control_flow.py 167.9 KB
Newer Older
1
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
14 15

from __future__ import print_function
S
rename  
sneaxiy 已提交
16
from ..wrapped_decorator import signature_safe_contextmanager
D
dzhwinter 已提交
17

18
from .layer_function_generator import autodoc, templatedoc
19
from .tensor import assign, cast, fill_constant
20
from .. import core
H
hong 已提交
21
from ..framework import Program, Variable, Operator, _non_static_mode, static_only, _in_legacy_dygraph, in_dygraph_mode
22
from ..layer_helper import LayerHelper, unique_name
M
minqiyang 已提交
23
from .nn import logical_and, logical_not, logical_or
24
from .utils import assert_same_structure, map_structure, hold_mutable_vars, copy_mutable_vars, padding_to_same_structure, is_sequence, pack_sequence_as, flatten, to_sequence
Y
yuyang18 已提交
25
import numpy
26
import warnings
27
import six
L
liym27 已提交
28
from functools import reduce, partial
29
from ..data_feeder import convert_dtype, check_variable_and_dtype, check_type, check_dtype
30 31
from ... import compat as cpt
from ..backward import _infer_var_data_type_shape_
W
wanghuancoder 已提交
32
from paddle import _C_ops
D
dzhwinter 已提交
33

Q
QI JUN 已提交
34
__all__ = [
W
Wu Yi 已提交
35
    'While', 'Switch', 'increment', 'array_write', 'create_array', 'less_than',
Z
zhoukunsheng 已提交
36
    'less_equal', 'greater_than', 'greater_equal', 'equal', 'not_equal',
37
    'array_read', 'array_length', 'cond', 'IfElse', 'DynamicRNN', 'StaticRNN',
H
Huihuang Zheng 已提交
38 39
    'reorder_lod_tensor_by_rank', 'Print', 'Assert', 'is_empty', 'case',
    'switch_case', 'while_loop'
D
dzhwinter 已提交
40 41
]

Y
Yu Yang 已提交
42

43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
def select_output(input, outputs, mask):
    """
    **select_output**    
    This API takes in one input and multiple outputs and an integer mask. It
    selects the output specified by the mask and copy the input to selected
    output. It is useful in control flow.

    Args:
        input(Variable): The input variable
        outputs(tuple|list): The output variables
        mask(Variable): A tensor containing 1 integer number selecting which
            output to be copied with input

    Returns:
        Variable: The outputs variables
    """
    helper = LayerHelper('select_output', **locals())
60 61 62 63
    check_type(input, 'input', (Variable), 'select_output')
    check_variable_and_dtype(mask, 'mask', ['int32'], 'select_output')
    check_type(outputs, 'outputs', (list, tuple), 'select_output')

64 65 66 67 68 69
    helper.append_op(type='select_output',
                     inputs={
                         'X': input,
                         'Mask': mask
                     },
                     outputs={'Out': outputs})
70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
    return outputs


def select_input(inputs, mask):
    """
    **select_input**
    
    This API takes in multiple inputs and uses an integer mask to select one
    input to output. It is useful in control flow.

    Args:
        inputs(tuple|list): The input variables
        mask(Variable): A tensor containing 1 integer number selecting which
            input to output

    Returns:
        Variable: The selected input variable
    """
    helper = LayerHelper('select_input', **locals())
89 90 91
    check_type(inputs, 'inputs', (list, tuple), 'select_input')
    check_variable_and_dtype(mask, 'mask', ['int32'], 'select_input')

92 93 94
    input_dtype = inputs[1].dtype
    input_shape = inputs[1].shape
    input_type = inputs[1].type
95

96 97 98 99 100 101 102 103 104
    out = helper.create_variable(dtype=input_dtype,
                                 shape=input_shape,
                                 type=input_type)
    helper.append_op(type='select_input',
                     inputs={
                         'X': inputs,
                         'Mask': mask
                     },
                     outputs={'Out': out})
105 106 107
    return out


108 109
def select_input_with_buildin_type(inputs, mask):
    from paddle.fluid.dygraph.dygraph_to_static.variable_trans_func import to_static_variable
110
    from paddle.fluid.dygraph.dygraph_to_static.utils import UndefinedVar, create_undefined_var_like
111 112
    false_var, true_var = inputs

113 114 115 116 117 118
    if isinstance(false_var, UndefinedVar) and isinstance(
            true_var, UndefinedVar):
        """ None -> UndefinedVar, so the real value is a [None, UndefinedVar] or [None, None], we just return None.
        """
        return None

119 120 121
    if isinstance(false_var, Variable) and isinstance(true_var, Variable):
        return select_input(inputs, mask)

122 123
    elif (isinstance(false_var, (support_ret_buildin_type))
          and isinstance(false_var, type(true_var))):
124 125 126 127
        if false_var == true_var:
            return false_var
        else:
            inputs = [
128 129
                to_static_variable(false_var),
                to_static_variable(true_var)
130 131
            ]
    # Deal with the situations like this: false_var is int and true_var is Variable
132 133 134 135
    elif ((isinstance(false_var, support_ret_buildin_type)
           and isinstance(true_var, Variable))
          or (isinstance(true_var, support_ret_buildin_type)
              and isinstance(false_var, Variable))):
136 137 138 139 140
        inputs = [to_static_variable(false_var), to_static_variable(true_var)]
        warnings.warn(
            "Return results from different branches in cond are not same type: "
            "false_var returned by fasle_fn is '{}' and true_var of true_fn is "
            "'{}'".format(type(false_var), type(true_var)))
141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
    elif ((isinstance(false_var, UndefinedVar)
           and isinstance(true_var, (Variable, ) + support_ret_buildin_type))
          or (isinstance(true_var, UndefinedVar)
              and isinstance(false_var,
                             (Variable, ) + support_ret_buildin_type))):

        def create_var_if_not_undefined_var(a):
            if isinstance(a, UndefinedVar): return a
            return to_static_variable(a)

        def create_like_if_undefined_var(a, b):
            if isinstance(a, UndefinedVar): return create_undefined_var_like(b)
            return a

        # TODO(xiongkun): add warning here.
        true_var, false_var = create_var_if_not_undefined_var(
            true_var), create_var_if_not_undefined_var(false_var)
        inputs = [
            create_like_if_undefined_var(false_var, true_var),
            create_like_if_undefined_var(true_var, false_var)
        ]
162 163 164 165 166 167 168 169 170
    else:
        raise TypeError(
            "Unsupported return type of true_fn and false_fn in cond: false_var "
            "returned by fasle_fn is '{}' and true_var of true_fn is '{}'".
            format(type(false_var), type(true_var)))

    return select_input(inputs, mask)


171
def split_lod_tensor(input, mask, level=0):
172 173 174 175
    """
    This function takes in an input that contains the complete lod information,
    and takes in a mask which is used to mask certain parts of the input.
    The output is the true branch and the false branch with the mask applied to
Q
qiaolongfei 已提交
176 177
    the input at a certain level in the tensor. Mainly used in IfElse to split
    data into two parts.
178 179

    Args:
180
        input(Variable|tuple|list|None): The input tensor that contains complete
181
                                lod information needed to construct the output.
182
        mask(Variable|list): A bool column vector which masks the input.
Q
qiaolongfei 已提交
183
        level(int): The specific lod level to split.
184 185

    Returns:
Q
qiaolongfei 已提交
186 187 188 189
        tuple(Variable, Variable):
        The true branch of tensor as per the mask applied to input.

        The false branch of tensor as per the mask applied to input.
190 191 192 193

    Examples:
        .. code-block:: python

194
          import paddle.fluid as fluid
Q
qiaolongfei 已提交
195
          x = fluid.layers.data(name='x', shape=[1])
196 197
          x.persistable = True

Q
qiaolongfei 已提交
198
          y = fluid.layers.data(name='y', shape=[1])
199 200
          y.persistable = True

Q
qiaolongfei 已提交
201
          out_true, out_false = fluid.layers.split_lod_tensor(
202
                input=x, mask=y, level=level)
203

204
    """
205 206 207 208
    check_type(input, 'input', (Variable, list, tuple, type(None)),
               'fluid.layers.split_lod_tensor')
    check_type(mask, 'mask', (Variable, list), 'fluid.layers.split_lod_tensor')
    check_type(level, 'level', int, 'fluid.layers.split_lod_tensor')
209
    helper = LayerHelper('split_lod_tensor', **locals())
X
Xin Pan 已提交
210 211
    out_true = helper.create_variable_for_type_inference(dtype=input.dtype)
    out_false = helper.create_variable_for_type_inference(dtype=input.dtype)
212 213 214 215 216 217 218 219 220 221
    helper.append_op(type='split_lod_tensor',
                     inputs={
                         'X': input,
                         'Mask': mask,
                     },
                     outputs={
                         'OutTrue': out_true,
                         'OutFalse': out_false
                     },
                     attrs={'level': level})
222 223 224
    return out_true, out_false


225
def merge_lod_tensor(in_true, in_false, x, mask, level=0):
226 227 228 229 230
    """
    **merge_lod_tensor**

    This function takes in an input :math:`x`, the True branch, the False
    branch and a binary :math:`mask`. Using this information, this function
Q
qiaolongfei 已提交
231 232 233
    merges the True and False branches of the tensor into a single tensor as
    output at a certain lod level indicated by :math:`level`. Used in IfElse
    to merge the output if True block and False Block.
234 235

    Args:
236 237 238
        in_true(Variable|tuple|list|None): The True branch to be merged.
        in_false(Variable|tuple|list|None): The False branch to be merged.
        x(Variable|tuple|list|None): The input tensor that contains complete
239
                            lod information needed to construct the output.
240
        mask(Variable|list): A bool column vector which masks the input.
Q
qiaolongfei 已提交
241
        level(int): The specific lod level to merge.
242 243 244 245 246 247 248

    Returns:
        Variable: The merged output tensor.

    Examples:
        .. code-block:: python

249
          import paddle.fluid as fluid
250 251 252 253 254 255 256 257 258 259 260 261
          x = layers.data(
                      name='x', shape=[1], dtype='float32', stop_gradient=False)
          y = layers.data(
                name='y', shape=[1], dtype='bool', stop_gradient=False)

          level = 0

          out_true, out_false = layers.split_lod_tensor(
                input=x, mask=y, level=level)
          out = layers.merge_lod_tensor(
                in_true=out_true, in_false=out_false, mask=y, x=x, level=level)
    """
262
    helper = LayerHelper('merge_lod_tensor', **locals())
263 264 265 266 267 268 269
    check_type(x, 'x', (Variable, list, tuple, type(None)),
               'fluid.layers.merge_lod_tensor')
    check_type(mask, 'mask', (Variable, list), 'fluid.layers.merge_lod_tensor')
    check_type(in_true, 'in_true', (Variable, list, tuple, type(None)),
               'fluid.layers.merge_lod_tensor')
    check_type(in_false, 'in_false', (Variable, list, tuple, type(None)),
               'fluid.layers.merge_lod_tensor')
X
Xin Pan 已提交
270
    out = helper.create_variable_for_type_inference(dtype=in_true.dtype)
271 272 273 274 275 276 277 278 279
    helper.append_op(type='merge_lod_tensor',
                     inputs={
                         'X': x,
                         'Mask': mask,
                         'InTrue': in_true,
                         'InFalse': in_false
                     },
                     outputs={'Out': out},
                     attrs={'level': level})
280 281 282
    return out


283
@static_only
Y
Yan Chunwei 已提交
284 285 286
def Print(input,
          first_n=-1,
          message=None,
287
          summarize=20,
Y
Yan Chunwei 已提交
288 289 290
          print_tensor_name=True,
          print_tensor_type=True,
          print_tensor_shape=True,
291
          print_tensor_layout=True,
Y
yangyaming 已提交
292 293
          print_tensor_lod=True,
          print_phase='both'):
Y
Yan Chunwei 已提交
294
    '''
295 296
    :api_attr: Static Graph

Y
Yan Chunwei 已提交
297 298 299 300 301 302 303 304 305
    **Print operator**

    This creates a print op that will print when a tensor is accessed.

    Wraps the tensor passed in so that whenever that a tensor is accessed,
    the message `message` is printed, along with the current value of the
    tensor `t`.

    Args:
Y
yangyaming 已提交
306
        input (Variable): A Tensor to print.
307
        summarize (int): Number of elements in the tensor to be print. If it's
T
tianshuo78520a 已提交
308
                value is -1, then all elements in the tensor will be print.
Y
yangyaming 已提交
309 310
        message (str): A string message to print as a prefix.
        first_n (int): Only log `first_n` number of times.
311 312 313
        print_tensor_name (bool, optional): Print the tensor name. Default: True.
        print_tensor_type (bool, optional): Print the tensor type. Defaultt: True.
        print_tensor_shape (bool, optional): Print the tensor shape. Default: True.
314
        print_tensor_layout (bool, optional): Print the tensor layout. Default: True.
315
        print_tensor_lod (bool, optional): Print the tensor lod. Default: True.
316
        print_phase (str): Which phase to displace, including 'forward',
317 318 319
                'backward' and 'both'. Default: 'both'. If set to 'backward', will 
                only print the gradients of input tensor; If set to 'both', will
                both print the input tensor itself and the gradients of input tensor.
Y
Yan Chunwei 已提交
320 321

    Returns:
322
        Variable: Output tensor.
Y
Yan Chunwei 已提交
323

324 325 326 327
    NOTES:
        The input and output are two different variables, and in the
        following process, you should use the output variable but not the input,
        otherwise, the print layer doesn't have backward.
Y
Yan Chunwei 已提交
328

Y
Yan Chunwei 已提交
329 330
    Examples:
        .. code-block:: python
331
           
332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349
           import paddle

           paddle.enable_static()
        
           x = paddle.full(shape=[2, 3], fill_value=3, dtype='int64')
           out = paddle.static.Print(x, message="The content of input layer:")

           main_program = paddle.static.default_main_program()
           exe = paddle.static.Executor(place=paddle.CPUPlace())
           res = exe.run(main_program, fetch_list=[out])
           # Variable: fill_constant_1.tmp_0
           #   - message: The content of input layer:
           #   - lod: {}
           #   - place: CPUPlace
           #   - shape: [2, 3]
           #   - layout: NCHW
           #   - dtype: long
           #   - data: [3 3 3 3 3 3]
Y
Yan Chunwei 已提交
350
    '''
351 352 353
    check_variable_and_dtype(input, 'input',
                             ['float32', 'float64', 'int32', 'int64', 'bool'],
                             'fluid.layers.Print')
354

355 356
    helper = LayerHelper('print' + "_" + input.name, **locals())
    output = helper.create_variable_for_type_inference(input.dtype)
357 358 359 360 361 362 363 364 365 366 367 368 369 370
    helper.append_op(type='print',
                     inputs={'In': input},
                     outputs={'Out': output},
                     attrs={
                         'first_n': first_n,
                         'summarize': summarize,
                         'message': message or "",
                         'print_tensor_name': print_tensor_name,
                         'print_tensor_type': print_tensor_type,
                         'print_tensor_shape': print_tensor_shape,
                         'print_tensor_layout': print_tensor_layout,
                         'print_tensor_lod': print_tensor_lod,
                         'print_phase': print_phase.upper()
                     })
371
    return output
Y
Yan Chunwei 已提交
372 373


H
Huihuang Zheng 已提交
374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436
def Assert(cond, data=None, summarize=20, name=None):
    '''
    This API creates an op that asserts the given condition is true. If the
    condition is false, prints the tensors in data. ``summarize`` specifies the
    number of the elements in the tensors to print.

    Args:
        cond (Variable): The boolean condition tensor whose numel should be 1.
        data (list|tuple, optional): list or tuple of tensors to print when
            condition is not true. If it's ``None``, no tensor will be printed.
            The default value is ``None``.
        summarize (int, optional): Number of elements in the tensor to be
            printed. If its value is -1, then all elements in the tensor will
            be printed. The default value is 20.
        name (str, optional): The default value is ``None`` . Normally users
            don't have to set this parameter. For more information, please
            refer to :ref:`api_guide_Name` .

    Returns:
        Operator: the created operation.

    Raises:
        TypeError: If ``cond`` is not boolean Variable.
        TypeError: If ``data`` is not a list or tuple or ``None``.
        TypeError: If ``summarize`` is not int.
        TypeError: If ``name`` is not a string or ``None`` .
        fluid.core.EnforceNotMet: If the condition is False in running time.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            import paddle.fluid.layers as layers

            x = layers.fill_constant(shape=[2, 3], dtype='float32', value=2.0)
            condition = layers.reduce_max(x) < 1.0 # False
            layers.Assert(condition, [x], 10, "example_assert_layer")

            exe = fluid.Executor()
            try:
                exe.run(fluid.default_main_program())
                # Print x and throws paddle.fluid.core.EnforceNotMet exception
                # Example printed message for x:
                #
                # Variable: fill_constant_0.tmp_0
                #   - lod: {}
                #   - place: CPUPlace()
                #   - shape: [2, 3]
                #   - layout: NCHW
                #   - dtype: float
                #   - data: [2 2 2 2 2 2]
            except fluid.core.EnforceNotMet as e:
                print("Assert Exception Example")

    '''
    check_variable_and_dtype(cond, "cond", ["bool"], "fluid.layers.Assert")
    check_type(data, "data", (list, tuple, type(None)), "fluid.layers.Assert")
    check_type(summarize, "summarize", int, "fluid.layers.Assert")
    check_type(name, "name", (str, type(None)), "fluid.layers.Assert")

    layer_name = name if name else ('assert_' + cond.name)
    helper = LayerHelper(layer_name, **locals())

437 438 439 440 441 442
    op = helper.append_op(type="assert",
                          inputs={
                              "Cond": cond,
                              "Data": [] if data is None else list(data)
                          },
                          attrs={"summarize": summarize})
H
Huihuang Zheng 已提交
443 444 445 446

    return op


Y
Yu Yang 已提交
447 448
class BlockGuard(object):
    """
449 450 451 452
    BlockGuard class.

    BlockGuard class is used to create a sub-block in a program by
    using the Python `with` keyword.
Y
Yu Yang 已提交
453 454
    """

455 456
    def __init__(self, main_program):
        if not isinstance(main_program, Program):
Y
Yu Yang 已提交
457
            raise TypeError("BlockGuard takes a program")
458
        self.main_program = main_program
Y
Yu Yang 已提交
459 460

    def __enter__(self):
W
Wu Yi 已提交
461
        self.main_program._create_block()
Y
Yu Yang 已提交
462 463

    def __exit__(self, exc_type, exc_val, exc_tb):
W
Wu Yi 已提交
464
        self.main_program._rollback()
Y
Yu Yang 已提交
465 466 467 468 469
        if exc_type is not None:
            return False  # re-raise exception
        return True


Y
Yang Yang 已提交
470 471 472 473 474
class BlockGuardWithCompletion(BlockGuard):
    """
    BlockGuardWithCompletion class.

    BlockGuardWithCompletion class is used to create an op with a block in a program.
475 476
    """

Y
Yu Yang 已提交
477
    def __init__(self, rnn):
X
Xin Pan 已提交
478
        if not isinstance(rnn, StaticRNN):
X
Xin Pan 已提交
479
            raise TypeError("BlockGuardWithCompletion takes a StaticRNN")
Y
Yang Yang 已提交
480
        super(BlockGuardWithCompletion, self).__init__(rnn.helper.main_program)
Y
Yu Yang 已提交
481 482 483 484
        self.rnn = rnn

    def __enter__(self):
        self.rnn.status = StaticRNN.IN_RNN_BLOCK
Y
Yang Yang 已提交
485
        return super(BlockGuardWithCompletion, self).__enter__()
Y
Yu Yang 已提交
486 487

    def __exit__(self, exc_type, exc_val, exc_tb):
Y
Yu Yang 已提交
488 489
        if exc_type is not None:
            return False
Y
Yu Yang 已提交
490
        self.rnn.status = StaticRNN.AFTER_RNN_BLOCK
491
        self.rnn._complete_op()
492 493
        return super(BlockGuardWithCompletion,
                     self).__exit__(exc_type, exc_val, exc_tb)
Y
Yu Yang 已提交
494 495 496 497


class StaticRNNMemoryLink(object):
    """
498 499 500 501
    StaticRNNMemoryLink class.

    StaticRNNMemoryLink class is used to create a link between two
    memory cells of a StaticRNN.
Y
yuyang18 已提交
502 503 504 505 506 507 508 509 510


    NOTE: This is a internal data structure of a very low-level API.
    Please use StaticRNN instead.

    Args:
        init(Variable): the initial variable for Memory.
        pre_mem(Variable): the memory variable in previous time step.
        mem(Variable): the memory variable in current time step.
Y
Yu Yang 已提交
511 512 513 514 515 516 517 518 519
    """

    def __init__(self, init, pre_mem, mem=None):
        self.init = init
        self.pre_mem = pre_mem
        self.mem = mem


class StaticRNN(object):
520
    """
521 522
    :api_attr: Static Graph

523 524
    StaticRNN class.

525 526 527 528 529 530 531
    The StaticRNN can process a batch of sequence data. The first dimension of inputs
    represents sequence length, the length of each input sequence must be equal.
    StaticRNN will unfold sequence into time steps, user needs to define how to process
    each time step during the :code:`with` step.

