framework.py 72.1 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.

15 16
from __future__ import print_function

Y
Yu Yang 已提交
17
import collections
Q
qiaolongfei 已提交
18
import contextlib
F
fengjiayi 已提交
19
import re
20
import six
21

Y
Yu Yang 已提交
22
import numpy as np
Q
qiaolongfei 已提交
23

M
minqiyang 已提交
24
from .. import compat as cpt
25
from .proto import framework_pb2
26 27
try:
    from . import core
28
except ImportError as e:
29 30 31 32
    raise ImportError(
        """NOTE: You may need to run \"export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH\"
    if you encounters \"libmkldnn.so not found\" errors. If you have python
    installed in other directory, replace \"/usr/local/lib\" with your own
M
minqiyang 已提交
33
    directory. The original error is: \n""" + cpt.get_exception_message(e))
34
except Exception as e:
35
    raise e
36
from . import unique_name
Y
Yu Yang 已提交
37

38
__all__ = [
39 40 41 42
    'Program',
    'default_startup_program',
    'default_main_program',
    'program_guard',
43
    'name_scope',
44
]
Y
Yu Yang 已提交
45

Q
qiaolongfei 已提交
46 47 48 49
EMPTY_VAR_NAME = core.kEmptyVarName()
TEMP_VAR_NAME = core.kTempVarName()
GRAD_VAR_SUFFIX = core.kGradVarSuffix()
ZERO_VAR_SUFFIX = core.kZeroVarSuffix()
W
Wu Yi 已提交
50 51 52
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()


53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116
class NameScope(object):
    def __init__(self, name="", parent=None):
        self._children = dict()
        self._name = name
        self._parent = parent

    def child(self, prefix):
        if prefix not in self._children:
            new_child = NameScope(prefix, self)
            self._children[prefix] = [new_child]
        else:
            new_child = NameScope(prefix + "_%d" % len(self._children[prefix]),
                                  self)
            self._children[prefix].append(new_child)
        return new_child

    def parent(self):
        return self._parent

    def name(self):
        return self._name


_name_scope = NameScope()


@contextlib.contextmanager
def name_scope(prefix=None):
    """
    Generate hierarchical name prefix for the operators.

    Note: This should only used for debugging and visualization purpose.
    Don't use it for serious analysis such as graph/program transformations.

    Args:
        prefix(str): prefix.

    Examples:
        .. code-block:: python
          with name_scope("encoder"):
             ...
          with name_scope("decoder"):
             ...
             with name_scope("attention"):
                ...
    """
    # TODO(panyx0718): Only [0-9a-z].
    assert prefix, "namescope prefix cannot be empty."
    global _name_scope
    _name_scope = _name_scope.child(prefix)
    yield
    _name_scope = _name_scope.parent()


def _full_name_scope():
    global _name_scope
    scope = _name_scope
    name = ""
    while scope:
        name = scope.name() + "/" + name
        scope = scope.parent()
    return name


W
Wu Yi 已提交
117 118 119
def generate_control_dev_var_name():
    import random
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
120 121 122 123


def grad_var_name(var_name):
    """
124 125
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
126 127 128
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
129

130
def convert_np_dtype_to_dtype_(np_dtype):
131 132
    """
    Convert the data type in numpy to the data type in Paddle
133

134
    Args:
135
        np_dtype(np.dtype): the data type in numpy.
136

137 138
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
139 140

    """
141 142
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
143
        return core.VarDesc.VarType.FP32
144
    elif dtype == np.float64:
145
        return core.VarDesc.VarType.FP64
146
    elif dtype == np.float16:
147
        return core.VarDesc.VarType.FP16
148
    elif dtype == np.int32:
149
        return core.VarDesc.VarType.INT32
150
    elif dtype == np.int16:
151
        return core.VarDesc.VarType.INT16
152
    elif dtype == np.int64:
153
        return core.VarDesc.VarType.INT64
154
    elif dtype == np.bool:
155
        return core.VarDesc.VarType.BOOL
156 157
    elif dtype == np.uint16:
        return core.VarDesc.VarType.INT16
158 159
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
160 161
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
162
    else:
M
minqiyang 已提交
163
        raise ValueError("Not supported numpy dtype %s" % dtype)
164 165 166


def dtype_is_floating(dtype):
167 168 169
    """
    Check the data type is floating or not.
    Args:
170
        dtype(np.dtype|core.VarDesc.VarType): data type.
171 172 173 174 175
            Could be numpy format or Paddle format

    Returns(bool): True if data type is a float value

    """
176
    if not isinstance(dtype, core.VarDesc.VarType):
177 178
        dtype = convert_np_dtype_to_dtype_(dtype)

179 180 181 182
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
183 184


Y
Yang Yang(Tony) 已提交
185
def _debug_string_(proto, throw_on_error=True):
186 187 188 189 190 191 192 193 194 195 196
    """
    Get the debug string of a protobuf message. The message could be not
    initialized.
    Args:
        proto(google.protobuf.message.Message): The protobuf message
        throw_on_error(bool): True if raise an error when the protobuf message
            is not initialized.

