io.py 39.4 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
14 15

from __future__ import print_function
F
fengjiayi 已提交
16
import contextlib
17
import multiprocessing
M
minqiyang 已提交
18
import six
Y
yuyang18 已提交
19
import threading
D
dzhwinter 已提交
20

Y
yuyang18 已提交
21
from ..data_feeder import DataFeeder
22 23
from .control_flow import BlockGuard
from .layer_function_generator import templatedoc
Y
yuyang18 已提交
24
from .. import core
Y
Refine  
Yu Yang 已提交
25
from ..executor import global_scope
Y
yuyang18 已提交
26
from ..framework import convert_np_dtype_to_dtype_, default_main_program, \
27
    default_startup_program, program_guard, Program, Variable
Y
yuyang18 已提交
28 29
from ..layer_helper import LayerHelper
from ..unique_name import generate as unique_name
Y
Yu Yang 已提交
30

Y
Yu Yang 已提交
31
__all__ = [
Y
yuyang 已提交
32
    'data', 'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer',
Q
qiaolongfei 已提交
33 34
    'random_data_generator', 'py_reader', 'py_reader_by_data', 'Preprocessor',
    'load'
Y
Yu Yang 已提交
35
]
Y
Yu Yang 已提交
36 37 38 39 40 41 42 43 44 45


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
46
    **Data Layer**
Y
Yu Yang 已提交
47

K
kavyasrinet 已提交
48
    This function takes in the input and based on whether data has
C
caoying03 已提交
49
    to be returned back as a minibatch, it creates the global variable by using
Y
Yu Yang 已提交
50
    the helper functions. The global variables can be accessed by all the
C
caoying03 已提交
51
    following operators in the graph.
Y
Yu Yang 已提交
52 53 54 55

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

K
kavyasrinet 已提交
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
    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
72 73 74
    """
    helper = LayerHelper('data', **locals())
    shape = list(shape)
M
minqiyang 已提交
75
    for i in six.moves.range(len(shape)):
Y
Yu Yang 已提交
76 77 78 79 80 81 82 83 84
        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
85
    data_var = helper.create_global_variable(
Y
Yu Yang 已提交
86 87 88 89 90
        name=name,
        shape=shape,
        dtype=dtype,
        type=type,
        stop_gradient=stop_gradient,
F
fengjiayi 已提交
91 92
        lod_level=lod_level,
        is_data=True)
Y
Yu Yang 已提交
93
    return data_var
T
typhoonzero 已提交
94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118


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
119
    **ListenAndServ Layer**
T
typhoonzero 已提交
120

Y
yi.wu 已提交
121 122 123 124 125 126 127 128 129
    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
130

Y
yi.wu 已提交
131 132 133 134 135 136 137 138 139 140 141 142 143 144 145
    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
146 147
            exe = fluid.Executor(place)
            exe.run(main)
T
typhoonzero 已提交
148 149
    """

Y
Yancey1989 已提交
150
    def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True):
151
        self.helper = LayerHelper("listen_and_serv")
Y
Yancey1989 已提交
152
        self.inputs = inputs
T
typhoonzero 已提交
153 154 155
        self.outputs = []
        self.endpoint = endpoint
        self.fan_in = fan_in
T
typhoonzero 已提交
156 157
        # FIXME(typhoonzero): add optimizer_mode is stupid, should make it more
        # general.
T
WIP  
typhoonzero 已提交
158
        self.optimizer_mode = optimizer_mode
T
typhoonzero 已提交
159 160 161 162 163 164 165 166 167 168 169 170 171

    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
172 173 174 175 176 177 178 179
            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
180 181
                        params.append(parent_block.var(in_var_name))
                        grads.append(parent_block.var(in_var_name))
T
typhoonzero 已提交
182 183 184

        return params, grads

T
typhoonzero 已提交
185 186 187 188 189 190 191
    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
192 193 194 195 196 197
    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(
198
            type='listen_and_serv',
Y
Yancey1989 已提交
199
            inputs={"X": self.inputs},
T
typhoonzero 已提交
200 201 202 203
            outputs={},
            attrs={
                'endpoint': self.endpoint,
                'Fanin': self.fan_in,
Y
Yancey1989 已提交
204 205 206
                'optimize_blocks': [
                    current_block
                ],  # did not support multiple optimize blocks in layers
207
                'sync_mode': True,  # did not support async now in layers
Q
qiaolongfei 已提交
208
                'grad_to_block_id': [""]
T
typhoonzero 已提交
209 210 211
            })