    Args:
        name (str, optional): Please refer to :ref:`api_guide_Name`, Default None.
C
chengduo 已提交
532 533

    Examples:
534 535 536 537 538 539
        .. code-block:: python

            import paddle.fluid as fluid
            import paddle.fluid.layers as layers

            vocab_size, hidden_size=10000, 200
540 541
            x = fluid.data(name="x", shape=[None, 1, 1], dtype='int64')
            # create word sequence
542 543 544 545 546
            x_emb = layers.embedding(
                input=x,
                size=[vocab_size, hidden_size],
                dtype='float32',
                is_sparse=False)
547
            # transform batch size to dim 1
548 549 550 551
            x_emb = layers.transpose(x_emb, perm=[1, 0, 2])

            rnn = fluid.layers.StaticRNN()
            with rnn.step():
552
                # mark created x_emb as input, each step process a word
553
                word = rnn.step_input(x_emb)
554
                # create prev memory parameter, batch size comes from word
555 556
                prev = rnn.memory(shape=[-1, hidden_size], batch_ref = word)
                hidden = fluid.layers.fc(input=[word, prev], size=hidden_size, act='relu')
557 558 559
                # use hidden to update prev
                rnn.update_memory(prev, hidden)
                # mark hidden as output 
560
                rnn.step_output(hidden)
561
            # get StaticrNN final output
562
            result = rnn()
C
chengduo 已提交
563

564
    """
Y
Yu Yang 已提交
565 566 567 568
    BEFORE_RNN_BLOCK = 0
    IN_RNN_BLOCK = 1
    AFTER_RNN_BLOCK = 2

569
    def __init__(self, name=None):
570
        check_type(name, "name", (str, type(None)), "fluid.layers.StaticRNN")
571
        self.helper = LayerHelper("static_rnn", name=name)
Y
Yu Yang 已提交
572 573 574 575 576 577 578 579
        self.memories = {}  # memory map, from pre_mem.name --> MemoryLink
        self.inputs = []  # input variable list in current block
        self.outputs = []  # output variable list in parent block
        self.status = StaticRNN.BEFORE_RNN_BLOCK  # status flag.
        # sequence length, since it is a static RNN, sequence length are fixed.
        self.seq_len = None

    def step(self):
C
chengduo 已提交
580
        """
581 582
        Define operators in each step. step is used in :code:`with` block, OP in :code:`with` block
        will be executed sequence_len times (sequence_len is the length of input)
C
chengduo 已提交
583
        """
Y
Yang Yang 已提交
584
        return BlockGuardWithCompletion(self)
Y
Yu Yang 已提交
585 586 587 588 589

    def _assert_in_rnn_block_(self, method):
        if self.status != StaticRNN.IN_RNN_BLOCK:
            raise ValueError("You must invoke {0} in rnn block".format(method))

590 591 592 593 594 595 596
    def memory(self,
               init=None,
               shape=None,
               batch_ref=None,
               init_value=0.0,
               init_batch_dim_idx=0,
               ref_batch_dim_idx=1):
597
        """
C
chengduo 已提交
598 599 600
        Create a memory variable for static rnn.
        If the :code:`init` is not None, :code:`memory` will be initialized by
        this Variable. If the :code:`init` is None, :code:`shape` and :code:`batch_ref`
601 602
        must be set, and this function will create a new variable with shape and batch_ref
        to initialize :code:`init` Variable.
C
chengduo 已提交
603

604
        Args:
605
            init(Variable, optional): Tensor used to init memory. If it is not set,
C
chengduo 已提交
606 607
                :code:`shape` and :code:`batch_ref` must be provided.
                Default: None.
608 609 610 611 612 613 614
            shape(list|tuple): When :code:`init` is None use this arg to initialize memory shape.
            NOTE the shape does not contain batch_size. Default: None.
            batch_ref(Variable, optional): When :code:`init` is None, memory's batch size will
            be set as batch_ref's ref_batch_dim_idx value. Default: None.
            init_value(float, optional): When :code:`init` is None, used to init memory's value. Default: 0.0.
            init_batch_dim_idx(int, optional): the batch_size axis of the :code:`init` Variable. Default: 0.
            ref_batch_dim_idx(int, optional): the batch_size axis of the :code:`batch_ref` Variable. Default: 1.
C
chengduo 已提交
615 616

        Returns:
617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
            Variable: The memory variable.

        Examples 1:
            .. code-block:: python

            	import paddle.fluid as fluid
            	import paddle.fluid.layers as layers

            	vocab_size, hidden_size=10000, 200
            	x = fluid.data(name="x", shape=[None, 1, 1], dtype='int64')
            	# create word sequence
            	x_emb = layers.embedding(
                	input=x,
                	size=[vocab_size, hidden_size],
                	dtype='float32',
                	is_sparse=False)
            	# transform batch size to dim 1
            	x_emb = layers.transpose(x_emb, perm=[1, 0, 2])

            	rnn = fluid.layers.StaticRNN()
            	with rnn.step():
                	# mark created x_emb as input, each step process a word
                	word = rnn.step_input(x_emb)
                	# create prev memory parameter, batch size comes from word
                	prev = rnn.memory(shape=[-1, hidden_size], batch_ref = word)
                	hidden = fluid.layers.fc(input=[word, prev], size=hidden_size, act='relu')
                	# use hidden to update prev
                	rnn.update_memory(prev, hidden)


        Examples 2:
648 649
            .. code-block:: python

650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
            	import paddle.fluid as fluid
            	import paddle.fluid.layers as layers
            	vocab_size, hidden_size=10000, 200
            	x = fluid.data(name="x", shape=[None, 1, 1], dtype='int64')
            	# create word sequence
            	x_emb = layers.embedding(
                	input=x,
                	size=[vocab_size, hidden_size],
                	dtype='float32',
                	is_sparse=False)
            	# transform batch size to dim 1
            	x_emb = layers.transpose(x_emb, perm=[1, 0, 2])
            	boot_memory = fluid.layers.data(name='boot', shape=[hidden_size], dtype='float32', lod_level=1)
            	rnn = fluid.layers.StaticRNN()
            	with rnn.step():
            		# mark created x_emb as input, each step process a word
            		word = rnn.step_input(x_emb)
            		# init memory
            		prev = rnn.memory(init=boot_memory)
            		hidden = fluid.layers.fc(input=[word, prev], size=hidden_size, act='relu')
            		# update hidden with prev
            		rnn.update_memory(prev, hidden)

673
        """
Y
Yu Yang 已提交
674
        self._assert_in_rnn_block_('memory')
675 676 677 678 679 680
        check_type(init, "init", (Variable, type(None)),
                   "fluid.layers.StaticRNN.memory")
        check_type(shape, "shape", (list, tuple, type(None)),
                   "fluid.layers.StaticRNN.memory")
        check_type(batch_ref, "batch_ref", (Variable, type(None)),
                   "fluid.layers.StaticRNN.memory")
Y
Yu Yang 已提交
681
        if init is None:
682
            if shape is None or batch_ref is None:
Y
Yu Yang 已提交
683
                raise ValueError(
684
                    "if init is None, memory at least need shape and batch_ref")
685
            parent_block = self._parent_block()
686
            var_name = unique_name.generate_with_ignorable_key("@".join(
Y
Yu Yang 已提交
687
                [self.helper.name, "memory_boot"]))
688 689 690 691 692 693 694 695 696 697 698 699 700 701 702
            boot_var = parent_block.create_var(name=var_name,
                                               shape=shape,
                                               dtype=batch_ref.dtype,
                                               persistable=False)

            parent_block.append_op(type="fill_constant_batch_size_like",
                                   inputs={'Input': [batch_ref]},
                                   outputs={'Out': [boot_var]},
                                   attrs={
                                       'value': init_value,
                                       'shape': boot_var.shape,
                                       'dtype': boot_var.dtype,
                                       'input_dim_idx': ref_batch_dim_idx,
                                       'output_dim_idx': init_batch_dim_idx
                                   })
Y
Yu Yang 已提交
703 704 705 706

            return self.memory(init=boot_var)
        else:
            pre_mem = self.helper.create_variable(
707 708
                name=unique_name.generate_with_ignorable_key("@".join(
                    [self.helper.name, "mem"])),
F
fengjiayi 已提交
709
                dtype=init.dtype,
Y
Yu Yang 已提交
710
                shape=init.shape)
711 712
            self.memories[pre_mem.name] = StaticRNNMemoryLink(init=init,
                                                              pre_mem=pre_mem)
Y
Yu Yang 已提交
713 714 715
            return pre_mem

    def step_input(self, x):
C
chengduo 已提交
716 717 718 719 720 721 722 723
        """
        Mark a sequence as a StaticRNN input.

        Args:
            x(Variable): The input sequence, the shape of x
                should be [seq_len, ...].

        Returns:
724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752
            Variable: The current time step data in the input sequence.

        Examples:
            .. code-block:: python

            	import paddle.fluid as fluid
            	import paddle.fluid.layers as layers

            	vocab_size, hidden_size=10000, 200
            	x = fluid.data(name="x", shape=[None, 1, 1], dtype='int64')
            	# create word sequence
            	x_emb = layers.embedding(
                	input=x,
                	size=[vocab_size, hidden_size],
                	dtype='float32',
                	is_sparse=False)
            	# transform batch size to dim 1
            	x_emb = layers.transpose(x_emb, perm=[1, 0, 2])

            	rnn = fluid.layers.StaticRNN()
            	with rnn.step():
                	# mark created x_emb as input, each step process a word
                	word = rnn.step_input(x_emb)
                	# create prev memory parameter, batch size comes from word
                	prev = rnn.memory(shape=[-1, hidden_size], batch_ref = word)
                	hidden = fluid.layers.fc(input=[word, prev], size=hidden_size, act='relu')
                	# use hidden to update prev
                	rnn.update_memory(prev, hidden)

C
chengduo 已提交
753
        """
Y
Yu Yang 已提交
754
        self._assert_in_rnn_block_('step_input')
755
        check_type(x, "x", Variable, "fluid.layers.StaticRNN.step_input")
Y
Yu Yang 已提交
756
        if self.seq_len is None:
Y
Yu Yang 已提交
757
            self.seq_len = x.shape[0]
758
        elif x.shape[0] != -1 and self.seq_len != x.shape[0]:
Y
Yu Yang 已提交
759 760
            raise ValueError("Static RNN only take fix seq_len input")

761 762 763 764
        ipt = self.helper.create_variable(name=x.name,
                                          dtype=x.dtype,
                                          shape=list(x.shape[1:]),
                                          type=x.type)
Y
Yu Yang 已提交
765 766 767 768
        self.inputs.append(ipt)
        return ipt

    def step_output(self, o):
C
chengduo 已提交
769 770 771 772 773 774 775 776
        """
        Mark a sequence as a StaticRNN output.

        Args:
            o(Variable): The output sequence.

        Returns:
            None.
777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807

        Examples:
            .. code-block:: python

            	import paddle.fluid as fluid
            	import paddle.fluid.layers as layers

            	vocab_size, hidden_size=10000, 200
            	x = fluid.data(name="x", shape=[None, 1, 1], dtype='int64')
            	# create word sequence
            	x_emb = layers.embedding(
                	input=x,
                	size=[vocab_size, hidden_size],
               		dtype='float32',
                	is_sparse=False)
            	# transform batch size to dim 1
            	x_emb = layers.transpose(x_emb, perm=[1, 0, 2])

            	rnn = fluid.layers.StaticRNN()
            	with rnn.step():
                	# mark created x_emb as input, each step process a word
               		word = rnn.step_input(x_emb)
                	# create prev memory parameter, batch size comes from word
                	prev = rnn.memory(shape=[-1, hidden_size], batch_ref = word)
                	hidden = fluid.layers.fc(input=[word, prev], size=hidden_size, act='relu')
                	# use hidden to update prev
                	rnn.update_memory(prev, hidden)
                	rnn.step_output(hidden)

            	result = rnn()

C
chengduo 已提交
808
        """
Y
Yu Yang 已提交
809
        self._assert_in_rnn_block_('step_output')
810
        check_type(o, "o", Variable, "fluid.layers.StaticRNN.step_output")
Y
Yu Yang 已提交
811

X
Xin Pan 已提交
812
        tmp_o = self.helper.create_variable_for_type_inference(dtype=o.dtype)
813 814 815 816
        self.helper.append_op(type='rnn_memory_helper',
                              inputs={'X': [o]},
                              outputs={'Out': tmp_o},
                              attrs={'dtype': o.dtype})
Y
Yu Yang 已提交
817

818 819 820 821
        out_var = self._parent_block().create_var(name=tmp_o.name,
                                                  shape=[self.seq_len] +
                                                  list(tmp_o.shape),
                                                  dtype=tmp_o.dtype)
Y
Yu Yang 已提交
822 823 824 825

        self.outputs.append(out_var)

    def output(self, *outputs):
C
chengduo 已提交
826 827 828 829
        """
        Mark the StaticRNN output variables.

        Args:
830
            outputs: The output Tensor, can mark multiple variables as output
C
chengduo 已提交
831 832 833

        Returns:
            None
834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864

        Examples:
            .. code-block:: python

            	import paddle.fluid as fluid
            	import paddle.fluid.layers as layers

            	vocab_size, hidden_size=10000, 200
            	x = fluid.data(name="x", shape=[None, 1, 1], dtype='int64')
            	# create word sequence
            	x_emb = layers.embedding(
                	input=x,
                	size=[vocab_size, hidden_size],
                	dtype='float32',
                	is_sparse=False)
            	# transform batch size to dim 1
            	x_emb = layers.transpose(x_emb, perm=[1, 0, 2])

            	rnn = fluid.layers.StaticRNN()
            	with rnn.step():
                	# mark created x_emb as input, each step process a word
                	word = rnn.step_input(x_emb)
                	# create prev memory parameter, batch size comes from word
                	prev = rnn.memory(shape=[-1, hidden_size], batch_ref = word)
                	hidden = fluid.layers.fc(input=[word, prev], size=hidden_size, act='relu')
                	# use hidden to update prev
                	rnn.update_memory(prev, hidden)
                	# mark each step's hidden and word as output
                	rnn.output(hidden, word)

            	result = rnn()
C
chengduo 已提交
865
        """
Y
Yu Yang 已提交
866 867 868 869
        for each in outputs:
            self.step_output(each)

    def update_memory(self, mem, var):
C
chengduo 已提交
870
        """
871
        Update the memory from :code:`mem` to :code:`var`.
C
chengduo 已提交
872 873 874

        Args:
            mem(Variable): the memory variable.
875
            var(Variable): the plain variable generated in RNN block, used to update memory.
T
tianshuo78520a 已提交
876
                           var and mem should have same dims and data type.
C
chengduo 已提交
877 878 879

        Returns:
            None
880

C
chengduo 已提交
881
        """
882 883
        check_type(mem, "mem", Variable, "fluid.layers.StaticRNN.update_memory")
        check_type(var, "var", Variable, "fluid.layers.StaticRNN.update_memory")
Y
Yu Yang 已提交
884 885
        self.memories[mem.name].mem = var

886
    def _parent_block(self):
887
        prog = self.helper.main_program
Y
Yu Yang 已提交
888 889 890 891 892 893 894 895 896 897 898 899 900 901 902
        parent_idx = prog.current_block().parent_idx
        assert parent_idx >= 0
        parent_block = prog.block(parent_idx)
        return parent_block

    def __call__(self, *args, **kwargs):
        if self.status != StaticRNN.AFTER_RNN_BLOCK:
            raise ValueError("RNN output can only be retrieved after rnn block")
        if len(self.outputs) == 0:
            raise ValueError("RNN has no output")
        elif len(self.outputs) == 1:
            return self.outputs[0]
        else:
            return self.outputs

903
    def _complete_op(self):
904 905
        main_program = self.helper.main_program
        rnn_block = main_program.current_block()
906
        parent_block = self._parent_block()
Y
Yu Yang 已提交
907 908 909 910 911 912 913 914 915 916 917 918 919 920

        local_inputs = set()

        for op in rnn_block.ops:
            assert isinstance(op, Operator)
            for oname in op.output_names:
                for out_var_name in op.output(oname):
                    local_inputs.add(out_var_name)

        for var in self.inputs:
            local_inputs.add(var.name)
        for m in self.memories:
            local_inputs.add(m)

C
chengduo 已提交
921 922 923
        # NOTE(zcd): the params have two categories of variables.
        #   - the variables that are the out of StaticRnn.
        #   - the variables that are the parameters of some layers, for example, conv2d.
Y
Yu Yang 已提交
924 925 926 927 928 929 930 931
        params = list()
        for op in rnn_block.ops:
            assert isinstance(op, Operator)
            for iname in op.input_names:
                for in_var_name in op.input(iname):
                    if in_var_name not in local_inputs:
                        params.append(in_var_name)

932 933 934
        parameters = [
            parent_block._find_var_recursive(name) for name in set(params)
        ]
Y
Yu Yang 已提交
935 936 937 938 939 940 941

        step_scope = parent_block.create_var(
            type=core.VarDesc.VarType.STEP_SCOPES)

        inlinks = [parent_block.var(i.name) for i in self.inputs]
        outlinks = self.outputs

C
chengduo 已提交
942
        # NOTE(zcd): the states maybe empty in some case.
Y
Yu Yang 已提交
943 944 945
        boot_memories = []
        pre_memories = []
        memories = []
M
minqiyang 已提交
946
        for _, mem in six.iteritems(self.memories):
Y
Yu Yang 已提交
947 948
            boot_memories.append(mem.init)
            pre_memories.append(mem.pre_mem.name)
C
chengduo 已提交
949 950
            assert mem.mem is not None, "%s should be updated in every step." % (
                mem.init.name)
Y
Yu Yang 已提交
951 952
            mem_var = rnn_block.var(mem.mem.name)
            assert isinstance(mem_var, Variable)
X
Xin Pan 已提交
953 954
            new_mem = self.helper.create_variable_for_type_inference(
                dtype=mem_var.dtype)
955 956 957 958
            rnn_block.append_op(type='rnn_memory_helper',
                                inputs={'X': [mem_var]},
                                outputs={'Out': [new_mem]},
                                attrs={'dtype': mem_var.dtype})
Y
Yu Yang 已提交
959 960 961

            memories.append(new_mem.name)

962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977
        parent_block.append_op(type='recurrent',
                               inputs={
                                   'inputs': inlinks,
                                   'initial_states': boot_memories,
                                   'parameters': parameters
                               },
                               outputs={
                                   'outputs': outlinks,
                                   'step_scopes': [step_scope]
                               },
                               attrs={
                                   'has_states': len(pre_memories) > 0,
                                   'ex_states': pre_memories,
                                   'states': memories,
                                   'sub_block': rnn_block
                               })
Y
Yu Yang 已提交
978 979


Y
Yang Yang(Tony) 已提交
980
class WhileGuard(BlockGuard):
981

Y
Yang Yang(Tony) 已提交
982 983 984 985 986 987 988 989 990 991 992 993 994 995
    def __init__(self, while_op):
        if not isinstance(while_op, While):
            raise TypeError("WhileGuard takes a while op")
        super(WhileGuard, self).__init__(while_op.helper.main_program)
        self.while_op = while_op

    def __enter__(self):
        self.while_op.status = While.IN_WHILE_BLOCK
        return super(WhileGuard, self).__enter__()

    def __exit__(self, exc_type, exc_val, exc_tb):
        if exc_type is not None:
            return False
        self.while_op.status = While.AFTER_WHILE_BLOCK
996
        self.while_op._complete()
Y
Yang Yang(Tony) 已提交
997 998 999
        return super(WhileGuard, self).__exit__(exc_type, exc_val, exc_tb)


1000 1001 1002 1003 1004 1005 1006 1007 1008 1009
def get_inputs_outputs_in_block(current_block, inner_inputs, inner_outputs,
                                helper):
    """
    Find inputs and outputs in current control flow block.
    :param current_block: Current control flow block.
    :param inner_inputs: Input var name of ops in current block.
    :param inner_outputs: Output var name of ops in current block.
    :return: inner_inputs, inner_outputs
    """

1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022
    def is_ignore_vars(op, var_name):
        # NOTE(dev): There are some persistable var created in some non-standard API
        # such as "contrib.layers.shuffle_batch". It create a "Seed" used both in
        # Input and Output. This var shall not be considered as a loop_var in
        # control_flow.
        IGNORE_VAR_NAMES = {"shuffle_batch": ["shuffle_batch_seed"]}
        if op.type in IGNORE_VAR_NAMES:
            var_names = IGNORE_VAR_NAMES[op.type]
            for name in var_names:
                if name in var_name:
                    return True
        return False

1023 1024 1025 1026 1027 1028 1029 1030
    # Step1: update inner_inputs and inner_outputs
    # NOTE: Here assumes that all variables are input or output of Ops,
    # but some variables are created without appendding a real op.
    # For example, in `arr = create_array(dtype)`, `arr` is not a output of a op.
    for op in current_block.ops:
        assert isinstance(op, Operator)
        for iname in op.input_names:
            for in_var_name in op.input(iname):
1031 1032
                if in_var_name not in inner_outputs and not is_ignore_vars(
                        op, in_var_name):
1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056
                    inner_inputs.add(in_var_name)

        for oname in op.output_names:
            for out_var_name in op.output(oname):
                inner_outputs.add(out_var_name)

    # Step2: Remove LOD_TENSOR_ARRAY created in current control flow block.
    remove_inner_inputs = set()
    parent_block = helper.main_program.block(current_block.parent_idx)

    for in_var_name in inner_inputs:
        parent_block_var = parent_block._find_var_recursive(in_var_name)
        current_block_var = None
        if current_block.has_var(in_var_name):
            current_block_var = current_block.var(in_var_name)
        if not parent_block_var and current_block_var and \
                current_block_var.type == core.VarDesc.VarType.LOD_TENSOR_ARRAY:
            remove_inner_inputs.add(in_var_name)

    inner_inputs = inner_inputs - remove_inner_inputs

    return inner_inputs, inner_outputs


Y
Yang Yang(Tony) 已提交
1057
class While(object):
X
Xin Pan 已提交
1058
    """
1059 1060
    :api_attr: Static Graph
    
1061
    while loop control flow. Repeat while body until cond is False.
X
Xin Pan 已提交
1062

1063 1064 1065 1066
    Note:
        A new OP :ref:`api_fluid_layers_while_loop` is highly recommended instead of ``While`` if the shape of parameter ``cond`` is [1].
        OP :ref:`api_fluid_layers_while_loop` is easier to use and is called with less code but does the same thing as ``While`` .

1067 1068 1069 1070 1071 1072
    Notice:
        Local variables created in ``While`` are similar to that created in while of C++, and cannot be referenced externally.
        As a result, they cannot be obtained through ``fetch_list`` of ``Executor``. If you would like to access the variable
        out of ``while`` , PaddlePaddle provides ``assign`` API to assign local variables to external. Please refer to example
        code 2 or refer to `issue#22724 <https://github.com/PaddlePaddle/Paddle/issues/22724>`_.