    Returns(str): The debug string of the protobuf message

    """
Y
Yu Yang 已提交
197
    error_fields = list()
Y
Yang Yang(Tony) 已提交
198
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
199 200
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
201 202 203
    return proto.__str__()


Y
Yu Yang 已提交
204
class Variable(object):
205
    """
206 207 208
    In Fluid, every input and output of an operator is a variable. In most
    cases, variables are used for holding different kinds of data or training
    labels. A variable belongs to a block. All variable has its own name and
209
    two variables in different blocks could have the same name.
210

211 212
    There are many kinds of variables. Each kind of them has its own attributes
    and usages. Please reference the framework.proto for details.
213

214
    Most of a Variable's member variables can be setted to be None. It mean
215
    it is not available or will be specified later.
216 217

    Args:
218
        block(Block): The block that the variable belongs to.
219 220
        type(core.VarDesc.VarType): Variable type. Please reference the
            framework.proto for details.
221 222
        name(str|None): The name of the variable. If setted None, it will be
            generated automatically. Default: None
223
        shape(tuple|list|None): The shape of the variable. -1 means the batch size.
224
            Some kinds of variable do not contain shape, just set it to None.
225 226 227
            Default: None
        dtype(np.dtype|core.VarDesc.VarType|str|None): The data type of variable.
            Default: None
228
        lod_level (int|None): The level of lod tensor. 0 means it is not a time
229
            series data.
230
            Default: None
231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252
        capacity (int|None): The capacity of Channel variable. Ignored for other
            types. Default: None
        persistable (bool|None): True if the variable is persistable. A persistable
            variable will not be deleted after an iteration ending. Defaults: None.
        error_clip (BaseErrorClipAttr|None): The error clip attributes of the
            corresponding gradient variable. Default: None
        stop_gradient (bool): True if the variable will stop to calculate its
            gradients when backward. Default: False.
        is_data (bool): True if the variable is an input data. Default: False

    Notes:
        The constructor of Variable should not be invoked directly. Please
        use `Block.create_var` to create a variable.

    Examples:
        .. code-block:: python

            cur_program = Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
253 254
    """

Y
Yu Yang 已提交
255 256
    def __init__(self,
                 block,
Y
Yu Yang 已提交
257
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
258 259 260 261
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
262
                 capacity=None,
Q
QI JUN 已提交
263
                 persistable=None,
F
fengjiayi 已提交
264
                 error_clip=None,
Y
Yu Yang 已提交
265
                 stop_gradient=False,
F
fengjiayi 已提交
266
                 is_data=False,
Y
Yu Yang 已提交
267
                 **kwargs):
Y
Yu Yang 已提交
268
        self.block = block
F
fengjiayi 已提交
269
        self.error_clip = error_clip
Y
Yu Yang 已提交
270 271

        if name is None:
Y
Yu Yang 已提交
272
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
273
        is_new_var = False
M
minqiyang 已提交
274
        name = cpt.to_text(name)
M
minqiyang 已提交
275
        self.desc = self.block.desc.find_var(cpt.to_bytes(name))
D
Dong Zhihong 已提交
276 277

        if self.desc is None:
M
minqiyang 已提交
278
            self.desc = self.block.desc.var(cpt.to_bytes(name))
Y
Yu Yang 已提交
279
            is_new_var = True
Y
Yu Yang 已提交
280

Y
Yu Yang 已提交
281 282 283 284 285 286 287 288
        if is_new_var:
            self.desc.set_type(type)
        elif self.desc.type() != type:
            raise ValueError("Variable {0} has been created before. The "
                             "previous type is {1}; the new type is {2}. They"
                             " are not matched".format(self.name,
                                                       self.desc.type(), type))

Y
Yu Yang 已提交
289
        if shape is not None:
Y
Yu Yang 已提交
290
            if is_new_var:
291
                self.desc.set_shape(shape)
Y
Yu Yang 已提交
292 293 294 295 296 297 298 299
            else:
                old_shape = self.shape
                shape = tuple(shape)
                if shape != old_shape:
                    raise ValueError(
                        "Variable {0} has been created before. the previous "
                        "shape is {1}; the new shape is {2}. They are not "
                        "matched.".format(self.name, old_shape, shape))
Y
Yu Yang 已提交
300
        if dtype is not None:
301
            if not isinstance(dtype, core.VarDesc.VarType):
302
                dtype = convert_np_dtype_to_dtype_(dtype)
Y
Yu Yang 已提交
303
            if is_new_var:
F
fengjiayi 已提交
304
                self.desc.set_dtype(dtype)
Y
Yu Yang 已提交
305
            else:
F
fengjiayi 已提交
306
                old_dtype = self.dtype
Q
QI JUN 已提交
307
                if dtype != old_dtype:
Y
Yu Yang 已提交
308 309 310 311 312
                    raise ValueError("Variable {0} has been created before. "
                                     "The previous data type is {1}; the new "
                                     "data type is {2}. They are not "
                                     "matched.".format(self.name, old_dtype,
                                                       dtype))
Y
Yu Yang 已提交
313 314