212
def Send(endpoints, send_vars, dummy_output=None, sync=True):
T
typhoonzero 已提交
213
    """
Y
yi.wu 已提交
214 215
    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
216 217

    Args:
Y
yi.wu 已提交
218
        endpoints (str): comma seperated IP:PORT pairs in the order
T
typhoonzero 已提交
219
                   of send_vars to send
Y
yi.wu 已提交
220 221
        send_vars (list): variables to send to server
        sync (bool): whether to wait the request finish
T
typhoonzero 已提交
222 223 224 225

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

226 227 228 229 230 231 232
    if dummy_output is None:
        dummy_output = []
    elif isinstance(dummy_output, Variable):
        dummy_output = [dummy_output]

    assert (type(dummy_output) == list)

T
typhoonzero 已提交
233
    epmap = endpoints.split(",")
T
typhoonzero 已提交
234
    endpoints = list(set(epmap))
T
typhoonzero 已提交
235 236

    helper = LayerHelper("Send", **locals())
Y
Yancey1989 已提交
237
    rpc_op_role_name = core.op_proto_and_checker_maker.kOpRoleAttrName()
Y
Yancey1989 已提交
238

T
typhoonzero 已提交
239 240 241
    helper.append_op(
        type="send",
        inputs={"X": send_vars},
242
        outputs={"Out": dummy_output},
Y
Yancey1989 已提交
243 244 245 246 247
        attrs={
            "endpoints": endpoints,
            "epmap": epmap,
            rpc_op_role_name: core.op_proto_and_checker_maker.OpRole.RPC
        })
Y
yi.wu 已提交
248
    if sync:
W
Wu Yi 已提交
249 250 251 252 253
        helper.append_op(
            type="send_barrier",
            inputs={"X": dummy_output},
            outputs={"Out": []},
            attrs={"endpoints": endpoints})
254 255


256
def Recv(endpoints, get_vars, dummy_input=None, sync=True):
257
    """
Y
yi.wu 已提交
258
    Receive variables from server side
259 260

    Args:
Y
yi.wu 已提交
261
        endpoints (str): comma seperated IP:PORT pairs in the order
262
                   of send_vars to send
Y
yi.wu 已提交
263 264
        get_vars (list): vars to get from server after send completes.
        sync (bool): whether to wait the request finish
265

Y
yi.wu 已提交
266 267
    Returns:
        list: list of received variables
268 269 270
    """
    assert (type(get_vars) == list)

271 272 273 274 275 276 277
    if dummy_input is None:
        dummy_input = []
    elif isinstance(dummy_input, Variable):
        dummy_input = [dummy_input]

    assert (type(dummy_input) == list)

278 279 280 281 282 283
    epmap = endpoints.split(",")
    endpoints = list(set(epmap))

    helper = LayerHelper("Recv", **locals())
    helper.append_op(
        type="recv",
284
        inputs={"X": dummy_input},
285 286 287
        outputs={"Out": get_vars},
        attrs={"endpoints": endpoints,
               "epmap": epmap})
Y
yi.wu 已提交
288
    if sync:
W
Wu Yi 已提交
289 290 291 292
        helper.append_op(
            type="fetch_barrier",
            outputs={"Out": get_vars},
            attrs={"endpoints": endpoints})
Y
yi.wu 已提交
293
    return get_vars
Y
Yu Yang 已提交
294 295


Y
Refine  
Yu Yang 已提交
296 297 298 299 300 301 302 303 304 305
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
306 307
    reader.stop_gradient = True
    reader.persistable = True
Y
Refine  
Yu Yang 已提交
308 309 310
    return reader


