io.py 30.9 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.
F
fengjiayi 已提交
14
import contextlib
D
dzhwinter 已提交
15

T
WIP  
typhoonzero 已提交
16
from control_flow import BlockGuard
Y
yuyang18 已提交
17 18
from layer_function_generator import templatedoc
from .. import core
Y
Refine  
Yu Yang 已提交
19
from ..executor import global_scope
Y
yuyang18 已提交
20 21 22 23
from ..framework import convert_np_dtype_to_dtype_, default_main_program, \
    default_startup_program
from ..layer_helper import LayerHelper
from ..unique_name import generate as unique_name
Y
Yu Yang 已提交
24

Y
Yu Yang 已提交
25
__all__ = [
S
sneaxiy 已提交
26 27 28 29
    'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'Recv',
    'open_recordio_file', 'open_files', 'read_file', 'shuffle', 'batch',
    'double_buffer', 'random_data_generator', 'py_reader', 'Preprocessor',
    'load'
Y
Yu Yang 已提交
30
]
Y
Yu Yang 已提交
31 32 33 34 35 36 37 38 39 40


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
41
    **Data Layer**
Y
Yu Yang 已提交
42

K
kavyasrinet 已提交
43
    This function takes in the input and based on whether data has
C
caoying03 已提交
44
    to be returned back as a minibatch, it creates the global variable by using
Y
Yu Yang 已提交
45
    the helper functions. The global variables can be accessed by all the
C
caoying03 已提交
46
    following operators in the graph.
Y
Yu Yang 已提交
47 48 49 50

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

K
kavyasrinet 已提交
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66
    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
67 68 69 70 71 72 73 74 75 76 77 78 79
    """
    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
80
    data_var = helper.create_global_variable(
Y
Yu Yang 已提交
81 82 83 84 85
        name=name,
        shape=shape,
        dtype=dtype,
        type=type,
        stop_gradient=stop_gradient,
F
fengjiayi 已提交
86 87
        lod_level=lod_level,
        is_data=True)
Y
Yu Yang 已提交
88
    return data_var
T
typhoonzero 已提交
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


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):
    """
Y
yi.wu 已提交
114
    **ListenAndServ Layer**
T
typhoonzero 已提交
115

Y
yi.wu 已提交
116 117 118 119 120 121 122 123 124
    ListenAndServ is used to create a rpc server bind and listen
    on specific TCP port, this server will run the sub-block when
    received variables from clients.

    Args:
        endpoint(string): IP:port string which the server will listen on.
        inputs(list): a list of variables that the server will get from clients.
        fan_in(int): how many client are expected to report to this server, default: 1.
        optimizer_mode(bool): whether to run the server as a parameter server, default: True.
Y
update  
yi.wu 已提交
125

Y
yi.wu 已提交
126 127 128 129 130 131 132 133 134 135 136 137 138 139 140
    Examples:
        .. code-block:: python

            with fluid.program_guard(main):
                serv = layers.ListenAndServ(
                    "127.0.0.1:6170", ["X"], optimizer_mode=False)
                with serv.do():
                    x = layers.data(
                        shape=[32, 32],
                        dtype='float32',
                        name="X",
                        append_batch_size=False)
                    fluid.initializer.Constant(value=1.0)(x, main.global_block())
                    layers.scale(x=x, scale=10.0, out=out_var)

Y
yi.wu 已提交
141 142
            exe = fluid.Executor(place)
            exe.run(main)
T
typhoonzero 已提交
143 144
    """

Y
Yancey1989 已提交
145
    def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True):
146
        self.helper = LayerHelper("listen_and_serv")
Y
Yancey1989 已提交
147
        self.inputs = inputs
T
typhoonzero 已提交
148 149 150
        self.outputs = []
        self.endpoint = endpoint
        self.fan_in = fan_in
T
typhoonzero 已提交
151 152
        # FIXME(typhoonzero): add optimizer_mode is stupid, should make it more
        # general.
T
WIP  
typhoonzero 已提交
153
        self.optimizer_mode = optimizer_mode
T
typhoonzero 已提交
154 155 156 157 158 159 160 161 162 163 164 165 166

