io.py 16.8 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 14
# 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.

Y
Yu Yang 已提交
15
from .. import core
Y
Yu Yang 已提交
16 17
from ..framework import convert_np_dtype_to_dtype_, default_main_program, default_startup_program
from ..unique_name import generate as unique_name
T
WIP  
typhoonzero 已提交
18 19
from control_flow import BlockGuard
from ..layer_helper import LayerHelper
Y
Refine  
Yu Yang 已提交
20
from ..executor import global_scope
Y
Yu Yang 已提交
21

Y
Yu Yang 已提交
22 23
__all__ = [
    'data', 'BlockGuardServ', 'ListenAndServ', 'Send', 'open_recordio_file',
J
JiayiFeng 已提交
24
    'open_files', 'read_file', 'shuffle', 'double_buffer'
Y
Yu Yang 已提交
25
]
Y
Yu Yang 已提交
26 27 28 29 30 31 32 33 34 35


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
36
    **Data Layer**
Y
Yu Yang 已提交
37

K
kavyasrinet 已提交
38
    This function takes in the input and based on whether data has
C
caoying03 已提交
39
    to be returned back as a minibatch, it creates the global variable by using
Y
Yu Yang 已提交
40
    the helper functions. The global variables can be accessed by all the
C
caoying03 已提交
41
    following operators in the graph.
Y
Yu Yang 已提交
42 43 44 45

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

K
kavyasrinet 已提交
46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63
    Args:
       name(str): The name/alias of the function
       shape(list): Tuple declaring the shape.
       append_batch_size(bool): Whether or not to append the data as a batch.
       dtype(int|float): The type of data : float32, float_16, int etc
       type(VarType): The output type. By default it is LOD_TENSOR.
       lod_level(int): The LoD Level. 0 means the input data is not a sequence.
       main_program(Program): Name of the main program that calls this
       startup_program(Program): Name of the startup program
       stop_gradient(bool): A boolean that mentions whether gradient should flow.

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

    Examples:
        .. code-block:: python

          data = fluid.layers.data(name='x', shape=[784], dtype='float32')
Y
Yu Yang 已提交
64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83
    """
    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

    return helper.create_global_variable(
        name=name,
        shape=shape,
        dtype=dtype,
        type=type,
        stop_gradient=stop_gradient,
        lod_level=lod_level)
T
typhoonzero 已提交
84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114


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

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

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

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

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


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

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

Y
Yancey1989 已提交
115
    def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True):
116
        self.helper = LayerHelper("listen_and_serv")
Y
Yancey1989 已提交
117
        self.inputs = inputs
T
typhoonzero 已提交
118 119 120
        self.outputs = []
        self.endpoint = endpoint
        self.fan_in = fan_in
T
typhoonzero 已提交
121 122
        # FIXME(typhoonzero): add optimizer_mode is stupid, should make it more
        # general.
T
WIP  
typhoonzero 已提交
123
        self.optimizer_mode = optimizer_mode
T
typhoonzero 已提交
124 125 126 127 128 129 130 131 132 133 134 135 136

    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
137 138 139 140 141 142 143 144
            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
145 146
                        params.append(parent_block.var(in_var_name))
                        grads.append(parent_block.var(in_var_name))
T
typhoonzero 已提交
147 148 149

        return params, grads

T
typhoonzero 已提交
150 151 152 153 154 155 156
    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
157 158 159 160 161 162
    def complete_op(self):
        main_program = self.helper.main_program
        current_block = main_program.current_block()
        parent_block = self.parent_block()

        parent_block.append_op(
163
            type='listen_and_serv',
Y
Yancey1989 已提交
164
            inputs={"X": self.inputs},
T
typhoonzero 已提交
165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189
            outputs={},
            attrs={
                'endpoint': self.endpoint,
                'Fanin': self.fan_in,
                'OptimizeBlock': current_block
            })


def Send(endpoints, send_vars, get_vars):
    """
    Send layer

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

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

    epmap = endpoints.split(",")
T
typhoonzero 已提交
190
    endpoints = list(set(epmap))
T
typhoonzero 已提交
191 192

    helper = LayerHelper("Send", **locals())
Y
Yancey1989 已提交
193 194 195
    rpc_client_var = default_main_program().global_block().create_var(
        name="RPC_CLIENT_VAR", persistable=True, type=core.VarDesc.VarType.RAW)

