io.py 31.5 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
15
import multiprocessing
D
dzhwinter 已提交
16

T
WIP  
typhoonzero 已提交
17
from control_flow import BlockGuard
Y
yuyang18 已提交
18 19
from layer_function_generator import templatedoc
from .. import core
Y
Refine  
Yu Yang 已提交
20
from ..executor import global_scope
Y
yuyang18 已提交
21 22 23 24
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
25

Y
Yu Yang 已提交
26
__all__ = [
S
sneaxiy 已提交
27 28 29 30
    '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
31
]
Y
Yu Yang 已提交
32 33 34 35 36 37 38 39 40 41


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
42
    **Data Layer**
Y
Yu Yang 已提交
43

K
kavyasrinet 已提交
44
    This function takes in the input and based on whether data has
C
caoying03 已提交
45
    to be returned back as a minibatch, it creates the global variable by using
Y
Yu Yang 已提交
46
    the helper functions. The global variables can be accessed by all the
C
caoying03 已提交
47
    following operators in the graph.
Y
Yu Yang 已提交
48 49 50 51

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

K
kavyasrinet 已提交
52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
    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
68 69 70 71 72 73 74 75 76 77 78 79 80
    """
    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
81
    data_var = helper.create_global_variable(
Y
Yu Yang 已提交
82 83 84 85 86
        name=name,
        shape=shape,
        dtype=dtype,
        type=type,
        stop_gradient=stop_gradient,
F
fengjiayi 已提交
87 88
        lod_level=lod_level,
        is_data=True)
Y
Yu Yang 已提交
89
    return data_var
T
typhoonzero 已提交
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


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
115
    **ListenAndServ Layer**
T
typhoonzero 已提交
116

Y
yi.wu 已提交
117 118 119 120 121 122 123 124 125
    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
126

Y
yi.wu 已提交
127 128 129 130 131 132 133 134 135 136 137 138 139 140 141
    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
142 143
            exe = fluid.Executor(place)
            exe.run(main)
T
typhoonzero 已提交
144 145
    """

Y
Yancey1989 已提交
146
    def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True):
147
        self.helper = LayerHelper("listen_and_serv")
Y
Yancey1989 已提交
148
        self.inputs = inputs
T
typhoonzero 已提交
149 150 151
        self.outputs = []
        self.endpoint = endpoint
        self.fan_in = fan_in
T
typhoonzero 已提交
152 153
        # FIXME(typhoonzero): add optimizer_mode is stupid, should make it more
        # general.
T
WIP  
typhoonzero 已提交
154
        self.optimizer_mode = optimizer_mode
T
typhoonzero 已提交
155 156 157 158 159 160 161 162 163 164 165 166 167

    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
168 169 170 171 172 173 174 175
            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
176 177
                        params.append(parent_block.var(in_var_name))
                        grads.append(parent_block.var(in_var_name))
T
typhoonzero 已提交
178 179 180

        return params, grads

T
typhoonzero 已提交
181 182 183 184 185 186 187
    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
188 189 190 191 192 193
    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(
194
            type='listen_and_serv',
Y
Yancey1989 已提交
195
            inputs={"X": self.inputs},
T
typhoonzero 已提交
196 197 198 199
            outputs={},
            attrs={
                'endpoint': self.endpoint,
                'Fanin': self.fan_in,
Y
Yancey1989 已提交
200 201 202
                'optimize_blocks': [
                    current_block
                ],  # did not support multiple optimize blocks in layers
203
                'sync_mode': True,  # did not support async now in layers
Q
qiaolongfei 已提交
204
                'grad_to_block_id': [""]
T
typhoonzero 已提交
205 206 207
            })


Y
yi.wu 已提交
208
def Send(endpoints, send_vars, sync=True):
T
typhoonzero 已提交
209
    """
Y
yi.wu 已提交
210 211
    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
212 213

    Args:
Y
yi.wu 已提交
214
        endpoints (str): comma seperated IP:PORT pairs in the order
T
typhoonzero 已提交
215
                   of send_vars to send
Y
yi.wu 已提交
216 217
        send_vars (list): variables to send to server
        sync (bool): whether to wait the request finish
T
typhoonzero 已提交
218 219 220 221 222

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

    epmap = endpoints.split(",")
T
typhoonzero 已提交
223
    endpoints = list(set(epmap))
T
typhoonzero 已提交
224 225

    helper = LayerHelper("Send", **locals())
Y
Yancey1989 已提交
226
    rpc_op_role_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
Y
Yancey1989 已提交
227