X
Xin Pan 已提交
1073
    Args:
1074
        cond(Variable): A Tensor whose data type is bool controlling whether to continue looping.
G
guofei 已提交
1075
        is_test(bool, optional): A flag indicating whether execution is in test phase. Default value is False.
1076
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` .
X
Xin Pan 已提交
1077

1078
    Examples 1:
X
Xin Pan 已提交
1079
          .. code-block:: python
1080 1081
            
            import paddle.fluid as fluid
1082 1083 1084 1085 1086
            import numpy as np

            i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=0)           # loop counter

            loop_len = fluid.layers.fill_constant(shape=[1],dtype='int64', value=10)    # loop length
1087

1088
            cond = fluid.layers.less_than(x=i, y=loop_len)
1089
            while_op = fluid.layers.While(cond=cond)
1090
            with while_op.block():
1091
                i = fluid.layers.increment(x=i, value=1, in_place=True)
1092
                fluid.layers.less_than(x=i, y=loop_len, cond=cond)
1093 1094 1095 1096 1097

            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            res = exe.run(fluid.default_main_program(), feed={}, fetch_list=[i])
1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126
            print(res) # [array([10])]


    Examples 2:
          .. code-block:: python

            import paddle.fluid as fluid
            import numpy as np

            i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=0)
            loop_len = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10)
            one = fluid.layers.fill_constant(shape=[1], dtype='float32', value=1)
            data = fluid.data(name='data', shape=[1], dtype='float32')
            sums = fluid.layers.fill_constant(shape=[1], dtype='float32', value=0)  # Define the variable to be obtained ouside of While, which name should be different from the variable inside the While to be obtained

            cond = fluid.layers.less_than(x=i, y=loop_len)
            while_op = fluid.layers.While(cond=cond)
            with while_op.block():
                sums_tensor = fluid.layers.elementwise_add(x=data, y=data)
                fluid.layers.assign(sums_tensor, sums)  # Update the value of sums_tensor defined in While to the sums which defined outside of While through layers.assign
                i = fluid.layers.increment(x=i, value=1, in_place=True)
                data = fluid.layers.elementwise_add(x=data, y=one)
                fluid.layers.less_than(x=i, y=loop_len, cond=cond)

            feed_data = np.ones(1).astype('float32')
            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())
            res = exe.run(fluid.default_main_program(), feed={'data': feed_data}, fetch_list=sums)
            print(res[0])  # [2.]    # Because the data in While does not update the value outside the While, the value of sums is [2.] after the loop
X
Xin Pan 已提交
1127 1128
    """

Y
Yang Yang(Tony) 已提交
1129 1130 1131 1132
    BEFORE_WHILE_BLOCK = 0
    IN_WHILE_BLOCK = 1
    AFTER_WHILE_BLOCK = 2

C
chengduo 已提交
1133
    def __init__(self, cond, is_test=False, name=None):
1134
        self.helper = LayerHelper("while", name=name)
Y
Yang Yang(Tony) 已提交
1135
        self.status = While.BEFORE_WHILE_BLOCK
1136
        check_variable_and_dtype(cond, 'cond', ['bool'], 'fluid.layers.While')
Y
Yang Yang(Tony) 已提交
1137
        if reduce(lambda a, b: a * b, cond.shape, 1) != 1:
1138
            raise TypeError(
1139
                "condition expected shape as [1], but given shape as {0}.".
1140
                format(list(cond.shape)))
Y
Yang Yang(Tony) 已提交
1141
        self.cond_var = cond
C
chengduo 已提交
1142
        self.is_test = is_test
Y
Yang Yang(Tony) 已提交
1143 1144 1145 1146

    def block(self):
        return WhileGuard(self)

1147
    def _complete(self):
Y
Yang Yang(Tony) 已提交
1148 1149
        main_program = self.helper.main_program
        while_block = main_program.current_block()
1150 1151
        parent_block = main_program.block(
            main_program.current_block().parent_idx)
Y
Yang Yang(Tony) 已提交
1152 1153 1154

        inner_outputs = {self.cond_var.name}
        x_name_list = set()
1155 1156
        x_name_list, inner_outputs = get_inputs_outputs_in_block(
            while_block, x_name_list, inner_outputs, self.helper)
Y
Yang Yang(Tony) 已提交
1157 1158 1159

        out_vars = []
        for inner_out_name in inner_outputs:
X
Xin Pan 已提交
1160 1161 1162
            inner_var = parent_block._find_var_recursive(inner_out_name)
            if inner_var:
                out_vars.append(inner_var)
Y
Yang Yang(Tony) 已提交
1163

1164 1165
        x_name_list |= set(map(lambda x: x.name, out_vars))

Y
Yang Yang(Tony) 已提交
1166 1167 1168 1169 1170 1171
        step_scope = parent_block.create_var(
            type=core.VarDesc.VarType.STEP_SCOPES)

        parent_block.append_op(
            type='while',
            inputs={
1172 1173
                'X':
                [parent_block._var_recursive(x_name) for x_name in x_name_list],
Y
Yang Yang(Tony) 已提交
1174 1175
                'Condition': [self.cond_var]
            },
1176 1177 1178 1179 1180 1181 1182 1183
            outputs={
                'Out': out_vars,
                'StepScopes': [step_scope]
            },
            attrs={
                'sub_block': while_block,
                "is_test": self.is_test
            })
Y
Yang Yang(Tony) 已提交
1184 1185


1186 1187 1188
support_ret_buildin_type = (bool, float, six.integer_types)


1189
def assign_skip_lod_tensor_array(input, output):
1190
    """
1191
    Assign input to output, but skip the process of copying LoDTensorArray unless it's created in while_block.
1192
    """
1193 1194 1195 1196 1197 1198 1199

    def has_shape_diff(x_var, y_var):
        if len(x_var.shape) != len(y_var.shape): return True
        for x_dim, y_dim in zip(x_var.shape, y_var.shape):
            if x_dim != y_dim and -1 not in [x_dim, y_dim]: return True
        return False

1200
    if not isinstance(input, (Variable, core.VarBase)):
1201 1202
        if isinstance(output, Variable) and isinstance(
                input, support_ret_buildin_type):
1203 1204 1205
            assign(input, output)
        else:
            output = input
1206 1207
        return

1208 1209
    if input.type == core.VarDesc.VarType.LOD_TENSOR_ARRAY:
        main_program = input.block.program
1210 1211
        parent_block = main_program.block(
            main_program.current_block().parent_idx)
1212 1213 1214
        if parent_block and not parent_block._find_var_recursive(input.name):
            assign(input, output)
    else:
1215 1216 1217 1218 1219
        if isinstance(output, Variable) and isinstance(
                input, Variable) and has_shape_diff(input, output):
            warnings.warn(
                "In dy2static mode, we attemp to assign a variable with shape {} into a variable with shape{}, which is not always right."
                .format(input.shape, output.shape))
1220
        assign(input, output)
1221 1222


G
guofei 已提交
1223
def while_loop(cond, body, loop_vars, is_test=False, name=None):
G
guofei 已提交
1224
    """
1225 1226
    :api_attr: Static Graph

G
guofei 已提交
1227 1228
    while_loop is one of the control flows. Repeats while_loop `body` until `cond` returns False.

1229 1230 1231 1232
    Notice:
        Local variables defined in ``body`` cannot be obtained through ``fetch_list`` of ``Executor`` , variables should
        be defined outside ``body`` and placed in ``loop_vars`` for looping, then these variables can be fetched by ``fetch_list`` .

G
guofei 已提交
1233
    Args:
1234 1235 1236 1237 1238
        cond(Callable): A callable returning a boolean tensor controlling whether to continue looping. And ``cond`` takes
	    as many arguments as ``loop_vars`` .
        body(Callable): A callable returning a tuple or list of tensors or LoDTensorArrays of the same arity
            (length and structure) and types as ``loops_vars`` . And ``body`` takes as many arguments as ``loop_vars`` .
        loop_vars(list|tuple): A list or tuple of tensors or LoDTensorArrays that is passed to both ``cond`` and ``body`` .
G
guofei 已提交
1239
        is_test(bool, optional): A flag indicating whether execution is in test phase. Default value is False.
G
guofei 已提交
1240 1241
        name(str, optional): Normally there is no need for users to set this property. For more information, please
            refer to :ref:`api_guide_Name`. Default is None.
1242

G
guofei 已提交
1243
    Returns:
C
Chen Long 已提交
1244
        A list or tuple of Tensors or LoDTensorArrays which returned by ``body`` .
G
guofei 已提交
1245 1246 1247 1248

    Examples:
        .. code-block:: python

1249 1250 1251
            import paddle
            paddle.enable_static()

1252 1253
            def cond(i, ten):
                return i < ten
G
guofei 已提交
1254

1255 1256 1257
            def body(i, ten):
                i = i + 1
                return [i, ten]
G
guofei 已提交
1258

C
Chen Long 已提交
1259 1260 1261 1262 1263 1264
            main_program = paddle.static.default_main_program()
            startup_program = paddle.static.default_startup_program()
            with paddle.static.program_guard(main_program, startup_program):
                i = paddle.full(shape=[1], fill_value=0, dtype='int64')     # loop counter
                ten = paddle.full(shape=[1], fill_value=10, dtype='int64')  # loop length
                i, ten = paddle.static.nn.while_loop(cond, body, [i, ten])
G
guofei 已提交
1265
                
C
Chen Long 已提交
1266
                exe = paddle.static.Executor(paddle.CPUPlace())
1267
                res = exe.run(main_program, feed={}, fetch_list=[i])
G
guofei 已提交
1268 1269 1270 1271 1272 1273 1274 1275
                print(res) # [array([10])]
    """
    helper = LayerHelper('while_loop', **locals())

    if not callable(cond):
        raise TypeError("cond in while_loop should be callable")
    if not callable(body):
        raise TypeError("body in while_loop should be callable")
1276
    check_type(loop_vars, 'loop_vars', (list, tuple), 'fluid.layers.while_loop')
G
guofei 已提交
1277 1278 1279 1280
    if len(loop_vars) == 0:
        raise ValueError("loop_vars in while_loop should not be empty")

    pre_cond = cond(*loop_vars)
1281 1282
    check_variable_and_dtype(pre_cond, 'var of cond returned', ['bool'],
                             'fluid.layers.while_loop')
G
guofei 已提交
1283 1284
    if reduce(lambda a, b: a * b, pre_cond.shape, 1) != 1:
        raise TypeError(
1285
            "the shape of the variable returned by cond should be [1],"
G
guofei 已提交
1286 1287
            "but given shape as {0}.".format(list(pre_cond.shape)))

J
Jiabin Yang 已提交
1288
    if _non_static_mode():
1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
        now_cond = pre_cond.numpy()[0]
        while (now_cond):
            output_vars = body(*loop_vars)
            if not isinstance(output_vars, (list, tuple)):
                output_vars = [output_vars]
            if len(output_vars) != len(loop_vars):
                raise ValueError(
                    "body in while_loop should return the same arity "
                    "(length and structure) and types as loop_vars")
            now_cond = cond(*output_vars).numpy()[0]
1299
            map_structure(assign_skip_lod_tensor_array, output_vars, loop_vars)
1300 1301
        return loop_vars

G
guofei 已提交
1302
    while_loop_block = While(pre_cond, is_test, name)
1303
    has_mutable_vars_in_loop = hold_mutable_vars(loop_vars)
G
guofei 已提交
1304
    with while_loop_block.block():
1305 1306 1307 1308 1309 1310 1311 1312 1313
        # If a variable with mutable type is included in loop_vars, like `dict/list`,
        # modifying it in the body function will cause origin variable to be modified
        # synchronously. This will raise an assignment error out of while block.
        # Here we make a copy of the mutable vars to avoid this problem.
        if has_mutable_vars_in_loop:
            new_loop_vars = copy_mutable_vars(loop_vars)
            output_vars = body(*new_loop_vars)
        else:
            output_vars = body(*loop_vars)
1314 1315
        if not isinstance(output_vars, (list, tuple)):
            output_vars = [output_vars]
1316
        try:
1317
            loop_vars = _deal_with_undefined_var(output_vars, loop_vars)
1318 1319
            assert_same_structure(output_vars, loop_vars, check_types=False)
        except ValueError as e:
1320 1321 1322
            raise ValueError(
                "body in while_loop should return the same arity "
                "(length and structure) as loop_vars: {0}".format(e))
1323
        now_cond = cond(*output_vars)
1324
        map_structure(assign_skip_lod_tensor_array, output_vars, loop_vars)
G
guofei 已提交
1325 1326 1327 1328
        assign(now_cond, pre_cond)
    return loop_vars


1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343
def _deal_with_undefined_var(output_vars, loop_vars):
    """ Deal with undefined var cases, We create undefined variable based on the results of body().
        In Dy2Static, we use undefined var to represent the var created in control flow. This function
        expand the loop_vars and replace original loop_vars.
        1. UndefinedVar = Variable      # create a variable
        2. UndefinedVar = None          # create a undefined var with RETURN_NO_VALUE_MAGIC_NUM
        3. UndefinedVar = List(int)     # create a list of variable
        4. UndefinedVar = value         # create a variable
    """
    from paddle.fluid.dygraph.dygraph_to_static.utils import UndefinedVar, create_undefined_variable

    def create_var_like(o_var):
        if isinstance(o_var,
                      (Variable, ) + support_ret_buildin_type) or o_var is None:
            return create_undefined_variable()
1344 1345 1346 1347 1348
        if is_sequence(o_var):
            """ 
            Create a complex container class inside the body of while, including Python list and python Dict
            """
            return map_structure(lambda x: create_undefined_variable(), o_var)
1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361

    if len(output_vars) != len(loop_vars):
        raise ValueError("The length of loop_vars should be the same.")

    results = []
    for o_var, l_var in zip(output_vars, loop_vars):
        if isinstance(l_var, UndefinedVar) or l_var is None:
            results.append(create_var_like(o_var))
        else:
            results.append(l_var)
    return results


1362
def lod_rank_table(x, level=0):
1363 1364
    """
    LoD Rank Table Operator. Given an input variable **x** and a level number
Y
yangyaming 已提交
1365 1366
    of LoD, this layer creates a LodRankTable object. A LoDRankTable object
    contains a list of bi-element tuples. Each tuple consists of an index and
1367
    a length, both of which are int type. Refering to specified level of LoD,
T
tianshuo78520a 已提交
1368
    the index is the sequence index number and the length represents the
Y
yangyaming 已提交
1369 1370
    sequence length. Please note that the list is ranked in descending order by
    the length. The following is an example:
Y
yangyaming 已提交
1371 1372 1373 1374

        .. code-block:: text

            x is a LoDTensor:
1375 1376
                x.lod = [[2,                1],
                         [5,             1, 1]]
Y
yangyaming 已提交
1377 1378
                x.data = [a, b, c, d, e, f, g]

Y
yangyaming 已提交
1379 1380 1381
            1. set level to 0:
                Create lod rank table:
                    lod_rank_table_obj = lod_rank_table(x, level=0)
Y
yangyaming 已提交
1382

Y
yangyaming 已提交
1383 1384 1385 1386 1387 1388 1389 1390 1391
                Get:
                    lod_rank_table_obj.items() = [(0, 2), (1, 1)]

            2. set level to 1:
                Create lod rank table:
                    lod_rank_table_obj = lod_rank_table(x, level=1)

                Get:
                    lod_rank_table_obj.items() = [(0, 5), (1, 1), (2, 1)]
Y
yangyaming 已提交
1392 1393 1394 1395

    Args:
        x (Variable): Input variable, a LoDTensor based which to create the lod
            rank table.
Y
yangyaming 已提交
1396 1397
        level (int): Specify the LoD level, on which to create the lod rank
            table.
Y
yangyaming 已提交
1398 1399 1400 1401 1402 1403 1404

    Returns:
        Variable: The created LoDRankTable object.

    Examples:
        .. code-block:: python

1405
            import paddle.fluid as fluid
Y
yangyaming 已提交
1406
            x = fluid.layers.data(name='x', shape=[10],
1407
                                  dtype='float32', lod_level=1)
Y
yangyaming 已提交
1408
            out = layers.lod_rank_table(x=x, level=0)
1409
    """
1410 1411 1412 1413 1414 1415
    check_type(x, 'x', (Variable, list), 'lod_rank_table')
    if isinstance(x, (list)):
        for i, input_x in enumerate(x):
            check_type(input_x, 'input[' + str(i) + ']', Variable,
                       'lod_rank_table')

Y
Yu Yang 已提交
1416
    helper = LayerHelper("lod_rank_table", **locals())
1417 1418 1419 1420 1421 1422
    table = helper.create_variable(type=core.VarDesc.VarType.LOD_RANK_TABLE,
                                   name=unique_name.generate("lod_rank_table"))
    helper.append_op(type='lod_rank_table',
                     inputs={'X': x},
                     outputs={'Out': table},
                     attrs={'level': level})
Y
Yu Yang 已提交
1423
    return table
Y
Yu Yang 已提交
1424 1425


Y
yuyang18 已提交
1426
@templatedoc()
1427
def max_sequence_len(rank_table):
Y
yuyang18 已提交
1428 1429 1430 1431 1432 1433 1434 1435
    """
    ${comment}

    >>> import paddle.fluid as fluid
    >>> x = fluid.layers.data(name='x', shape=[10], dtype='float32',
    >>>                       lod_level=1)
    >>> rank_table = layers.lod_rank_table(x=x, level=0)
    >>> max_seq_len = layers.max_sequence_len(rank_table)
Y
yangyaming 已提交
1436 1437

    Args:
Y
yuyang18 已提交
1438
        rank_table(${rank_table_type}): ${rank_table_comment}.
Y
yangyaming 已提交
1439 1440

    Returns:
Y
yuyang18 已提交
1441
        ${out_comment}.
F
fengjiayi 已提交
1442 1443
    """
    helper = LayerHelper("max_seqence_len", **locals())
X
Xin Pan 已提交
1444
    res = helper.create_variable_for_type_inference(dtype="int64")
1445 1446 1447
    helper.append_op(type="max_sequence_len",
                     inputs={"RankTable": rank_table},
                     outputs={"Out": res})
F
fengjiayi 已提交
1448 1449 1450
    return res


1451
def lod_tensor_to_array(x, table):
1452
    """
F
fengjiayi 已提交
1453 1454
    Convert a LoDTensor to a LoDTensorArray.

1455 1456 1457 1458 1459
    This function split a LoDTesnor to a LoDTensorArray according to its LoD
    information. LoDTensorArray is an alias of C++ std::vector<LoDTensor> in
    PaddlePaddle. The generated LoDTensorArray of this function can be further read
    or written by `read_from_array()` and `write_to_array()` operators. However,
    this function is generally an internal component of PaddlePaddle `DynamicRNN`.
F
fengjiayi 已提交
1460
    Users should not use it directly.
1461 1462

    Args:
F
fengjiayi 已提交
1463
        x (Variable|list): The LoDTensor to be converted to a LoDTensorArray.
1464 1465
        table (ParamAttr|list): The variable that stores the level of lod
                                which is ordered by sequence length in
1466
                                descending order. It is generally generated
F
fengjiayi 已提交
1467
                                by `layers.lod_rank_table()` API.
1468 1469

    Returns:
F
fengjiayi 已提交
1470
        Variable: The LoDTensorArray that has been converted from the input tensor.
1471 1472 1473 1474

    Examples:
        .. code-block:: python

1475
          import paddle.fluid as fluid
1476 1477 1478
          x = fluid.layers.data(name='x', shape=[10])
          table = fluid.layers.lod_rank_table(x, level=0)
          array = fluid.layers.lod_tensor_to_array(x, table)
1479
    """
1480 1481 1482 1483 1484 1485 1486 1487 1488 1489
    check_type(x, 'x', (Variable, list), 'lod_tensor_to_array')
    if isinstance(x, (list)):
        for i, input_x in enumerate(x):
            check_type(input_x, 'input[' + str(i) + ']', Variable,
                       'lod_tensor_to_array')
    check_type(table, 'table', (Variable, list), 'lod_tensor_to_array')
    if isinstance(table, (list)):
        for i, table_x in enumerate(table):
            check_type(table_x, 'table[' + str(i) + ']', Variable,
                       'lod_tensor_to_array')
1490 1491
    helper = LayerHelper("lod_tensor_to_array", **locals())
    array = helper.create_variable(
Y
Yu Yang 已提交
1492
        name=unique_name.generate("lod_tensor_to_array"),
1493
        type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
F
fengjiayi 已提交
1494
        dtype=x.dtype)
1495 1496 1497 1498 1499 1500
    helper.append_op(type='lod_tensor_to_array',
                     inputs={
                         'X': x,
                         'RankTable': table
                     },
                     outputs={'Out': array})
1501 1502 1503
    return array


1504
def array_to_lod_tensor(x, table):
1505
    """Convert a LoD_Tensor_Aarry to an LoDTensor.
1506 1507

    Args:
1508
        x (Variable|list): The lod tensor array to be converted to a tensor.
1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519
        table (ParamAttr|list): The variable that stores the level of lod
                                which is ordered by sequence length in
                                descending order.

    Returns:
        Variable: The variable of type tensor that has been converted
                  from an array.