    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
167 168 169 170 171 172 173 174
            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
175 176
                        params.append(parent_block.var(in_var_name))
                        grads.append(parent_block.var(in_var_name))
T
typhoonzero 已提交
177 178 179

        return params, grads

T
typhoonzero 已提交
180 181 182 183 184 185 186
    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
187 188 189 190 191 192
    def complete_op(self):
        main_program = self.helper.main_program
        current_block = main_program.current_block()
        parent_block = self.parent_block()

        parent_block.append_op(
193
            type='listen_and_serv',
Y
Yancey1989 已提交
194
            inputs={"X": self.inputs},
T
typhoonzero 已提交
195 196 197 198
            outputs={},
            attrs={
                'endpoint': self.endpoint,
                'Fanin': self.fan_in,
Y
Yancey1989 已提交
199 200 201
                'optimize_blocks': [
                    current_block
                ],  # did not support multiple optimize blocks in layers
202
                'sync_mode': True,  # did not support async now in layers
Q
qiaolongfei 已提交
203
                'grad_to_block_id': [""]
T
typhoonzero 已提交
204 205 206
            })


Y
yi.wu 已提交
207
def Send(endpoints, send_vars, sync=True):
T
typhoonzero 已提交
208
    """
Y
yi.wu 已提交
209 210
    Send variables to the server side, and get vars from server
    side when server have finished running server side program.
T
typhoonzero 已提交
211 212

    Args:
Y
yi.wu 已提交
213
        endpoints (str): comma seperated IP:PORT pairs in the order
T
typhoonzero 已提交
214
                   of send_vars to send
Y
yi.wu 已提交
215 216
        send_vars (list): variables to send to server
        sync (bool): whether to wait the request finish
T
typhoonzero 已提交
217 218 219 220 221

    """
    assert (type(send_vars) == list)

    epmap = endpoints.split(",")
T
typhoonzero 已提交
222
    endpoints = list(set(epmap))
T
typhoonzero 已提交
223 224

    helper = LayerHelper("Send", **locals())
Y
Yancey1989 已提交
225
    rpc_op_role_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
Y
Yancey1989 已提交
226

T
typhoonzero 已提交
227 228 229
    helper.append_op(
        type="send",
        inputs={"X": send_vars},
Y
Yancey1989 已提交
230 231 232 233 234
        attrs={
            "endpoints": endpoints,
            "epmap": epmap,
            rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC
        })
Y
yi.wu 已提交
235 236
    if sync:
        helper.append_op(type="send_barrier", attrs={"endpoints": endpoints})
237 238


Y
yi.wu 已提交
239
def Recv(endpoints, get_vars, sync=True):
240
    """
Y
yi.wu 已提交
241
    Receive variables from server side
242 243

    Args:
Y
yi.wu 已提交
244
        endpoints (str): comma seperated IP:PORT pairs in the order
245
                   of send_vars to send
Y
yi.wu 已提交
246 247
        get_vars (list): vars to get from server after send completes.
        sync (bool): whether to wait the request finish
248

Y
yi.wu 已提交
249 250
    Returns:
        list: list of received variables
251 252 253 254 255 256 257 258 259 260 261 262 263
    """
    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
yi.wu 已提交
264 265 266
    if sync:
        helper.append_op(type="fetch_barrier", attrs={"endpoints": endpoints})
    return get_vars
Y
Yu Yang 已提交
267 268


Y
Refine  
Yu Yang 已提交
269 270 271 272 273 274 275 276 277 278
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
279 280
    reader.stop_gradient = True
    reader.persistable = True
Y
Refine  
Yu Yang 已提交
281 282 283
    return reader


Y
Yu Yang 已提交
284 285 286 287 288
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
289 290 291 292
    return new_var


def _copy_reader_create_op_(block, op):
F
fengjiayi 已提交
293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308
    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
309
    new_op = block.append_op(
F
fengjiayi 已提交
310 311 312
        type=op.type,
        inputs=new_input_map,
        outputs=new_output_map,
J
JiayiFeng 已提交
313
        attrs=op.all_attrs())
F
fengjiayi 已提交
314
    return new_op
Y
Yu Yang 已提交
315 316