T
typhoonzero 已提交
196 197 198
    helper.append_op(
        type="send",
        inputs={"X": send_vars},
Y
Yancey1989 已提交
199 200
        outputs={"Out": get_vars,
                 "RPCClient": rpc_client_var},
T
typhoonzero 已提交
201 202
        attrs={"endpoints": endpoints,
               "epmap": epmap})
203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230


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

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

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

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

    helper = LayerHelper("Recv", **locals())
    helper.append_op(
        type="recv",
        inputs={"X": get_vars},
        outputs={"Out": get_vars},
        attrs={"endpoints": endpoints,
               "epmap": epmap})
Y
Yu Yang 已提交
231 232


Y
Refine  
Yu Yang 已提交
233 234 235 236 237 238 239 240 241 242 243 244 245 246
def monkey_patch_reader_methods(reader):
    def __get_reader__():
        scope = global_scope()
        var = scope.find_var(reader.name)
        return var.get_reader()

    def eof():
        return not __get_reader__().has_next()

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

    reader.eof = eof
    reader.reset = reset
Y
Yu Yang 已提交
247 248
    reader.stop_gradient = True
    reader.persistable = True
Y
Refine  
Yu Yang 已提交
249 250 251
    return reader


Y
Yu Yang 已提交
252 253 254 255 256
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
257 258 259 260
    return new_var


def _copy_reader_create_op_(block, op):
F
fengjiayi 已提交
261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276
    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
277
    new_op = block.append_op(
F
fengjiayi 已提交
278 279 280
        type=op.type,
        inputs=new_input_map,
        outputs=new_output_map,
J
JiayiFeng 已提交
281
        attrs=op.all_attrs())
F
fengjiayi 已提交
282
    return new_op
Y
Yu Yang 已提交
283 284


F
fengjiayi 已提交
285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321
def open_recordio_file(filename,
                       shapes,
                       lod_levels,
                       dtypes,
                       pass_num=1,
                       for_parallel=False):
    """
    Open a RecordIO file

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

    Args:
       filename(str): The RecordIO file's name.
       shapes(list): List of tuples which declaring data shapes.
       lod_levels(list): List of ints which declaring data lod_level.
       dtypes(list): List of strs which declaring data type.
       pass_num(int): Number of passes to run. After completing the
            given number of passes, 'has_next()' will return False.
       for_parallel(Bool): Set it as True if you are going to run
            subsequent operators in parallel.

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

    Examples:
       .. code-block:: python

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

         # Via the reader, we can use 'read_file' layer to get data:
         image, label = fluid.layers.read_file(reader)
    """
Y
Yu Yang 已提交
322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
    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
346 347
    main_prog_var = _copy_reader_var_(default_main_program().current_block(),
                                      startup_var)
F
fengjiayi 已提交
348 349 350 351 352

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

    if for_parallel:
J
JiayiFeng 已提交
353
        main_prog_var = parallelize(reader=main_prog_var)
F
fengjiayi 已提交
354

F
fengjiayi 已提交
355
    return monkey_patch_reader_methods(main_prog_var)
Y
Yu Yang 已提交
356 357


358 359 360 361 362
def open_files(filenames,
               shapes,
               lod_levels,
               dtypes,
               thread_num,
F
fengjiayi 已提交
363 364 365
               buffer_size=None,
               pass_num=1,
               for_parallel=False):
F
fengjiayi 已提交
366 367 368
    """
    Open files

F
fengjiayi 已提交
369 370 371
    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
372 373 374 375 376 377 378 379

    Args:
       filenames(list): The list of file names.
       shapes(list): List of tuples which declaring data shapes.
       lod_levels(list): List of ints which declaring data lod_level.
       dtypes(list): List of strs which declaring data type.
       thread_num(int): The maximal concurrent prefetch thread number.
       buffer_size(int): The size of prefetch buffer.
F
fengjiayi 已提交
380 381 382 383
       pass_num(int): Number of passes to run. After completing the 
            given number of passes, 'has_next()' will return False.
       for_parallel(Bool): Set it as True if you are going to run 
            subsequent operators in parallel.
F
fengjiayi 已提交
384 385 386 387 388 389 390

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

    Examples:
       .. code-block:: python

F
fengjiayi 已提交
391
         reader = fluid.layers.io.open_files(filenames=['./data1.recordio',
F
fengjiayi 已提交
392
                                                     './data2.recordio'],
F
fengjiayi 已提交
393 394 395 396 397
                                             shapes=[(3,224,224), (1)],
                                             lod_levels=[0, 0],
                                             dtypes=['float32', 'int64'],
                                             thread_num=2,
                                             buffer_size=2)
F
fengjiayi 已提交
398 399