        if lod_level is not None:
Y
Yu Yang 已提交
315
            if is_new_var:
316
                self.desc.set_lod_level(lod_level)
Y
Yu Yang 已提交
317 318 319 320 321 322 323
            else:
                if lod_level != self.lod_level:
                    raise ValueError("Variable {0} has been created before. "
                                     "The previous lod_level is {1}; the new "
                                     "lod_level is {2}. They are not "
                                     "matched".format(self.name, self.lod_level,
                                                      lod_level))
324 325 326 327 328 329 330 331 332 333 334
        if persistable is not None:
            if is_new_var:
                self.desc.set_persistable(persistable)
            else:
                if persistable != self.persistable:
                    raise ValueError(
                        "Variable {0} has been created before."
                        "The previous persistable is {1}; the new "
                        "persistable is {2}. They are not matched".format(
                            self.name, self.persistable, persistable))

335 336 337 338 339 340 341 342
        if capacity is not None:
            if is_new_var:
                self.desc.set_capacity(capacity)
            else:
                # TODO(abhinavarora) : Compare with set capacity once,
                # get_capacity is implemented
                pass

Y
Yu Yang 已提交
343
        self.block.vars[name] = self
Y
Yu Yang 已提交
344
        self.op = None
Y
Yu Yang 已提交
345
        self.stop_gradient = stop_gradient
F
fengjiayi 已提交
346
        self.is_data = is_data
Y
Yu Yang 已提交
347

348
    def __str__(self):
Y
Yang Yang(Tony) 已提交
349 350
        return self.to_string(True)

F
update  
fengjiayi 已提交
351
    def to_string(self, throw_on_error, with_details=False):
352 353 354 355
        """
        Get debug string.

        Args:
356 357
            throw_on_error(bool): True if raise an exception when self is
                not initialized.
F
update  
fengjiayi 已提交
358
            with_details(bool): more details about variables and parameters
359 360
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False;
361

362 363
        Returns:
            str: The debug string.
364
        """
F
update  
fengjiayi 已提交
365 366
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
367
        protostr = self.desc.serialize_to_string()
368
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
369 370 371 372
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
373 374
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
375
        return res_str
376 377 378

    __repr__ = __str__

W
Wu Yi 已提交
379
    def _set_desc(self, input):
380 381 382 383 384 385 386 387 388
        """
        Set the variable description.

        Args:
            input(core.VarDesc): The new VarDesc.

        Returns:
            None
        """
389 390
        self.desc = input

391 392 393 394
    @property
    def persistable(self):
        return self.desc.persistable()

Y
Yu Yang 已提交
395 396 397 398
    @persistable.setter
    def persistable(self, p):
        self.desc.set_persistable(p)

Y
Yu Yang 已提交
399 400
    @property
    def name(self):
M
minqiyang 已提交
401
        return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
402

T
typhoonzero 已提交
403 404 405 406
    @name.setter
    def name(self, new_name):
        self.desc.set_name(new_name)

Y
Yu Yang 已提交
407 408 409
    @property
    def shape(self):
        # convert to tuple, make it as same as numpy API.
410
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
411 412

    @property
F
fengjiayi 已提交
413 414
    def dtype(self):
        return self.desc.dtype()
Y
Yu Yang 已提交
415 416 417

    @property
    def lod_level(self):
418
        return self.desc.lod_level()
Y
Yu Yang 已提交
419

Y
Yu Yang 已提交
420 421 422 423
    @property
    def type(self):
        return self.desc.type()

W
Wu Yi 已提交
424
    def _set_error_clip(self, error_clip):
425 426 427 428 429 430 431 432 433
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
434 435
        self.error_clip = error_clip

Y
Yu Yang 已提交
436

F
fengjiayi 已提交
437 438 439
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
440

441 442
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
443 444 445 446
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
447
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
448 449 450 451 452
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
453 454 455 456
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
457 458 459 460 461 462 463 464 465
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
466
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
467 468 469 470 471 472
        op_protos = get_all_op_protos()
        self.op_proto_map = {}
        for proto in op_protos:
            self.op_proto_map[proto.type] = proto

    def get_op_proto(self, type):
473 474 475 476 477 478 479 480
        """
        Get OpProto by a type string.
        Args:
            type(str): The type that operator registered in C++ side.

