io.py 22.3 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.
F
fengjiayi 已提交
14
import contextlib
D
dzhwinter 已提交
15

Y
Yu Yang 已提交
16
from .. import core
T
typhoonzero 已提交
17
from ..framework import convert_np_dtype_to_dtype_, default_main_program, default_startup_program, Program
Y
Yu Yang 已提交
18
from ..unique_name import generate as unique_name
T
WIP  
typhoonzero 已提交
19 20
from control_flow import BlockGuard
from ..layer_helper import LayerHelper
Y
Refine  
Yu Yang 已提交
21
from ..executor import global_scope
Y
Yu Yang 已提交
22

Y
Yu Yang 已提交
23 24
__all__ = [
    'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'open_recordio_file',
F
fengjiayi 已提交
25
    'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer',
W
Wu Yi 已提交
26
    'random_data_generator', 'Preprocessor'
Y
Yu Yang 已提交
27
]
Y
Yu Yang 已提交
28 29 30 31 32 33 34 35 36 37


def data(name,
         shape,
         append_batch_size=True,
         dtype='float32',
         lod_level=0,
         type=core.VarDesc.VarType.LOD_TENSOR,
         stop_gradient=True):
    """
K
kavyasrinet 已提交
38
    **Data Layer**
Y
Yu Yang 已提交
39

K
kavyasrinet 已提交
40
    This function takes in the input and based on whether data has
C
caoying03 已提交
41
    to be returned back as a minibatch, it creates the global variable by using
Y
Yu Yang 已提交
42
    the helper functions. The global variables can be accessed by all the
C
caoying03 已提交
43
    following operators in the graph.
Y
Yu Yang 已提交
44 45 46 47

    All the input variables of this function are passed in as local variables
    to the LayerHelper constructor.

K
kavyasrinet 已提交
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63
    Args:
       name(str): The name/alias of the function
       shape(list): Tuple declaring the shape.
       append_batch_size(bool): Whether or not to append the data as a batch.
       dtype(int|float): The type of data : float32, float_16, int etc
       type(VarType): The output type. By default it is LOD_TENSOR.
       lod_level(int): The LoD Level. 0 means the input data is not a sequence.
       stop_gradient(bool): A boolean that mentions whether gradient should flow.

    Returns:
        Variable: The global variable that gives access to the data.

    Examples:
        .. code-block:: python

          data = fluid.layers.data(name='x', shape=[784], dtype='float32')
Y
Yu Yang 已提交
64 65 66 67 68 69 70 71 72 73 74 75 76
    """
    helper = LayerHelper('data', **locals())
    shape = list(shape)
    for i in xrange(len(shape)):
        if shape[i] is None:
            shape[i] = -1
            append_batch_size = False
        elif shape[i] < 0:
            append_batch_size = False

    if append_batch_size:
        shape = [-1] + shape  # append batch size as -1

Y
Yu Yang 已提交
77
    data_var = helper.create_global_variable(
Y
Yu Yang 已提交
78 79 80 81 82
        name=name,
        shape=shape,
        dtype=dtype,
        type=type,
        stop_gradient=stop_gradient,
F
fengjiayi 已提交
83 84
        lod_level=lod_level,
        is_data=True)
Y
Yu Yang 已提交
85
    return data_var
T
typhoonzero 已提交
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116


class BlockGuardServ(BlockGuard):
    """
    BlockGuardServ class.

    BlockGuardServ class is used to create an op with a block in a program.
    """

    def __init__(self, server):
        if not (isinstance(server, ListenAndServ)):
            raise TypeError("BlockGuardServ takes a ListenAndServ")
        super(BlockGuardServ, self).__init__(server.helper.main_program)
        self.server = server

    def __exit__(self, exc_type, exc_val, exc_tb):
        if exc_type is not None:
            return False

        self.server.complete_op()
        return super(BlockGuardServ, self).__exit__(exc_type, exc_val, exc_tb)


class ListenAndServ(object):
    """
    ListenAndServ class.

