         self.assertEqual(out1.shape, [])
-        np.testing.assert_allclose(out1, 0.0)
+        np.testing.assert_allclose(out1, 0)
 
         self.assertEqual(out2.shape, [])
-        np.testing.assert_allclose(out2, 0.0)
+        np.testing.assert_allclose(out2, 0)
 
         self.assertEqual(out3.shape, [])
-        np.testing.assert_allclose(out3, 0.0)
+        np.testing.assert_allclose(out3, 0)
+
+        # 2) x is 1D
+        x = paddle.rand([5])
+        x.stop_gradient = False
+        out = paddle.argmin(x, 0)
+        out.backward()
+        self.assertEqual(out.shape, [])
+
+        # 3) x is ND
+        x = paddle.rand([3, 5])
+        x.stop_gradient = False
+        out = paddle.argmin(x)
+        out.backward()
+        self.assertEqual(out.shape, [])
+
+        # 4) x is ND, keepdim=True
+        x = paddle.rand([3, 5])
+        x.stop_gradient = False
+        out = paddle.argmin(x, keepdim=True)
+        out.backward()
+        self.assertEqual(out.shape, [1, 1])
 
     def test_argmax(self):
+        # 1) x is 0D
         x = paddle.rand([])
         out1 = paddle.argmax(x, 0)
         out2 = paddle.argmax(x, -1)
         out3 = paddle.argmax(x, None)
+
         self.assertEqual(out1.shape, [])
-        np.testing.assert_allclose(out1, 0.0)
+        np.testing.assert_allclose(out1, 0)
 
         self.assertEqual(out2.shape, [])
-        np.testing.assert_allclose(out2, 0.0)
+        np.testing.assert_allclose(out2, 0)
 
         self.assertEqual(out3.shape, [])
-        np.testing.assert_allclose(out3, 0.0)
+        np.testing.assert_allclose(out3, 0)
+
+        # 2) x is 1D
+        x = paddle.rand([5])
+        out = paddle.argmax(x, 0)
+        self.assertEqual(out.shape, [])
+
+        # 3) x is ND
+        x = paddle.rand([3, 5])
+        out = paddle.argmax(x)
+        self.assertEqual(out.shape, [])
+
+        # 4) x is ND, keepdim=True
+        x = paddle.rand([3, 5])
+        out = paddle.argmax(x, keepdim=True)
+        self.assertEqual(out.shape, [1, 1])
 
     def test_median(self):
         x = paddle.rand([])
@@ -575,15 +629,29 @@ def test_numpy(self):
         np.testing.assert_array_equal(x.numpy(), np.array(0.5))
 
     def test_numel(self):
+        # 1) x is 0D
         out = paddle.numel(self.x)
         self.assertEqual(out.shape, [])
         np.testing.assert_array_equal(out.numpy(), np.array(1))
 
+        # 2) x is ND
+        x = paddle.full([3, 5], 0.5)
+        out = paddle.numel(x)
+        self.assertEqual(out.shape, [])
+        np.testing.assert_array_equal(out.numpy(), np.array(15))
+
     def test_rank(self):
+        # 1) x is 0D
         out = paddle.rank(self.x)
         self.assertEqual(out.shape, [])
         np.testing.assert_array_equal(out.numpy(), np.array(0))
 
+        # 1) x is ND
+        x = paddle.full([3, 5], 0.5)
+        out = paddle.rank(x)
+        self.assertEqual(out.shape, [])
+        np.testing.assert_array_equal(out.numpy(), np.array(2))
+
     def test_shape(self):
         out = paddle.shape(self.x)
         self.assertEqual(out.shape, [0])