Y
Yu Yang 已提交
311 312 313 314 315
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
316 317 318 319
    return new_var


def _copy_reader_create_op_(block, op):
F
fengjiayi 已提交
320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335
    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
336
    new_op = block.append_op(
F
fengjiayi 已提交
337 338 339
        type=op.type,
        inputs=new_input_map,
        outputs=new_output_map,
J
JiayiFeng 已提交
340
        attrs=op.all_attrs())
F
fengjiayi 已提交
341
    return new_op
Y
Yu Yang 已提交
342 343


Y
yuyang18 已提交
344
@templatedoc(op_type='create_recordio_file_reader')
F
fengjiayi 已提交
345 346 347 348 349
def open_recordio_file(filename,
                       shapes,
                       lod_levels,
                       dtypes,
                       pass_num=1,
F
fengjiayi 已提交
350
                       for_parallel=True):
F
fengjiayi 已提交
351
    """
Y
yuyang18 已提交
352
    ${comment}
F
fengjiayi 已提交
353 354

    Args:
Y
yuyang18 已提交
355
       filename(${filename_type}): ${filename_comment}.
F
fengjiayi 已提交
356
       shapes(list): List of tuples which declaring data shapes.
Y
yuyang18 已提交
357
       lod_levels(${lod_levels_type}): ${lod_levels_comment}.
F
fengjiayi 已提交
358
       dtypes(list): List of strs which declaring data type.
F
fengjiayi 已提交
359
       pass_num(int): Number of passes to run.
F
fengjiayi 已提交
360 361 362 363
       for_parallel(Bool): Set it as True if you are going to run
            subsequent operators in parallel.

    Returns:
Y
yuyang18 已提交
364
       ${out_comment}.
F
fengjiayi 已提交
365 366 367

    Examples:

Y
yuyang18 已提交
368 369 370 371 372 373 374 375
        >>> 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
376
    """
Y
Yu Yang 已提交
377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400
    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
401 402
    main_prog_var = _copy_reader_var_(default_main_program().current_block(),
                                      startup_var)
F
fengjiayi 已提交
403 404 405 406

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

F
fengjiayi 已提交
407
    return monkey_patch_reader_methods(main_prog_var)
Y
Yu Yang 已提交
408 409


F
fengjiayi 已提交
410 411 412 413 414
def random_data_generator(low, high, shapes, lod_levels, for_parallel=True):
    """
    Create a uniform random data generator

    This layer returns a Reader Variable.
415 416 417
    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
F
fengjiayi 已提交
418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
    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:

433
        .. code-block:: python
F
fengjiayi 已提交
434

435 436 437 438 439 440 441
            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
442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473
    """
    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)

    return monkey_patch_reader_methods(main_prog_var)


Q
qiaolongfei 已提交
474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620
def _py_reader(capacity,
               shapes,
               dtypes,
               lod_levels=None,
               name=None,
               use_double_buffer=True,
               feed_list=None):
    if feed_list is not None:
        assert isinstance(feed_list, list)
        lod_levels = []
        dtypes = []
        shape_concat = []
        ranks = []
        shapes = []

        for data in feed_list:
            dtypes.append(data.dtype)
            shape_concat.extend(data.shape)
            ranks.append(len(data.shape))
            shapes.append(data.shape)
            lod_levels.append(data.lod_level)
    else:
        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))

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

    if name is None:
        queue_name = unique_name('lod_tensor_blocking_queue')
        reader_name = unique_name('create_py_reader')
        double_buffer_name = unique_name('double_buffer')
    else:
        queue_name = "_".join([name, "queue"])
        reader_name = "_".join([name, "reader"])
        double_buffer_name = "_".join([name, "double_buffer"])

    var = global_scope().var(queue_name)
    feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes)

    startup_blk = default_startup_program().current_block()
    startup_var = startup_blk.create_var(name=reader_name)
    startup_blk.append_op(
        type='create_py_reader',
        inputs={'blocking_queue': [queue_name]},
        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)

    reader = monkey_patch_reader_methods(main_prog_var)
    if use_double_buffer:
        double_buffer_reader = double_buffer(reader, name=double_buffer_name)
        # we return a double buffer reader. However, the reset method comes from
        # py_reader.
        double_buffer_reader.reset = reader.reset
        reader = double_buffer_reader