        Returns(framework_pb2.OpProto): The OpProto

        """
Y
Yu Yang 已提交
481 482
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
483 484
        return self.op_proto_map[type]

485 486 487 488
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
489
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
490
            core.op_proto_and_checker_maker.kOpNameScopeAttrName()
491 492
        }

F
fengjiayi 已提交
493

Y
Yu Yang 已提交
494
class Operator(object):
495
    """
496 497 498 499 500 501 502
    In Fluid, all the operation are represented by Operator, and Operator
    is regarded as a build in an instruction of a Block. Users can use the
    build in instructions to describe their neural network.

    Args:
        block(Block): The block has the current operator.
        desc(core.OpDesc): The protobuf description of Operator.
C
chengduoZH 已提交
503
        type(str): The type of operator. Default None.
504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
        inputs(dict): The input of this Operator. it is a dictionary, for every
            element, key is the input parameter name, and value is a list of
            variables. Default None.
        outputs(dict): The output of this Operator. it is a dictionary, for
            every element, key is the input parameter name, and value is a list
            of variables. Default None.
        attrs(dict): The attributes of this Operator. it is a dictionary, for
            every element, key is attribute name, and value is the attribute value.
            The attribute type should be as same as the type registered in C++ side.
            Default None.

    Returns:
        Operator: The initialized Operator.

    Raises:
        ValueError: If the passed input, output and attrs doesn't match the
            initializing Operator's that registered in C++ side.

    Notes:
        The constructor of operator should not be invoked directly. Use
W
Wu Yi 已提交
524
        Block.append_op or Block._prepend_op instead.
525 526 527 528 529 530 531 532 533 534

    Examples:
        .. code-block:: python

            cur_program = Program()
            cur_block = cur_program.current_block()
            # var1 += var2 + var3
            cur_block.append_op(type="sum",
                                inputs={"X": [var1, var2, var3]},
                                outputs={"Out": [var1]})
535
    """
536 537 538 539
    OP_WITHOUT_KERNEL_SET = {
        'feed', 'fetch', 'save', 'load', 'recurrent', 'go',
        'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv',
        'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine',
X
Xin Pan 已提交
540
        'ncclInit', 'select', 'checkpoint_notify', 'gen_nccl_id'
541
    }
542

Y
Yu Yang 已提交
543 544
    def __init__(self,
                 block,
Y
Yu Yang 已提交
545
                 desc,
Y
Yu Yang 已提交
546 547 548 549 550
                 type=None,
                 inputs=None,
                 outputs=None,
                 attrs=None):
        self.block = block
Y
Yu Yang 已提交
551
        self.desc = desc
G
gongweibao 已提交
552 553 554 555 556
        # note: not add self.attrs here:
        # https://github.com/PaddlePaddle/Paddle/pull/12583#pullrequestreview-145093173
        op_attrs = attrs
        if op_attrs is None:
            op_attrs = dict()
Y
yuyang18 已提交
557 558 559 560
        del attrs

        op_maker = core.op_proto_and_checker_maker

G
gongweibao 已提交
561 562
        if op_maker.kOpRoleAttrName() not in op_attrs:
            op_attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role
Y
yuyang18 已提交
563 564 565

        role_var_name = op_maker.kOpRoleVarAttrName()
        if len(self.block.program.
G
gongweibao 已提交
566 567
               op_role_var) != 0 and role_var_name not in op_attrs:
            op_attrs[role_var_name] = self.block.program.op_role_var
Y
yuyang18 已提交
568

G
gongweibao 已提交
569 570
        if role_var_name in op_attrs and len(op_attrs[role_var_name]) == 0:
            del op_attrs[role_var_name]
Y
yuyang18 已提交
571

F
fengjiayi 已提交
572 573 574 575 576
        if len(self.desc.type()) != 0:
            return
        if type is None:
            raise ValueError(
                "`type` to initilized an Operator can not be None.")
F
Update  
fengjiayi 已提交
577
        self.desc.set_type(type)
F
fengjiayi 已提交
578
        proto = OpProtoHolder.instance().get_op_proto(type)
579

580 581 582
        namescope_var_name = op_maker.kOpNameScopeAttrName()
        op_attrs[namescope_var_name] = _full_name_scope()

Y
Yang Yang(Tony) 已提交
583 584
        def find_name(var_list, name):
            for var_name in var_list:
Q
Qiao Longfei 已提交
585
                if var_list[var_name] is not None and var_name == name:
Y
Yang Yang(Tony) 已提交
586 587
                    return True
            return False
Q
QI JUN 已提交
588

