[Zero-Dim] correct some code to adapt to 0D Tensor (#51562)

python/paddle/fluid/dygraph/learning_rate_scheduler.py

@@ -92,12 +92,10 @@ class LearningRateDecay:
                 continue
             value = self.__dict__[key]
             if isinstance(value, Variable):
-                assert value.shape == [
-                    1
-                ], "shape of Variable in state_dict must be [1] {}".format(
-                    value.shape
-                )
-                value = value.numpy()[0]
+                assert (
+                    value.size == 1
+                ), "size of Variable in state_dict must be 1"
+                value = float(value)
             state_dict[key] = value
 
         return state_dict
@@ -857,7 +855,7 @@ class ReduceLROnPlateau(LearningRateDecay):
                 # adjust learning rate according to avg_loss
                 reduce_lr.step(avg_loss)
                 lr = adam.current_step_lr()
-                print("current avg_loss is %s, current lr is %s" % (avg_loss.numpy()[0], lr))
+                print("current avg_loss is %s, current lr is %s" % (float(avg_loss), lr))
 
     """
 
@@ -979,14 +977,11 @@ class ReduceLROnPlateau(LearningRateDecay):
                 )
                 if self.learning_rate - new_lr > self.eps:
                     if self.verbose:
-                        old_lr = (
-                            self.learning_rate.numpy()[0]
-                            if isinstance(self.learning_rate, Variable)
-                            else self.learning_rate
-                        )
                         print(
                             'Epoch {}: reducing learning rate from {} to {}.'.format(
-                                self.epoch_num, old_lr, new_lr.numpy()[0]
+                                self.epoch_num,
+                                float(self.learning_rate),
+                                float(new_lr),
                             )
                         )
                     self.learning_rate = new_lr

python/paddle/fluid/layers/control_flow.py

@@ -1150,7 +1150,7 @@ def while_loop(cond, body, loop_vars, is_test=False, name=None):
         )
 
     if in_dygraph_mode():
-        now_cond = pre_cond.numpy()[0]
+        now_cond = pre_cond.numpy().item()
         while now_cond:
             output_vars = body(*loop_vars)
             if not isinstance(output_vars, (list, tuple)):
@@ -1160,7 +1160,7 @@ def while_loop(cond, body, loop_vars, is_test=False, name=None):
                     "body in while_loop should return the same arity "
                     "(length and structure) and types as loop_vars"
                 )
-            now_cond = cond(*output_vars).numpy()[0]
+            now_cond = cond(*output_vars).numpy().item()
             map_structure(assign_skip_lod_tensor_array, output_vars, loop_vars)
         return loop_vars
     else:

python/paddle/fluid/optimizer.py

@@ -596,19 +596,19 @@ class Optimizer:
         """
         current_lr = self._global_learning_rate()
         if isinstance(current_lr, framework.Variable):
-            return self._global_learning_rate().numpy()[0]
+            return float(current_lr)
 
         if isinstance(self._learning_rate, float):
             return self._learning_rate
         elif isinstance(self._learning_rate, _LearningRateEpochDecay):
             step_lr = self._learning_rate()
-            return step_lr.numpy()[0]
+            return float(step_lr)
         else:
             step_lr = self._learning_rate.step()
             if isinstance(step_lr, (float, int)):
                 return step_lr
             else:
-                return step_lr.numpy()[0]
+                return float(step_lr)
 
     def _global_learning_rate(self, program=None):
         """

python/paddle/fluid/tests/unittests/dygraph_to_static/ifelse_simple_func.py

@@ -41,7 +41,7 @@ def dyfunc_empty_nonlocal(x):
 