    Examples:
        .. code-block:: python

1520
          import paddle.fluid as fluid
1521 1522 1523 1524
          x = fluid.layers.data(name='x', shape=[10])
          table = fluid.layers.lod_rank_table(x, level=0)
          array = fluid.layers.lod_tensor_to_array(x, table)
          lod_tensor = fluid.layers.array_to_lod_tensor(array, table)
1525
    """
1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536
    check_type(x, 'x', (Variable, list), 'array_to_lod_tensor')
    if isinstance(x, (list)):
        for i, input_x in enumerate(x):
            check_type(input_x, 'input[' + str(i) + ']', Variable,
                       'array_to_lod_tensor')
    check_type(table, 'table', (Variable, list), 'array_to_lod_tensor')
    if isinstance(table, (list)):
        for i, table_x in enumerate(table):
            check_type(table_x, 'table[' + str(i) + ']', Variable,
                       'array_to_lod_tensor')

1537
    helper = LayerHelper("array_to_lod_tensor", **locals())
X
Xin Pan 已提交
1538
    tmp = helper.create_variable_for_type_inference(dtype=x.dtype)
1539 1540 1541 1542 1543 1544
    helper.append_op(type="array_to_lod_tensor",
                     inputs={
                         'X': x,
                         'RankTable': table
                     },
                     outputs={'Out': tmp})
1545 1546 1547
    return tmp


1548
def increment(x, value=1.0, in_place=True):
1549
    """
1550 1551
    The OP is usually used for control flow to increment the data of :attr:`x` by an amount :attr:`value`.
    Notice that the number of elements in :attr:`x` must be equal to 1.
1552

1553
    Parameters:
T
tianshuo78520a 已提交
1554
        x (Variable): A tensor that must always contain only one element, its data type supports
1555 1556 1557
            float32, float64, int32 and int64.
        value (float, optional): The amount to increment the data of :attr:`x`. Default: 1.0.
        in_place (bool, optional): Whether the OP should be performed in-place. Default: True.
1558 1559

    Returns:
1560
        Variable: The elementwise-incremented tensor with the same shape and data type as :attr:`x`.
1561 1562 1563 1564

    Examples:
        .. code-block:: python

1565
          import paddle.fluid as fluid
1566 1567
          counter = fluid.layers.zeros(shape=[1], dtype='float32') # [0.]
          fluid.layers.increment(counter) # [1.]
1568
    """
H
hong 已提交
1569 1570 1571
    if in_dygraph_mode():
        return _C_ops.final_state_increment_(x, value)

1572 1573
    check_variable_and_dtype(x, 'x', ['float32', 'float64', 'int32', 'int64'],
                             'increment')
Y
Yu Yang 已提交
1574
    helper = LayerHelper("increment", **locals())
Y
Yang Yang(Tony) 已提交
1575
    if not in_place:
X
Xin Pan 已提交
1576
        out = helper.create_variable_for_type_inference(dtype=x.dtype)
Y
Yang Yang(Tony) 已提交
1577 1578
    else:
        out = x
1579 1580 1581 1582
    helper.append_op(type='increment',
                     inputs={'X': [x]},
                     outputs={'Out': [out]},
                     attrs={'step': float(value)})
Y
Yang Yu 已提交
1583
    return out
Y
Yu Yang 已提交
1584 1585


1586
def array_write(x, i, array=None):
1587
    """
1588 1589 1590 1591
    This OP writes the input ``x`` into the i-th position of the ``array``
    :ref:`api_fluid_LoDTensorArray` and returns the modified array.
    If ``array`` is none, a new LoDTensorArray will be created and returned.
    This OP is often used together with :ref:`api_fluid_layers_array_read` OP.
1592 1593

    Args:
1594 1595 1596 1597 1598 1599 1600
        x (Variable): The input data to be written into array. It's multi-dimensional
            Tensor or LoDTensor. Data type: float32, float64, int32, int64.
        i (Variable): 1-D Tensor with shape [1], which represents the position into which
            ``x`` is written. Data type: int64.
        array (LoDTensorArray, optional): The LoDTensorArray into which ``x`` is written. 
            The default value is None, when a new LoDTensorArray will be created and returned 
            as a result.
1601

1602
    Returns:
1603
        Variable: The input ``array`` after ``x`` is written into.
1604 1605

    Examples:
D
dzhwinter 已提交
1606
        .. code-block:: python
1607

1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634
            import paddle.fluid as fluid
            tmp = fluid.layers.fill_constant(shape=[3, 2], dtype='int64', value=5)
            i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10)
            # Write tmp into the position of arr with subscript 10 and return arr.
            arr = fluid.layers.array_write(tmp, i=i)

            # Now, arr is a LoDTensorArray with length 11. We can use array_read OP to read
            # the data at subscript 10 and print it out.
            item = fluid.layers.array_read(arr, i=i)
            input = fluid.layers.Print(item, message="The content of i-th LoDTensor:")
            main_program = fluid.default_main_program()
            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(main_program)

            # The printed result is:
            # 1570533133    The content of i-th LoDTensor:  The place is:CPUPlace
            # Tensor[array_read_0.tmp_0]
            #    shape: [3,2,]
            #    dtype: l
            #    data: 5,5,5,5,5,5,

            # the output is 2-D Tensor with shape [3,2], which is tmp above.
            # dtype is the corresponding C++ data type, which may vary in different environments.
            # Eg: if the data type of tensor is int64, then the corresponding C++ data type is int64_t, 
            #       so the dtype value is typeid(int64_t).Name(), which is 'x' on MacOS, 'l' on Linux, 
            #       and '__int64' on Windows. They both represent 64-bit integer variables.

1635
    """
J
Jiabin Yang 已提交
1636
    if _non_static_mode():
1637 1638 1639 1640 1641 1642 1643 1644 1645
        assert isinstance(
            x, Variable
        ), "The input data 'x' in array_write must be Variable in dygraph mode"
        assert isinstance(
            i, Variable
        ), "The index 'i' in array_write must be Variable in dygraph mode"
        assert i.shape == [
            1
        ], "The shape of index 'i' should be [1] in dygraph mode"
1646
        i = i.numpy().item(0)
1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660
        if array is None:
            array = create_array(x.dtype)
        assert isinstance(
            array,
            list), "The 'array' in array_write must be a list in dygraph mode"
        assert i <= len(
            array
        ), "The index 'i' should not be greater than the length of 'array' in dygraph mode"
        if i < len(array):
            array[i] = x
        else:
            array.append(x)
        return array

1661 1662
    check_variable_and_dtype(i, 'i', ['int64'], 'array_write')
    check_type(x, 'x', (Variable), 'array_write')
Y
Yu Yang 已提交
1663
    helper = LayerHelper('array_write', **locals())
1664 1665
    if array is not None:
        if not isinstance(
1666 1667
                array, Variable
        ) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY:
1668 1669
            raise TypeError(
                "array should be tensor array vairable in array_write Op")
Y
Yu Yang 已提交
1670 1671 1672 1673
    if array is None:
        array = helper.create_variable(
            name="{0}.out".format(helper.name),
            type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
F
fengjiayi 已提交
1674
            dtype=x.dtype)
1675 1676 1677 1678 1679 1680
    helper.append_op(type='write_to_array',
                     inputs={
                         'X': [x],
                         'I': [i]
                     },
                     outputs={'Out': [array]})
Y
Yu Yang 已提交
1681 1682 1683
    return array


1684
def create_array(dtype, initialized_list=None):
1685
    """
1686 1687 1688 1689
    This OP creates an LOD_TENSOR_ARRAY. It is used as
    the input of :ref:`api_fluid_layers_array_read` and 
    :ref:`api_fluid_layers_array_write`. Also it can be used
    with  :ref:`api_fluid_layers_While` to create RNN network.
1690 1691

    Args:
1692 1693
        dtype (str): The data type of the elements in the lod_tensor_array.
                     Support data type: float32, float64, int32, int64.
1694 1695
        initialized_list(list): Used to initialize as default value for created array.
                    All values in initialized list should be a Tensor.
1696 1697

    Returns:
1698
        Variable: The empty lod_tensor_array. The data type of elements in Tensor is ``dtype``.
1699 1700 1701 1702

    Examples:
        .. code-block:: python

1703
          import paddle.fluid as fluid
1704
          data = fluid.layers.create_array(dtype='float32') # Create a float32 LoDTensorArray.
1705 1706

    """
1707 1708 1709 1710
    array = []
    if initialized_list is not None:
        if not isinstance(initialized_list, (list, tuple)):
            raise TypeError(
1711 1712
                "Require type(initialized_list) should be list/tuple, but received {}"
                .format(type(initialized_list)))
1713 1714 1715 1716 1717 1718
        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(
1719 1720
                "All values in `initialized_list` should be Variable, but recevied {}."
                .format(type(val)))
1721

J
Jiabin Yang 已提交
1722
    if _non_static_mode():
1723
        return array
1724

Y
Yang Yang(Tony) 已提交
1725
    helper = LayerHelper("array", **locals())
1726
    tensor_array = helper.create_variable(
Y
Yang Yang(Tony) 已提交
1727 1728 1729 1730
        name="{0}.out".format(helper.name),
        type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
        dtype=dtype)

1731 1732 1733 1734 1735
    for val in array:
        array_write(x=val, i=array_length(tensor_array), array=tensor_array)

    return tensor_array

Y
Yang Yang(Tony) 已提交
1736

Y
yuyang18 已提交
1737
@templatedoc()
W
wawltor 已提交
1738
def less_than(x, y, force_cpu=None, cond=None, name=None):
1739
    """
1740

Y
yuyang18 已提交
1741
    ${comment}
1742 1743

    Args:
N
Noel 已提交
1744 1745
        x(Tensor): ${x_comment}.
        y(Tensor): ${y_comment}.
Y
yuyang18 已提交
1746
        force_cpu(${force_cpu_type}): ${force_cpu_comment}.
N
Noel 已提交
1747
        cond(Tensor, optional): Optional output which can be any created Tensor
1748
            that meets the requirements to store the result of *less_than*.
N
Noel 已提交
1749
            if cond is None, a new Tensor will be created to store the result.
W
wawltor 已提交
1750 1751
        name(str, optional): The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
1752
    Returns:
Y
yuyang18 已提交
1753
        ${out_comment}.
1754 1755 1756 1757

    Examples:
        .. code-block:: python

N
Noel 已提交
1758 1759 1760 1761 1762 1763 1764
            import paddle

            x = paddle.to_tensor([1, 2, 3, 4], dtype='float32')
            y = paddle.to_tensor([2, 2, 1, 3], dtype='float32')
            result = paddle.less_than(x, y)
            print(result) # [True, False, False, False]

1765
    """
1766 1767 1768 1769 1770 1771 1772 1773 1774
    check_variable_and_dtype(x, "x", ["float32", "float64", "int32", "int64"],
                             "less_than")
    check_variable_and_dtype(y, "y", ["float32", "float64", "int32", "int64"],
                             "less_than")
    if cond is not None:
        check_type(cond, "cond", Variable, "less_than")
    if force_cpu != None:
        check_type(force_cpu, "force_cpu", bool, "less_than")

Y
Yang Yang(Tony) 已提交
1775 1776
    helper = LayerHelper("less_than", **locals())
    if cond is None:
X
Xin Pan 已提交
1777
        cond = helper.create_variable_for_type_inference(dtype='bool')
Y
Yang Yang(Tony) 已提交
1778 1779
        cond.stop_gradient = True

Y
yuyang18 已提交
1780 1781 1782 1783
    attrs = dict()
    if force_cpu is not None:
        attrs['force_cpu'] = force_cpu

1784 1785 1786 1787 1788 1789 1790
    helper.append_op(type='less_than',
                     inputs={
                         'X': [x],
                         'Y': [y]
                     },
                     outputs={'Out': [cond]},
                     attrs=attrs)
Y
Yang Yang(Tony) 已提交
1791 1792 1793
    return cond


Z
zhoukunsheng 已提交
1794
@templatedoc()
W
wawltor 已提交
1795
def less_equal(x, y, cond=None, name=None):
Z
zhoukunsheng 已提交
1796
    """
1797 1798 1799 1800
    :alias_main: paddle.less_equal
	:alias: paddle.less_equal,paddle.tensor.less_equal,paddle.tensor.logic.less_equal
	:old_api: paddle.fluid.layers.less_equal

1801
    This OP returns the truth value of :math:`x <= y` elementwise, which is equivalent function to the overloaded operator `<=`.
Z
zhoukunsheng 已提交
1802 1803

    Args:
1804 1805
        x(Variable): First input to compare which is N-D tensor. The input data type should be float32, float64, int32, int64. 
        y(Variable): Second input to compare which is N-D tensor. The input data type should be float32, float64, int32, int64.
1806 1807
        cond(Variable, optional): Optional output which can be any created Variable that meets the requirements to store the result of *less_equal*.
            if cond is None, a new Varibale will be created to store the result.
W
wawltor 已提交
1808 1809
        name(str, optional): The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
Z
zhoukunsheng 已提交
1810 1811

    Returns:
1812
        Variable, the output data type is bool: The tensor variable storing the output, the output shape is same as input :attr:`x`.
Z
zhoukunsheng 已提交
1813 1814 1815 1816

    Examples:
        .. code-block:: python

1817
          import paddle.fluid as fluid
1818 1819 1820 1821 1822 1823
          import numpy as np
          label = fluid.layers.assign(np.array([1, 3], dtype='int32'))
          limit = fluid.layers.assign(np.array([1, 2], dtype='int32'))
          out = fluid.layers.less_equal(x=label, y=limit) #out=[True, False]
          out1 = label<= limit #out1=[True, False]

Z
zhoukunsheng 已提交
1824
    """
1825 1826 1827 1828 1829
    check_variable_and_dtype(x, "x", ["float32", "float64", "int32", "int64"],
                             "less_equal")
    check_variable_and_dtype(y, "y", ["float32", "float64", "int32", "int64"],
                             "less_equal")
    if cond is not None:
1830
        check_type(cond, "cond", Variable, "less_equal")
1831

Z
zhoukunsheng 已提交
1832 1833 1834 1835 1836 1837 1838
    helper = LayerHelper("less_equal", **locals())
    if cond is None:
        cond = helper.create_variable_for_type_inference(dtype='bool')
        cond.stop_gradient = True

    attrs = dict()

1839 1840 1841 1842 1843 1844 1845
    helper.append_op(type='less_equal',
                     inputs={
                         'X': [x],
                         'Y': [y]
                     },
                     outputs={'Out': [cond]},
                     attrs=attrs)
Z
zhoukunsheng 已提交
1846 1847 1848 1849
    return cond


@templatedoc()
W
wawltor 已提交
1850
def greater_than(x, y, cond=None, name=None):
Z
zhoukunsheng 已提交
1851
    """
1852 1853 1854 1855
    :alias_main: paddle.greater_than
	:alias: paddle.greater_than,paddle.tensor.greater_than,paddle.tensor.logic.greater_than
	:old_api: paddle.fluid.layers.greater_than

1856
    This OP returns the truth value of :math:`x > y` elementwise, which is equivalent function to the overloaded operator `>`.
Z
zhoukunsheng 已提交
1857 1858

    Args:
1859 1860
        x(Variable): First input to compare which is N-D tensor. The input data type should be float32, float64, int32, int64. 
        y(Variable): Second input to compare which is N-D tensor. The input data type should be float32, float64, int32, int64.
1861 1862
        cond(Variable, optional): Optional output which can be any created Variable that meets the requirements to store the result of *greater_than*.
            if cond is None, a new Varibale will be created to store the result.
W
wawltor 已提交
1863 1864
        name(str, optional): The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
Z
zhoukunsheng 已提交
1865 1866

    Returns:
1867
        Variable, the output data type is bool: The tensor variable storing the output, the output shape is same as input :attr:`x` .
Z
zhoukunsheng 已提交
1868 1869 1870 1871

    Examples:
        .. code-block:: python

1872
          import paddle.fluid as fluid
1873 1874 1875 1876 1877
          import numpy as np
          label = fluid.layers.assign(np.array([2, 3], dtype='int32'))
          limit = fluid.layers.assign(np.array([3, 2], dtype='int32'))
          out = fluid.layers.greater_than(x=label, y=limit) #out=[False, True]
          out1 = label > limit #out1=[False, True]
Z
zhoukunsheng 已提交
1878
    """
1879 1880 1881 1882 1883
    check_variable_and_dtype(x, "x", ["float32", "float64", "int32", "int64"],
                             "greater_than")
    check_variable_and_dtype(y, "y", ["float32", "float64", "int32", "int64"],
                             "greater_than")
    if cond is not None:
1884
        check_type(cond, "cond", Variable, "greater_than")
1885

Z
zhoukunsheng 已提交
1886 1887 1888 1889 1890 1891 1892
    helper = LayerHelper("greater_than", **locals())
    if cond is None:
        cond = helper.create_variable_for_type_inference(dtype='bool')
        cond.stop_gradient = True

    attrs = dict()

1893 1894 1895
    if in_dygraph_mode():
        return _C_ops.final_state_greater_than(x, y, -1)
    else:
1896 1897 1898 1899 1900 1901 1902
        helper.append_op(type='greater_than',
                         inputs={
                             'X': [x],
                             'Y': [y]
                         },
                         outputs={'Out': [cond]},
                         attrs=attrs)
1903
        return cond
Z
zhoukunsheng 已提交
1904 1905 1906


@templatedoc()
W
wawltor 已提交
1907
def greater_equal(x, y, cond=None, name=None):
Z
zhoukunsheng 已提交
1908
    """
1909 1910 1911 1912
    :alias_main: paddle.greater_equal
	:alias: paddle.greater_equal,paddle.tensor.greater_equal,paddle.tensor.logic.greater_equal
	:old_api: paddle.fluid.layers.greater_equal

1913
    This OP returns the truth value of :math:`x >= y` elementwise, which is equivalent function to the overloaded operator `>=`.
Z
zhoukunsheng 已提交
1914 1915

    Args:
1916 1917
        x(Variable): First input to compare which is N-D tensor. The input data type should be float32, float64, int32, int64. 
        y(Variable): Second input to compare which is N-D tensor. The input data type should be float32, float64, int32, int64.
1918 1919
        cond(Variable, optional): Optional output which can be any created Variable that meets the requirements to store the result of *greater_equal*.
            if cond is None, a new Varibale will be created to store the result.
W
wawltor 已提交
1920 1921
        name(str, optional): The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
Z
zhoukunsheng 已提交
1922 1923

    Returns:
1924
        Variable, the output data type is bool: The tensor variable storing the output, the output shape is same as input :attr:`x`.
Z
zhoukunsheng 已提交
1925 1926 1927 1928

    Examples:
        .. code-block:: python

1929
          import paddle.fluid as fluid
1930 1931 1932 1933 1934 1935
          import numpy as np

          label = fluid.layers.assign(np.array([2, 2], dtype='int32'))
          limit = fluid.layers.assign(np.array([2, 3], dtype='int32'))
          out = fluid.layers.greater_equal(x=label, y=limit) #out=[True, False]
          out_1 = label >= limit #out1=[True, False]
1936

Z
zhoukunsheng 已提交
1937
    """
1938 1939 1940 1941 1942
    check_variable_and_dtype(x, "x", ["float32", "float64", "int32", "int64"],
                             "greater_equal")
    check_variable_and_dtype(y, "y", ["float32", "float64", "int32", "int64"],
                             "greater_equal")
    if cond is not None:
1943
        check_type(cond, "cond", Variable, "greater_equal")
1944

Z
zhoukunsheng 已提交
1945 1946 1947 1948 1949 1950 1951
    helper = LayerHelper("greater_equal", **locals())
    if cond is None:
        cond = helper.create_variable_for_type_inference(dtype='bool')
        cond.stop_gradient = True

    attrs = dict()

1952 1953 1954 1955 1956 1957 1958
    helper.append_op(type='greater_equal',
                     inputs={
                         'X': [x],
                         'Y': [y]
                     },
                     outputs={'Out': [cond]},
                     attrs=attrs)
Z
zhoukunsheng 已提交
1959 1960 1961
    return cond


W
wawltor 已提交
1962
def equal(x, y, cond=None, name=None):
1963 1964 1965 1966
    """
    This layer returns the truth value of :math:`x == y` elementwise.

    Args:
W
wangchaochaohu 已提交
1967 1968 1969 1970 1971
        x(Variable): Tensor, data type is float32, float64, int32, int64.
        y(Variable): Tensor, data type is float32, float64, int32, int64.
        cond(Variable, optional): Optional output which can be any created 
            Variable that meets the requirements to store the result of *equal*.
            if cond is None, a new Varibale will be created to store the result.
W
wawltor 已提交
1972 1973
        name(str, optional): The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
1974 1975

    Returns:
W
wangchaochaohu 已提交
1976 1977
        Variable: output Tensor, it's shape is the same as the input's Tensor,
        and the data type is bool.
1978 1979 1980 1981

    Examples:
        .. code-block:: python

1982
          import paddle.fluid as fluid
W
wangchaochaohu 已提交
1983 1984 1985 1986 1987 1988 1989
          import numpy as np
          out_cond =fluid.data(name="input1", shape=[2], dtype='bool')
          label = fluid.layers.assign(np.array([3, 3], dtype="int32"))
          limit = fluid.layers.assign(np.array([3, 2], dtype="int32"))
          label_cond = fluid.layers.assign(np.array([1, 2], dtype="int32"))
          out1 = fluid.layers.equal(x=label,y=limit) #out1=[True, False]
          out2 = fluid.layers.equal(x=label_cond,y=limit, cond=out_cond) #out2=[False, True] out_cond=[False, True]
1990
    """
H
hong 已提交
1991 1992 1993 1994
    if in_dygraph_mode():
        default_axis = -1
        return _C_ops.final_state_equal(x, y, default_axis)

1995 1996 1997 1998 1999
    check_variable_and_dtype(x, "x", ["float32", "float64", "int32", "int64"],
                             "equal")
    check_variable_and_dtype(y, "y", ["float32", "float64", "int32", "int64"],
                             "equal")
    if cond is not None:
2000
        check_type(cond, "cond", Variable, "equal")
2001

2002 2003
    helper = LayerHelper("equal", **locals())
    if cond is None:
X
Xin Pan 已提交
2004
        cond = helper.create_variable_for_type_inference(dtype='bool')
2005 2006
        cond.stop_gradient = True

2007 2008 2009 2010 2011 2012
    helper.append_op(type='equal',
                     inputs={
                         'X': [x],
                         'Y': [y]
                     },
                     outputs={'Out': [cond]})
2013 2014 2015
    return cond


W
wawltor 已提交
2016
def not_equal(x, y, cond=None, name=None):
Z
zhoukunsheng 已提交
2017
    """
2018 2019 2020 2021
    :alias_main: paddle.not_equal
	:alias: paddle.not_equal,paddle.tensor.not_equal,paddle.tensor.logic.not_equal
	:old_api: paddle.fluid.layers.not_equal

2022
    This OP returns the truth value of :math:`x != y` elementwise, which is equivalent function to the overloaded operator `!=`.
Z
zhoukunsheng 已提交
2023 2024

    Args:
2025 2026
        x(Variable): First input to compare which is N-D tensor. The input data type should be float32, float64, int32, int64. 
        y(Variable): Second input to compare which is N-D tensor. The input data type should be float32, float64, int32, int64.
2027 2028
        cond(Variable, optional): Optional output which can be any created Variable that meets the requirements to store the result of *not_equal*.
            if cond is None, a new Varibale will be created to store the result.
W
wawltor 已提交
2029 2030
        name(str, optional): The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
Z
zhoukunsheng 已提交
2031 2032

    Returns:
2033
        Variable, the output data type is bool: The tensor variable storing the output, the output shape is same as input :attr:`x`.
Z
zhoukunsheng 已提交
2034 2035 2036 2037

    Examples:
        .. code-block:: python

2038 2039 2040 2041
          import paddle.fluid as fluid
          
          label = fluid.layers.data(name='label', shape=[1], dtype='int64')
          limit = fluid.layers.fill_constant(shape=[1], value=1, dtype='int64')
Z
zhoukunsheng 已提交
2042 2043
          out = fluid.layers.not_equal(x=label, y=limit)
    """
2044 2045 2046 2047 2048
    check_variable_and_dtype(x, "x", ["float32", "float64", "int32", "int64"],
                             "not_equal")
    check_variable_and_dtype(y, "y", ["float32", "float64", "int32", "int64"],
                             "not_equal")
    if cond is not None:
2049
        check_type(cond, "cond", Variable, "not_equal")
2050

Z
zhoukunsheng 已提交
2051 2052 2053 2054 2055
    helper = LayerHelper("not_equal", **locals())
    if cond is None:
        cond = helper.create_variable_for_type_inference(dtype='bool')
        cond.stop_gradient = True

2056 2057 2058 2059 2060 2061
    helper.append_op(type='not_equal',
                     inputs={
                         'X': [x],
                         'Y': [y]
                     },
                     outputs={'Out': [cond]})
Z
zhoukunsheng 已提交
2062 2063 2064
    return cond


2065
def array_read(array, i):
2066
    """
2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081
    This OP is used to read data at the specified position from the input array 
    :ref:`api_fluid_LoDTensorArray` . ``array`` is the input array and ``i``
    is the specified read position. This OP is often used together with 
    :ref:`api_fluid_layers_array_write` OP.

