support cast, dropout, gather, mean prim ops zero dim (#54966)

Charles-hit · web-flow · commit 24b2cc8d0880 · 2023-06-29T15:02:11.000+08:00
diff --git a/paddle/fluid/prim/api/composite_backward/composite_backward_api.h b/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
@@ -150,11 +150,20 @@ void gather_grad(const Tensor& x,
   }
 
   // transpose out_grad and zero grad to target rank.
-  auto tmp_zero_x_grad = transpose<T>(zero_tensor, tmp_perm);
-  auto tmp_out_grad = transpose<T>(out_grad, tmp_perm);
+  auto tmp_zero_x_grad = zero_tensor;
+  auto tmp_out_grad = out_grad;
+  if (zero_tensor.dims().size() > 0) {
+    tmp_zero_x_grad = transpose<T>(zero_tensor, tmp_perm);
+  }
+  if (out_grad.dims().size() > 0) {
+    tmp_out_grad = transpose<T>(out_grad, tmp_perm);
+  }
   // scatter grad to grad_x
   auto tmp_grad_x = scatter<T>(tmp_zero_x_grad, index, tmp_out_grad, false);
-  auto tmp_grad_x_tranposed = transpose<T>(tmp_grad_x, reverse_perm);
+  auto tmp_grad_x_tranposed = tmp_grad_x;
+  if (tmp_grad_x.dims().size() > 0) {
+    tmp_grad_x_tranposed = transpose<T>(tmp_grad_x, reverse_perm);
+  }
   set_output<T>(tmp_grad_x_tranposed, grad_x);
 }
 
diff --git a/python/paddle/incubate/autograd/composite_rules.py b/python/paddle/incubate/autograd/composite_rules.py
@@ -255,9 +255,11 @@ def mean_composite(x, axis, keepdim):
     axes = axis or list(range(0, len(x.shape)))
     axes = [axes] if isinstance(axes, int) else axes
     sum_x = sum(x, axis=axes, keepdim=keepdim)
-    value_to_fill = functools.reduce(
-        operator.mul, [x.shape[axis] for axis in axes]
-    )
+    ele_nums_list = [x.shape[axis] for axis in axes]
+    if ele_nums_list == []:
+        value_to_fill = 1
+    else:
+        value_to_fill = functools.reduce(operator.mul, ele_nums_list)
     norm = fill_constant(
         shape=[],
         value=value_to_fill,
diff --git a/test/legacy_test/test_cast_op.py b/test/legacy_test/test_cast_op.py
@@ -34,7 +34,8 @@ def cast_wrapper(x, out_dtype=None):
 
 class TestCastOpFp32ToFp64(OpTest):
     def setUp(self):
-        ipt = np.random.random(size=[10, 10])
+        self.init_shapes()
+        ipt = np.random.random(size=self.input_shape)
         self.inputs = {'X': ipt.astype('float32')}
         self.outputs = {'Out': ipt.astype('float64')}
         self.attrs = {
@@ -46,13 +47,21 @@ def setUp(self):
         self.python_api = cast_wrapper
         self.public_python_api = cast_wrapper
 
+    def init_shapes(self):
+        self.input_shape = [10, 10]
+
     def test_check_output(self):
         self.check_output()
 
     def test_grad(self):
         self.check_grad(['X'], ['Out'], check_prim=True)
 
 
+class TestCastOpFp32ToFp64_ZeroDim(TestCastOpFp32ToFp64):
+    def init_shapes(self):
+        self.input_shape = ()
+
+
 class TestCastOpFp16ToFp32(OpTest):
     def setUp(self):
         ipt = np.random.random(size=[10, 10])
diff --git a/test/legacy_test/test_dropout_op.py b/test/legacy_test/test_dropout_op.py
@@ -87,6 +87,24 @@ def test_check_grad_normal(self):
         self.check_grad(['X'], 'Out', check_prim=False)
 