T
typhoonzero 已提交
228 229 230
    helper.append_op(
        type="send",
        inputs={"X": send_vars},
Y
Yancey1989 已提交
231 232 233 234 235
        attrs={
            "endpoints": endpoints,
            "epmap": epmap,
            rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC
        })
Y
yi.wu 已提交
236 237
    if sync:
        helper.append_op(type="send_barrier", attrs={"endpoints": endpoints})
238 239


Y
yi.wu 已提交
240
def Recv(endpoints, get_vars, sync=True):
241
    """
Y
yi.wu 已提交
242
    Receive variables from server side
243 244

    Args:
Y
yi.wu 已提交
245
        endpoints (str): comma seperated IP:PORT pairs in the order
246
                   of send_vars to send
Y
yi.wu 已提交
247 248
        get_vars (list): vars to get from server after send completes.
        sync (bool): whether to wait the request finish
249

Y
yi.wu 已提交
250 251
    Returns:
        list: list of received variables
252 253 254 255 256 257 258 259 260 261 262 263 264
    """
    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
265 266 267
    if sync:
        helper.append_op(type="fetch_barrier", attrs={"endpoints": endpoints})
    return get_vars
Y
Yu Yang 已提交
268 269


Y
Refine  
Yu Yang 已提交
270 271 272 273 274 275 276 277 278 279
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
280 281
    reader.stop_gradient = True
    reader.persistable = True
Y
Refine  
Yu Yang 已提交
282 283 284
    return reader


Y
Yu Yang 已提交
285 286 287 288 289
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
290 291 292 293
    return new_var


def _copy_reader_create_op_(block, op):
F
fengjiayi 已提交
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
    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
310
    new_op = block.append_op(
F
fengjiayi 已提交
311 312 313
        type=op.type,
        inputs=new_input_map,
        outputs=new_output_map,
J
JiayiFeng 已提交
314
        attrs=op.all_attrs())
F
fengjiayi 已提交
315
    return new_op
Y
Yu Yang 已提交
316 317


Y
yuyang18 已提交
318
@templatedoc(op_type='create_recordio_file_reader')
F
fengjiayi 已提交
319 320 321 322 323
def open_recordio_file(filename,
                       shapes,
                       lod_levels,
                       dtypes,
                       pass_num=1,
F
fengjiayi 已提交
324
                       for_parallel=True):
F
fengjiayi 已提交
325
    """
Y
yuyang18 已提交
326
    ${comment}
F
fengjiayi 已提交
327 328

    Args:
Y
yuyang18 已提交
329
       filename(${filename_type}): ${filename_comment}.
F
fengjiayi 已提交
330
       shapes(list): List of tuples which declaring data shapes.
Y
yuyang18 已提交
331
       lod_levels(${lod_levels_type}): ${lod_levels_comment}.
F
fengjiayi 已提交
332
       dtypes(list): List of strs which declaring data type.
F
fengjiayi 已提交
333
       pass_num(int): Number of passes to run.
F
fengjiayi 已提交
334 335 336 337
       for_parallel(Bool): Set it as True if you are going to run
            subsequent operators in parallel.

    Returns:
Y
yuyang18 已提交
338
       ${out_comment}.
F
fengjiayi 已提交
339 340 341

    Examples:

Y
yuyang18 已提交
342 343 344 345 346 347 348 349
        >>> 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
350
    """
Y
Yu Yang 已提交
351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374
    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
375 376
    main_prog_var = _copy_reader_var_(default_main_program().current_block(),
                                      startup_var)
F
fengjiayi 已提交
377 378 379 380

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

F
fengjiayi 已提交
381
    return monkey_patch_reader_methods(main_prog_var)
Y
Yu Yang 已提交
382 383


F
fengjiayi 已提交
384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406
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:

407
        .. code-block:: python
F
fengjiayi 已提交
408

409 410 411 412 413 414 415
            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
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450
    """
    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
451 452 453 454 455 456
def py_reader(capacity,
              shapes,
              dtypes,
              lod_levels=None,
              name=None,
              use_double_buffer=True):
S
sneaxiy 已提交
457 458
    """
    Create a reader and blocking queue for data feeding in Python
S
sneaxiy 已提交
459
    
S
sneaxiy 已提交
460
    This layer returns a Reader Variable and a BlockingQueue.
S
sneaxiy 已提交
461 462 463 464 465
    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 
466
    using `close()` method when unused.
S
sneaxiy 已提交
467 468

    Args:
Y
yuyang18 已提交
469
       use_double_buffer(bool): Whether use double buffer or not.
S
sneaxiy 已提交
470
       capacity(int): The maximum capacity of the BlockingQueue.
Y
yuyang18 已提交
471 472 473 474 475
       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
476 477

    Returns:
S
sneaxiy 已提交
478 479 480 481
       tuple(Variable, BlockingQueue):
       A Reader Variable from which we can get feeding data.
       