    Case 1:
    ::
        Input:
            The shape of first three tensors are [1], and that of the last one is [1,2]:
                array = ([0.6], [0.1], [0.3], [0.4, 0.2])
            And:
                i = [3]

        Output:
            output = [0.4, 0.2]
2082

K
kavyasrinet 已提交
2083
    Args:
2084 2085 2086
        array (LoDTensorArray): The input LoDTensorArray.
        i (Variable): 1-D Tensor, whose shape is [1] and dtype is int64. It represents the
            specified read position of ``array``.
2087

K
kavyasrinet 已提交
2088
    Returns:
2089
        Variable: The LoDTensor or Tensor that is read at the specified position of ``array``.
2090

K
kavyasrinet 已提交
2091
    Examples:
2092 2093
        .. code-block:: python

2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124
            # First we're going to create a LoDTensorArray, then we're going to write the Tensor into
            # the specified position, and finally we're going to read the Tensor at that position.
            import paddle.fluid as fluid
            arr = fluid.layers.create_array(dtype='float32')
            tmp = fluid.layers.fill_constant(shape=[3, 2], dtype='int64', value=5)
            i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10)
            # tmp is the Tensor with shape [3,2], and if we write it into the position with subscript 10
            # of the empty-array: arr, then the length of arr becomes 11.
            arr = fluid.layers.array_write(tmp, i, array=arr)
            # Read the data of the position with subscript 10.
            item = fluid.layers.array_read(arr, i)

            # You can print out the data via executor.
            input = fluid.layers.Print(item, message="The LoDTensor of the i-th position:")
            main_program = fluid.default_main_program()
            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(main_program)

            # The printed result is:

            # 1569588169  The LoDTensor of the i-th position: The place is:CPUPlace
            # Tensor[array_read_0.tmp_0]
            #    shape: [3,2,]
            #    dtype: l
            #    data: 5,5,5,5,5,5,

            # the output is 2-D Tensor with shape [3,2].
            # dtype is the corresponding C++ data type, which may vary in different environments.
            # Eg: if the data type of tensor is int64, then the corresponding C++ data type is int64_t, 
            #       so the dtype value is typeid(int64_t).Name(), which is 'x' on MacOS, 'l' on Linux, 
            #       and '__int64' on Windows. They both represent 64-bit integer variables.
2125
    """
J
Jiabin Yang 已提交
2126
    if _non_static_mode():
2127 2128 2129 2130 2131 2132 2133 2134 2135
        assert isinstance(
            array,
            list), "The 'array' in array_read must be list in dygraph mode"
        assert isinstance(
            i, Variable
        ), "The index 'i' in array_read must be Variable in dygraph mode"
        assert i.shape == [
            1
        ], "The shape of index 'i' should be [1] in dygraph mode"
2136
        i = i.numpy().item(0)
2137 2138
        return array[i]

2139
    check_variable_and_dtype(i, 'i', ['int64'], 'array_read')
Y
Yu Yang 已提交
2140 2141 2142 2143 2144
    helper = LayerHelper('array_read', **locals())
    if not isinstance(
            array,
            Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY:
        raise TypeError("array should be tensor array vairable")
X
Xin Pan 已提交
2145
    out = helper.create_variable_for_type_inference(dtype=array.dtype)
2146 2147 2148 2149 2150 2151
    helper.append_op(type='read_from_array',
                     inputs={
                         'X': [array],
                         'I': [i]
                     },
                     outputs={'Out': [out]})
Y
Yu Yang 已提交
2152
    return out
Y
Yang Yu 已提交
2153 2154


2155
def shrink_memory(x, i, table):
2156
    """
Y
yuyang18 已提交
2157
    This function creates an operator to shrink rnn memory using the RankTable
2158
    as mentioned in the input parameter.
Y
yuyang18 已提交
2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178

    NOTE: This API is very low-level API. It is used by DynamicRNN only.

    Since the Dynamic RNN uses no-padding way to implement RNN. The sequence
    will be sorted by order, and the length of valid memory will be shrink after
    each time step.

    Args:
        x(Variable): The memory object in the previous time step.
        i(Variable): The step count variable. A int scalar as LoDTensor.
        table(Variable): The RNNRankTable object.

    Returns:
        the memory variable after shrink.

    Examples:

        Since this API is very low level API. The example is not provided.
        Please reference the implementation of class DynamicRNN for detail
        usage.
2179
    """
Y
Yang Yu 已提交
2180
    helper = LayerHelper('shrink_memory', **locals())
2181 2182 2183
    check_type(x, 'x', Variable, 'shrink_memory')
    check_type(i, 'i', Variable, 'shrink_memory')
    check_type(table, 'table', Variable, 'shrink_memory')
X
Xin Pan 已提交
2184
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
2185 2186 2187 2188 2189 2190 2191 2192
    helper.append_op(type='shrink_rnn_memory',
                     inputs={
                         'X': [x],
                         'I': [i],
                         'RankTable': [table]
                     },
                     outputs={'Out': [out]},
                     attrs={})
Y
Yang Yu 已提交
2193
    return out
Y
Yang Yu 已提交
2194 2195


2196
def array_length(array):
2197
    """
2198 2199
    This OP is used to get the length of the input array :ref:`api_fluid_LoDTensorArray` .
    It can be used together with :ref:`api_fluid_layers_array_read` , :ref:`api_fluid_layers_array_write` , 
T
tianshuo78520a 已提交
2200
    :ref:`api_fluid_layers_While` OP to traverse, read and write LoDTensorArray.
2201

K
kavyasrinet 已提交
2202
    Args:
2203
        array (LoDTensorArray): The input array that will be used to compute the length.
K
kavyasrinet 已提交
2204 2205

    Returns:
2206
        Variable: 1-D Tensor with shape [1], which is the length of array. Datatype: int64.
K
kavyasrinet 已提交
2207 2208

    Examples:
Q
qiaolongfei 已提交
2209
        .. code-block:: python
K
kavyasrinet 已提交
2210

2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226
            import paddle.fluid as fluid
            tmp = fluid.layers.zeros(shape=[10], dtype='int32')
            i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10)
            # tmp is 1-D Tensor with shape [10]. We write tmp into arr on subscript 10,
            # then the length of arr becomes 11.
            arr = fluid.layers.array_write(tmp, i=i)
            # return the length of arr
            arr_len = fluid.layers.array_length(arr)

            # You can use executor to print out the length of LoDTensorArray.
            input = fluid.layers.Print(arr_len, message="The length of LoDTensorArray:")
            main_program = fluid.default_main_program()
            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(main_program)

            # The printed result is:
Q
qiaolongfei 已提交
2227

2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239
            # 1569576542  The length of LoDTensorArray:   The place is:CPUPlace
            # Tensor[array_length_0.tmp_0]
            #    shape: [1,]
            #    dtype: l
            #    data: 11,
            
            # 1-D Tensor with shape [1], whose value is 11. It means that the length of LoDTensorArray
            # is 11.
            # dtype is the corresponding C++ data type, which may vary in different environments.
            # Eg: if the data type of tensor is int64, then the corresponding C++ data type is int64_t, 
            #       so the dtype value is typeid(int64_t).Name(), which is 'x' on MacOS, 'l' on Linux, 
            #       and '__int64' on Windows. They both represent 64-bit integer variables.
2240
    """
2241

J
Jiabin Yang 已提交
2242
    if _non_static_mode():
2243 2244 2245 2246 2247
        assert isinstance(
            array,
            list), "The 'array' in array_write must be a list in dygraph mode"
        return len(array)

2248 2249 2250 2251 2252 2253
    if not isinstance(
            array,
            Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY:
        raise TypeError(
            "array should be tensor array vairable in array_length Op")

Y
Yang Yu 已提交
2254
    helper = LayerHelper('array_length', **locals())
X
Xin Pan 已提交
2255
    tmp = helper.create_variable_for_type_inference(dtype='int64')
Y
Yang Yu 已提交
2256
    tmp.stop_gradient = True
2257 2258 2259
    helper.append_op(type='lod_array_length',
                     inputs={'X': [array]},
                     outputs={'Out': [tmp]})
Y
Yang Yu 已提交
2260
    return tmp
Y
Yu Yang 已提交
2261 2262 2263


class ConditionalBlockGuard(BlockGuard):
F
fengjiayi 已提交
2264
    """
2265 2266 2267
    ConditionalBlockGuard is derived from BlockGuard. It is dedicated for
    holding a ConditionalBlock, and helping users entering and exiting the
    ConditionalBlock via Python's 'with' keyword. However, ConditionalBlockGuard
F
fengjiayi 已提交
2268 2269 2270
    is generally an internal component of IfElse, users should not use it directly.
    """

Y
Yu Yang 已提交
2271
    def __init__(self, block):
2272
        check_type(block, "block", ConditionalBlock, "ConditionalBlockGuard")
Y
Yu Yang 已提交
2273 2274 2275 2276 2277 2278 2279 2280
        super(ConditionalBlockGuard, self).__init__(block.helper.main_program)
        self.block = block

    def __enter__(self):
        return super(ConditionalBlockGuard, self).__enter__()

    def __exit__(self, exc_type, exc_val, exc_tb):
        self.block.complete()
2281 2282
        return super(ConditionalBlockGuard,
                     self).__exit__(exc_type, exc_val, exc_tb)
Y
Yu Yang 已提交
2283 2284 2285


class ConditionalBlock(object):
Y
Yan Chunwei 已提交
2286 2287 2288 2289 2290 2291 2292 2293
    '''
    **ConditionalBlock**

    ConditionalBlock is an operator that bind a block to a specific condition,
    if the condition matches, the corresponding block will be executed.

    Args:
        inputs (Variable): bool conditions.
T
tianshuo78520a 已提交
2294
        is_scalar_condition (bool): whether the branch is controlled by a scalar.
Y
Yan Chunwei 已提交
2295 2296 2297 2298 2299
        name(str): name of this ConditionalBlock.

    Examples:
        .. code-block:: python

2300
             import paddle.fluid as fluid
Y
Yan Chunwei 已提交
2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311
             cond = layers.less_than(x=label, y=limit)
             true_image, false_image = layers.split_lod_tensor(
                 input=image, mask=cond)
             true_cond = layers.ConditionalBlock([true_image])

             with true_cond.block():
                 ...
             with false_cond.block():
                 ...
    '''

2312
    def __init__(self, inputs, is_scalar_condition=False, name=None):
Y
Yu Yang 已提交
2313
        for each_input in inputs:
2314
            check_type(each_input, "input", Variable, "ConditionalBlock")
Y
Yu Yang 已提交
2315
        self.inputs = inputs
2316
        self.is_scalar_condition = is_scalar_condition
2317
        self.helper = LayerHelper('conditional_block', name=name)
Y
Yu Yang 已提交
2318 2319 2320 2321 2322 2323 2324 2325 2326 2327

    def block(self):
        return ConditionalBlockGuard(self)

    def complete(self):
        inside_block = self.helper.main_program.current_block()
        parent_block = self.helper.main_program.block(inside_block.parent_idx)

        intermediate = set()
        params = set()
2328 2329 2330 2331
        params, intermediate = get_inputs_outputs_in_block(inside_block,
                                                           params,
                                                           intermediate,
                                                           helper=self.helper)
Y
Yu Yang 已提交
2332

2333 2334 2335
        # Todo(liym27) Here assume that all params are in recursive parent block
        # but when minimize() called in control flow, some params may be in
        # conditional grad block
Y
Yu Yang 已提交
2336
        param_list = [
W
Wu Yi 已提交
2337
            parent_block._var_recursive(each_name) for each_name in params
Y
Yu Yang 已提交
2338 2339
        ]

X
Xin Pan 已提交
2340 2341 2342 2343 2344
        out_list = []
        for inner_out_name in intermediate:
            inner_var = parent_block._find_var_recursive(inner_out_name)
            if inner_var:
                out_list.append(inner_var)
Y
Yu Yang 已提交
2345 2346

        step_scope = parent_block.create_var(
2347
            type=core.VarDesc.VarType.STEP_SCOPES)
2348
        conditional_block_op = parent_block.append_op(
Y
Yu Yang 已提交
2349 2350
            type='conditional_block',
            inputs={
2351 2352
                'Cond': self.inputs,
                'Input': param_list,
Y
Yu Yang 已提交
2353
            },
2354 2355 2356 2357
            outputs={
                'Out': out_list,
                'Scope': [step_scope]
            },
2358 2359 2360 2361 2362
            attrs={
                'sub_block': inside_block,
                'is_scalar_condition': self.is_scalar_condition
            })

2363 2364 2365 2366 2367 2368
        if self.need_append_conditional_block_grad(inside_block):
            self.append_conditional_block_grad(parent_block, inside_block,
                                               conditional_block_op)

    def need_append_conditional_block_grad(self, inside_block):
        grad_sub_block_idx = inside_block.backward_block_idx
2369
        inside_block_idx = inside_block.idx
2370

2371 2372 2373
        # if inside_block have grad_block and grad_block is not itself,
        # we will append conditional block grad.
        return grad_sub_block_idx != -1 and grad_sub_block_idx != inside_block_idx
2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414

    def append_conditional_block_grad(self, parent_block, inside_block,
                                      conditional_block_op):
        '''
        Append op `conditional_block_grad` manually.
        When `optimizer.minimize/append_backward` is called in Paddle control flow,
        grad ops will be appended before appending op `conditional_block` so that
        op `conditional_block_grad` can't be appended when calling
        `optimizer.minimize/append_backward`. After appending op `conditional_block`,
        `conditional_block_grad` is appended manually.

        Args:
            parent_block (Block): The block that `conditional_block_op` blongs to.
            inside_block (Block): The sub block of `conditional_block_op`.
            conditional_block_op (Operator): The forward op conditional_block.
        '''

        grad_sub_block_idx = inside_block.backward_block_idx
        grad_sub_block = self.helper.main_program.block(grad_sub_block_idx)

        intermediate = set()
        params = set()

        for each_op in grad_sub_block.ops:
            assert isinstance(each_op, Operator)
            for iname in each_op.input_names:
                for in_var_name in each_op.input(iname):
                    if in_var_name not in intermediate:
                        params.add(in_var_name)

            for oname in each_op.output_names:
                for out_var_name in each_op.output(oname):
                    intermediate.add(out_var_name)

        param_list = []
        for inner_input_name in params:
            inner_var = parent_block._find_var_recursive(inner_input_name)
            if inner_var:
                param_list.append(cpt.to_text(inner_var.name))

        grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
2415 2416
            conditional_block_op.desc, cpt.to_text(set()),
            [grad_sub_block.desc])
2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430

        # append op_desc in grad_op_descs to target_block
        op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
        backward = core.op_proto_and_checker_maker.OpRole.Backward
        new_op_desc = parent_block.desc.append_op()
        new_op_desc.copy_from(grad_op_desc[0])
        new_op_desc._set_attr(op_role_attr_name, backward)
        # set input and output manually
        new_op_desc.set_input('Input', param_list)
        new_op_desc.set_output('Input@GRAD',
                               [param + "@GRAD" for param in param_list])

        new_vars = set()
        for grad_var_name in new_op_desc.output_arg_names():
2431 2432
            if grad_sub_block.desc.has_var_recursive(cpt.to_bytes(
                    grad_var_name)) or grad_var_name == core.empty_var_name():
2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448
                continue
            grad_sub_block.desc.var(cpt.to_bytes(grad_var_name))
            new_vars.add(grad_var_name)
            if grad_var_name not in op_grad_to_var:
                continue

        # infer_shape and infer_type
        new_op_desc.infer_var_type(grad_sub_block.desc)
        new_op_desc.infer_shape(grad_sub_block.desc)

        for arg in new_op_desc.output_arg_names():
            if arg in new_vars:
                _infer_var_data_type_shape_(arg, grad_sub_block)

        self.helper.main_program._sync_with_cpp()

2449

2450
def copy_var_to_parent_block(var, layer_helper):
2451 2452
    if not isinstance(var, Variable):
        return var
2453 2454 2455 2456 2457
    prog = layer_helper.main_program
    parent_idx = prog.current_block().parent_idx
    assert parent_idx >= 0, "Got wrong parent block index when assigning var to parent scope in control_flow"
    parent_block = prog.block(parent_idx)

2458 2459 2460 2461
    if var.type == core.VarDesc.VarType.LOD_TENSOR_ARRAY \
            and parent_block._find_var_recursive(var.name):
        parent_block_var = var
    else:
2462 2463 2464
        parent_block_var = parent_block.create_var(dtype=var.dtype,
                                                   shape=var.shape,
                                                   type=var.type)
2465
        assign(var, parent_block_var)
2466 2467 2468
    return parent_block_var


2469
def cond(pred, true_fn=None, false_fn=None, name=None, return_names=None):
2470
    """
2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481
    This API returns ``true_fn()`` if the predicate ``pred`` is true else
    ``false_fn()`` . Users could also set ``true_fn`` or ``false_fn`` to
    ``None`` if do nothing and this API will treat the callable simply returns
    ``None`` in this case.

    ``true_fn`` and ``false_fn`` should return same nest structure of tensors
    or both return ``None`` if user doens't like to return anything. A nest
    structure of tensors in PaddlePaddle is tensor(s), or tuple of tensors, or
    list of tensors.
    
    Note: 
2482 2483 2484 2485
        1. The tuples or lists returned by ``true_fn`` and ``false_fn`` must have
        the same shape because of dataflow model of PaddlePaddle while the
        tensors in the tuples or the lists can have different shapes.

2486 2487 2488 2489 2490 2491 2492
        2. This API could be used under both static mode or dygraph mode. If it
        is in dygraph mode, the API only runs one branch based on condition.