 
 def dyfunc_with_if_else(x_v, label=None):
-    if paddle.mean(x_v).numpy()[0] > 5:
+    if paddle.mean(x_v).numpy() > 5:
         x_v = x_v - 1
     else:
         x_v = x_v + 1
@@ -61,7 +61,7 @@ def dyfunc_with_if_else2(x, col=100):
         #  `x` is Tensor, `col` is not Tensor, and `col` is the return value of `true_fn` after transformed.
         # col = -1
         col = fluid.layers.fill_constant(shape=[1], value=-1, dtype="int64")
-    if paddle.mean(x).numpy()[0] > x.numpy()[row][col]:
+    if paddle.mean(x).numpy() > x.numpy()[row][col]:
         y = paddle.nn.functional.relu(x)
     else:
         x_pow = paddle.pow(x, 2)
@@ -89,14 +89,14 @@ def dyfunc_with_if_else3(x):
         m = x + 2
         n = x + 3
         return q, x, y, z
-    q, x, y, z = paddle.static.nn.cond(paddle.mean(x)[0] < 5, lambda :
+    q, x, y, z = paddle.static.nn.cond(paddle.mean(x) < 5, lambda :
         paddle.jit.dy2static.convert_call(true_fn_0)(q, x, y),
         lambda : paddle.jit.dy2static.convert_call(false_fn_0)(q,
         x, y))
     """
     y = x + 1
     # NOTE: x_v[0] < 5 is True
-    if paddle.mean(x).numpy()[0] < 5:
+    if paddle.mean(x).numpy() < 5:
         x = x + 1
         z = x + 2
         q = x + 3
@@ -164,7 +164,7 @@ def nested_if_else(x_v):
         if y.numpy()[0] < 10:
             tmp = y * w
             y = paddle.nn.functional.relu(tmp)
-            if paddle.mean(y).numpy()[0] < batch_size:
+            if paddle.mean(y).numpy() < batch_size:
                 y = paddle.abs(y)
             else:
                 tmp = fluid.layers.fill_constant(
@@ -264,7 +264,7 @@ class NetWithControlFlowIf(fluid.dygraph.Layer):
         )
         # Control flow `if` statement
         fc_out = self.fc(input)
-        if paddle.mean(fc_out).numpy()[0] < 0:
+        if paddle.mean(fc_out).numpy() < 0:
             y = fc_out + self.constant_vars['bias']
             self.constant_vars['w'] = fluid.layers.fill_constant(
                 [5], dtype='float32', value=10
@@ -297,7 +297,7 @@ def if_with_and_or(x_v, label=None):
     batch_size = paddle.shape(x_v)
     if (
         x_v is not None
-        and (paddle.mean(x_v).numpy()[0] > 0 or label is not None)
+        and (paddle.mean(x_v).numpy() > 0 or label is not None)
         and batch_size[0] > 1
         and True
     ):
@@ -338,10 +338,10 @@ def if_with_and_or_3(x, y=None):
         x is not None
         and batch_size[0] > 1
         and y is not None
-        and mean_res.numpy()[0] > 0
+        and mean_res.numpy() > 0
     ):
         x = x + 1
-    if mean_res.numpy()[0] > 0 and (x is not None and batch_size[0] > 1) and y:
+    if mean_res.numpy() > 0 and (x is not None and batch_size[0] > 1) and y:
         x = x - 1
     return x
 
@@ -350,11 +350,11 @@ def if_with_and_or_4(x, y=None):
     batch_size = paddle.shape(x)
     mean_res = paddle.mean(x)
     if (x is not None and batch_size[0] > 1) or (
-        y is not None and mean_res.numpy()[0] > 0
+        y is not None and mean_res.numpy() > 0
     ):
         x = x + 1
     if (x is not None or batch_size[0] > 1) and (
-        y is not None or mean_res.numpy()[0] > 0
+        y is not None or mean_res.numpy() > 0
     ):
         x = x - 1
     return x

python/paddle/fluid/tests/unittests/dygraph_to_static/test_bmn.py

@@ -620,19 +620,19 @@ def val_bmn(model, args):
         avg_loss = paddle.mean(loss)
 
         loss_data += [
-            avg_loss.numpy()[0],
-            tem_loss.numpy()[0],
-            pem_reg_loss.numpy()[0],
-            pem_cls_loss.numpy()[0],
+            float(avg_loss),
+            float(tem_loss),
+            float(pem_reg_loss),
+            float(pem_cls_loss),
         ]
 
         print(
             '[VALID] iter {} '.format(batch_id)
             + '\tLoss = {}, \ttem_loss = {}, \tpem_reg_loss = {}, \tpem_cls_loss = {}'.format(
-                '%f' % avg_loss.numpy()[0],
-                '%f' % tem_loss.numpy()[0],
-                '%f' % pem_reg_loss.numpy()[0],
-                '%f' % pem_cls_loss.numpy()[0],
+                '%f' % float(avg_loss),
+                '%f' % float(tem_loss),
+                '%f' % float(pem_reg_loss),
+                '%f' % float(pem_cls_loss),
             )
         )
 