    # monkey patch py_reader special methods
    reader.queue = feed_queue
    current_reset_method = reader.reset
    reader.thread = None
    reader.tensor_provider = None
    reader.exited = False

    def start_provide_thread(func):
        def __provider_thread__():
            for tensors in func():
                array = core.LoDTensorArray()
                for item in tensors:
                    if not isinstance(item, core.LoDTensor):
                        tmp = core.LoDTensor()
                        tmp.set(item, core.CPUPlace())
                        item = tmp

                    array.append(item)

                if reader.exited:
                    break
                feed_queue.push(array)
                if reader.exited:
                    break
            feed_queue.close()

        reader.thread = threading.Thread(target=__provider_thread__)
        reader.thread.daemon = True
        reader.thread.start()

    def __set_tensor_provider__(func):
        reader.tensor_provider = func

    def __set_paddle_reader__(paddle_reader):
        with program_guard(Program(), Program()):
            if feed_list is None:
                feed_list = []
                counter = 0
                for dtype, shape, lod_level in zip(dtypes, shapes, lod_levels):
                    name = str(counter)
                    feed_list.append(
                        data(
                            name=name,
                            dtype=dtype,
                            shape=shape,
                            lod_level=lod_level))
                    counter += 1

            feeder = DataFeeder(feed_list=feed_list, place=core.CPUPlace())
            paddle_reader = feeder.decorate_reader(
                paddle_reader, multi_devices=False)

        def __tensor_provider__():
            for slots in paddle_reader():
                yield [slots[str(idx)] for idx in six.moves.xrange(counter)]

        __set_tensor_provider__(__tensor_provider__)

    def __reset__():
        current_reset_method()
        if reader.thread is not None and reader.tensor_provider is not None:
            reader.exited = True
            reader.thread.join()
            reader.exited = False

    def __start__():
        start_provide_thread(reader.tensor_provider)

    reader.reset = __reset__
    reader.decorate_tensor_provider = __set_tensor_provider__
    reader.decorate_paddle_reader = __set_paddle_reader__
    reader.start = __start__

    return reader


Y
yuyang18 已提交
621 622 623 624 625 626
def py_reader(capacity,
              shapes,
              dtypes,
              lod_levels=None,
              name=None,
              use_double_buffer=True):
S
sneaxiy 已提交
627
    """
628
    Create a Python reader for data feeding in Python
F
fengjiayi 已提交
629

630
    This layer returns a Reader Variable.
631 632
    The Reader provides :code:`decorate_paddle_reader()` and
    :code:`decorate_tensor_provider()` to set a Python generator as the data
633 634 635 636 637 638 639 640
    source in Python side. When :code:`Executor::Run()` is invoked in C++
    side, the data from the generator would be read automatically. Unlike
    :code:`DataFeeder.feed()`, the data reading process and
    :code:`Executor::Run()` process can run in parallel using
    :code:`py_reader`. The :code:`start()` method of the Reader should be
    called when each pass begins, while the :code:`reset()` method should be
    called when the pass ends and :code:`fluid.core.EOFException` raises.
    Note that :code:`Program.clone()` method cannot clone :code:`py_reader`.
S
sneaxiy 已提交
641 642

    Args:
643
       capacity(int): The buffer capacity maintained by :code:`py_reader`.
Y
yuyang18 已提交
644 645 646 647 648
       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.
649
       use_double_buffer(bool): Whether use double buffer or not.
S
sneaxiy 已提交
650 651

    Returns:
652
       Variable: A Reader from which we can get feeding data.
S
sneaxiy 已提交
653 654 655

    Examples:

656
        1. The basic usage of :code:`py_reader` is as follows:
S
sneaxiy 已提交
657

658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732
        >>> import paddle.v2
        >>> import paddle.fluid as fluid
        >>> import paddle.dataset.mnist as mnist
        >>>
        >>> reader = fluid.layers.py_reader(capacity=64,
        >>>                                 shapes=[(-1,3,224,224), (-1,1)],
        >>>                                 dtypes=['float32', 'int64'])
        >>> reader.decorate_paddle_reader(
        >>>     paddle.v2.reader.shuffle(paddle.batch(mnist.train())
        >>>
        >>> img, label = fluid.layers.read_file(reader)
        >>> loss = network(img, label) # some network definition
        >>>
        >>> fluid.Executor(fluid.CUDAPlace(0)).run(fluid.default_startup_program())
        >>>
        >>> exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name)
        >>> for epoch_id in range(10):
        >>>     reader.start()
        >>>     try:
        >>>         while True:
        >>>             exe.run(fetch_list=[loss.name])
        >>>     except fluid.core.EOFException:
        >>>         reader.reset()