        3. If it is in static mode, any tensors or operations created outside 
        or inside of ``true_fn`` and ``false_fn`` will be in net building
        regardless of which branch is selected at runtime. This has frequently
        surprised users who expected a lazy semantics. For example:
2493 2494

        .. code-block:: python
2495 2496 2497 2498 2499

            import paddle

            a = paddle.zeros((1, 1))
            b = paddle.zeros((1, 1))
2500
            c = a * b
2501
            out = paddle.static.nn.cond(a < b, lambda: a + c, lambda: b * b)
2502

2503 2504 2505
        No matter whether ``a < b`` , ``c = a * b`` will be in net building and
        run. ``a + c`` and ``b * b`` will be in net building, but only one
        branch will be executed during runtime.
2506 2507

    Args:
2508
        pred(Tensor): A boolean tensor whose numel should be 1. The boolean
2509
            value determines whether to return the result of ``true_fn`` or
2510 2511 2512 2513 2514 2515
            ``false_fn`` .
        true_fn(callable, optional): A callable to be performed if ``pred`` is
            true. The default value is ``None`` .
        false_fn(callable, optional): A callable to be performed if ``pred`` is
            false. The default value is ``None`` .
        name(str, optional): The default value is ``None`` . Normally users
2516
             don't have to set this parameter. For more information, please
2517
             refer to :ref:`api_guide_Name` .
2518 2519 2520 2521
        return_names(sequence of string, optional): The default value is ``None`` . 
             Normally users don't have to set this parameters.  A sequence of strings 
             to represents the name of returned vars.  The structure of sequence must 
             be same with return values of true_fn and false_fn.
2522 2523

    Returns:
2524
        Tensor|list(Tensor)|tuple(Tensor): returns ``true_fn()`` if the
2525
        predicate ``pred`` is true else ``false_fn()`` .
2526 2527 2528

    Raises:
        TypeError: if ``true_fn`` or ``false_fn`` is not callable.
2529 2530
        ValueError: if ``true_fn`` and ``false_fn`` don't return the same nest
            structure of tensors.
2531 2532 2533 2534

    Examples:
        .. code-block:: python

2535
            import paddle
2536 2537 2538 2539 2540 2541 2542 2543 2544 2545

            #
            # pseudocode:
            # if 0.1 < 0.23:
            #     return 1, True
            # else:
            #     return 3, 2
            #

            def true_func():
2546 2547 2548 2549
                return paddle.full(shape=[1, 2], dtype='int32',
                                   fill_value=1), paddle.full(shape=[2, 3],
                                                              dtype='bool',
                                                              fill_value=True)
2550

2551 2552

            def false_func():
2553 2554 2555 2556 2557
                return paddle.full(shape=[3, 4], dtype='float32',
                                   fill_value=3), paddle.full(shape=[4, 5],
                                                              dtype='int64',
                                                              fill_value=2)

2558

2559 2560
            x = paddle.full(shape=[1], dtype='float32', fill_value=0.1)
            y = paddle.full(shape=[1], dtype='float32', fill_value=0.23)
2561
            pred = paddle.less_than(x=x, y=y, name=None)
2562
            ret = paddle.static.nn.cond(pred, true_func, false_func)
2563
            # ret is a tuple containing 2 tensors
2564 2565
            # ret[0] = [[1 1]]
            # ret[1] = [[ True  True  True]
2566
            #           [ True  True  True]]            
2567

2568
    """
J
Jiabin Yang 已提交
2569
    if _non_static_mode():
2570
        assert isinstance(pred, Variable), "The pred in cond must be Variable"
C
crystal 已提交
2571
        assert pred.size == 1, "condition input's numel should be 1"
2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583
        pred = pred.numpy()[0]
        if pred:
            if true_fn is not None:
                if not callable(true_fn):
                    raise TypeError(
                        "The true_fn in cond must be callable, but received {}".
                        format(type(true_fn).__name__))
                return true_fn()
        else:
            if false_fn is not None:
                if not callable(false_fn):
                    raise TypeError(
2584 2585
                        "The false_fn in cond must be callable, but received {}"
                        .format(type(false_fn).__name__))
2586 2587 2588
                return false_fn()
        return None

2589 2590
    check_variable_and_dtype(pred, "pred", ['bool'], "fluid.layers.cond")
    check_type(name, "name", (str, type(None)), "fluid.layers.cond")
2591 2592 2593
    helper = LayerHelper('cond', **locals())
    true_output = None
    false_output = None
2594
    copy_to_parent_func = lambda var: copy_var_to_parent_block(var, helper)
2595 2596
    if true_fn is not None:
        if not callable(true_fn):
2597 2598 2599
            raise TypeError(
                "The true_fn in cond must be callable, but received {}".format(
                    type(true_fn).__name__))
2600 2601 2602 2603
        true_cond_block = ConditionalBlock([pred], is_scalar_condition=True)
        with true_cond_block.block():
            origin_true_output = true_fn()
            if origin_true_output is not None:
2604
                true_output = map_structure(copy_to_parent_func,
2605 2606 2607
                                            origin_true_output)
    if false_fn is not None:
        if not callable(false_fn):
2608 2609 2610
            raise TypeError(
                "The false_fn in cond must be callable, but received {}".format(
                    type(false_fn).__name__))
2611 2612
        false_cond_block = ConditionalBlock([logical_not(pred)],
                                            is_scalar_condition=True)
2613 2614 2615
        with false_cond_block.block():
            origin_false_output = false_fn()
            if origin_false_output is not None:
2616
                false_output = map_structure(copy_to_parent_func,
2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631
                                             origin_false_output)

    if true_output is None and false_output is None:
        return None

    if true_output is None:
        raise ValueError(
            "Incompatible return values of true_fn and false_fn in cond: "
            "true_fn returns None while false_fn returns non-None")
    if false_output is None:
        raise ValueError(
            "Incompatible return values of true_fn and false_fn in cond: "
            "true_fn returns non-None while false_fn returns None")

    # Merge ture and false output if they are not None
2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642
    if return_names is None:
        return_names = ["no name"] * len(to_sequence(true_output))
    else:
        """ 
        dy2static will set the return_names and expand the return values to UndefinedVar.
        """
        true_output, false_output = expand_undefined_var(
            true_output, false_output, return_names)
        true_output, false_output = change_none_to_undefinedvar(
            true_output, false_output)
    if len(to_sequence(true_output)) != len(to_sequence(false_output)):
2643
        raise ValueError(
2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655
            "true fn returns {} vars, but false fn returns {} vars, which is not equals"
            .format(len(to_sequence(true_output)),
                    len(to_sequence(false_output))))
    for true_out, false_out, return_name in zip(to_sequence(true_output),
                                                to_sequence(false_output),
                                                to_sequence(return_names)):
        try:
            assert_same_structure(true_out, false_out, check_types=False)
        except ValueError as e:
            raise ValueError(
                "Incompatible return values of `{}` in true_fn and false_fn in cond: {}"
                .format(return_name, e))
2656 2657

    mask = cast(pred, dtype='int32')
2658 2659
    merge_func = lambda false_var, true_var: select_input_with_buildin_type(
        [false_var, true_var], mask)
2660 2661 2662 2663
    merged_output = map_structure(merge_func, false_output, true_output)
    return merged_output


2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676
def change_none_to_undefinedvar(nest1, nest2):
    from paddle.fluid.dygraph.dygraph_to_static.utils import UndefinedVar

    def map_fn(x):
        if x is None: return UndefinedVar("padding")
        return x

    nest1_out = pack_sequence_as(nest1, list(map(map_fn, flatten(nest1))))
    nest2_out = pack_sequence_as(nest2, list(map(map_fn, flatten(nest2))))
    return nest1_out, nest2_out


def expand_undefined_var(nest1, nest2, names):
2677 2678 2679 2680 2681
    """ TODO: make this function recursively.
        nest1: Var1, (UndefinedVar, [1,2,3])
        nest2: Var2, ([1,2,3,4], UndefinedVar)
        In this case, we should not expand recursively.
    """
2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703
    from paddle.fluid.dygraph.dygraph_to_static.utils import UndefinedVar
    from paddle.fluid.dygraph.dygraph_to_static.return_transformer import RETURN_VALUE_PREFIX

    def pack_undefined_var_as(seq):
        return pack_sequence_as(seq,
                                [UndefinedVar("padding") for i in flatten(seq)])

    def map_fn(n1, n2, name):
        if not name.startswith(RETURN_VALUE_PREFIX) and (isinstance(
                n1, UndefinedVar) or n1 is None):
            return pack_undefined_var_as(n2)
        return n1

    nest1_out = list(
        map(map_fn, to_sequence(nest1), to_sequence(nest2), to_sequence(names)))
    nest2_out = list(
        map(map_fn, to_sequence(nest2), to_sequence(nest1), to_sequence(names)))
    if not is_sequence(nest1): nest1_out = nest1_out[0]
    if not is_sequence(nest2): nest2_out = nest2_out[0]
    return nest1_out, nest2_out


L
liym27 已提交
2704
def _error_message(what, arg_name, op_name, right_value, error_value):
2705
    error_message = "{what} of '{arg_name}' in {op_name} must be " \
L
liym27 已提交
2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717
        "{right_value}, but received: {error_value}.".format(
        what=what,
        arg_name=arg_name,
        op_name=op_name,
        right_value=right_value,
        error_value=error_value)

    return error_message


def case(pred_fn_pairs, default=None, name=None):
    '''
2718 2719
    :api_attr: Static Graph

L
liym27 已提交
2720 2721 2722 2723 2724 2725 2726 2727
    This operator works like an if-elif-elif-else chain.

    Args:
        pred_fn_pairs(list|tuple): A list or tuple of (pred, fn) pairs. ``pred`` is a boolean Tensor with shape [1], ``fn`` is a callable. All callables return the same structure of Tensors.
        default(callable, optional): Callable that returns a structure of Tensors.
        name(str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`.

    Returns:
2728
        Tensor|list(Tensor): Tensors returned by the callable from the first pair whose pred is True,
L
liym27 已提交
2729 2730 2731 2732 2733 2734 2735
        or Tensors returned by ``default`` if no pred in ``pred_fn_pairs`` is True and ``default`` is not None,
        or Tensors returned by the last callable in ``pred_fn_pairs``  if no pred in ``pred_fn_pairs`` is True and ``default`` is None.

    Raises:
        TypeError: If the type of ``pred_fn_pairs`` is not list or tuple.
        TypeError: If the type of elements in ``pred_fn_pairs`` is not tuple.
        TypeError: If the size of tuples in ``pred_fn_pairs`` is not 2.
2736
        TypeError: If the first element of 2-tuple in ``pred_fn_pairs`` is not a Tensor.
L
liym27 已提交
2737 2738 2739 2740 2741 2742
        TypeError: If the second element of 2-tuple in ``pred_fn_pairs`` is not callable.
        TypeError: If ``default`` is not None but it is not callable.

    Examples:
        .. code-block:: python

2743 2744 2745
            import paddle

            paddle.enable_static()
L
liym27 已提交
2746 2747

            def fn_1():
2748
                return paddle.full(shape=[1, 2], dtype='float32', fill_value=1)
L
liym27 已提交
2749 2750

            def fn_2():
2751
                return paddle.full(shape=[2, 2], dtype='int32', fill_value=2)
L
liym27 已提交
2752 2753

            def fn_3():
2754
                return paddle.full(shape=[3], dtype='int32', fill_value=3)
L
liym27 已提交
2755

2756 2757 2758 2759
            main_program = paddle.static.default_startup_program()
            startup_program = paddle.static.default_main_program()

            with paddle.static.program_guard(main_program, startup_program):
2760 2761 2762
                x = paddle.full(shape=[1], dtype='float32', fill_value=0.3)
                y = paddle.full(shape=[1], dtype='float32', fill_value=0.1)
                z = paddle.full(shape=[1], dtype='float32', fill_value=0.2)
L
liym27 已提交
2763

2764 2765 2766
                pred_1 = paddle.less_than(z, x)  # true: 0.2 < 0.3
                pred_2 = paddle.less_than(x, y)  # false: 0.3 < 0.1
                pred_3 = paddle.equal(x, y)      # false: 0.3 == 0.1
L
liym27 已提交
2767 2768

                # Call fn_1 because pred_1 is True
2769
                out_1 = paddle.static.nn.case(
L
liym27 已提交
2770 2771 2772 2773
                    pred_fn_pairs=[(pred_1, fn_1), (pred_2, fn_2)], default=fn_3)

                # Argument default is None and no pred in pred_fn_pairs is True. fn_3 will be called.
                # because fn_3 is the last callable in pred_fn_pairs.
2774
                out_2 = paddle.static.nn.case(pred_fn_pairs=[(pred_2, fn_2), (pred_3, fn_3)])
L
liym27 已提交
2775

2776
                exe = paddle.static.Executor(paddle.CPUPlace())
L
liym27 已提交
2777 2778 2779 2780 2781 2782 2783 2784 2785 2786
                res_1, res_2 = exe.run(main_program, fetch_list=[out_1, out_2])
                print(res_1)  # [[1. 1.]]
                print(res_2)  # [3 3 3]
    '''
    helper = LayerHelper('case', **locals())

    def _case_check_args(pred_fn_pairs, default):
        '''
        Check arguments pred_fn_pairs and default. Return canonical pre_fn_pairs and default.
        '''
2787
        check_type(pred_fn_pairs, 'pred_fn_pairs', (list, tuple), 'case')
L
liym27 已提交
2788 2789 2790 2791 2792

        for pred_fn in pred_fn_pairs:
            if not isinstance(pred_fn, tuple):
                raise TypeError(
                    _error_message("The elements' type", "pred_fn_pairs",
2793
                                   "case", tuple, type(pred_fn)))
L
liym27 已提交
2794 2795 2796
            if len(pred_fn) != 2:
                raise TypeError(
                    _error_message("The tuple's size", "pred_fn_pairs", "case",
2797 2798
                                   "2",
                                   str(len(pred_fn)) + "-tuple"))
L
liym27 已提交
2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830
            pred, fn = pred_fn

            if not isinstance(pred, Variable):
                raise TypeError(
                    _error_message("The pred's type", "pred_fn_pairs", "case",
                                   "boolean Variable", type(pred)))

            if not callable(fn):
                raise TypeError(
                    "The fn for {} of pred_fn_pairs in Op(case) must"
                    " be callable.".format(pred.name))

        if default is None:
            default_index = len(pred_fn_pairs) - 1  # pick the last one
            default = pred_fn_pairs[default_index][1]
            pred_fn_pairs = pred_fn_pairs[:default_index]
        elif not callable(default):
            raise TypeError("The default in Op(case) must be callable.")

        return pred_fn_pairs, default

    pred_fn_pairs, default = _case_check_args(pred_fn_pairs, default)

    false_fn = default
    for pred, true_fn in reversed(pred_fn_pairs):
        false_fn = partial(cond, pred=pred, true_fn=true_fn, false_fn=false_fn)

    final_fn = false_fn

    return final_fn()


2831
class Switch(object):
Q
qiaolongfei 已提交
2832
    """
2833
    :api_attr: Static Graph
Q
qiaolongfei 已提交
2834

2835 2836 2837 2838 2839 2840 2841
    This class is used to implement Switch branch control function. 
    Switch branch contains several case branches and one default branch. 
    Switch control flow checks whether the case branch conditions are satisfied in turn, 
    and only executes the statement after the first case branch that satisfies the conditions. 
    If there is no case branch that satisfies the condition, 
    only the statement following the default branch is executed.

2842 2843 2844 2845
    Note:
        A new OP :ref:`api_fluid_layers_case` is highly recommended instead of ``Switch`` if the shape of parameter ``cond`` is [1].
        OP :ref:`api_fluid_layers_case` is easier to use and is called with less code but does the same thing as ``Switch`` .

2846
    Member Functions:
2847
        case(condition): The case branch of Switch whose parameter cond is a scalar Variable of bool type. Only if the cond of the current case branch is True and the cond of the previous case branch is False, the statement after the case branch will be executed, and the statement after the case branch will not be executed.
2848 2849 2850 2851 2852 2853
        
        default(): The default branch of Switch. When cond of all case branches is False, the statement after default branch is executed.

    Case and default functions can only be used inside the scope of Switch, as shown below:

    .. code-block:: python
2854

2855 2856 2857 2858 2859 2860 2861 2862 2863
        '''
        with fluid.layers.Switch() as switch:
            with switch.case(cond1):
                i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=1)
            with switch.case(cond2):
                i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=2)
            with switch.default():
                i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=0)
        '''
Q
qiaolongfei 已提交
2864

2865 2866
    Args:
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` .
Q
qiaolongfei 已提交
2867 2868 2869

    Examples:
        .. code-block:: python
2870 2871
            
            import paddle.fluid as fluid
Q
qiaolongfei 已提交
2872

2873
            lr = fluid.layers.create_global_var(
Q
qiaolongfei 已提交
2874 2875 2876 2877 2878
                shape=[1],
                value=0.0,
                dtype='float32',
                persistable=True,
                name="learning_rate")
2879
            zero_var = fluid.layers.fill_constant(
2880
                shape=[1], dtype='float32', value=0.0)
2881
            one_var = fluid.layers.fill_constant(
Q
qiaolongfei 已提交
2882
                shape=[1], dtype='float32', value=1.0)
2883
            two_var = fluid.layers.fill_constant(
2884
                shape=[1], dtype='float32', value=2.0)
2885

2886
            global_step = fluid.layers.autoincreased_step_counter(counter_name='@LR_DECAY_COUNTER@', begin=0, step=1)
Q
qiaolongfei 已提交
2887 2888

            with fluid.layers.control_flow.Switch() as switch:
Q
qiaolongfei 已提交
2889
                with switch.case(global_step == zero_var):
2890
                    fluid.layers.assign(input=one_var, output=lr)
Q
qiaolongfei 已提交
2891
                with switch.default():
2892
                    fluid.layers.assign(input=two_var, output=lr)
Q
qiaolongfei 已提交
2893

2894 2895 2896 2897 2898
            exe = fluid.Executor(fluid.CPUPlace())
            exe.run(fluid.default_startup_program())

            res = exe.run(fluid.default_main_program(), feed={}, fetch_list=[lr])
            print(res) # [array([1.], dtype=float32)]
Q
qiaolongfei 已提交
2899 2900
    """

2901 2902 2903 2904 2905 2906 2907 2908 2909
    def __init__(self, name=None):
        self.helper = LayerHelper('switch', name=name)
        self.inside_scope = False
        self.pre_not_conditions = []

    def case(self, condition):
        if not self.inside_scope:
            raise ValueError("case should be called inside with")

2910 2911 2912 2913
        check_variable_and_dtype(
            condition, 'condition', ['bool'],
            'the member function case of fluid.layers.Switch')

2914 2915 2916 2917 2918 2919 2920
        if len(self.pre_not_conditions) == 0:
            cond_block = ConditionalBlock([condition], is_scalar_condition=True)
            not_cond = logical_not(x=condition)
            self.pre_not_conditions.append(not_cond)
        else:
            pre_cond_num = len(self.pre_not_conditions)
            pre_not_cond = self.pre_not_conditions[pre_cond_num - 1]
2921 2922
            new_not_cond = logical_and(x=pre_not_cond,
                                       y=logical_not(x=condition))
2923 2924
            self.pre_not_conditions.append(new_not_cond)
            cond_block = ConditionalBlock(
2925
                [logical_and(x=pre_not_cond, y=condition)],
2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952
                is_scalar_condition=True)

        return ConditionalBlockGuard(cond_block)

    def default(self):
        pre_cond_num = len(self.pre_not_conditions)
        if pre_cond_num == 0:
            raise ValueError("there should be at least one condition")
        cond_block = ConditionalBlock(
            [self.pre_not_conditions[pre_cond_num - 1]],
            is_scalar_condition=True)
        return ConditionalBlockGuard(cond_block)

    def __enter__(self):
        """
        set flag that now is inside switch.block {}
        :return:
        """
        self.inside_scope = True
        return self

    def __exit__(self, exc_type, exc_val, exc_tb):
        self.inside_scope = False
        if exc_type is not None:
            return False  # re-raise exception

        return True
Y
Yu Yang 已提交
2953 2954 2955


class IfElseBlockGuard(object):
2956

Y
Yu Yang 已提交
2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989
    def __init__(self, is_true, ifelse):
        if not isinstance(ifelse, IfElse):
            raise TypeError("ifelse must be an instance of IfElse class")

        if ifelse.status != IfElse.OUT_IF_ELSE_BLOCKS:
            raise ValueError("You cannot invoke IfElse.block() inside a block")

        self.is_true = is_true
        self.ie = ifelse
        if is_true:
            self.cond_block = ifelse.conditional_true_block
        else:
            self.cond_block = ifelse.conditional_false_block

        if not isinstance(self.cond_block, ConditionalBlock):
            raise TypeError("Unexpected situation")

        self.cond_block = self.cond_block.block()

    def __enter__(self):
        self.ie.status = IfElse.IN_IF_ELSE_TRUE_BLOCKS if self.is_true else IfElse.IN_IF_ELSE_FALSE_BLOCKS
        self.cond_block.__enter__()

    def __exit__(self, exc_type, exc_val, exc_tb):
        if not self.cond_block.__exit__(exc_type, exc_val, exc_tb):
            # re-raise inside exception
            return False
        if len(self.ie.output_table[1 if self.is_true else 0]) == 0:
            raise ValueError("Must set output inside block")
        self.ie.status = IfElse.OUT_IF_ELSE_BLOCKS


class IfElse(object):
X
Xin Pan 已提交
2990
    """
2991 2992
    :api_attr: Static Graph

2993 2994 2995 2996
    This class is used to implement IfElse branch control function. IfElse contains two blocks, true_block and false_block. IfElse will put data satisfying True or False conditions into different blocks to run.

    Cond is a 2-D Tensor with shape [N, 1] and data type bool, representing the execution conditions of the corresponding part of the input data.

2997 2998 2999 3000
    Note:
        A new OP :ref:`api_fluid_layers_cond` is highly recommended instead of ``IfElse``. if the shape of parameter ``cond`` is [1].
        OP :ref:`api_fluid_layers_cond` is easier to use and is called with less code but does the same thing as ``IfElse`` .

3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041
    IfElse OP is different from other OPs in usage, which may cause some users confusion. Here is a simple example to illustrate this OP.

    .. code-block:: python
        
        # The following code completes the function: subtract 10 from the data greater than 0 in x, add 10 to the data less than 0 in x, and sum all the data.
        import numpy as np
        import paddle.fluid as fluid

        x = fluid.layers.data(name='x', shape=[4, 1], dtype='float32', append_batch_size=False)
        y = fluid.layers.data(name='y', shape=[4, 1], dtype='float32', append_batch_size=False)

        x_d = np.array([[3], [1], [-2], [-3]]).astype(np.float32)
        y_d = np.zeros((4, 1)).astype(np.float32)
        
        # Compare the size of x, y pairs of elements, output cond, cond is shape [4, 1], data type bool 2-D tensor.
        # Based on the input data x_d, y_d, it can be inferred that the data in cond are [[true], [true], [false], [false]].
        cond = fluid.layers.greater_than(x, y)
        # Unlike other common OPs, ie below returned by the OP is an IfElse OP object
        ie = fluid.layers.IfElse(cond)

        with ie.true_block():
            # In this block, according to cond condition, the data corresponding to true dimension in X is obtained and subtracted by 10.
            out_1 = ie.input(x)
            out_1 = out_1 - 10
            ie.output(out_1)
        with ie.false_block():
            # In this block, according to cond condition, get the data of the corresponding condition in X as false dimension, and add 10
            out_1 = ie.input(x)
            out_1 = out_1 + 10
            ie.output(out_1)

        # According to cond condition, the data processed in the two blocks are merged. The output here is output, the type is List, and the element type in List is Variable.
        output = ie() #  [array([[-7.], [-9.], [ 8.], [ 7.]], dtype=float32)] 

        # Get the first Variable in the output List and add all elements.
        out = fluid.layers.reduce_sum(output[0])

        exe = fluid.Executor(fluid.CPUPlace())
        exe.run(fluid.default_startup_program())

        res = exe.run(fluid.default_main_program(), feed={"x":x_d, "y":y_d}, fetch_list=[out])
3042
        print(res)
3043
        # [array([-1.], dtype=float32)] 
X
Xin Pan 已提交
3044 3045

    Args:
3046 3047
        cond (Variable): cond is a 2-D Tensor with shape [N, 1] and data type bool, representing the corresponding execution conditions of N input data. The data type is bool.
        name(str, optional): The default value is None.  Normally there is no need for user to set this property.  For more information, please refer to :ref:`api_guide_Name` .
X
Xin Pan 已提交
3048

3049 3050
    Returns:
        Unlike other common OPs, the OP call returns an IfElse OP object (e.g. ie in the example), which branches the input data by calling the internal functions of the object ``true_block ()``, ``false_block ()``, ``input ()``, ``output ()``, and integrates the data processed by different branches as the overall output by calling the internal ``call ()`` function. The output type is a list, and the type of each element in the list is Variable.
X
Xin Pan 已提交
3051

3052 3053 3054 3055 3056 3057 3058 3059 3060 3061
    Internal Functions:
        The block is constructed by calling the ``with ie. true_block()`` function in the object, and the computational logic under condition true is put into the block. If no corresponding block is constructed, the input data in the corresponding conditional dimension is unchanged.
 
        The block is constructed by calling the ``with ie. false_block()`` function in the object, and the computational logic under condition false is put into the block. If no corresponding block is constructed, the input data in the corresponding conditional dimension is unchanged.