@@ -716,10 +716,10 @@ class TestTrain(unittest.TestCase):
                     bmn.clear_gradients()
                     # log loss data to verify correctness
                     loss_data += [
-                        avg_loss.numpy()[0],
-                        tem_loss.numpy()[0],
-                        pem_reg_loss.numpy()[0],
-                        pem_cls_loss.numpy()[0],
+                        float(avg_loss),
+                        float(tem_loss),
+                        float(pem_reg_loss),
+                        float(pem_cls_loss),
                     ]
 
                     if args.log_interval > 0 and (
@@ -728,10 +728,10 @@ class TestTrain(unittest.TestCase):
                         print(
                             '[TRAIN] Epoch {}, iter {} '.format(epoch, batch_id)
                             + '\tLoss = {}, \ttem_loss = {}, \tpem_reg_loss = {}, \tpem_cls_loss = {}'.format(
-                                '%f' % avg_loss.numpy()[0],
-                                '%f' % tem_loss.numpy()[0],
-                                '%f' % pem_reg_loss.numpy()[0],
-                                '%f' % pem_cls_loss.numpy()[0],
+                                '%f' % float(avg_loss),
+                                '%f' % float(tem_loss),
+                                '%f' % float(pem_reg_loss),
+                                '%f' % float(pem_cls_loss),
                             )
                         )
 

python/paddle/fluid/tests/unittests/dygraph_to_static/test_convert_call.py

@@ -32,7 +32,7 @@ np.random.seed(SEED)
 # Use a decorator to test exception
 @paddle.jit.to_static
 def dyfunc_with_if(x_v):
-    if paddle.mean(x_v).numpy()[0] > 5:
+    if paddle.mean(x_v).numpy() > 5:
         x_v = x_v - 1
     else:
         x_v = x_v + 1
@@ -53,7 +53,7 @@ def nested_func(x_v):
 @paddle.jit.to_static
 def dyfunc_with_third_library_logging(x_v):
     logging.info('test dyfunc_with_third_library_logging')
-    if paddle.mean(x_v).numpy()[0] > 5:
+    if paddle.mean(x_v).numpy() > 5:
         x_v = x_v - 1
     else:
         x_v = x_v + 1

python/paddle/fluid/tests/unittests/dygraph_to_static/test_cycle_gan.py

@@ -669,7 +669,7 @@ def train(args, to_static):
                     cyc_B_loss,
                     idt_loss_B,
                 ]
-                cur_batch_loss = [x.numpy()[0] for x in cur_batch_loss]
+                cur_batch_loss = [float(x) for x in cur_batch_loss]
 
                 batch_time = time.time() - s_time
                 t_time += batch_time

python/paddle/fluid/tests/unittests/dygraph_to_static/test_layer_hook.py

@@ -75,12 +75,12 @@ class TestNestLayerHook(unittest.TestCase):
         if to_static:
             paddle.jit.save(net, self.path)
 
-        return out.numpy()[0]
+        return float(out)
 
     def load_train(self):
         net = paddle.jit.load(self.path)
         out = net(self.x)
-        return out.numpy()[0]
+        return float(out)
 
     def test_hook(self):
         dy_out = self.train_net(to_static=False)

python/paddle/fluid/tests/unittests/dygraph_to_static/test_mnist.py

@@ -219,7 +219,7 @@ class TestMNISTWithToStatic(TestMNIST):
                     avg_loss.backward()
 
                     adam.minimize(avg_loss)
-                    loss_data.append(avg_loss.numpy()[0])
+                    loss_data.append(float(avg_loss))
                     # save checkpoint
                     mnist.clear_gradients()
                     if batch_id % 10 == 0:
@@ -236,7 +236,7 @@ class TestMNISTWithToStatic(TestMNIST):
                     if batch_id == 50:
                         mnist.eval()
                         prediction, acc, avg_loss = mnist(img, label)
-                        loss_data.append(avg_loss.numpy()[0])
+                        loss_data.append(float(avg_loss))
                         # new save load check
                         self.check_jit_save_load(
                             mnist, [dy_x_data], [img], to_static, prediction

python/paddle/fluid/tests/unittests/dygraph_to_static/test_mnist_amp.py

@@ -86,7 +86,7 @@ class TestAMP(TestMNIST):
                 scaled.backward()
                 scaler.minimize(adam, scaled)
 