        2. When training and testing are both performed, two different
        :code:`py_reader` should be created with different names, e.g.:

        >>> import paddle.v2
        >>> import paddle.fluid as fluid
        >>> import paddle.dataset.mnist as mnist
        >>>
        >>> def network(reader):
        >>>     img, label = fluid.layers.read_file(reader)
        >>>     # Here, we omitted the network definition
        >>>     return loss
        >>>
        >>> train_reader = fluid.layers.py_reader(capacity=64,
        >>>                                       shapes=[(-1,3,224,224), (-1,1)],
        >>>                                       dtypes=['float32', 'int64'],
        >>>                                       name='train_reader')
        >>> train_reader.decorate_paddle_reader(
        >>>     paddle.v2.reader.shuffle(paddle.batch(mnist.train())
        >>>
        >>> test_reader = fluid.layers.py_reader(capacity=32,
        >>>                                      shapes=[(-1,3,224,224), (-1,1)],
        >>>                                      dtypes=['float32', 'int64'],
        >>>                                      name='test_reader')
        >>> test_reader.decorate_paddle_reader(paddle.batch(mnist.test(), 512))
        >>>
        >>> # Create train_main_prog and train_startup_prog
        >>> train_main_prog = fluid.Program()
        >>> train_startup_prog = fluid.Program()
        >>> with fluid.program_guard(train_main_prog, train_startup_prog):
        >>>     # Use fluid.unique_name.guard() to share parameters with test program
        >>>     with fluid.unique_name.guard():
        >>>         train_loss = network(train_reader) # some network definition
        >>>         adam = fluid.optimizer.Adam(learning_rate=0.01)
        >>>         adam.minimize(loss)
        >>>
        >>> # Create test_main_prog and test_startup_prog
        >>> test_main_prog = fluid.Program()
        >>> test_startup_prog = fluid.Program()
        >>> with fluid.program_guard(test_main_prog, test_startup_prog):
        >>>     # Use fluid.unique_name.guard() to share parameters with train program
        >>>     with fluid.unique_name.guard():
        >>>         test_loss = network(test_reader)
        >>>
        >>> fluid.Executor(fluid.CUDAPlace(0)).run(train_startup_prog)
        >>> fluid.Executor(fluid.CUDAPlace(0)).run(test_startup_prog)
        >>>
        >>> train_exe = fluid.ParallelExecutor(use_cuda=True,
        >>>                 loss_name=train_loss.name, main_program=train_main_prog)
        >>> test_exe = fluid.ParallelExecutor(use_cuda=True,
        >>>                 loss_name=test_loss.name, main_program=test_main_prog)
        >>> for epoch_id in range(10):
733
        >>>     train_reader.start()
734 735 736 737 738 739
        >>>     try:
        >>>         while True:
        >>>             train_exe.run(fetch_list=[train_loss.name])
        >>>     except fluid.core.EOFException:
        >>>         train_reader.reset()
        >>>
740
        >>>     test_reader.start()
741 742 743 744 745
        >>>     try:
        >>>         while True:
        >>>             test_exe.run(fetch_list=[test_loss.name])
        >>>     except fluid.core.EOFException:
        >>>         test_reader.reset()
S
sneaxiy 已提交
746
    """
Q
qiaolongfei 已提交
747 748 749 750 751 752 753
    return _py_reader(
        capacity=capacity,
        shapes=shapes,
        dtypes=dtypes,
        lod_levels=lod_levels,
        name=name,
        use_double_buffer=use_double_buffer)
Y
yuyang18 已提交
754