    ListenAndServ class is used to wrap listen_and_serv op to create a server
    which can receive variables from clients and run a block.
    """

Y
Yancey1989 已提交
117
    def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True):
118
        self.helper = LayerHelper("listen_and_serv")
Y
Yancey1989 已提交
119
        self.inputs = inputs
T
typhoonzero 已提交
120 121 122
        self.outputs = []
        self.endpoint = endpoint
        self.fan_in = fan_in
T
typhoonzero 已提交
123 124
        # FIXME(typhoonzero): add optimizer_mode is stupid, should make it more
        # general.
T
WIP  
typhoonzero 已提交
125
        self.optimizer_mode = optimizer_mode
T
typhoonzero 已提交
126 127 128 129 130 131 132 133 134 135 136 137 138

    def do(self):
        return BlockGuardServ(self)

    def get_params_and_grads(self):
        main_program = self.helper.main_program
        current_block = main_program.current_block()
        parent_block = self.parent_block()
        # params and grads in the same order.
        params = list()
        grads = list()
        for op in current_block.ops:
            # FIXME(typhoonzero): op.inputs is None if it's cloned.
T
WIP  
typhoonzero 已提交
139 140 141 142 143 144 145 146
            if self.optimizer_mode:
                if "Grad" in op.inputs and "Param" in op.inputs:
                    params.append(op.inputs["Param"].name)
                    grads.append(op.inputs["Grad"].name)
            else:
                # simple recv mode, recv operators inputs.
                for iname in op.input_names:
                    for in_var_name in op.input(iname):
T
typhoonzero 已提交
147 148
                        params.append(parent_block.var(in_var_name))
                        grads.append(parent_block.var(in_var_name))
T
typhoonzero 已提交
149 150 151

        return params, grads

T
typhoonzero 已提交
152 153 154 155 156 157 158
    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

T
typhoonzero 已提交
159 160 161 162
    def complete_op(self):
        main_program = self.helper.main_program
        current_block = main_program.current_block()
        parent_block = self.parent_block()
T
typhoonzero 已提交
163
        empty_block = Program().global_block()
T
typhoonzero 已提交
164 165

        parent_block.append_op(
166
            type='listen_and_serv',
Y
Yancey1989 已提交
167
            inputs={"X": self.inputs},
T
typhoonzero 已提交
168 169 170 171
            outputs={},
            attrs={
                'endpoint': self.endpoint,
                'Fanin': self.fan_in,
T
typhoonzero 已提交
172
                'OptimizeBlock': current_block,
173 174
                'PrefetchBlock': empty_block,
                'sync_mode': True,  # did not support async now in layers
Q
qiaolongfei 已提交
175
                'grad_to_block_id': [""]
T
typhoonzero 已提交
176 177 178
            })


T
typhoonzero 已提交
179
def Send(endpoints, send_vars, get_vars=None):
T
typhoonzero 已提交
180 181 182 183 184 185 186 187 188 189 190 191 192 193 194
    """
    Send layer

    Args:
        endpoints: comma seperated IP:PORT pairs in the order
                   of send_vars to send
        send_vars: vars to send
        get_vars: vars to get from server after send completes.

    Send variables to the server side, and get vars from server
    side when server have finished running server side program.
    """
    assert (type(send_vars) == list)

    epmap = endpoints.split(",")
T
typhoonzero 已提交
195
    endpoints = list(set(epmap))
T
typhoonzero 已提交
196 197

    helper = LayerHelper("Send", **locals())
T
typhoonzero 已提交
198 199 200 201 202
    if not get_vars:
        get_vars = []
        for s in send_vars:
            v = helper.create_tmp_variable(dtype=s.dtype, stop_gradient=True)
            get_vars.append(v)
Y
Yancey1989 已提交
203
    rpc_op_role_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
Y
Yancey1989 已提交
204

T
typhoonzero 已提交
205 206 207
    helper.append_op(
        type="send",
        inputs={"X": send_vars},
Y
Yancey1989 已提交
208 209 210 211 212 213 214
        outputs={"Out": get_vars},
        attrs={
            "endpoints": endpoints,
            "epmap": epmap,
            rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC
        })

T
typhoonzero 已提交
215
    return get_vars
216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243


def Recv(endpoints, get_vars):
    """
    Recv layer

    Args:
        endpoints: comma seperated IP:PORT pairs in the order
                   of send_vars to send
        send_vars: vars to send
        get_vars: vars to get from server after send completes.