-                loss_data.append(avg_loss.numpy()[0])
+                loss_data.append(float(avg_loss))
                 # save checkpoint
                 mnist.clear_gradients()
                 if batch_id % 10 == 0:

python/paddle/fluid/tests/unittests/dygraph_to_static/test_tsm.py

@@ -346,25 +346,25 @@ def train(args, fake_data_reader, to_static):
                 optimizer.minimize(avg_loss)
                 video_model.clear_gradients()
 
-                total_loss += avg_loss.numpy()[0]
-                total_acc1 += acc_top1.numpy()[0]
-                total_acc5 += acc_top5.numpy()[0]
+                total_loss += float(avg_loss)
+                total_acc1 += float(acc_top1)
+                total_acc5 += float(acc_top5)
                 total_sample += 1
 
                 print(
                     'TRAIN Epoch {}, iter {}, loss = {}, acc1 {}, acc5 {}'.format(
                         epoch,
                         batch_id,
-                        avg_loss.numpy()[0],
-                        acc_top1.numpy()[0],
-                        acc_top5.numpy()[0],
+                        float(avg_loss),
+                        float(acc_top1),
+                        float(acc_top5),
                     )
                 )
                 ret.extend(
                     [
-                        avg_loss.numpy()[0],
-                        acc_top1.numpy()[0],
-                        acc_top5.numpy()[0],
+                        float(avg_loss),
+                        float(acc_top1),
+                        float(acc_top5),
                     ]
                 )
 

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

@@ -25,10 +25,7 @@ from paddle.fluid.backward import _append_grad_suffix_, _as_list
 
 
 def _product(t):
-    if isinstance(t, int):
-        return t
-    else:
-        return np.product(t)
+    return int(np.product(t))
 
 
 def dtype_to_np_dtype(dtype):

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

@@ -1546,7 +1546,19 @@ class TestSundryAPI(unittest.TestCase):
         self.assertEqual(x2.grad.numpy(), 0)
 
     def test_lerp(self):
-        # 0D + 0D
+        # 0D + 0D, weight is float scalar
+        x = paddle.rand([])
+        y = paddle.rand([])
+        x.stop_gradient = False
+        y.stop_gradient = False
+        out = paddle.lerp(x, y, 0.5)
+        out.backward()
+
+        self.assertEqual(out.shape, [])
+        self.assertEqual(x.grad.shape, [])
+        self.assertEqual(y.grad.shape, [])
+
+        # 0D + 0D, weigh is 0D
         x0 = paddle.rand([])
         y0 = paddle.rand([])
         w0 = paddle.rand([])
@@ -2896,11 +2908,15 @@ class TestSundryAPIStatic(unittest.TestCase):
             [(), (), (), ()],
             [(), (64, 64), (), (64, 64)],
             [(64, 64), (), (), (64, 64)],
+            [(64, 64), (), 0.5, (64, 64)],
         ]
         for shape in shapes:
             x = paddle.rand(shape[0])
             y = paddle.rand(shape[1])
-            w = paddle.rand(shape[2])
+            if isinstance(shape[2], float):
+                w = shape[2]
+            else:
+                w = paddle.rand(shape[2])
 
             x.stop_gradient = False
             y.stop_gradient = False

python/paddle/nn/functional/pooling.py

@@ -706,7 +706,7 @@ def _unpool_output_size(x, kernel_size, stride, padding, output_size):
         else:
             for i, var in enumerate(output_size):
                 if isinstance(var, Variable):
-                    output_size[i] = var.numpy()[0]
+                    output_size[i] = var.numpy().item()
 
     if len(output_size) == len(kernel_size) + 2:
         output_size = output_size[2:]

python/paddle/optimizer/lr.py

@@ -156,12 +156,10 @@ class LRScheduler:
                 continue
             value = self.__dict__[key]
             if isinstance(value, Tensor):
-                assert value.shape == [
-                    1
-                ], "shape of Tensor in state_dict must be [1] {}".format(
-                    value.shape
-                )
-                value = value.numpy()[0]
+                assert (
+                    value.size == 1
+                ), "numel of Tensor in state_dict must be 1"
+                value = float(value)
             state_dict[key] = value
 
         return state_dict
@@ -1236,7 +1234,7 @@ class ReduceOnPlateau(LRScheduler):
     Reduce learning rate when ``metrics`` has stopped descending. Models often benefit from reducing the learning rate
     by 2 to 10 times once model performance has no longer improvement.
 