Y
yuyang18 已提交
317
@templatedoc(op_type='create_recordio_file_reader')
F
fengjiayi 已提交
318 319 320 321 322
def open_recordio_file(filename,
                       shapes,
                       lod_levels,
                       dtypes,
                       pass_num=1,
F
fengjiayi 已提交
323
                       for_parallel=True):
F
fengjiayi 已提交
324
    """
Y
yuyang18 已提交
325
    ${comment}
F
fengjiayi 已提交
326 327

    Args:
Y
yuyang18 已提交
328
       filename(${filename_type}): ${filename_comment}.
F
fengjiayi 已提交
329
       shapes(list): List of tuples which declaring data shapes.
Y
yuyang18 已提交
330
       lod_levels(${lod_levels_type}): ${lod_levels_comment}.
F
fengjiayi 已提交
331
       dtypes(list): List of strs which declaring data type.
F
fengjiayi 已提交
332
       pass_num(int): Number of passes to run.
F
fengjiayi 已提交
333 334 335 336
       for_parallel(Bool): Set it as True if you are going to run
            subsequent operators in parallel.

    Returns:
Y
yuyang18 已提交
337
       ${out_comment}.
F
fengjiayi 已提交
338 339 340

    Examples:

Y
yuyang18 已提交
341 342 343 344 345 346 347 348
        >>> import paddle.fluid as fluid
        >>> 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:
        >>> image, label = fluid.layers.io.read_file(reader)
F
fengjiayi 已提交
349
    """
Y
Yu Yang 已提交
350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373
    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
374 375
    main_prog_var = _copy_reader_var_(default_main_program().current_block(),
                                      startup_var)
F
fengjiayi 已提交
376 377 378 379

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

F
fengjiayi 已提交
380
    return monkey_patch_reader_methods(main_prog_var)
Y
Yu Yang 已提交
381 382


F
fengjiayi 已提交
383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405
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:

406
        .. code-block:: python
F
fengjiayi 已提交
407

408 409 410 411 412 413 414
            reader = fluid.layers.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.read_file(reader)
F
fengjiayi 已提交
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
    """
    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)


Y
yuyang18 已提交
450
def py_reader(capacity, shapes, dtypes, lod_levels=None, name=None):
S
sneaxiy 已提交
451 452
    """
    Create a reader and blocking queue for data feeding in Python
S
sneaxiy 已提交
453
    
S
sneaxiy 已提交
454
    This layer returns a Reader Variable and a BlockingQueue.
S
sneaxiy 已提交
455 456 457 458 459
    The BlockingQueue provides `push()` method to push a `LoDTensorArray` 
    object into the queue in Python side. In C++ side, the Reader 
    Variable would invoke `pop()` method of the queue to retrieve the 
    feeding data. The process of feeding data in Python side and fetching 
    data in C++ side can run in parallel. The BlockingQueue should be closed 
460
    using `close()` method when unused.
S
sneaxiy 已提交
461 462 463

    Args:
       capacity(int): The maximum capacity of the BlockingQueue.
Y
yuyang18 已提交
464 465 466 467 468
       shapes(list|tuple): List of tuples which declaring data shapes.
       dtypes(list|tuple): List of strs which declaring data type.
       lod_levels(list|tuple): List of ints which declaring data lod_level.
       name(basestring): The prefix Python queue name and Reader name. None will
            be generated automatically.
S
sneaxiy 已提交
469 470

    Returns:
S
sneaxiy 已提交
471 472 473 474
       tuple(Variable, BlockingQueue):
       A Reader Variable from which we can get feeding data.
       