 
+class TestDropoutOp_ZeroDim(TestDropoutOp):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.prim_op_type = "comp"
+        self.python_api = dropout_wapper
+        self.public_python_api = prim_dropout_wrapper
+        self.inputs = {'X': np.random.random(()).astype("float32")}
+        self.attrs = {'dropout_prob': 0.0, 'fix_seed': True, 'is_test': False}
+        self.outputs = {
+            'Out': self.inputs['X'],
+            'Mask': np.ones(()).astype('uint8'),
+        }
+        # Because prim op compare res with dygraph
+        # when p = 0 dropout api return x,in dygraph mode x_grad = out_grad,
+        # but in static mode x_grad = []
+        self.enable_check_static_comp = False
+
+
 class TestDropoutOpInput1d(OpTest):
     def setUp(self):
         self.op_type = "dropout"
@@ -126,6 +144,20 @@ def setUp(self):
         }
 
 
+class TestDropoutOp2_ZeroDim(TestDropoutOp2):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.python_api = dropout_wapper
+        self.public_python_api = prim_dropout_wrapper
+        self.prim_op_type = "comp"
+        self.inputs = {'X': np.random.random(()).astype("float32")}
+        self.attrs = {'dropout_prob': 1.0, 'fix_seed': True, 'is_test': False}
+        self.outputs = {
+            'Out': np.zeros(()).astype('float32'),
+            'Mask': np.zeros(()).astype('uint8'),
+        }
+
+
 class TestDropoutOp3(TestDropoutOp):
     def setUp(self):
         self.op_type = "dropout"
diff --git a/test/legacy_test/test_gather_op.py b/test/legacy_test/test_gather_op.py
@@ -70,6 +70,20 @@ def if_enable_cinn(self):
         pass
 
 
+class TestGatherOp_ZeroDim(TestGatherOp):
+    def config(self):
+        """
+        For multi-dimension input
+        """
+        self.x_shape = 100
+        self.config_dtype()
+        self.index = 2
+        self.index_type = "int32"
+
+    def if_enable_cinn(self):
+        self.enable_cinn = False
+
+
 class TestGatherOpFP16(TestGatherOp):
     def config_dtype(self):
         self.x_type = "float16"
diff --git a/test/legacy_test/test_mean_op.py b/test/legacy_test/test_mean_op.py
@@ -152,17 +152,20 @@ def setUp(self):
         self.public_python_api = reduce_mean_wrapper
         self.prim_op_type = "comp"
         self.dtype = 'float64'
-        self.shape = [2, 3, 4, 5]
+        self.init_shapes()
         self.axis = [0]
+        if self.shape == []:
+            self.axis = []
         self.keepdim = False
         self.set_attrs()
         self.if_enable_cinn()
 
         np.random.seed(10)
         x_np = np.random.uniform(-1, 1, self.shape).astype(self.dtype)
-        if not hasattr(self, "reduce_all"):
+        if not hasattr(self, "reduce_all") and not x_np.shape == ():
             self.reduce_all = (not self.axis) or len(self.axis) == len(x_np)
-
+        if x_np.shape == ():
+            self.reduce_all = True
         out_np = ref_reduce_mean(x_np, self.axis, self.keepdim, self.reduce_all)
         self.inputs = {'X': x_np}
         self.outputs = {'Out': out_np}
@@ -172,6 +175,9 @@ def setUp(self):
             'reduce_all': self.reduce_all,
         }
 
+    def init_shapes(self):
+        self.shape = [2, 3, 4, 5]
+
     def set_attrs(self):
         pass
 
@@ -195,6 +201,12 @@ def test_check_grad(self):
             )
 
 
+class TestReduceMeanOp_ZeroDim(TestReduceMeanOp):
+    def init_shapes(self):
+        self.shape = []
+        self.enable_cinn = False
+
+
 @unittest.skipIf(
     not core.is_compiled_with_cuda()
     or not core.is_bfloat16_supported(core.CUDAPlace(0)),