-    The ``metrics`` is the one which has been pass into ``step`` , it must be 1-D Tensor with shape [1]. When ``metrics``
+    The ``metrics`` is the one which has been pass into ``step`` , it's shape must [] or [1]. When ``metrics``
     stop descending for a ``patience`` number of epochs, the learning rate will be reduced to ``learning_rate * factor`` .
     (Specially, ``mode`` can also be set to ``'max`` , in this case, when ``metrics`` stop ascending for a ``patience``
     number of epochs, the learning rate will be reduced.)
@@ -1390,7 +1388,7 @@ class ReduceOnPlateau(LRScheduler):
         Args:
             metrics (Tensor|numpy.ndarray|float): Which will be monitored to determine whether the learning rate will reduce.
                 If it stop descending for a ``patience`` number of epochs, the learning rate will reduce. If it's 'Tensor' or
-                'numpy.ndarray', its shape must be [1].
+                'numpy.ndarray', its numel must be 1.
             epoch (int, None): specify current epoch. Default: None. Auto-increment from last_epoch=-1.
 
         Returns:
@@ -1404,13 +1402,12 @@ class ReduceOnPlateau(LRScheduler):
         else:
             self.last_epoch = epoch
 
-        # loss must be float, numpy.ndarray or 1-D Tensor with shape [1]
+        # loss must be float, numpy.ndarray or 1-D Tensor with numel 1
         if isinstance(metrics, (core.eager.Tensor, numpy.ndarray)):
-            assert len(metrics.shape) == 1 and metrics.shape[0] == 1, (
-                "the metrics.shape "
-                "should be (1L,), but the current metrics.shape is {}. Maybe that "
+            assert metrics.size == 1, (
+                "the size of metrics must be 1, but the current metrics.size is {}. Maybe that "
                 "you should call paddle.mean to process it first.".format(
-                    metrics.shape
+                    metrics.size
                 )
             )
         elif not isinstance(

python/paddle/optimizer/optimizer.py

@@ -437,7 +437,7 @@ class Optimizer:
                 self._learning_rate._var_name = lr_name
                 lr_var = self.helper.create_global_variable(
                     name=lr_name,
-                    shape=[1],
+                    shape=[],
                     persistable=True,
                     stop_gradient=True,
                     dtype=_lr_dtype,
@@ -465,7 +465,7 @@ class Optimizer:
                     framework.default_main_program()
                 ] = paddle.static.create_global_var(
                     name=unique_name.generate("learning_rate"),
-                    shape=[1],
+                    shape=[],
                     value=float(self._learning_rate),
                     dtype=_lr_dtype,
                     persistable=True,

python/paddle/tensor/math.py

@@ -4200,15 +4200,12 @@ def lerp(x, y, weight, name=None):
             # out: [5.5, 6., 6.5, 7.]
 
     """
-    if in_dygraph_mode():
-        if isinstance(weight, float):
-            weight = paddle.to_tensor(weight, dtype=x.dtype)
+    if isinstance(weight, float):
+        weight = paddle.full(shape=[], fill_value=weight, dtype=x.dtype)
 
+    if in_dygraph_mode():
         return _C_ops.lerp(x, y, weight)
     else:
-        if isinstance(weight, float):
-            weight = paddle.full(shape=[1], fill_value=weight, dtype=x.dtype)
-
         check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'lerp')
         check_variable_and_dtype(y, 'y', ['float32', 'float64'], 'lerp')
         check_variable_and_dtype(

python/paddle/vision/transforms/functional_tensor.py

@@ -457,7 +457,7 @@ def rotate(
         nh = npos[0][1]
 
         if paddle.in_dynamic_mode():
-            ow, oh = int(nw.numpy()[0]), int(nh.numpy()[0])
+            ow, oh = int(nw), int(nh)
         else:
             ow, oh = nw.astype("int32"), nh.astype("int32")