        ``Out = ie. input (x)`` will take out the data of the corresponding conditional dimension in X and put it into out, supporting the internal processing of multiple inputs in block.

        ``ie. output (out)`` writes the result to the output of the corresponding condition.

        There is a ``call ()`` function inside the object, that is, by calling ``output = ie ()``, all the outputs inside the block of False are fused as the whole output, the output type is a list, and the type of each element in the list is Variable.
3062

X
Xin Pan 已提交
3063
    """
Y
Yu Yang 已提交
3064 3065 3066 3067
    OUT_IF_ELSE_BLOCKS = 0
    IN_IF_ELSE_TRUE_BLOCKS = 1
    IN_IF_ELSE_FALSE_BLOCKS = 2

3068
    def __init__(self, cond, name=None):
3069 3070
        check_type(cond, "cond", Variable, "fluid.layers.IfElse")
        check_type(name, "name", (str, type(None)), "fluid.layers.IfElse")
3071
        self.helper = LayerHelper('ifelse', name=name)
Y
Yu Yang 已提交
3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082
        self.cond = cond
        self.input_table = {}
        self.status = IfElse.OUT_IF_ELSE_BLOCKS
        self.conditional_true_block = ConditionalBlock(inputs=[self.cond])
        self.conditional_false_block = ConditionalBlock(inputs=[self.cond])
        self.output_table = ([], [])  # (true_outs, false_outs)

    def input(self, x):
        if self.status == IfElse.OUT_IF_ELSE_BLOCKS:
            raise ValueError("input must in true/false blocks")
        if id(x) not in self.input_table:
3083
            parent_block = self._parent_block()
Y
Yu Yang 已提交
3084
            out_true = parent_block.create_var(
3085 3086
                name=unique_name.generate_with_ignorable_key('ifelse_input' +
                                                             self.helper.name),
F
fengjiayi 已提交
3087
                dtype=x.dtype)
Y
Yu Yang 已提交
3088 3089

            out_false = parent_block.create_var(
3090 3091
                name=unique_name.generate_with_ignorable_key('ifelse_input' +
                                                             self.helper.name),
F
fengjiayi 已提交
3092
                dtype=x.dtype)
3093 3094 3095 3096 3097 3098 3099 3100 3101 3102
            parent_block.append_op(type='split_lod_tensor',
                                   inputs={
                                       'X': x,
                                       'Mask': self.cond,
                                   },
                                   outputs={
                                       'OutTrue': out_true,
                                       'OutFalse': out_false
                                   },
                                   attrs={'level': 0})
Y
Yu Yang 已提交
3103 3104 3105 3106 3107 3108 3109 3110 3111
            self.input_table[id(x)] = (out_true, out_false)
        else:
            out_true, out_false = self.input_table[id(x)]

        if self.status == IfElse.IN_IF_ELSE_TRUE_BLOCKS:
            return out_true
        else:
            return out_false

3112
    def _parent_block(self):
Y
Yu Yang 已提交
3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127
        current_block = self.helper.main_program.current_block()
        return self.helper.main_program.block(current_block.parent_idx)

    def true_block(self):
        return IfElseBlockGuard(True, self)

    def false_block(self):
        return IfElseBlockGuard(False, self)

    def output(self, *outs):
        if self.status == self.OUT_IF_ELSE_BLOCKS:
            raise ValueError("output can only be invoked in the sub-block")

        out_table = self.output_table[1 if self.status ==
                                      self.IN_IF_ELSE_TRUE_BLOCKS else 0]
3128
        parent_block = self._parent_block()
Y
Yu Yang 已提交
3129
        for each_out in outs:
3130 3131
            check_type(each_out, "each output", Variable,
                       "fluid.layers.IfElse.output")
Y
Yu Yang 已提交
3132 3133
            # create outside tensor
            outside_out = parent_block.create_var(
3134
                name=unique_name.generate_with_ignorable_key("_".join(
Y
Yu Yang 已提交
3135
                    [self.helper.name, 'output'])),
F
fengjiayi 已提交
3136
                dtype=each_out.dtype)
Y
Yu Yang 已提交
3137 3138 3139
            out_table.append(outside_out)

            # assign local var to outside
3140
            assign(input=each_out, output=outside_out)
Y
Yu Yang 已提交
3141 3142 3143 3144

    def __call__(self):
        if self.status != self.OUT_IF_ELSE_BLOCKS:
            raise ValueError("IfElse::__call__ must be out of sub-block")
3145
        false_len, true_len = list(map(len, self.output_table))
Y
Yu Yang 已提交
3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158
        if false_len == 0 and true_len == 0:
            raise ValueError("Must invoke true_block/false_block before "
                             "__call__")
        elif false_len != true_len and false_len != 0 and true_len != 0:
            raise ValueError("The output side must be same")
        elif false_len == 0 or true_len == 0:
            return self.output_table[0 if false_len != 0 else 1]

        # else none of false_len/true_len is zero
        # merge together
        rlist = []
        for false_var, true_var in zip(*self.output_table):
            rlist.append(
3159 3160 3161 3162 3163
                merge_lod_tensor(in_true=true_var,
                                 in_false=false_var,
                                 mask=self.cond,
                                 x=self.cond,
                                 level=0))
Y
Yu Yang 已提交
3164
        return rlist
3165 3166 3167


class DynamicRNN(object):
Y
yuyang18 已提交
3168
    """
3169 3170
    :api_attr: Static Graph

3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182
    **Note: the input of this class should be LoDTensor which holds the
    information of variable-length sequences. If the input is fixed-length Tensor,
    please use StaticRNN (fluid.layers.** :ref:`api_fluid_layers_StaticRNN` **) for
    better performance.**

    DynamicRNN can process a minibatch of variable-length sequences.
    The length of each sample can be different and is recorded in LoD.
    In DynamicRNN, an input sequence will be unfolded into time steps and users
    can define how to process each time step in :code:`block()` .
    The total number of time steps is determined by the longest sequence.
    DynamicRNN will not pad all sequences to the same length, instead it will
    sort the sequences internally by the sequence length in descending order.
T
tianshuo78520a 已提交
3183
    The input sequences will be shrank because only sequences of which the
3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195
    length is larger than the time step will participate the remaining calculation.

    If defined :code:`drnn = DynamicRNN()`, then users can call :code:`drnn()`
    to obtain the result sequences. It is a LoDTensor gained by merging all
    time steps's output. When RNN's input sequence x meets :code:`x.lod_level == 1`,
    the output LoDTensor will have the same LoD with x. The result of :code:`drnn()`
    includes RNN's outputs of all time steps, users can call
    :ref:`api_fluid_layers_sequence_last_step` to extract the data of the last time step.

    Warning:
        Currently it is not supported to set :code:`is_sparse = True` of any
        layers defined within DynamicRNN's :code:`block` function.
Y
yuyang18 已提交
3196

3197 3198 3199 3200
    Args:
        name (str, optional): The default value is None.  Normally there is no
            need for user to set this property.  For more information,
            please refer to :ref:`api_guide_Name` .
3201 3202 3203 3204

    Examples:
        .. code-block:: python

3205
            import paddle.fluid as fluid
3206

3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232
            sentence = fluid.data(name='sentence', shape=[None, 32], dtype='float32', lod_level=1)
            encoder_proj = fluid.data(name='encoder_proj', shape=[None, 32], dtype='float32', lod_level=1)
            decoder_boot = fluid.data(name='boot', shape=[None, 10], dtype='float32')

            drnn = fluid.layers.DynamicRNN()
            with drnn.block():
                # Set sentence as RNN's input, each time step processes a word from the sentence
                current_word = drnn.step_input(sentence)
                # Set encode_proj as RNN's static input
                encoder_word = drnn.static_input(encoder_proj)
                # Initialize memory with boot_memory, which need reorder according to RNN's input sequences
                memory = drnn.memory(init=decoder_boot, need_reorder=True)
                fc_1 = fluid.layers.fc(input=encoder_word, size=30)
                fc_2 = fluid.layers.fc(input=current_word, size=30)
                decoder_inputs = fc_1 + fc_2
                hidden, _, _ = fluid.layers.gru_unit(input=decoder_inputs, hidden=memory, size=30)
                # Update memory with hidden
                drnn.update_memory(ex_mem=memory, new_mem=hidden)
                out = fluid.layers.fc(input=hidden, size=10, bias_attr=True, act='softmax')
                # Set hidden and out as RNN's outputs
                drnn.output(hidden, out)

            # Get RNN's result
            hidden, out = drnn()
            # Get RNN's result of the last time step
            last = fluid.layers.sequence_last_step(out)
Y
yuyang18 已提交
3233
    """
3234 3235 3236 3237
    BEFORE_RNN = 0
    IN_RNN = 1
    AFTER_RNN = 2

3238 3239
    def __init__(self, name=None):
        self.helper = LayerHelper('dynamic_rnn', name=name)
3240 3241 3242 3243
        self.status = DynamicRNN.BEFORE_RNN
        self.lod_rank_table = None
        self.max_seq_len = None
        self.step_idx = None
3244
        self.zero_idx = None
3245 3246 3247
        self.mem_dict = dict()
        self.output_array = []
        self.outputs = []
X
Xin Pan 已提交
3248
        self.cond = self.helper.create_variable_for_type_inference(dtype='bool')
3249 3250 3251 3252 3253
        self.cond.stop_gradient = False
        self.while_op = While(self.cond)
        self.input_array = []
        self.mem_link = []

3254
    def step_input(self, x, level=0):
3255
        r"""
3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298
        This function is used to set sequence x as DynamicRNN's input.
        The maximum sequence length in x determines the number of time steps
        the RNN unit will be executed. DynamicRNN can take multiple inputs.
        When all inputs' :code:`lod_level` are 1, all inputs should hold the
        same LoD. When :code:`x.lod_level >= 2` , the input sequence will be
        unfold along specified level, and the slice of each time step is a
        LoDTensor whose lod_level is :code:`x.lod_level - level - 1` .
        In this case, the specified LoD level of multiple inputs should be the same.

        - Case 1:

        .. code-block:: text

            # input, where Si is slice data of shape [1, N]
            level = 0
            x.lod = [[2, 1, 3]]
            x.shape = [6, N]
            x.data = [[S0],
                      [S0],
                      [S1],
                      [S2],
                      [S2],
                      [S2]]

            # output
            # step 0, time step data of 3 sequences
            out.lod = [[]]
            out.shape = [3, N]
            out.data = [[S2],
                        [S0],
                        [S1]]

            # step 1, time step data of 2 sequences
            out.lod = [[]]
            out.shape = [2, N]
            out.data = [[S2],
                        [S0]]

            # step 2, time step data of 1 sequences
            out.lod = [[]]
            out.shape = [1, N]
            out.data = [[S2]]

H
haowang101779990 已提交
3299

Y
yuyang18 已提交
3300
        Args:
3301 3302 3303 3304 3305 3306 3307
            x (Variable): The input LoDTensor which holds information of a
                minibatch of variable-length sequences and should meet :code:`x.lod_level >= 1` .
                When RNN has multiple inputs, the first dimension should match
                across all inputs, but other shape components may differ.
                Optional data types are: bool, float16, float32, float64, int8, int16, int32, int64, uint8.
            level (int, optional): The level of lod used to split steps.
                It should be in range :math:`[0, x.lod\_level)` . The default value is 0.
Y
yuyang18 已提交
3308 3309

        Returns:
3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343
            Variable: The current time step in the input sequence. If there are :code:`num_sequences` \
                sequences in x whose length is larger than :code:`step_idx` , the returned Variable \
                will only hold the :code:`step_idx` -th time step of those `num_sequences` sequences. \
                The data type is the same as input. If :code:`x.lod_level == 1` , the return value is \
                a Tensor of shape :math:`\{num\_sequences, x.shape[1], ...\}` , or it will \
                be a variable-length LoDTensor.

        Raises:
            ValueError: When :code:`step_input()` is called outside :code:`block()` .
            TypeError: When x is not a Variable.

        Examples:
            ..  code-block:: python

                import paddle.fluid as fluid

                sentence = fluid.data(name='sentence', shape=[None, 1], dtype='int64', lod_level=1)
                embedding = fluid.layers.embedding(input=sentence, size=[65536, 32], is_sparse=True)

                drnn = fluid.layers.DynamicRNN()
                with drnn.block():
                    # Set embedding as RNN's input, each time step processes a word from the sentence
                    word = drnn.step_input(embedding)
                    # Initialize memory to a Tensor whose value is 0, shape=[batch_size, 200],
                    # where batch_size is the number of sequences in embedding.
                    memory = drnn.memory(shape=[200])
                    hidden = fluid.layers.fc(input=[word, memory], size=200, act='relu')
                    # Update memory to hidden
                    drnn.update_memory(ex_mem=memory, new_mem=hidden)
                    # Set hidden as RNN's output
                    drnn.output(hidden)

                # Get RNN's result
                rnn_output = drnn()
Y
yuyang18 已提交
3344
        """
3345
        self._assert_in_rnn_block_("step_input")
3346
        check_type(x, 'x', Variable, 'fluid.layers.DynamicRNN.step_input()')
3347 3348 3349
        parent_block = self._parent_block_()
        if self.lod_rank_table is None:
            self.lod_rank_table = parent_block.create_var(
Y
Yu Yang 已提交
3350
                name=unique_name.generate('lod_rank_table'),
3351 3352
                type=core.VarDesc.VarType.LOD_RANK_TABLE)
            self.lod_rank_table.stop_gradient = True
3353 3354 3355 3356
            parent_block.append_op(type='lod_rank_table',
                                   inputs={"X": x},
                                   outputs={"Out": self.lod_rank_table},
                                   attrs={"level": level})
3357
            self.max_seq_len = parent_block.create_var(
Y
Yu Yang 已提交
3358 3359
                name=unique_name.generate('dynamic_rnn_max_seq_len'),
                dtype='int64')
3360
            self.max_seq_len.stop_gradient = False
3361 3362 3363
            parent_block.append_op(type='max_sequence_len',
                                   inputs={'RankTable': self.lod_rank_table},
                                   outputs={"Out": self.max_seq_len})
3364
            self.cond.stop_gradient = True
3365 3366 3367 3368 3369 3370 3371
            parent_block.append_op(type='less_than',
                                   inputs={
                                       'X': self.step_idx,
                                       'Y': self.max_seq_len
                                   },
                                   outputs={'Out': self.cond},
                                   attrs={'force_cpu': True})
3372 3373

        input_array = parent_block.create_var(
Y
Yu Yang 已提交
3374
            name=unique_name.generate('dynamic_rnn_input_array'),
3375 3376 3377
            type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
            dtype=x.dtype)
        self.input_array.append((input_array, x.dtype))
3378 3379 3380 3381 3382 3383
        parent_block.append_op(type='lod_tensor_to_array',
                               inputs={
                                   'X': x,
                                   'RankTable': self.lod_rank_table
                               },
                               outputs={'Out': input_array})
3384
        return array_read(array=input_array, i=self.step_idx)
3385

Y
yangyaming 已提交
3386
    def static_input(self, x):
3387
        r"""
3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460
        This function is used to set x as DynamicRNN's static input. It is optional.

        - Case 1, set static input with LoD

        .. code-block:: text

            # RNN's input is the same as the case listed in step_input
            # static input, where Si is slice data of shape [1, M]
            x.lod = [[3, 1, 2]]
            x.shape = [6, M]
            x.data = [[S0],
                      [S0],
                      [S0],
                      [S1],
                      [S2],
                      [S2]]

            # step 0, batch data corresponding to the 3 input sequences
            out.lod = [[2, 3, 1]]
            out.shape = [6, M]
            out.data = [[S2],
                        [S2],
                        [S0],
                        [S0],
                        [S0],
                        [S1]]

            # step 1, batch data corresponding to the 2 input sequences
            out.lod = [[2, 3]]
            out.shape = [5, M]
            out.data = [[S2],
                        [S2],
                        [S0],
                        [S0],
                        [S0]]

            # step 2, batch data corresponding to the 1 input sequences
            out.lod = [[2]]
            out.shape = [2, M]
            out.data = [[S2],
                        [S2]]


        - Case 2, set static input without LoD

        .. code-block:: text

            # RNN's input is the same as the case listed in step_input
            # static input, where Si is slice data of shape [1, M]
            x.lod = [[]]
            x.shape = [3, M]
            x.data = [[S0],
                      [S1],
                      [S2]]

            # step 0, batch data corresponding to the 3 input sequences
            out.lod = [[]]
            out.shape = [3, M]
            out.data = [[S2],
                        [S0],
                        [S1]]

            # step 1, batch data corresponding to the 2 input sequences
            out.lod = [[]]
            out.shape = [2, M]
            out.data = [[S2],
                        [S0]]

            # step 2, batch data corresponding to the 1 input sequences
            out.lod = [[]]
            out.shape = [1, M]
            out.data = [[S2]]

H
haowang101779990 已提交
3461

Y
yuyang18 已提交
3462
        Args:
3463 3464 3465 3466
            x (Variable): The static input LoDTensor which should hold the same number of sequences
                as RNN's input (the input LoDTensor set by :code:`step_input()` ). If the LoD is None,
                the input x will be treated as a minibatch with :code:`x.shape[0]` sequences of length 1.
                Optional data types are: bool, float16, float32, float64, int8, int16, int32, int64, uint8.
Y
yuyang18 已提交
3467 3468

        Returns:
T
tianshuo78520a 已提交
3469
            Variable: The input LoDTensor after sorted and shrank. If there are :code:`num_sequences` \
3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480
                sequences in RNN's input LoDTensor whose length is larger than :code:`step_idx` , \
                the static input Tensor will be sorted to the same order as RNN's input and \
                will only retain data corresponding to those :code:`num_sequences` sequences. \
                The data type is the same as input. If :code:`x.lod == None` , the return value is \
                a Tensor of shape :math:`\{num\_sequences, x.shape[1], ...\}` , or it will \
                be a variable-length LoDTensor.

        Raises:
            ValueError: When :code:`static_input()` is called outside :code:`block()` .
            TypeError: When x is not a Variable.
            RuntimeError: When :code:`static_input()` is called before :code:`step_input()` .
3481 3482 3483 3484

        Examples:
            .. code-block:: python

3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510
                import paddle.fluid as fluid

                sentence = fluid.data(name='sentence', shape=[None, 32], dtype='float32', lod_level=1)
                encoder_proj = fluid.data(name='encoder_proj', shape=[None, 32], dtype='float32', lod_level=1)
                decoder_boot = fluid.data(name='boot', shape=[None, 10], dtype='float32')

                drnn = fluid.layers.DynamicRNN()
                with drnn.block():
                    # Set sentence as RNN's input, each time step processes a word from the sentence
                    current_word = drnn.step_input(sentence)
                    # Set encode_proj as RNN's static input
                    encoder_word = drnn.static_input(encoder_proj)
                    # Initialize memory with boot_memory, which need reorder according to RNN's input sequences
                    memory = drnn.memory(init=decoder_boot, need_reorder=True)
                    fc_1 = fluid.layers.fc(input=encoder_word, size=30)
                    fc_2 = fluid.layers.fc(input=current_word, size=30)
                    decoder_inputs = fc_1 + fc_2
                    hidden, _, _ = fluid.layers.gru_unit(input=decoder_inputs, hidden=memory, size=30)
                    # Update memory with hidden
                    drnn.update_memory(ex_mem=memory, new_mem=hidden)
                    out = fluid.layers.fc(input=hidden, size=10, bias_attr=True, act='softmax')
                    # Set out as RNN's output
                    drnn.output(out)

                # Get RNN's result
                rnn_output = drnn()
Y
yuyang18 已提交
3511
        """
Y
yangyaming 已提交
3512
        self._assert_in_rnn_block_("static_input")
3513
        check_type(x, 'x', Variable, 'fluid.layers.DynamicRNN.static_input()')
Y
yangyaming 已提交
3514 3515 3516 3517 3518
        if self.lod_rank_table is None:
            raise RuntimeError(
                "static_input() must be called after step_input().")
        parent_block = self._parent_block_()
        x_reordered = parent_block.create_var(
Y
Yu Yang 已提交
3519
            name=unique_name.generate("dynamic_rnn_static_input_reordered"),
Y
yangyaming 已提交
3520 3521
            type=core.VarDesc.VarType.LOD_TENSOR,
            dtype=x.dtype)
3522 3523 3524 3525 3526 3527
        parent_block.append_op(type='reorder_lod_tensor_by_rank',
                               inputs={
                                   'X': [x],
                                   'RankTable': [self.lod_rank_table]
                               },
                               outputs={'Out': [x_reordered]})
Y
yangyaming 已提交
3528 3529
        return shrink_memory(x_reordered, self.step_idx, self.lod_rank_table)

S
rename  
sneaxiy 已提交
3530
    @signature_safe_contextmanager
3531
    def block(self):
Y
yuyang18 已提交
3532
        """
3533 3534 3535 3536 3537 3538
        The function is used to list the operations executed during
        each time step in RNN. The operation list will be executed :code:`max_sequence_len`
        times (where :code:`max_sequence_len` is the maximum length of RNN's input sequences).

        Raises:
            ValueError: When :code:`block()` is called multi-times.
Y
yuyang18 已提交
3539
        """
3540 3541
        if self.status != DynamicRNN.BEFORE_RNN:
            raise ValueError("rnn.block() can only be invoke once")
3542 3543 3544 3545
        self.step_idx = fill_constant(shape=[1],
                                      dtype='int64',
                                      value=0,
                                      force_cpu=True)
3546 3547 3548 3549
        self.step_idx.stop_gradient = False
        self.status = DynamicRNN.IN_RNN
        with self.while_op.block():
            yield
3550
            increment(x=self.step_idx, value=1.0, in_place=True)
3551 3552

            for new_mem, mem_array in self.mem_link:
3553 3554
                array_write(x=new_mem, i=self.step_idx, array=mem_array)

3555 3556 3557 3558
            less_than(x=self.step_idx,
                      y=self.max_seq_len,
                      force_cpu=True,
                      cond=self.cond)
3559 3560 3561 3562

        self.status = DynamicRNN.AFTER_RNN
        for each_array in self.output_array:
            self.outputs.append(
3563
                array_to_lod_tensor(x=each_array, table=self.lod_rank_table))
3564 3565

    def __call__(self, *args, **kwargs):
Y
yuyang18 已提交
3566
        """
T
tianshuo78520a 已提交
3567
        This function is used to get the output  sequences of DynamicRNN.
3568 3569 3570 3571 3572 3573 3574 3575 3576

        Args:
            None

        Returns:
            Variable or Variable list: RNN's output sequences.