Y
yuyang18 已提交
755

Q
qiaolongfei 已提交
756
def py_reader_by_data(capacity, feed_list, name=None, use_double_buffer=True):
Q
qiaolongfei 已提交
757 758 759 760
    """
    Works much like py_reader except that it's input is feed_list
    instead of shapes, dtypes, lod_levels
    """
Q
qiaolongfei 已提交
761 762 763 764 765 766 767 768
    return _py_reader(
        capacity=capacity,
        shapes=None,
        dtypes=None,
        lod_levels=None,
        name=name,
        use_double_buffer=use_double_buffer,
        feed_list=feed_list)
S
sneaxiy 已提交
769 770


771 772 773 774
def open_files(filenames,
               shapes,
               lod_levels,
               dtypes,
Y
yuyang18 已提交
775
               thread_num=None,
F
fengjiayi 已提交
776 777
               buffer_size=None,
               pass_num=1,
Y
yuyang18 已提交
778
               is_test=None):
F
fengjiayi 已提交
779 780 781
    """
    Open files

782 783 784
    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
785 786 787 788 789 790

    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
791 792 793
       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
794
       pass_num(int): Number of passes to run.
Y
yuyang18 已提交
795 796 797 798
       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
799 800 801 802 803 804 805

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

    Examples:
       .. code-block:: python

F
fengjiayi 已提交
806
         reader = fluid.layers.io.open_files(filenames=['./data1.recordio',
F
fengjiayi 已提交
807
                                                     './data2.recordio'],
F
fengjiayi 已提交
808 809
                                             shapes=[(3,224,224), (1)],
                                             lod_levels=[0, 0],
Y
yuyang18 已提交
810
                                             dtypes=['float32', 'int64'])
F
fengjiayi 已提交
811 812

         # Via the reader, we can use 'read_file' layer to get data:
F
fengjiayi 已提交
813
         image, label = fluid.layers.io.read_file(reader)
F
fengjiayi 已提交
814
    """
Y
yuyang18 已提交
815 816 817 818 819 820 821 822 823
    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
824

M
minqiyang 已提交
825
    if isinstance(filenames, six.string_types):
F
fengjiayi 已提交
826
        filenames = [filenames]
F
fengjiayi 已提交
827 828 829 830 831 832 833 834
    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
835
    multi_file_reader_name = unique_name('multi_file_reader')
F
fengjiayi 已提交
836
    startup_blk = default_startup_program().current_block()
F
fengjiayi 已提交
837
    startup_reader = startup_blk.create_var(name=multi_file_reader_name)
Y
yuyang18 已提交
838 839 840 841
    attrs = {
        'shape_concat': shape_concat,
        'lod_levels': lod_levels,
        'ranks': ranks,
Y
yuyang18 已提交
842 843 844
        'file_names': filenames,
        'thread_num': thread_num,
        'buffer_size': buffer_size
Y
yuyang18 已提交
845 846 847
    }
    if is_test is not None:
        attrs['is_test'] = is_test
F
fengjiayi 已提交
848
    startup_blk.append_op(
Y
yuyang18 已提交
849
        type='open_files', outputs={'Out': [startup_reader]}, attrs=attrs)
F
fengjiayi 已提交
850

F
fengjiayi 已提交
851 852 853 854 855 856 857
    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
858

F
fengjiayi 已提交
859 860 861
    return monkey_patch_reader_methods(main_prog_reader)


J
JiayiFeng 已提交
862
def __create_shared_decorated_reader__(op_type, reader, attrs):
Y
Yu Yang 已提交
863 864 865
    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
866
    startop_op = startup_blk.append_op(
Y
Yu Yang 已提交
867 868 869 870 871
        type=op_type,
        inputs={'UnderlyingReader': reader},
        outputs={'Out': [startup_var]},
        attrs=attrs)
    startup_var.persistable = True
F
fengjiayi 已提交
872 873 874 875
    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
876 877


878 879
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)
880 881 882 883 884 885 886 887 888 889
    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
890
def shuffle(reader, buffer_size):
891 892 893
    """
    Shuffle the reader.
    """
894 895
    return __create_unshared_decorated_reader__(
        'create_shuffle_reader', reader, {'buffer_size': int(buffer_size)})
Y
Yu Yang 已提交
896 897