    Send variables to the server side, and get vars from server
    side when server have finished running server side program.
    """
    assert (type(send_vars) == list)
    assert (type(get_vars) == list)

    epmap = endpoints.split(",")
    endpoints = list(set(epmap))

    helper = LayerHelper("Recv", **locals())
    helper.append_op(
        type="recv",
        inputs={"X": get_vars},
        outputs={"Out": get_vars},
        attrs={"endpoints": endpoints,
               "epmap": epmap})
Y
Yu Yang 已提交
244 245


Y
Refine  
Yu Yang 已提交
246 247 248 249 250 251 252 253 254 255
def monkey_patch_reader_methods(reader):
    def __get_reader__():
        scope = global_scope()
        var = scope.find_var(reader.name)
        return var.get_reader()

    def reset():
        return __get_reader__().reset()

    reader.reset = reset
Y
Yu Yang 已提交
256 257
    reader.stop_gradient = True
    reader.persistable = True
Y
Refine  
Yu Yang 已提交
258 259 260
    return reader


Y
Yu Yang 已提交
261 262 263 264 265
def _copy_reader_var_(block, var):
    new_var = block.create_var(name=var.name, type=core.VarDesc.VarType.READER)
    new_var.desc.set_shapes(var.desc.shapes())
    new_var.desc.set_dtypes(var.desc.dtypes())
    new_var.persistable = True
F
fengjiayi 已提交
266 267 268 269
    return new_var


def _copy_reader_create_op_(block, op):
F
fengjiayi 已提交
270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285
    input_param_names = op.input_names
    new_input_map = {}
    for param_name in input_param_names:
        new_input_map[param_name] = []
        arg_names = op.input(param_name)
        for arg_name in arg_names:
            new_input_map[param_name].append(block.var(arg_name))

    output_param_names = op.output_names
    new_output_map = {}
    for param_name in output_param_names:
        new_output_map[param_name] = []
        arg_names = op.output(param_name)
        for arg_name in arg_names:
            new_output_map[param_name].append(block.var(arg_name))

F
fengjiayi 已提交
286
    new_op = block.append_op(
F
fengjiayi 已提交
287 288 289
        type=op.type,
        inputs=new_input_map,
        outputs=new_output_map,
J
JiayiFeng 已提交
290
        attrs=op.all_attrs())
F
fengjiayi 已提交
291
    return new_op
Y
Yu Yang 已提交
292 293


F
fengjiayi 已提交
294 295 296 297 298
def open_recordio_file(filename,
                       shapes,
                       lod_levels,
                       dtypes,
                       pass_num=1,
F
fengjiayi 已提交
299
                       for_parallel=True):
F
fengjiayi 已提交
300 301 302 303 304 305 306 307 308 309 310
    """
    Open a RecordIO file

    This layer takes a RecordIO file to read from and returns a Reader Variable.
    Via the Reader Variable, we can get data from the given RecordIO file.

    Args:
       filename(str): The RecordIO file's name.
       shapes(list): List of tuples which declaring data shapes.
       lod_levels(list): List of ints which declaring data lod_level.
       dtypes(list): List of strs which declaring data type.
F
fengjiayi 已提交
311
       pass_num(int): Number of passes to run.
F
fengjiayi 已提交
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
       for_parallel(Bool): Set it as True if you are going to run
            subsequent operators in parallel.

    Returns:
       Variable: A Reader Variable via which we can get RecordIO file data.

    Examples:
       .. code-block:: python

         reader = fluid.layers.io.open_recordio_file(
                                          filename='./data.recordio',
                                          shapes=[(3,224,224), (1)],
                                          lod_levels=[0, 0],
                                          dtypes=['float32', 'int64'])

         # Via the reader, we can use 'read_file' layer to get data:
F
fengjiayi 已提交
328
         image, label = fluid.layers.io.read_file(reader)
F
fengjiayi 已提交
329
    """
Y
Yu Yang 已提交
330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353
    dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes]
    shape_concat = []
    ranks = []

    for shape in shapes:
        shape_concat.extend(shape)
        ranks.append(len(shape))

    var_name = unique_name('open_recordio_file')

    startup_blk = default_startup_program().current_block()
    startup_var = startup_blk.create_var(name=var_name)
    startup_blk.append_op(
        type='create_recordio_file_reader',
        outputs={'Out': [startup_var]},
        attrs={
            'shape_concat': shape_concat,
            'lod_levels': lod_levels,
            'filename': filename,
            'ranks': ranks
        })

    startup_var.desc.set_dtypes(dtypes)
    startup_var.persistable = True
F
fengjiayi 已提交
354 355
    main_prog_var = _copy_reader_var_(default_main_program().current_block(),
                                      startup_var)
F
fengjiayi 已提交
356 357 358 359 360

    if pass_num > 1:
        main_prog_var = multi_pass(reader=main_prog_var, pass_num=pass_num)

    if for_parallel:
J
JiayiFeng 已提交
361
        main_prog_var = parallel(reader=main_prog_var)
F
fengjiayi 已提交
362