       A BlockingQueue object for data feeding.
S
sneaxiy 已提交
475 476 477 478 479 480 481 482 483 484 485

    Examples:

        .. code-block:: python

            reader, queue = fluid.layers.py_reader(
                                             capacity=10,
                                             shapes=[[-1,3,224,224], [-1,1]],
                                             dtypes=['float32', 'int64'])
            # Via the reader, we can use 'read_file' layer to get data:
            image, label = fluid.layers.read_file(reader)
S
sneaxiy 已提交
486
            
S
sneaxiy 已提交
487 488 489 490 491 492 493 494 495
            # Via the blocking queue, we can feed data using threads
            def feed_data(queue, feed_images, feed_labels):
                for feed_image, feed_label in zip(feed_images, feed_labels):
                    data = core.LoDTensorArray()
                    data.append(feed_image)
                    data.append(feed_label)
                    queue.push(data)
            
            thread = threading.Thread(target=feed_data, args=(queue, feed_images, feed_labels))
496
            thread.start()
S
sneaxiy 已提交
497 498 499 500 501 502 503 504 505
    """
    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))

506 507 508
    if lod_levels is None:
        lod_levels = [0] * len(shapes)

Y
yuyang18 已提交
509 510 511 512 513 514 515
    if name is None:
        queue_name = unique_name('lod_tensor_blocking_queue')
        reader_name = unique_name('create_py_reader')
    else:
        queue_name = "_".join([name, "queue"])
        reader_name = "_".join([name, "reader"])

S
sneaxiy 已提交
516 517 518 519
    var = global_scope().var(queue_name)
    feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes)

    startup_blk = default_startup_program().current_block()
Y
yuyang18 已提交
520
    startup_var = startup_blk.create_var(name=reader_name)
S
sneaxiy 已提交
521 522
    startup_blk.append_op(
        type='create_py_reader',
Y
yuyang18 已提交
523
        inputs={'blocking_queue': [queue_name]},
S
sneaxiy 已提交
524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539
        outputs={'Out': [startup_var]},
        attrs={
            '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)

    return monkey_patch_reader_methods(main_prog_var), feed_queue


540 541 542 543
def open_files(filenames,
               shapes,
               lod_levels,
               dtypes,
Y
yi.wu 已提交
544
               thread_num=1,
F
fengjiayi 已提交
545 546
               buffer_size=None,
               pass_num=1,
F
fengjiayi 已提交
547
               for_parallel=True):
F
fengjiayi 已提交
548 549 550
    """
    Open files

F
fengjiayi 已提交
551 552 553
    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
554 555 556 557 558 559 560

    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.
561 562 563
       buffer_size(int|None): The size of prefetch buffer. If it is setted None, 
            buffer size will be thread_num * 3.
            Default: None
F
fengjiayi 已提交
564
       pass_num(int): Number of passes to run.
F
fengjiayi 已提交
565 566
       for_parallel(Bool): Set it as True if you are going to run 
            subsequent operators in parallel.
567
            Default: True
F
fengjiayi 已提交
568 569 570 571 572 573 574

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

    Examples:
       .. code-block:: python

F
fengjiayi 已提交
575
         reader = fluid.layers.io.open_files(filenames=['./data1.recordio',
F
fengjiayi 已提交
576
                                                     './data2.recordio'],
F
fengjiayi 已提交
577 578 579 580 581
                                             shapes=[(3,224,224), (1)],
                                             lod_levels=[0, 0],
                                             dtypes=['float32', 'int64'],
                                             thread_num=2,
                                             buffer_size=2)
F
fengjiayi 已提交
582 583

         # Via the reader, we can use 'read_file' layer to get data:
F
fengjiayi 已提交
584
         image, label = fluid.layers.io.read_file(reader)
F
fengjiayi 已提交
585
    """
586
    if buffer_size is None:
587
        buffer_size = thread_num * 3
F
fengjiayi 已提交
588 589
    if isinstance(filenames, basestring):
        filenames = [filenames]
F
fengjiayi 已提交
590 591 592 593 594 595 596 597
    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
598
    multi_file_reader_name = unique_name('multi_file_reader')
F
fengjiayi 已提交
599
    startup_blk = default_startup_program().current_block()
F
fengjiayi 已提交
600
    startup_reader = startup_blk.create_var(name=multi_file_reader_name)
F
fengjiayi 已提交
601 602
    startup_blk.append_op(
        type='open_files',
F
fengjiayi 已提交
603
        outputs={'Out': [startup_reader]},
F
fengjiayi 已提交
604 605 606 607
        attrs={
            'shape_concat': shape_concat,
            'lod_levels': lod_levels,
            'ranks': ranks,
F
fengjiayi 已提交
608
            'file_names': filenames,
609 610
            'thread_num': thread_num,
            'buffer_size': buffer_size
F
fengjiayi 已提交
611 612
        })