       A BlockingQueue object for data feeding.
S
sneaxiy 已提交
482 483 484 485 486 487 488 489 490 491 492

    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
493
            
S
sneaxiy 已提交
494 495 496 497 498 499 500 501 502
            # 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))
503
            thread.start()
S
sneaxiy 已提交
504 505 506 507 508 509 510 511 512
    """
    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))

513 514 515
    if lod_levels is None:
        lod_levels = [0] * len(shapes)

Y
yuyang18 已提交
516 517 518
    if name is None:
        queue_name = unique_name('lod_tensor_blocking_queue')
        reader_name = unique_name('create_py_reader')
Y
yuyang18 已提交
519
        double_buffer_name = unique_name('double_buffer')
Y
yuyang18 已提交
520 521 522
    else:
        queue_name = "_".join([name, "queue"])
        reader_name = "_".join([name, "reader"])
Y
yuyang18 已提交
523
        double_buffer_name = "_".join([name, "double_buffer"])
Y
yuyang18 已提交
524

S
sneaxiy 已提交
525 526 527 528
    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
529
    startup_var = startup_blk.create_var(name=reader_name)
S
sneaxiy 已提交
530 531
    startup_blk.append_op(
        type='create_py_reader',
Y
yuyang18 已提交
532
        inputs={'blocking_queue': [queue_name]},
S
sneaxiy 已提交
533 534 535 536 537 538 539 540 541 542 543 544 545
        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)

Y
yuyang18 已提交
546 547 548 549
    reader = monkey_patch_reader_methods(main_prog_var)
    if use_double_buffer:
        reader = double_buffer(reader, name=double_buffer_name)
    return reader, feed_queue
S
sneaxiy 已提交
550 551


552 553 554 555
def open_files(filenames,
               shapes,
               lod_levels,
               dtypes,
Y
yuyang18 已提交
556
               thread_num=None,
F
fengjiayi 已提交
557 558
               buffer_size=None,
               pass_num=1,
Y
yuyang18 已提交
559
               is_test=None):
F
fengjiayi 已提交
560 561 562
    """
    Open files

F
fengjiayi 已提交
563 564 565
    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
566 567 568 569 570 571

    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.
Y
yuyang18 已提交
572 573 574
       thread_num(None): The number of thread to read files.
            Default: min(len(filenames), cpu_number).
       buffer_size(None): The buffer size of reader. Default: 3 * thread_num
F
fengjiayi 已提交
575
       pass_num(int): Number of passes to run.
Y
yuyang18 已提交
576 577 578 579
       is_test(bool|None): Whether `open_files` used for testing or not. If it
            is used for testing, the order of data generated is same as the file
            order. Otherwise, it is not guaranteed the order of data is same
            between every epoch. [Default: False].
F
fengjiayi 已提交
580 581 582 583 584 585 586

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

    Examples:
       .. code-block:: python

F
fengjiayi 已提交
587
         reader = fluid.layers.io.open_files(filenames=['./data1.recordio',
F
fengjiayi 已提交
588
                                                     './data2.recordio'],
F
fengjiayi 已提交
589 590
                                             shapes=[(3,224,224), (1)],
                                             lod_levels=[0, 0],
Y
yuyang18 已提交
591
                                             dtypes=['float32', 'int64'])
F
fengjiayi 已提交
592 593

         # Via the reader, we can use 'read_file' layer to get data:
F
fengjiayi 已提交
594
         image, label = fluid.layers.io.read_file(reader)
F
fengjiayi 已提交
595
    """
Y
yuyang18 已提交
596 597 598 599 600 601 602 603 604
    if thread_num is None:
        thread_num = min(len(filenames), multiprocessing.cpu_count())
    else:
        thread_num = int(thread_num)

    if buffer_size is None:
        buffer_size = 3 * thread_num
    else:
        buffer_size = int(buffer_size)
Y
yuyang18 已提交
605

F
fengjiayi 已提交
606 607
    if isinstance(filenames, basestring):
        filenames = [filenames]
F
fengjiayi 已提交
608 609 610 611 612 613 614 615
    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
616
    multi_file_reader_name = unique_name('multi_file_reader')
F
fengjiayi 已提交
617
    startup_blk = default_startup_program().current_block()
F
fengjiayi 已提交
618
    startup_reader = startup_blk.create_var(name=multi_file_reader_name)
Y
yuyang18 已提交
619 620 621 622
    attrs = {
        'shape_concat': shape_concat,
        'lod_levels': lod_levels,
        'ranks': ranks,
Y
yuyang18 已提交
623 624 625
        'file_names': filenames,
        'thread_num': thread_num,
        'buffer_size': buffer_size
Y
yuyang18 已提交
626 627 628
    }
    if is_test is not None:
        attrs['is_test'] = is_test
F
fengjiayi 已提交
629
    startup_blk.append_op(
Y
yuyang18 已提交
630
        type='open_files', outputs={'Out': [startup_reader]}, attrs=attrs)
F
fengjiayi 已提交
631