F
fengjiayi 已提交
363
    return monkey_patch_reader_methods(main_prog_var)
Y
Yu Yang 已提交
364 365


F
fengjiayi 已提交
366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
def random_data_generator(low, high, shapes, lod_levels, for_parallel=True):
    """
    Create a uniform random data generator

    This layer returns a Reader Variable.
    Instead of opening a file and reading data from it, this 
    Reader Variable generates float uniform random data by itself. 
    It can be used as a dummy reader to test a network without 
    opening a real file.

    Args:
       low(float): The lower bound of data's uniform distribution.
       high(float): The upper bound of data's uniform distribution.
       shapes(list): List of tuples which declaring data shapes.
       lod_levels(list): List of ints which declaring data lod_level.
       for_parallel(Bool): Set it as True if you are going to run
            subsequent operators in parallel.

    Returns:
       Variable: A Reader Variable from which we can get random data.

    Examples:
       .. code-block:: python

         reader = fluid.layers.io.random_data_generator(
                                          low=0.0,
                                          high=1.0,
                                          shapes=[(3,224,224), (1)],
                                          lod_levels=[0, 0])

         # Via the reader, we can use 'read_file' layer to get data:
         image, label = fluid.layers.io.read_file(reader)
    """
    dtypes = [core.VarDesc.VarType.FP32] * len(shapes)
    shape_concat = []
    ranks = []

    for shape in shapes:
        shape_concat.extend(shape)
        ranks.append(len(shape))

    var_name = unique_name('random_data_generator')

    startup_blk = default_startup_program().current_block()
    startup_var = startup_blk.create_var(name=var_name)
    startup_blk.append_op(
        type='create_random_data_generator',
        outputs={'Out': [startup_var]},
        attrs={
            'low': low,
            'high': high,
            'shape_concat': shape_concat,
            'lod_levels': lod_levels,
            'ranks': ranks
        })

    startup_var.desc.set_dtypes(dtypes)
    startup_var.persistable = True
    main_prog_var = _copy_reader_var_(default_main_program().current_block(),
                                      startup_var)

    if for_parallel:
        main_prog_var = parallel(reader=main_prog_var)

    return monkey_patch_reader_methods(main_prog_var)


433 434 435 436
def open_files(filenames,
               shapes,
               lod_levels,
               dtypes,
Y
yi.wu 已提交
437
               thread_num=1,
F
fengjiayi 已提交
438 439
               buffer_size=None,
               pass_num=1,
F
fengjiayi 已提交
440
               for_parallel=True):
F
fengjiayi 已提交
441 442 443
    """
    Open files

F
fengjiayi 已提交
444 445 446
    This layer takes a list of files to read from and returns a Reader Variable. 
    Via the Reader Variable, we can get data from given files. All files must 
    have name suffixs to indicate their formats, e.g., '*.recordio'. 
F
fengjiayi 已提交
447 448 449 450 451 452 453 454

    Args:
       filenames(list): The list of file names.
       shapes(list): List of tuples which declaring data shapes.
       lod_levels(list): List of ints which declaring data lod_level.
       dtypes(list): List of strs which declaring data type.
       thread_num(int): The maximal concurrent prefetch thread number.
       buffer_size(int): The size of prefetch buffer.
F
fengjiayi 已提交
455
       pass_num(int): Number of passes to run.
F
fengjiayi 已提交
456 457
       for_parallel(Bool): Set it as True if you are going to run 
            subsequent operators in parallel.
F
fengjiayi 已提交
458 459 460 461 462 463 464

    Returns:
       Variable: A Reader Variable via which we can get file data.