Y
Yang Yang(Tony) 已提交
589 590 591 592 593 594 595
        if inputs is not None:
            for in_proto in proto.inputs:
                found = find_name(inputs, in_proto.name)
                assert found or in_proto.dispensable, "Input {} not found".format(
                    in_proto.name)

                if found:
596 597 598 599
                    in_args = inputs[in_proto.name]
                    if not isinstance(in_args, list):
                        in_args = [in_args]
                    if not in_proto.duplicable and len(in_args) > 1:
Y
Yang Yang(Tony) 已提交
600 601
                        raise ValueError(
                            "Input %s expects only one input, but %d are given."
602 603 604
                            % (in_proto.name, len(in_args)))
                    in_arg_names = []
                    for arg in in_args:
605
                        if isinstance(arg, six.string_types):
Y
Yang Yu 已提交
606
                            in_arg_names.append(arg)
607 608
                        elif isinstance(arg, six.binary_type):
                            in_arg_names.append(arg.decode())
Y
Yang Yu 已提交
609
                        else:
M
minqiyang 已提交
610
                            in_arg_names.append(cpt.to_text(arg.name))
611
                    self.desc.set_input(in_proto.name, in_arg_names)
Y
Yang Yang(Tony) 已提交
612 613
                else:
                    self.desc.set_input(in_proto.name, [])
F
Update  
fengjiayi 已提交
614

Y
Yu Yang 已提交
615
        if outputs is not None:
616 617 618 619 620 621 622
            given = set()
            need = set()
            for n in outputs:
                given.add(n)
            for m in proto.outputs:
                need.add(m.name)
            if not given == need:
C
caoying03 已提交
623 624
                raise ValueError(("Incorrect setting for output(s) of "
                                  "operator \"%s\". Need: [%s] Given: [%s]") %
625 626 627
                                 (type,
                                  ", ".join(six.binary_type(e) for e in need),
                                  ", ".join(six.binary_type(e) for e in given)))
628

F
fengjiayi 已提交
629
            for out_proto in proto.outputs:
630 631 632 633
                out_args = outputs[out_proto.name]
                if not isinstance(out_args, list):
                    out_args = [out_args]
                if not out_proto.duplicable and len(out_args) > 1:
F
Update  
fengjiayi 已提交
634 635
                    raise ValueError(
                        "Output %s expects only one output, but %d are given." %
636 637 638
                        (out_proto.name, len(out_args)))
                out_arg_names = []
                for arg in out_args:
M
minqiyang 已提交
639
                    out_arg_names.append(cpt.to_text(arg.name))
640 641
                    arg.op = self
                self.desc.set_output(out_proto.name, out_arg_names)
F
Update  
fengjiayi 已提交
642

G
gongweibao 已提交
643 644
        if op_attrs is not None:
            if not isinstance(op_attrs, dict):
645
                raise TypeError("'attrs' should be a dict.")
F
fengjiayi 已提交
646
            for attr in proto.attrs:
F
Update  
fengjiayi 已提交
647
                attr_name = attr.name
G
gongweibao 已提交
648
                if (attr_name not in op_attrs) or (op_attrs[attr_name] is None):
F
Update  
fengjiayi 已提交
649
                    continue
G
gongweibao 已提交
650
                attr_val = op_attrs[attr_name]
G
gongweibao 已提交
651 652
                self._update_desc_attr(attr_name, attr_val)

653
        self.desc.check_attrs()
W
Wu Yi 已提交
654
        if self._has_kernel(type):
Q
QI JUN 已提交
655
            self.desc.infer_var_type(self.block.desc)
Y
Yu Yang 已提交
656
            self.desc.infer_shape(self.block.desc)
F
fengjiayi 已提交
657

W
Wu Yi 已提交
658
    def _has_kernel(self, op_type):
659 660
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
661
    def to_string(self, throw_on_error):
662
        """
663 664
        Get debug string.

665
        Args:
666 667
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
668

669 670
        Returns:
            str: The debug string.
671 672

        """
673
        protostr = self.desc.serialize_to_string()
674
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
675 676 677 678
        return _debug_string_(proto, throw_on_error)

    def __str__(self):
        return self.to_string(True)
679 680 681

    __repr__ = __str__

F
fengjiayi 已提交
682 683 684 685 686
    @property
    def type(self):
        return self.desc.type()

    def input(self, name):
687
        """
688
        Get the input arguments according to the input parameter name.
689

690 691
        Args:
            name(str): The input parameter name.
692

693 694 695
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
696
        """
F
fengjiayi 已提交
697 698
        return self.desc.input(name)

W
Wu Yi 已提交
699
    def _rename_input(self, old_name, new_name):
700 701 702 703 704 705 706 707 708 709
        """
        Rename the `old_name` to `new_name`.

        Args:
            old_name(str): The old name of the Operator's input.
            new_name(str): The new name of the Operator's input.

        Returns:
            None
        """
W
Wu Yi 已提交
710
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
711

W
Wu Yi 已提交
712
    def _rename_output(self, old_name, new_name):
713 714 715 716 717 718 719 720 721 722
        """
        Rename the `old_name` to `new_name`.

        Args:
            old_name(str): The old name of the Operator's output.
            new_name(str): The new name of the Operator's output.