F
fengjiayi 已提交
613 614 615 616 617 618 619
    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
620

F
fengjiayi 已提交
621 622 623
    return monkey_patch_reader_methods(main_prog_reader)


J
JiayiFeng 已提交
624
def __create_shared_decorated_reader__(op_type, reader, attrs):
Y
Yu Yang 已提交
625 626 627
    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
628
    startop_op = startup_blk.append_op(
Y
Yu Yang 已提交
629 630 631 632 633
        type=op_type,
        inputs={'UnderlyingReader': reader},
        outputs={'Out': [startup_var]},
        attrs=attrs)
    startup_var.persistable = True
F
fengjiayi 已提交
634 635 636 637
    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
638 639


640 641
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)
642 643 644 645 646 647 648 649 650 651
    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
652
def shuffle(reader, buffer_size):
653 654 655
    """
    Shuffle the reader.
    """
656 657
    return __create_unshared_decorated_reader__(
        'create_shuffle_reader', reader, {'buffer_size': int(buffer_size)})
Y
Yu Yang 已提交
658 659


J
JiayiFeng 已提交
660
def batch(reader, batch_size):
F
fengjiayi 已提交
661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695
    """
    This layer is a reader decorator. It takes a reader and adds 
    'batching' decoration on it. When reading with the result 
    decorated reader, output data will be automatically organized 
    to the form of batches.

    Args:
        reader(Variable): The reader to be decorated with 'batching'.
        batch_size(int): The batch size.

    Returns:
        Variable: The reader which has been decorated with 'batching'.

    Examples:
        .. code-block:: python

            raw_reader = fluid.layers.io.open_files(filenames=['./data1.recordio',
                                                           './data2.recordio'],
                                                    shapes=[(3,224,224), (1)],
                                                    lod_levels=[0, 0],
                                                    dtypes=['float32', 'int64'],
                                                    thread_num=2,
                                                    buffer_size=2)
            batch_reader = fluid.layers.batch(reader=raw_reader, batch_size=5)

            # If we read data with the raw_reader:
            #     data = fluid.layers.read_file(raw_reader)
            # We can only get data instance by instance.
            # 
            # However, if we read data with the batch_reader:
            #     data = fluid.layers.read_file(batch_reader)
            # Each 5 adjacent instances will be automatically combined together 
            # to become a batch. So what we get('data') is a batch data instead 
            # of an instance.
    """
J
JiayiFeng 已提交
696 697 698 699
    return __create_unshared_decorated_reader__(
        'create_batch_reader', reader, {'batch_size': int(batch_size)})


700
def double_buffer(reader, place=None, name=None):
Y
yuyang18 已提交
701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723
    """
    Wrap a double buffer reader. The data will copy to target place with a
    double buffer queue. If the target place is None, the place that executor
    perform on will be used.

    Args:
        reader(Variable): the reader variable need to be wrapped.
        place(Place): the place of target data. Default is the sample place of
            executor perform.

        name(str): Variable name. None if the user does not care.

    Returns:
        wrapped reader with double buffer.

    Examples:

        >>> reader = fluid.layers.open_files(filenames=['somefile'],
        >>>                                  shapes=[[-1, 784], [-1, 1]],
        >>>                                  dtypes=['float32', 'int64'])
        >>> reader = fluid.layers.double_buffer(reader)
        >>> img, label = fluid.layers.read_file(reader)
    """
Y
Yu Yang 已提交
724 725 726
    attrs = dict()
    if place is not None:
        attrs['place'] = str(place).upper()
727 728
    return __create_unshared_decorated_reader__(
        'create_double_buffer_reader', reader, attrs, name=name)
Y
Yu Yang 已提交
729 730


F
fengjiayi 已提交
731
def multi_pass(reader, pass_num):
732 733
    return __create_shared_decorated_reader__(
        'create_multi_pass_reader', reader, {'pass_num': int(pass_num)})
F
fengjiayi 已提交
734 735


F
fengjiayi 已提交
736
def read_file(reader):
F
fengjiayi 已提交
737
    """
F
fengjiayi 已提交
738
    Execute the given reader and get data via it.
F
fengjiayi 已提交
739

F
fengjiayi 已提交
740
    A reader is also a Variable. It can be a raw reader generated by 
F
fengjiayi 已提交
741 742 743 744 745
    `fluid.layers.open_files()` or a decorated one generated by 
    `fluid.layers.double_buffer()` and so on.