    Examples:
       .. code-block:: python

F
fengjiayi 已提交
465
         reader = fluid.layers.io.open_files(filenames=['./data1.recordio',
F
fengjiayi 已提交
466
                                                     './data2.recordio'],
F
fengjiayi 已提交
467 468 469 470 471
                                             shapes=[(3,224,224), (1)],
                                             lod_levels=[0, 0],
                                             dtypes=['float32', 'int64'],
                                             thread_num=2,
                                             buffer_size=2)
F
fengjiayi 已提交
472 473

         # Via the reader, we can use 'read_file' layer to get data:
F
fengjiayi 已提交
474
         image, label = fluid.layers.io.read_file(reader)
F
fengjiayi 已提交
475
    """
476 477
    if buffer_size is None:
        buffer_size = thread_num
F
fengjiayi 已提交
478 479
    if isinstance(filenames, basestring):
        filenames = [filenames]
F
fengjiayi 已提交
480 481 482 483 484 485 486 487
    dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes]
    shape_concat = []
    ranks = []

    for shape in shapes:
        shape_concat.extend(shape)
        ranks.append(len(shape))

F
fengjiayi 已提交
488
    multi_file_reader_name = unique_name('multi_file_reader')
F
fengjiayi 已提交
489
    startup_blk = default_startup_program().current_block()
F
fengjiayi 已提交
490
    startup_reader = startup_blk.create_var(name=multi_file_reader_name)
F
fengjiayi 已提交
491 492
    startup_blk.append_op(
        type='open_files',
F
fengjiayi 已提交
493
        outputs={'Out': [startup_reader]},
F
fengjiayi 已提交
494 495 496 497
        attrs={
            'shape_concat': shape_concat,
            'lod_levels': lod_levels,
            'ranks': ranks,
F
fengjiayi 已提交
498
            'file_names': filenames,
499 500
            'thread_num': thread_num,
            'buffer_size': buffer_size
F
fengjiayi 已提交
501 502
        })

F
fengjiayi 已提交
503 504 505 506 507 508 509
    startup_reader.desc.set_dtypes(dtypes)
    startup_reader.persistable = True
    main_prog_reader = _copy_reader_var_(default_main_program().current_block(),
                                         startup_reader)
    if pass_num > 1:
        main_prog_reader = multi_pass(
            reader=main_prog_reader, pass_num=pass_num)
F
fengjiayi 已提交
510

F
fengjiayi 已提交
511
    if for_parallel:
J
JiayiFeng 已提交
512
        main_prog_reader = parallel(reader=main_prog_reader)
F
fengjiayi 已提交
513

F
fengjiayi 已提交
514 515 516
    return monkey_patch_reader_methods(main_prog_reader)


J
JiayiFeng 已提交
517
def __create_shared_decorated_reader__(op_type, reader, attrs):
Y
Yu Yang 已提交
518 519 520
    var_name = unique_name(op_type)
    startup_blk = default_startup_program().current_block()
    startup_var = startup_blk.create_var(name=var_name)
F
fengjiayi 已提交
521
    startop_op = startup_blk.append_op(
Y
Yu Yang 已提交
522 523 524 525 526
        type=op_type,
        inputs={'UnderlyingReader': reader},
        outputs={'Out': [startup_var]},
        attrs=attrs)
    startup_var.persistable = True
F
fengjiayi 已提交
527 528 529 530
    main_prog_block = default_main_program().current_block()
    main_prog_var = _copy_reader_var_(main_prog_block, startup_var)
    _copy_reader_create_op_(main_prog_block, startop_op)
    return monkey_patch_reader_methods(main_prog_var)
Y
Yu Yang 已提交
531 532


533 534
def __create_unshared_decorated_reader__(op_type, reader, attrs, name=None):
    new_reader_name = name if name is not None else unique_name(op_type)
535 536 537 538 539 540 541 542 543 544
    main_blk = default_main_program().current_block()
    new_reader = main_blk.create_var(name=new_reader_name)
    main_blk.append_op(
        type=op_type,
        inputs={'UnderlyingReader': reader},
        outputs={'Out': [new_reader]},
        attrs=attrs)
    return monkey_patch_reader_methods(new_reader)


F
fengjiayi 已提交
545
def shuffle(reader, buffer_size):
546 547
    return __create_unshared_decorated_reader__(
        'create_shuffle_reader', reader, {'buffer_size': int(buffer_size)})
Y
Yu Yang 已提交
548 549


J
JiayiFeng 已提交
550 551 552 553 554
def batch(reader, batch_size):
    return __create_unshared_decorated_reader__(
        'create_batch_reader', reader, {'batch_size': int(batch_size)})


555
def double_buffer(reader, place=None, name=None):
Y
Yu Yang 已提交
556 557 558
    attrs = dict()
    if place is not None:
        attrs['place'] = str(place).upper()
559 560
    return __create_unshared_decorated_reader__(
        'create_double_buffer_reader', reader, attrs, name=name)
Y
Yu Yang 已提交
561 562


