[Dy2Stat]Support append method and initialized value for List in ControlFlow (#35212)

* Support append method and initialized value for List in ControlFlow

* polish error msg and en doc

* fix code style

python/paddle/fluid/dygraph/dygraph_to_static/list_transformer.py

@@ -93,7 +93,8 @@ class ListTransformer(gast.NodeTransformer):
         for child_node in gast.walk(node):
             if isinstance(child_node, gast.Assign):
                 if self._need_to_create_tensor_array(child_node):
-                    child_node.value = self._create_tensor_array()
+                    child_node.value = self._create_tensor_array(
+                        child_node.value)
 
     def _transform_list_append_in_control_flow(self, node):
         for child_node in gast.walk(node):
@@ -189,9 +190,11 @@ class ListTransformer(gast.NodeTransformer):
             return True
         return False
 
-    def _create_tensor_array(self):
+    def _create_tensor_array(self, value_node):
         # Although `dtype='float32'`, other types such as `int32` can also be supported
-        func_code = "paddle.tensor.create_array(dtype='float32')"
+        init_value = ast_to_source_code(value_node).strip()
+        func_code = "paddle.tensor.create_array('float32', {})".format(
+            init_value)
         func_node = gast.parse(func_code).body[0].value
         return func_node
 

python/paddle/fluid/layers/control_flow.py

@@ -1538,7 +1538,7 @@ def array_write(x, i, array=None):
     return array
 
 
-def create_array(dtype):
+def create_array(dtype, initialized_list=None):
     """
     This OP creates an LOD_TENSOR_ARRAY. It is used as
     the input of :ref:`api_fluid_layers_array_read` and 
@@ -1548,6 +1548,8 @@ def create_array(dtype):
     Args:
         dtype (str): The data type of the elements in the lod_tensor_array.
                      Support data type: float32, float64, int32, int64.
+        initialized_list(list): Used to initialize as default value for created array.
+                    All values in initialized list should be a Tensor.
 
     Returns:
         Variable: The empty lod_tensor_array. The data type of elements in Tensor is ``dtype``.
@@ -1559,15 +1561,35 @@ def create_array(dtype):
           data = fluid.layers.create_array(dtype='float32') # Create a float32 LoDTensorArray.
 
     """
+    array = []
+    if initialized_list is not None:
+        if not isinstance(initialized_list, (list, tuple)):
+            raise TypeError(
+                "Require type(initialized_list) should be list/tuple, but received {}".
+                format(type(initialized_list)))
+        array = list(initialized_list)
+
+    # NOTE: Only support plain list like [x, y,...], not support nested list in static mode.
+    for val in array:
+        if not isinstance(val, Variable):
+            raise TypeError(
+                "All values in `initialized_list` should be Variable, but recevied {}.".
+                format(type(val)))
+
     if in_dygraph_mode():
-        return []
+        return array
 
     helper = LayerHelper("array", **locals())
-    return helper.create_variable(
+    tensor_array = helper.create_variable(
         name="{0}.out".format(helper.name),
         type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
         dtype=dtype)
 
+    for val in array:
+        array_write(x=val, i=array_length(tensor_array), array=tensor_array)
+
+    return tensor_array
+
 
 @templatedoc()
 def less_than(x, y, force_cpu=None, cond=None, name=None):

python/paddle/fluid/layers/math_op_patch.py

@@ -18,8 +18,9 @@ import warnings
 import inspect
 
 from .. import core
-from ..framework import Variable, unique_name
+from ..framework import Variable, unique_name, static_only
 from .layer_function_generator import OpProtoHolder
+from .control_flow import array_write, array_length
 
 _supported_int_dtype_ = [
     core.VarDesc.VarType.BOOL,
@@ -182,6 +183,24 @@ def monkey_patch_variable():
         out.stop_gradient = self.stop_gradient
         return out
 
+    @static_only
+    def append(self, var):
+        """
+         **Notes**:
+            **The type variable must be LoD Tensor Array.
+        
+        """
+        if not isinstance(var, Variable):
+            raise TypeError(
+                "Required input var should be Variable, but received {}".format(
+                    type(var)))
+        if self.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY:
+            raise TypeError(
+                "Only Variable with VarType.LOD_TENSOR_ARRAY support `append` method, but received type: {}".
+                format(self.type))
+
+        array_write(x=var, i=array_length(self), array=self)
+
     def _scalar_op_(var, scale, bias):
         block = current_block(var)
         out = create_new_tmp_var(block, var.dtype)
@@ -344,6 +363,7 @@ def monkey_patch_variable():
         #   b=-a
         ('__neg__', _neg_),
         ('astype', astype),
+        ('append', append),
         ('dim', lambda x: len(x.shape)),
         ('ndimension', lambda x: len(x.shape)),
         ('ndim', _ndim_),