        Returns:
            None
        """
W
Wu Yi 已提交
723
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
724

F
fengjiayi 已提交
725 726 727 728
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
729 730 731 732 733 734 735 736
    @property
    def input_arg_names(self):
        return self.desc.input_arg_names()

    @property
    def output_arg_names(self):
        return self.desc.output_arg_names()

F
fengjiayi 已提交
737
    def output(self, name):
738
        """
739
        Get output arguments by the output parameter name.
740

741 742
        Args:
            name(str): The output parameter name.
743

744 745 746
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
747
        """
F
fengjiayi 已提交
748 749 750 751 752 753
        return self.desc.output(name)

    @property
    def output_names(self):
        return self.desc.output_names()

754 755 756 757 758 759 760 761
    @property
    def idx(self):
        for i, op in enumerate(self.block.ops):
            if op == self:
                return i
        raise ValueError(
            "Can't find op itself in it's block. It could be a bug of Paddle.")

F
fengjiayi 已提交
762
    def has_attr(self, name):
763
        """
764 765
        Whether this Operator has the attribute with name or not.

766
        Args:
767
            name(str): the attribute name.
768

769 770
        Returns:
            bool: True if has this attribute.
771 772

        """
F
fengjiayi 已提交
773 774 775
        return self.desc.has_attr(name)

    def attr_type(self, name):
776
        """
777
        Get the type of attribute by attribute's name.
778

779 780
        Args:
            name(str): the attribute name.
781

782 783
        Returns:
            core.AttrType: the attribute type.
784
        """
F
fengjiayi 已提交
785 786
        return self.desc.attr_type(name)

W
Wu Yi 已提交
787
    def _set_attr(self, name, val):
788 789 790 791 792 793 794 795 796 797
        """
        Set the value of attribute by attribute's name.

        Args:
            name(str): the attribute name.
            val(bool|int|str|float|list): the value of the attribute.

        Raises:
            ValueError: If the type of value doesn't match with desc.attr_type(name).
        """
G
gongweibao 已提交
798 799 800 801 802 803 804 805 806 807 808 809 810
        self._update_desc_attr(name, val)

    def _update_desc_attr(self, name, val):
        """
        Update the value of desc's attribute by attribute's name.

        Args:
            name(str): the attribute name.
            val(bool|int|str|float|list): the value of the attribute.

        Raises:
            ValueError: If the type of value doesn't match with desc.attr_type(name).
        """
Q
Qiyang Min 已提交
811 812
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
813 814
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
815
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
816 817 818 819
        elif isinstance(val, core.BlockDesc) or \
                isinstance(val, core.ProgramDesc):
            self.desc.set_serialized_attr(name, val.serialize_to_string())
        else:
W
Wu Yi 已提交
820
            self.desc._set_attr(name, val)
Y
yuyang18 已提交
821

F
fengjiayi 已提交
822 823 824 825 826
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
827
        """
828 829
        Get the attribute by name.

830
        Args:
831
            name(str): the attribute name.
832

833 834
        Returns:
            bool|int|str|float|list: The attribute value. The return value
835 836
            can be any valid attribute type.
        """
F
fengjiayi 已提交
837
        return self.desc.attr(name)
Y
Yu Yang 已提交
838

W
Wu Yi 已提交
839
    def _block_attr_id(self, name):
840
        """
G
gongweibao 已提交
841
        Get the block attribute's id by name.
842

843 844
        Args:
            name(str): the attribute name.
845

846 847
        Returns:
            int: the block index.
848
        """
W
Wu Yi 已提交
849
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
850

W
Wu Yi 已提交
851
    def _block_attr(self, name):
G
gongweibao 已提交
852 853 854 855 856 857 858 859 860 861
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
862
        id = self._block_attr_id(name)
G
gongweibao 已提交
863 864 865
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
866
    def _blocks_attr(self, name):
G
gongweibao 已提交
867 868 869 870 871 872 873 874 875 876
        """
        Get the blocks attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            list: list of the blocks attribute.
        """
        attrs = []
W
Wu Yi 已提交
877
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
878 879 880 881 882
            assert (i >= 0 and i < len(self.block.program.blocks))
            attrs.append(self.block.program.blocks[i])

        return attrs

W
Wu Yi 已提交
883
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
884 885 886 887 888 889 890 891 892 893
        """
        Get the blocks attribute's ids by name.

        Args:
            name(str): the attribute name.