F
fengjiayi 已提交
563
def multi_pass(reader, pass_num):
564 565
    return __create_shared_decorated_reader__(
        'create_multi_pass_reader', reader, {'pass_num': int(pass_num)})
F
fengjiayi 已提交
566 567


J
JiayiFeng 已提交
568
def parallel(reader):
J
JiayiFeng 已提交
569 570
    return __create_shared_decorated_reader__('create_threaded_reader', reader,
                                              {})
F
fengjiayi 已提交
571 572


Y
Yu Yang 已提交
573 574 575 576 577
def read_file(file_obj):
    helper = LayerHelper('read_file')
    out = [
        helper.create_tmp_variable(
            stop_gradient=True, dtype='float32')
Y
Yu Yang 已提交
578
        for _ in range(len(file_obj.desc.shapes()))
Y
Yu Yang 已提交
579 580 581 582 583 584 585
    ]
    helper.append_op(
        type='read', inputs={'Reader': [file_obj]}, outputs={'Out': out})
    if len(out) == 1:
        return out[0]
    else:
        return out
F
fengjiayi 已提交
586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630


class Preprocessor(object):
    BEFORE_SUB_BLOCK = 0
    IN_SUB_BLOCK = 1
    AFTER_SUB_BLOCK = 2

    def __init__(self, reader, name=None):
        self.underlying_reader = reader
        new_reader_name = name if name is not None else unique_name(
            "create_custom_reader")
        self.main_prog = default_main_program()
        self.reader = self.main_prog.current_block().create_var(
            name=new_reader_name)
        self.sub_block = None
        self.source_var_names = None
        self.sink_var_names = None
        self.status = Preprocessor.BEFORE_SUB_BLOCK

    def is_completed(self):
        return self.sub_block and self.source_var_names and self.sink_var_names

    @contextlib.contextmanager
    def block(self):
        self.status = Preprocessor.IN_SUB_BLOCK
        self.sub_block = self.main_prog.create_block()
        yield
        self.main_prog.rollback()
        self.status = Preprocessor.AFTER_SUB_BLOCK
        if not self.is_completed():
            raise RuntimeError(
                "The definition of preprocessor is incompleted! "
                "Please make sure that you have set input and output "
                "variables by invoking 'inputs' and 'outputs' in "
                "Preprocessor's sub-block.")

    def inputs(self):
        if self.status != Preprocessor.IN_SUB_BLOCK:
            raise RuntimeError(
                "Preprocessor.inputs() can only be invoked inside the sub-block."
            )

        source_shapes = self.underlying_reader.desc.shapes()
        source_dtypes = self.underlying_reader.desc.dtypes()
        source_lod_levels = self.underlying_reader.desc.lod_levels()
F
fengjiayi 已提交
631 632 633 634
        self.source_var_names = [
            unique_name("preprocessor_source")
            for _ in xrange(len(source_shapes))
        ]
F
fengjiayi 已提交
635
        source_vars = []
F
fengjiayi 已提交
636 637 638
        for var_name, shape, dtype, lod_level in zip(
                self.source_var_names, source_shapes, source_dtypes,
                source_lod_levels):
F
fengjiayi 已提交
639
            source_vars.append(self.main_prog.current_block().create_var(
F
fengjiayi 已提交
640
                name=var_name, shape=shape, dtype=dtype, lod_level=lod_level))
F
fengjiayi 已提交
641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664
        return source_vars

    def outputs(self, *outs):
        if self.status != Preprocessor.IN_SUB_BLOCK:
            raise RuntimeError(
                "Preprocessor.outputs() can only be invoked inside the sub-block."
            )
        self.sink_var_names = [var.name for var in outs]

    def __call__(self, *args, **kwargs):
        if self.status != Preprocessor.AFTER_SUB_BLOCK:
            raise RuntimeError(
                "Preprocessor output can only be retrieved after rnn block.")

        self.main_prog.current_block().append_op(
            type="create_custom_reader",
            inputs={'UnderlyingReader': self.underlying_reader},
            outputs={'Out': [self.reader]},
            attrs={
                "sub_block": self.sub_block,
                "source_var_names": self.source_var_names,
                "sink_var_names": self.sink_var_names
            })
        return monkey_patch_reader_methods(self.reader)