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

@@ -147,14 +147,17 @@ def test_list_pop_without_control_flow_2(x):
 def test_list_pop_in_if(x):
     x = fluid.dygraph.to_variable(x)
     a = []
+    b = [x * 2 + (x + 1)]
     if x.numpy()[0] > 0:
         a.append(x)
+        b.append(x + 1)
         a.append(fluid.layers.fill_constant(shape=[1], value=1, dtype="int64"))
     else:
         a.append(x + 1)
+        b.append(x - 1)
         a.append(fluid.layers.fill_constant(shape=[2], value=2, dtype="int64"))
     item1 = a.pop(1)
-    return item1
+    return item1, b[-1]
 
 
 def test_list_pop_in_for_loop(x, iter_num):
@@ -165,14 +168,16 @@ def test_list_pop_in_for_loop(x, iter_num):
     )  # TODO(liym27): Delete it if the type of parameter iter_num can be resolved
 
     a = []
+    b = [x - 1, x + 1]
     for i in range(iter_num):
         a.append(x + i)
+        b.append(x * 2)
 
     one = fluid.layers.ones(shape=[1], dtype="int32")
     for i in range(one.numpy()[0]):
         item = a.pop()
 
-    return a[0], item
+    return a[0], item, b[1]
 
 
 def test_list_pop_in_while_loop(x, iter_num):
@@ -180,14 +185,18 @@ def test_list_pop_in_while_loop(x, iter_num):
     iter_num = fluid.layers.fill_constant(
         shape=[1], value=iter_num, dtype="int32")
     a = []
+    b = [x]
+    b.append(x)
+    b.pop()
     i = 0
 
     while i < iter_num:
         a.append(x + i)
+        b.append(x - i)
         i += 1
         if i % 2 == 1:
             a.pop()
-    return a[0]
+    return a[0], b[2]
 
 
 class TestListWithoutControlFlow(unittest.TestCase):

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

@@ -15,6 +15,7 @@
 from __future__ import print_function
 
 import unittest
+import paddle
 import paddle.fluid.core as core
 import numpy
 
@@ -50,5 +51,49 @@ class TestLoDTensorArray(unittest.TestCase):
             self.assertEqual([[1]], t.recursive_sequence_lengths())
 
 
+class TestCreateArray(unittest.TestCase):
+    def setUp(self):
+        self.place = paddle.CPUPlace()
+        self.shapes = [[10, 4], [8, 12], [1]]
+
+    def test_initialized_list_and_error(self):
+        paddle.disable_static()
+        init_data = [
+            numpy.random.random(shape).astype('float32')
+            for shape in self.shapes
+        ]
+        array = paddle.tensor.create_array(
+            'float32', [paddle.to_tensor(x) for x in init_data])
+        for res, gt in zip(array, init_data):
+            self.assertTrue(numpy.array_equal(res, gt))
+
+        # test for None
+        array = paddle.tensor.create_array('float32')
+        self.assertTrue(isinstance(array, list))
+        self.assertEqual(len(array), 0)
+
+        # test error
+        with self.assertRaises(TypeError):
+            paddle.tensor.create_array('float32', 'str')
+
+    def test_static(self):
+        paddle.enable_static()
+        init_data = [paddle.ones(shape, dtype='int32') for shape in self.shapes]
+        array = paddle.tensor.create_array('float32', init_data)
+        for res, gt in zip(array, init_data):
+            self.assertTrue(res.shape, gt.shape)
+
+        # test error with nest list
+        with self.assertRaises(TypeError):
+            paddle.tensor.create_array('float32',
+                                       [init_data[0], [init_data[1]]])
+
+        # test error with not variable
+        with self.assertRaises(TypeError):
+            paddle.tensor.create_array('float32', ("str"))
+
+        paddle.enable_static()
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/tensor/array.py

@@ -122,13 +122,15 @@ def array_write(x, i, array=None):
     return layers.array_write(x, i, array)
 
 
-def create_array(dtype):
+def create_array(dtype, initialized_list=None):
     """
     This OP creates an array. It is used as the input of :ref:`api_paddle_tensor_array_array_read` and
     :ref:`api_paddle_tensor_array_array_write`.
 
     Args:
         dtype (str): The data type of the elements in the array. Support data type: float32, float64, int32, int64 and bool.
+        initialized_list(list): Used to initialize as default value for created array.
+                    All values in initialized list should be a Tensor.
 
     Returns:
         list|Tensor: An empty array. In dynamic mode, ``array`` is a Python list. But in static mode, array is a Tensor
@@ -149,4 +151,4 @@ def create_array(dtype):
             print(item)     # [[5., 5., 5.]]
 
     """
-    return layers.create_array(dtype)
+    return layers.create_array(dtype, initialized_list)