        Returns:
            list: list of the blocks ids.
        """

W
Wu Yi 已提交
894
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
895

J
JiayiFeng 已提交
896
    def all_attrs(self):
F
fengjiayi 已提交
897
        """
898 899 900
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
901
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
902 903 904 905
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
G
gongweibao 已提交
906 907
            attr_type = self.desc.attr_type(n)
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
908
                attr_map[n] = self._block_attr(n)
G
gongweibao 已提交
909 910 911
                continue

            if attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
912
                attr_map[n] = self._blocks_attr(n)
G
gongweibao 已提交
913 914 915 916
                continue

            attr_map[n] = self.attr(n)

F
fengjiayi 已提交
917 918
        return attr_map

Y
Yu Yang 已提交
919

Y
Yu Yang 已提交
920
class Block(object):
921 922 923 924 925 926 927 928 929 930 931 932 933 934
    """
    In Fluid, a Program is consistence of multi-Block, and Block stores
    VarDesc and OpDesc. In a specific Block, a VarDesc have a unique name.
    One block could have some child blocks, and child block's name scopes
    should inherit the parent's so that OpDesc in child block can reference
    a VarDesc that is stored in the parent block.
    Please reference the framework.proto for details.

    Args:
        program(Program): The Program that the Block belongs to.
        idx(int): The block's id in the Program.

    Notes:
        The constructor of Block should not be invoked directly. Please
W
Wu Yi 已提交
935
        use `Program._create_block()` to create a block.
936 937 938 939 940 941 942 943 944 945 946 947 948 949

    Examples:
        .. code-block:: python

            cur_program = Program()
            cur_block = cur_program.current_block()
            var = cur_block.create_var(name="X",
                                       shape=[-1, 23, 48],
                                       dtype='float32')
            cur_block.append_op(type="abs",
                                inputs={"X": [var]},
                                outputs={"Out": [var]})
    """

Y
Yu Yang 已提交
950
    def __init__(self, program, idx):
Y
Yu Yang 已提交
951
        self.desc = program.desc.block(idx)
952
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
953
        self.ops = list()  # operator list
Y
Yu Yang 已提交
954
        self.program = program
955
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
956

957
    def __str__(self):
Y
Yang Yang(Tony) 已提交
958 959
        return self.to_string(True)

F
fengjiayi 已提交
960 961
    def to_string(self, throw_on_error, with_details=False):
        """
962 963
        Get debug string.

F
fengjiayi 已提交
964 965
        Args:
            throw_on_error(bool): raise exception when self is not initialized
966
                when throw_on_error is True.
F
update  
fengjiayi 已提交
967
            with_details(bool): more details about variables and parameters
968 969
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
970

971 972
        Returns:
            str: The debug string.
F
fengjiayi 已提交
973 974 975 976
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
977
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
978 979
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
980
            for var in list(self.vars.values()):
F
fengjiayi 已提交
981
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
982
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
983
            for op in self.ops:
F
fengjiayi 已提交
984 985
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
986 987 988
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
989 990
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
991 992
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
993 994 995

    __repr__ = __str__

Y
Yu Yang 已提交
996 997
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
998
        return self.desc.parent
Y
Yu Yang 已提交
999

Y
Yu Yang 已提交
1000 1001 1002 1003
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
1004
    def _set_forward_block_idx(self, idx):
1005 1006 1007 1008 1009 1010 1011 1012 1013
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
1014
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
1015

Y
Yu Yang 已提交
1016 1017
    @property
    def idx(self):
Y
Yu Yang 已提交
1018
        return self.desc.id
Y
Yu Yang 已提交
1019

Q
Qiao Longfei 已提交
1020
    def var(self, name):
1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033
        """
        Get a Variable by name from this block.

        Args:
            name(str): the Variable's name.

        Raises:
            ValueError: The If input's type is not str, or this block
                doesn't have a Variable with the giving name.

        Returns:
            Variable: the Variable with the giving name.
        """
1034
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
1035 1036 1037
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
1038 1039
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
1040
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
1041
        return v
Q
Qiao Longfei 已提交
1042

W
Wu Yi 已提交
1043
    def _var_recursive(self, name):
1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056
        """
        Get a Variable by name from this block recursively.

        Args:
            name(str): the Variable's name.

        Raises:
            ValueError: this block and this parent block doesn't
                have a Variable with the giving name.

        Returns:
            Variable: the Variable with the giving name.
        """
Y
Yu Yang 已提交
1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082
        frontier = list()
        visited = set()

        frontier.append(self)

        prog = self.program

        while len(frontier) != 0:  # BFS
            cur = frontier[0]
            frontier = frontier[1:]

            if id(cur) in visited:
                continue

            if cur.has_var(name):
                return cur.var(name)

            if cur.parent_idx != -1:
                frontier.append(prog.block(cur.parent_idx))

            if cur.forward_block_idx != -1:
                frontier.append(prog.block(cur.forward_block_idx))

            visited.add(id(cur))

        raise ValueError("Var {0} is not found recursively".format(name))
F
fengjiayi 已提交
1083

Q
Qiao Longfei 已提交
1084
    def all_parameters(self):
1085
        return list(self.iter_parameters())
1086

1087
    def iter_parameters(self):
M
minqiyang 已提交
1088
        return (item[1] for item in six.iteritems(self.vars)
1089
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
1090

Y
Yu Yang 已提交
1091
    def create_var(self, *args, **kwargs):
1092
        var = Variable(block=self, *args, **kwargs)
1093 1094
        if 'initializer' in kwargs:
            kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
1095
        return var
Y
Yu Yang 已提交
1096

Q
Qiao Longfei 已提交
1097 1098 1099
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
1100
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
1101 1102
        """
        Rename variable in vars and ops' inputs and outputs
1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114

        Args:
            name(str): the name that need to be renamed.
            new_name(str): the name that need to rename to.