    Args:

F
fengjiayi 已提交
746
        reader(Variable): The reader to execute.
F
fengjiayi 已提交
747 748

    Returns:
F
fengjiayi 已提交
749
        Tuple[Variable]: Data read via the given reader.
F
fengjiayi 已提交
750 751 752 753 754 755 756 757 758 759 760 761 762

    Examples:
        .. code-block:: python

           data_file = fluid.layers.open_files(
                filenames=['mnist.recordio'],
                shapes=[(-1, 748), (-1, 1)],
                lod_levels=[0, 0],
                dtypes=["float32", "int64"])
            data_file = fluid.layers.double_buffer(
                fluid.layers.batch(data_file, batch_size=64))
            input, label = fluid.layers.read_file(data_file)
    """
Y
Yu Yang 已提交
763 764 765 766
    helper = LayerHelper('read_file')
    out = [
        helper.create_tmp_variable(
            stop_gradient=True, dtype='float32')
F
fengjiayi 已提交
767
        for _ in range(len(reader.desc.shapes()))
Y
Yu Yang 已提交
768 769
    ]
    helper.append_op(
F
fengjiayi 已提交
770
        type='read', inputs={'Reader': [reader]}, outputs={'Out': out})
Y
Yu Yang 已提交
771 772 773 774
    if len(out) == 1:
        return out[0]
    else:
        return out
F
fengjiayi 已提交
775 776 777


class Preprocessor(object):
X
Xin Pan 已提交
778 779 780 781 782 783 784 785 786
    """
    A block for data pre-processing in reader.

    Args:
        reader (Variable): A reader variable.
        name (str, default None): The name of the reader.

    Examples:
          .. code-block:: python
X
Xin Pan 已提交
787

X
Xin Pan 已提交
788 789 790 791 792 793 794 795 796 797
            preprocessor = fluid.layers.io.Preprocessor(reader=reader)
            with preprocessor.block():
                img, lbl = preprocessor.inputs()
                img_out = img / 2
                lbl_out = lbl + 1
                preprocessor.outputs(img_out, lbl_out)

            data_file = fluid.layers.io.double_buffer(preprocessor())

    """
F
fengjiayi 已提交
798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839
    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
840 841 842 843
        self.source_var_names = [
            unique_name("preprocessor_source")
            for _ in xrange(len(source_shapes))
        ]
F
fengjiayi 已提交
844
        source_vars = []
F
fengjiayi 已提交
845 846 847
        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
848
            source_vars.append(self.main_prog.current_block().create_var(
F
fengjiayi 已提交
849
                name=var_name, shape=shape, dtype=dtype, lod_level=lod_level))
F
fengjiayi 已提交
850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873
        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)
Y
yuyang18 已提交
874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899


@templatedoc()
def load(out, file_path, load_as_fp16=None):
    """
    ${comment}

    >>> import paddle.fluid as fluid
    >>> tmp_tensor = fluid.layers.create_tensor(dtype='float32')
    >>> fluid.layers.load(tmp_tensor, "./tmp_tensor.bin")

    Args:
        out(${out_type}): ${out_comment}.

        file_path(${file_path_type}): ${file_path_comment}.

        load_as_fp16(${load_as_fp16_type}): ${load_as_fp16_comment}.

    Returns:
        None
    """
    helper = LayerHelper("load", **locals())
    attrs = {"file_path": file_path}
    if load_as_fp16 is not None:
        attrs['load_as_fp16'] = load_as_fp16
    helper.append_op(type="load", inputs={}, output={"Out": out}, args=attrs)