        Raises:
            ValueError: When :code:`__call__()` is called before :code:`block()` .
Y
yuyang18 已提交
3577
        """
3578
        if self.status != DynamicRNN.AFTER_RNN:
3579 3580
            raise ValueError(("Output of the dynamic RNN can only be visited "
                              "outside the rnn block."))
3581 3582 3583 3584 3585
        if len(self.outputs) == 1:
            return self.outputs[0]
        else:
            return self.outputs

3586 3587 3588 3589 3590 3591
    def memory(self,
               init=None,
               shape=None,
               value=0.0,
               need_reorder=False,
               dtype='float32'):
3592
        r"""
3593 3594 3595
        Create a memory Variable for DynamicRNN to deliver data cross time steps.
        It can be initialized by an existing Tensor or a constant Tensor of given
        dtype and shape.
Y
yuyang18 已提交
3596

3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608
        Args:
            init (Variable, optional): LoDTensor used to initialize the memory.
                If init is not None, it should hold the same number of sequences
                as RNN's input (the input LoDTensor set by :code:`step_input()` )
                and the memory will be initialized to it. If init's LoD is None,
                it will be treated as a minibatch with :code:`init.shape[0]` sequences
                of length 1. The default value is None.
            shape (list|tuple, optional): When init is None, it is used to specify
                the memory's shape. Note that the shape does not include the batch_size.
                If setting shape to :math:`\{D_1, D_2, ...\}` , the shape of memory Tensor
                will be :math:`\{batch\_size, D_1, D_2, ...\}` , where batch_size is
                determined by RNN's input sequences. The default value is None.
T
tianshuo78520a 已提交
3609
            value (float, optional): When init is None, it is used as initialized value
3610 3611
                of memory. The default value is 0.0.
            need_reorder (bool, optional): When init is not None, it determines whether
T
tianshuo78520a 已提交
3612
                the memory needs to reorder like the RNN's input sequences. It should be
3613 3614 3615 3616 3617 3618 3619
                set to True when the initialized memory depends on the order of input samples.
                The default value is False.
            dtype (str|numpy.dtype, optional): When init is None, it is used to set the
                data type of memory. The default value is "float32". Optional data types
                are: "float32", "float64", "int32", "int64".

        Returns:
T
tianshuo78520a 已提交
3620
            Variable: The memory LoDTensor after shrank.  If there are :code:`num_sequences` \
3621
                sequences in RNN's input LoDTensor whose length is larger than :code:`step_idx` , \
T
tianshuo78520a 已提交
3622
                the memory Tensor also need to be shrank and will only retain data \
3623 3624 3625 3626 3627 3628
                corresponding to those :code:`num_sequences` sequences.

        Raises:
            ValueError: When :code:`memory()` is called outside :code:`block()` .
            TypeError: When init is set and is not a Variable.
            ValueError: When :code:`memory()` is called before :code:`step_input()` .
Y
yuyang18 已提交
3629

3630 3631 3632
        Examples:
            .. code-block:: python

3633
                import paddle.fluid as fluid
3634

3635 3636
                sentence = fluid.data(name='sentence', shape=[None, 32], dtype='float32', lod_level=1)
                boot_memory = fluid.data(name='boot', shape=[None, 10], dtype='float32')
3637

3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648
                drnn = fluid.layers.DynamicRNN()
                with drnn.block():
                    # Set sentence as RNN's input, each time step processes a word from the sentence
                    word = drnn.step_input(sentence)
                    # Initialize memory with boot_memory, which need reorder according to RNN's input sequences
                    memory = drnn.memory(init=boot_memory, need_reorder=True)
                    hidden = fluid.layers.fc(input=[word, memory], size=10, act='tanh')
                    # Update memory with hidden
                    drnn.update_memory(ex_mem=memory, new_mem=hidden)
                    # Set hidden as RNN's output
                    drnn.output(hidden)
Y
yuyang18 已提交
3649

3650 3651
                # Get RNN's result
                rnn_output = drnn()
Y
yuyang18 已提交
3652 3653


3654 3655
        Examples:
            .. code-block:: python
Y
yuyang18 已提交
3656

3657 3658 3659 3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675
                import paddle.fluid as fluid

                sentence = fluid.data(name='sentence', shape=[None, 32], dtype='float32', lod_level=1)

                drnn = fluid.layers.DynamicRNN()
                with drnn.block():
                    # Set sentence as RNN's input, each time step processes a word from the sentence
                    word = drnn.step_input(sentence)
                    # Initialize memory to a Tensor whose value is 0, shape=[batch_size, 10],
                    # where batch_size is the number of sequences in sentence.
                    memory = drnn.memory(shape=[10], dtype='float32', value=0)
                    hidden = fluid.layers.fc(input=[word, memory], size=10, act='tanh')
                    # Update memory with hidden
                    drnn.update_memory(ex_mem=memory, new_mem=hidden)
                    # Set hidden as RNN's output
                    drnn.output(hidden)

                # Get RNN's result
                rnn_output = drnn()
Y
yuyang18 已提交
3676
        """
3677
        self._assert_in_rnn_block_('memory')
3678
        self._init_zero_idx_()
3679 3680 3681
        if shape is not None:
            check_type(shape, 'shape', (list, tuple),
                       'fluid.layers.DynamicRNN.memory()')
3682
        if init is not None:
3683 3684
            check_type(init, 'init', Variable,
                       'fluid.layers.DynamicRNN.memory()')
3685
            parent_block = self._parent_block_()
3686 3687 3688 3689 3690 3691 3692 3693
            init_tensor = init
            if need_reorder == True:
                if self.lod_rank_table is None:
                    raise ValueError(
                        'If set need_reorder to True, make sure step_input be '
                        'invoked before '
                        'memory(init=init, need_reordered=True, ...).')
                init_reordered = parent_block.create_var(
Y
Yu Yang 已提交
3694
                    name=unique_name.generate('dynamic_rnn_mem_init_reordered'),
3695 3696
                    type=core.VarDesc.VarType.LOD_TENSOR,
                    dtype=init.dtype)
3697 3698 3699 3700 3701 3702
                parent_block.append_op(type='reorder_lod_tensor_by_rank',
                                       inputs={
                                           'X': [init_tensor],
                                           'RankTable': [self.lod_rank_table]
                                       },
                                       outputs={'Out': [init_reordered]})
3703
                init_tensor = init_reordered
3704
            mem_array = parent_block.create_var(
Y
Yu Yang 已提交
3705
                name=unique_name.generate('dynamic_rnn_mem_array'),
3706 3707
                type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
                dtype=init.dtype)
3708 3709 3710 3711 3712 3713
            parent_block.append_op(type='write_to_array',
                                   inputs={
                                       'X': init_tensor,
                                       'I': self.zero_idx
                                   },
                                   outputs={'Out': mem_array})
3714
            retv = array_read(array=mem_array, i=self.step_idx)
3715 3716 3717
            retv = shrink_memory(x=retv,
                                 i=self.step_idx,
                                 table=self.lod_rank_table)
3718 3719 3720 3721 3722 3723 3724 3725 3726
            self.mem_dict[retv.name] = mem_array
            return retv
        else:
            if len(self.input_array) == 0:
                raise ValueError(
                    "step_input should be invoked before memory(shape=..., value=...)"
                )
            parent_block = self._parent_block_()
            init = parent_block.create_var(
Y
Yu Yang 已提交
3727
                name=unique_name.generate('mem_init'), dtype=dtype)
3728
            arr, dtype = self.input_array[0]
3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744
            in0 = parent_block.create_var(name=unique_name.generate('in0'),
                                          dtype=dtype)
            parent_block.append_op(type='read_from_array',
                                   inputs={
                                       'X': [arr],
                                       'I': [self.zero_idx]
                                   },
                                   outputs={'Out': [in0]})
            parent_block.append_op(type='fill_constant_batch_size_like',
                                   inputs={'Input': [in0]},
                                   outputs={'Out': [init]},
                                   attrs={
                                       'shape': [-1] + shape,
                                       'value': float(value),
                                       'dtype': init.dtype
                                   })
3745 3746 3747
            return self.memory(init=init)

    def update_memory(self, ex_mem, new_mem):
Y
yuyang18 已提交
3748
        """
3749 3750
        Update the memory which need to be delivered across time steps.

Y
yuyang18 已提交
3751
        Args:
3752 3753 3754
            ex_mem (Variable): The memory data of previous time step.
            new_mem (Variable): The new memory data produced in current time step.
                The shape and data type of ex_mem and new_mem should be the same.
Y
yuyang18 已提交
3755 3756 3757

        Returns:
            None
3758 3759 3760 3761 3762 3763
        
        Raises:
            ValueError: When :code:`update_memory()` is called outside :code:`block()` .
            TypeError: When :code:`ex_mem` or :code:`new_mem` is not a Variable.
            ValueError: When :code:`ex_mem` is defined by :code:`memory()` .
            ValueError: When :code:`update_memory()` is called before :code:`step_input()` .
Y
yuyang18 已提交
3764
        """
3765
        self._assert_in_rnn_block_('update_memory')
3766 3767 3768 3769
        check_type(ex_mem, 'ex_mem', Variable,
                   'fluid.layers.DynamicRNN.update_memory()')
        check_type(new_mem, 'new_mem', Variable,
                   'fluid.layers.DynamicRNN.update_memory()')
3770 3771 3772 3773 3774 3775 3776 3777 3778 3779

        mem_array = self.mem_dict.get(ex_mem.name, None)
        if mem_array is None:
            raise ValueError("Please invoke memory before update_memory")
        if self.lod_rank_table is None:
            raise ValueError("Please invoke step_input before update_memory")

        self.mem_link.append((new_mem, mem_array))

    def output(self, *outputs):
Y
yuyang18 已提交
3780
        """
3781
        This function is used to set :code:`outputs` as RNN's output.
Y
yuyang18 已提交
3782 3783

        Args:
3784 3785
            *outputs (Variable ...): The output Tensor. DynamicRNN can mark multiple
                Variables as its output.
Y
yuyang18 已提交
3786 3787 3788

        Returns:
            None
3789 3790 3791

        Raises:
            ValueError: When :code:`output()` is called outside :code:`block()` .
Y
yuyang18 已提交
3792
        """
3793 3794 3795
        self._assert_in_rnn_block_('output')
        parent_block = self._parent_block_()
        for each in outputs:
3796 3797
            check_type(each, "outputs", Variable,
                       "fluid.layers.DynamicRNN.output")
3798
            outside_array = parent_block.create_var(
3799
                name=unique_name.generate_with_ignorable_key("_".join(
3800 3801 3802 3803 3804 3805
                    [self.helper.name, "output_array", each.name])),
                type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
                dtype=each.dtype)
            array_write(x=each, i=self.step_idx, array=outside_array)
            self.output_array.append(outside_array)

3806 3807 3808 3809 3810
    def _init_zero_idx_(self):
        if self.zero_idx is None:
            parent_block = self._parent_block_()
            self.zero_idx = parent_block.create_var(
                name=unique_name.generate('zero_idx'), dtype='int64')
3811 3812 3813 3814 3815 3816 3817 3818 3819
            parent_block.append_op(type='fill_constant',
                                   inputs={},
                                   outputs={'Out': [self.zero_idx]},
                                   attrs={
                                       'shape': [1],
                                       'dtype': self.zero_idx.dtype,
                                       'value': float(0),
                                       'force_cpu': True
                                   })
3820

3821 3822 3823 3824 3825 3826 3827 3828 3829 3830
    def _parent_block_(self):
        prog = self.helper.main_program
        parent_idx = prog.current_block().parent_idx
        assert parent_idx >= 0
        parent_block = prog.block(parent_idx)

        return parent_block

    def _assert_in_rnn_block_(self, method):
        if self.status != DynamicRNN.IN_RNN:
3831 3832
            raise ValueError(
                "{0} can only be invoked inside rnn block.".format(method))
Y
Yang Yu 已提交
3833 3834


L
liym27 已提交
3835 3836
def switch_case(branch_index, branch_fns, default=None, name=None):
    '''
3837 3838
    :api_attr: Static Graph

L
liym27 已提交
3839 3840 3841
    This operator is like a C++ switch/case statement.

    Args:
3842
        branch_index(Tensor): A Tensor with shape [1] to specify which branch to execute. The data type is ``int32``, ``int64`` or ``uint8``.
L
liym27 已提交
3843 3844 3845 3846 3847
        branch_fns(dict|list|tuple): If it's a list or tuple, the elements in it could be pairs of (int, callable) or simple callables whose actual index will be used as the index of callable. If it's a dict, its key is a python integer and the value is a callable. All callables return the same structure of Tensors.
        default(callable, optional): Callable that returns a structure of Tensors.
        name(str, optional): The default value is None. Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`.

    Returns:
3848
        Tensor|list(Tensor): Tensors returned by the callable specified by ``branch_index`` in ``branch_fns``,
L
liym27 已提交
3849 3850 3851 3852
        or Tensors returned by ``default`` if ``default`` is not None and no index matches in ``branch_fns``,
        or Tensors returned by the callable with the max index in ``branch_fns`` if ``default`` is None and no index matches in ``branch_fns``.

    Raises:
3853
        TypeError: If the type of ``branch_index`` is not Tensor.
L
liym27 已提交
3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864
        TypeError: If the data type of ``branch_index`` is not ``int32``, ``int64`` or ``uint8``.
        TypeError: If the type of ``branch_fns`` is not dict, list or tuple.
        TypeError: If the elements of ``branch_fns`` is not 2-tuple.
        TypeError: If the first element of 2-tuple in ``branch_fns`` is not integer.
        ValueError: If the first element of 2-tuple in ``branch_fns`` is not unique.
        TypeError: If the second element of 2-tuple in ``branch_fns`` is not callable.
        TypeError: If ``default`` is not None but it is not callable.

    Examples:
        .. code-block:: python

3865 3866 3867
            import paddle

            paddle.enable_static()
3868

L
liym27 已提交
3869
            def fn_1():
3870
                return paddle.full(shape=[1, 2], dtype='float32', fill_value=1)
L
liym27 已提交
3871 3872

            def fn_2():
3873
                return paddle.full(shape=[2, 2], dtype='int32', fill_value=2)
L
liym27 已提交
3874 3875

            def fn_3():
3876
                return paddle.full(shape=[3], dtype='int32', fill_value=3)
L
liym27 已提交
3877

3878 3879 3880
            main_program = paddle.static.default_startup_program()
            startup_program = paddle.static.default_main_program()
            with paddle.static.program_guard(main_program, startup_program):
3881 3882
                index_1 = paddle.full(shape=[1], dtype='int32', fill_value=1)
                index_2 = paddle.full(shape=[1], dtype='int32', fill_value=2)
L
liym27 已提交
3883

3884
                out_1 = paddle.static.nn.switch_case(
L
liym27 已提交
3885 3886 3887 3888
                    branch_index=index_1,
                    branch_fns={1: fn_1, 2: fn_2},
                    default=fn_3)

3889
                out_2 = paddle.static.nn.switch_case(
L
liym27 已提交
3890 3891 3892 3893 3894
                    branch_index=index_2,
                    branch_fns=[(1, fn_1), (2, fn_2)],
                    default=fn_3)

                # Argument default is None and no index matches. fn_3 will be called because of the max index 7.
3895
                out_3 = paddle.static.nn.switch_case(
L
liym27 已提交
3896 3897 3898
                    branch_index=index_2,
                    branch_fns=[(0, fn_1), (4, fn_2), (7, fn_3)])

3899
                exe = paddle.static.Executor(paddle.CPUPlace())
3900
                res_1, res_2, res_3 = exe.run(main_program, fetch_list=[out_1, out_2, out_3])
L
liym27 已提交
3901 3902 3903 3904 3905 3906 3907 3908
                print(res_1)  # [[1. 1.]]
                print(res_2)  # [[2 2] [2 2]]
                print(res_3)  # [3 3 3]
    '''
    helper = LayerHelper('switch_case', **locals())

    def _check_args(branch_index, branch_fns, default):

3909 3910
        check_variable_and_dtype(branch_index, 'branch_index',
                                 ['uint8', 'int32', 'int64'], 'switch_case')
L
liym27 已提交
3911 3912 3913 3914

        if convert_dtype(branch_index.dtype) != "int64":
            branch_index = cast(branch_index, "int64")

3915
        check_type(branch_fns, 'branch_fns', (list, tuple, dict), 'switch_case')
L
liym27 已提交
3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927

        branch_fns = branch_fns.items() if isinstance(branch_fns,
                                                      dict) else branch_fns

        branch_fns = list(enumerate(branch_fns)) if all(
            callable(fn) for fn in branch_fns) else branch_fns

        keys_of_fns = []
        for index_fn_pair in branch_fns:
            if not isinstance(index_fn_pair, tuple):
                raise TypeError(
                    _error_message("The elements' type", "branch_fns",
3928
                                   "switch_case", tuple, type(branch_fns)))
L
liym27 已提交
3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940

            if len(index_fn_pair) != 2:
                raise TypeError(
                    _error_message("The tuple's size", "branch_fns",
                                   "switch_case", "2",
                                   str(len(index_fn_pair)) + "-tuple"))

            key, fn = index_fn_pair

            if not isinstance(key, int):
                raise TypeError(
                    _error_message("The key's type", "branch_fns",
3941
                                   "switch_case", int, type(key)))
L
liym27 已提交
3942 3943 3944

            if key in keys_of_fns:
                raise ValueError(
3945 3946
                    "The key in 'branch_fns' must be unique, but '{}' appears more than once."
                    .format(key))
L
liym27 已提交
3947 3948 3949 3950 3951
            else:
                keys_of_fns.append(key)

            if not callable(fn):
                raise TypeError(
3952 3953 3954
                    _error_message(
                        "The type of function for key {}".format(key),
                        "branch_fns", "switch_case", "callable", type(fn)))
L
liym27 已提交
3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978

        if default is None:
            default = sorted(branch_fns)[-1][1]
            branch_fns = sorted(branch_fns)[:-1]
        elif not callable(default):
            raise TypeError("The default in Op(case) must be callable.")

        pred_fn_pairs = []
        for index, fn in branch_fns:
            new_index = fill_constant(shape=[1], dtype="int64", value=index)
            pred = equal(branch_index, new_index)
            pred_fn_pairs.append((pred, fn))

        return pred_fn_pairs, default

    pred_fn_pairs, default = _check_args(branch_index, branch_fns, default)
    false_fn = default
    for pred, true_fn in pred_fn_pairs:
        false_fn = partial(cond, pred=pred, true_fn=true_fn, false_fn=false_fn)

    final_fn = false_fn
    return final_fn()


3979
@templatedoc()
Y
Yang Yu 已提交
3980
def reorder_lod_tensor_by_rank(x, rank_table):
3981 3982 3983 3984
    """
    ${comment}

    Args:
3985 3986
        x(${x_type}): ${x_comment}.
        rank_table(${rank_table_type}): ${rank_table_comment}.
3987 3988
    
    Returns:
3989
        out(${out_type}): ${out_comment}.
3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000 4001 4002

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid
          data_desc = (['input', [9], 0], ['ref', [5], 1])
          data = fluid.layers.data(name=data_desc[0][0], shape=data_desc[0][1])
          rank_data = fluid.layers.data(name=data_desc[1][0], shape=data_desc[1][1])
          table = fluid.layers.control_flow.lod_rank_table(rank_data)
          new_data = fluid.layers.reorder_lod_tensor_by_rank(
                           x=data, rank_table=table)

    """
4003 4004 4005 4006 4007 4008 4009

    check_type(x, 'x', (Variable), 'reorder_lod_tensor_by_rank')
    check_type(rank_table, 'rank_table', (Variable),
               'reorder_lod_tensor_by_rank')
    if rank_table.type != core.VarDesc.VarType.LOD_RANK_TABLE:
        raise TypeError("The type of rank_table should be LOD_RANK_TABLE.")

Y
Yang Yu 已提交
4010 4011
    helper = LayerHelper('reorder_lod_tensor_by_rank', **locals())

X
Xin Pan 已提交
4012
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
4013 4014 4015 4016 4017 4018
    helper.append_op(type='reorder_lod_tensor_by_rank',
                     inputs={
                         'X': [x],
                         'RankTable': [rank_table]
                     },
                     outputs={'Out': [out]})
Y
Yang Yu 已提交
4019
    return out
4020 4021


4022
def is_empty(x, name=None):
4023
    """
4024

4025
    Test whether a Tensor is empty.
4026 4027

    Args:
4028 4029 4030 4031
        x (Tensor): The Tensor to be tested.
        name (str, optional): The default value is ``None`` . Normally users
                            don't have to set this parameter. For more information,
                            please refer to :ref:`api_guide_Name` .
4032 4033

    Returns:
4034
        Tensor: A bool scalar Tensor. True if 'x' is an empty Tensor.
4035 4036 4037 4038

    Examples:
        .. code-block:: python

4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049
            import paddle

            input = paddle.rand(shape=[4, 32, 32], dtype='float32')
            res = paddle.is_empty(x=input)
            print("res:", res)
            # ('res:', Tensor: eager_tmp_1
            #    - place: CPUPlace
            #    - shape: [1]
            #    - layout: NCHW
            #    - dtype: bool
            #    - data: [0])
4050

4051
    """
H
hong 已提交
4052 4053 4054
    if in_dygraph_mode():
        return _C_ops.final_state_is_empty(x)
    if _in_legacy_dygraph():
W
wanghuancoder 已提交
4055
        return _C_ops.is_empty(x)
4056

4057 4058
    check_variable_and_dtype(x, 'x', ['float32', 'float64', 'int32', 'int64'],
                             'is_empty')
4059 4060
    check_type(name, "name", (str, type(None)), "is_empty")

4061
    helper = LayerHelper("is_empty", **locals())
4062 4063
    cond = helper.create_variable_for_type_inference(dtype='bool')
    cond.stop_gradient = True
4064 4065 4066
    helper.append_op(type='is_empty',
                     inputs={'X': [x]},
                     outputs={'Out': [cond]})
4067
    return cond