F
fengjiayi 已提交
632 633 634 635 636 637 638
    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
639

F
fengjiayi 已提交
640 641 642
    return monkey_patch_reader_methods(main_prog_reader)


J
JiayiFeng 已提交
643
def __create_shared_decorated_reader__(op_type, reader, attrs):
Y
Yu Yang 已提交
644 645 646
    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
647
    startop_op = startup_blk.append_op(
Y
Yu Yang 已提交
648 649 650 651 652
        type=op_type,
        inputs={'UnderlyingReader': reader},
        outputs={'Out': [startup_var]},
        attrs=attrs)
    startup_var.persistable = True
F
fengjiayi 已提交
653 654 655 656
    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
657 658


659 660
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)
661 662 663 664 665 666 667 668 669 670
    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
671
def shuffle(reader, buffer_size):
672 673 674
    """
    Shuffle the reader.
    """
675 676
    return __create_unshared_decorated_reader__(
        'create_shuffle_reader', reader, {'buffer_size': int(buffer_size)})
Y
Yu Yang 已提交
677 678


J
JiayiFeng 已提交
679
def batch(reader, batch_size):
F
fengjiayi 已提交
680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714
    """
    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
715 716 717 718
    return __create_unshared_decorated_reader__(
        'create_batch_reader', reader, {'batch_size': int(batch_size)})


719
def double_buffer(reader, place=None, name=None):
Y
yuyang18 已提交
720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742
    """
    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
743 744 745
    attrs = dict()
    if place is not None:
        attrs['place'] = str(place).upper()
746 747
    return __create_unshared_decorated_reader__(
        'create_double_buffer_reader', reader, attrs, name=name)
Y
Yu Yang 已提交
748 749


F
fengjiayi 已提交
750
def multi_pass(reader, pass_num):
751 752
    return __create_shared_decorated_reader__(
        'create_multi_pass_reader', reader, {'pass_num': int(pass_num)})
F
fengjiayi 已提交
753 754


F
fengjiayi 已提交
755
def read_file(reader):
F
fengjiayi 已提交
756
    """
F
fengjiayi 已提交
757
    Execute the given reader and get data via it.
F
fengjiayi 已提交
758

F
fengjiayi 已提交
759
    A reader is also a Variable. It can be a raw reader generated by 
F
fengjiayi 已提交
760 761 762 763 764
    `fluid.layers.open_files()` or a decorated one generated by 
    `fluid.layers.double_buffer()` and so on.

    Args:

F
fengjiayi 已提交
765
        reader(Variable): The reader to execute.
F
fengjiayi 已提交
766 767

    Returns:
F
fengjiayi 已提交
768
        Tuple[Variable]: Data read via the given reader.
F
fengjiayi 已提交
769 770 771 772 773 774 775 776 777 778 779 780 781

    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
782 783 784 785
    helper = LayerHelper('read_file')
    out = [
        helper.create_tmp_variable(
            stop_gradient=True, dtype='float32')
F
fengjiayi 已提交
786
        for _ in range(len(reader.desc.shapes()))
Y
Yu Yang 已提交
787 788
    ]
    helper.append_op(
F
fengjiayi 已提交
789
        type='read', inputs={'Reader': [reader]}, outputs={'Out': out})
Y
Yu Yang 已提交
790 791 792 793
    if len(out) == 1:
        return out[0]
    else:
        return out
F
fengjiayi 已提交
794 795 796


class Preprocessor(object):
X
Xin Pan 已提交
797 798 799 800 801 802 803 804 805
    """
    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
806

X
Xin Pan 已提交
807 808 809 810 811 812 813 814 815 816
            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
817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858
    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
859 860 861 862
        self.source_var_names = [
            unique_name("preprocessor_source")
            for _ in xrange(len(source_shapes))
        ]
F
fengjiayi 已提交
863
        source_vars = []
F
fengjiayi 已提交
864 865 866
        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
867
            source_vars.append(self.main_prog.current_block().create_var(
F
fengjiayi 已提交
868
                name=var_name, shape=shape, dtype=dtype, lod_level=lod_level))
F
fengjiayi 已提交
869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892
        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
893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918


@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)