         # Via the reader, we can use 'read_file' layer to get data:
F
fengjiayi 已提交
400
         image, label = fluid.layers.io.read_file(reader)
F
fengjiayi 已提交
401
    """
402 403
    if buffer_size is None:
        buffer_size = thread_num
F
fengjiayi 已提交
404 405
    if isinstance(filenames, basestring):
        filenames = [filenames]
F
fengjiayi 已提交
406 407 408 409 410 411 412 413
    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
414
    multi_file_reader_name = unique_name('multi_file_reader')
F
fengjiayi 已提交
415
    startup_blk = default_startup_program().current_block()
F
fengjiayi 已提交
416
    startup_reader = startup_blk.create_var(name=multi_file_reader_name)
F
fengjiayi 已提交
417 418
    startup_blk.append_op(
        type='open_files',
F
fengjiayi 已提交
419
        outputs={'Out': [startup_reader]},
F
fengjiayi 已提交
420 421 422 423
        attrs={
            'shape_concat': shape_concat,
            'lod_levels': lod_levels,
            'ranks': ranks,
F
fengjiayi 已提交
424
            'file_names': filenames,
425 426
            'thread_num': thread_num,
            'buffer_size': buffer_size
F
fengjiayi 已提交
427 428
        })

F
fengjiayi 已提交
429 430 431 432 433 434 435
    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
436

F
fengjiayi 已提交
437
    if for_parallel:
J
JiayiFeng 已提交
438
        main_prog_reader = parallelize(reader=main_prog_reader)
F
fengjiayi 已提交
439

F
fengjiayi 已提交
440 441 442
    return monkey_patch_reader_methods(main_prog_reader)


J
JiayiFeng 已提交
443
def __create_shared_decorated_reader__(op_type, reader, attrs):
Y
Yu Yang 已提交
444 445 446
    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
447
    startop_op = startup_blk.append_op(
Y
Yu Yang 已提交
448 449 450 451 452
        type=op_type,
        inputs={'UnderlyingReader': reader},
        outputs={'Out': [startup_var]},
        attrs=attrs)
    startup_var.persistable = True
F
fengjiayi 已提交
453 454 455 456
    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
457 458


J
JiayiFeng 已提交
459
def __create_unshared_decorated_reader__(op_type, reader, attrs):
460 461 462 463 464 465 466 467 468 469 470 471 472
    new_reader_name = unique_name(op_type)
    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)
    new_reader.persistable = True
    new_reader.stop_gradient = True
    return monkey_patch_reader_methods(new_reader)


F
fengjiayi 已提交
473
def shuffle(reader, buffer_size):
474 475
    return __create_unshared_decorated_reader__(
        'create_shuffle_reader', reader, {'buffer_size': int(buffer_size)})
Y
Yu Yang 已提交
476 477


F
fengjiayi 已提交
478
def double_buffer(reader, place=None):
Y
Yu Yang 已提交
479 480 481
    attrs = dict()
    if place is not None:
        attrs['place'] = str(place).upper()
482 483
    return __create_unshared_decorated_reader__('create_double_buffer_reader',
                                                reader, attrs)
Y
Yu Yang 已提交
484 485


F
fengjiayi 已提交
486
def multi_pass(reader, pass_num):
487 488
    return __create_shared_decorated_reader__(
        'create_multi_pass_reader', reader, {'pass_num': int(pass_num)})
F
fengjiayi 已提交
489 490


J
JiayiFeng 已提交
491 492 493
def parallelize(reader):
    return __create_shared_decorated_reader__('create_threaded_reader', reader,
                                              {})
F
fengjiayi 已提交
494 495


Y
Yu Yang 已提交
496 497 498 499 500
def read_file(file_obj):
    helper = LayerHelper('read_file')
    out = [
        helper.create_tmp_variable(
            stop_gradient=True, dtype='float32')
Y
Yu Yang 已提交
501
        for _ in range(len(file_obj.desc.shapes()))
Y
Yu Yang 已提交
502 503 504 505 506 507 508
    ]
    helper.append_op(
        type='read', inputs={'Reader': [file_obj]}, outputs={'Out': out})
    if len(out) == 1:
        return out[0]
    else:
        return out