J
JiayiFeng 已提交
898
def batch(reader, batch_size):
F
fengjiayi 已提交
899
    """
900 901 902
    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
F
fengjiayi 已提交
903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926
    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.
927
            #
F
fengjiayi 已提交
928 929
            # However, if we read data with the batch_reader:
            #     data = fluid.layers.read_file(batch_reader)
930 931
            # Each 5 adjacent instances will be automatically combined together
            # to become a batch. So what we get('data') is a batch data instead
F
fengjiayi 已提交
932 933
            # of an instance.
    """
J
JiayiFeng 已提交
934 935 936 937
    return __create_unshared_decorated_reader__(
        'create_batch_reader', reader, {'batch_size': int(batch_size)})


938
def double_buffer(reader, place=None, name=None):
Y
yuyang18 已提交
939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961
    """
    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
962 963 964
    attrs = dict()
    if place is not None:
        attrs['place'] = str(place).upper()
965 966
    return __create_unshared_decorated_reader__(
        'create_double_buffer_reader', reader, attrs, name=name)
Y
Yu Yang 已提交
967 968


F
fengjiayi 已提交
969
def multi_pass(reader, pass_num):
970 971
    return __create_shared_decorated_reader__(
        'create_multi_pass_reader', reader, {'pass_num': int(pass_num)})
F
fengjiayi 已提交
972 973


F
fengjiayi 已提交
974
def read_file(reader):
F
fengjiayi 已提交
975
    """
F
fengjiayi 已提交
976
    Execute the given reader and get data via it.
F
fengjiayi 已提交
977

978 979
    A reader is also a Variable. It can be a raw reader generated by
    `fluid.layers.open_files()` or a decorated one generated by
F
fengjiayi 已提交
980 981 982 983
    `fluid.layers.double_buffer()` and so on.

    Args:

F
fengjiayi 已提交
984
        reader(Variable): The reader to execute.
F
fengjiayi 已提交
985 986

    Returns:
F
fengjiayi 已提交
987
        Tuple[Variable]: Data read via the given reader.
F
fengjiayi 已提交
988 989 990 991 992 993 994 995 996 997 998 999 1000

    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
1001 1002 1003 1004
    helper = LayerHelper('read_file')
    out = [
        helper.create_tmp_variable(
            stop_gradient=True, dtype='float32')
F
fengjiayi 已提交
1005
        for _ in range(len(reader.desc.shapes()))
Y
Yu Yang 已提交
1006 1007
    ]
    helper.append_op(
F
fengjiayi 已提交
1008
        type='read', inputs={'Reader': [reader]}, outputs={'Out': out})
Y
Yu Yang 已提交
1009 1010 1011 1012
    if len(out) == 1:
        return out[0]
    else:
        return out
F
fengjiayi 已提交
1013 1014 1015


class Preprocessor(object):
X
Xin Pan 已提交
1016 1017 1018 1019 1020 1021 1022 1023 1024
    """
    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
1025

X
Xin Pan 已提交
1026 1027 1028 1029 1030 1031 1032 1033 1034 1035
            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
1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051
    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

X
Xin Pan 已提交
1052
    def _is_completed(self):
F
fengjiayi 已提交
1053 1054 1055 1056 1057
        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
W
Wu Yi 已提交
1058
        self.sub_block = self.main_prog._create_block()
F
fengjiayi 已提交
1059
        yield
W
Wu Yi 已提交
1060
        self.main_prog._rollback()
F
fengjiayi 已提交
1061
        self.status = Preprocessor.AFTER_SUB_BLOCK
X
Xin Pan 已提交
1062
        if not self._is_completed():
F
fengjiayi 已提交
1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077
            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
1078 1079
        self.source_var_names = [
            unique_name("preprocessor_source")
M
minqiyang 已提交
1080
            for _ in six.moves.range(len(source_shapes))
F
fengjiayi 已提交
1081
        ]
F
fengjiayi 已提交
1082
        source_vars = []
F
fengjiayi 已提交
1083 1084 1085
        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
1086
            source_vars.append(self.main_prog.current_block().create_var(
F
fengjiayi 已提交
1087
                name=var_name, shape=shape, dtype=dtype, lod_level=lod_level))
F
fengjiayi 已提交
1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111
        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
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137


@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)