        Raises:
            ValueError: If this block doesn't have this the giving name,
                or the type of the var with the giving name is not Parameter
                or Variable.

        Returns:
            Variable: the Variable with the giving name.
T
typhoonzero 已提交
1115
        """
M
minqiyang 已提交
1116 1117
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
1118

T
typhoonzero 已提交
1119
        if not self.has_var(name):
1120
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
1121 1122
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
1123
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
1124 1125 1126 1127 1128 1129 1130
            stop_gradient = v.stop_gradient
            trainable = v.trainable
            optimize_attr = v.optimize_attr
            regularizer = v.regularizer
            gradient_clip_attr = v.gradient_clip_attr
            error_clip = v.error_clip
        elif type(v) == Variable:
T
typhoonzero 已提交
1131
            var_type = "Variable"
T
wip  
typhoonzero 已提交
1132 1133 1134 1135
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
1136
        orig_var_type = v.type
M
minqiyang 已提交
1137
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
1138
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
1139
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
1140
        if var_type == "Parameter":
T
wip  
typhoonzero 已提交
1141 1142 1143 1144
            var = Parameter(
                self,
                d.shape(),
                d.dtype(),
T
typhoonzero 已提交
1145
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1146 1147 1148 1149 1150 1151 1152
                name=new_name,
                stop_gradient=stop_gradient,
                trainable=trainable,
                optimize_attr=optimize_attr,
                regularizer=regularizer,
                gradient_clip_attr=gradient_clip_attr,
                error_clip=error_clip)
T
typhoonzero 已提交
1153
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
1154 1155
            var = Variable(
                self,
T
typhoonzero 已提交
1156
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1157 1158 1159 1160
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
1161
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
1162 1163 1164
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
1165
        self._sync_with_cpp()
1166
        return var
T
typhoonzero 已提交
1167

W
Wu Yi 已提交
1168 1169
    def _remove_var(self, name):
        self._sync_with_cpp()
M
minqiyang 已提交
1170
        self.desc._remove_var(cpt.to_bytes(name))
1171 1172
        del self.vars[name]

Y
Yu Yang 已提交
1173 1174
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
Q
Qiao Longfei 已提交
1175
        param = Parameter(global_block, *args, **kwargs)
1176
        if 'initializer' in kwargs:
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196

            def _is_inited_by(block, var):
                init_ops = []
                for op in block.ops:
                    if var.name in op.output_arg_names:
                        init_ops.append(op)
                return init_ops

            initializer = kwargs['initializer']
            init_ops = _is_inited_by(global_block, param)
            init_ops_len = len(init_ops)
            if init_ops_len > 1:
                raise RuntimeError("param " + param.name +
                                   " is inited by multiple init ops " + str(
                                       init_ops))
            elif init_ops_len == 1:
                #TODO already inited, do nothing, should log a warning
                pass
            else:
                initializer(param, self)
Q
Qiao Longfei 已提交
1197
        return param
Y
Yu Yang 已提交
1198

Y
Yu Yang 已提交
1199
    def append_op(self, *args, **kwargs):
1200 1201 1202 1203 1204 1205
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
Y
Yu Yang 已提交
1206
        op_desc = self.desc.append_op()
1207
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
Y
Yu Yang 已提交
1208 1209 1210
        self.ops.append(op)
        return op

W
Wu Yi 已提交
1211
    def _insert_op(self, index, *args, **kwargs):
1212 1213 1214 1215 1216 1217 1218 1219 1220
        """
        Insert a Operator according to the giving arguments.

        Args:
            index(int): the place that the operator to insert.

        Returns:
            Operator: the insert Operator.
        """
W
Wu Yi 已提交
1221 1222
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
1223 1224 1225 1226
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

W
Wu Yi 已提交
1227
    def _remove_op(self, index):
1228 1229 1230 1231 1232 1233 1234 1235 1236
        """
        Remove the specific position operator.

        Args:
            index(int): the position that the operator to insert.

        Returns:
            None
        """
W
Wu Yi 已提交
1237 1238
        self._sync_with_cpp()
        self.desc._remove_op(index, index + 1)
1239 1240
        del self.ops[index]

W
Wu Yi 已提交
1241
    def _slice_ops(self, start, end):
1242 1243 1244 1245 1246 1247 1248 1249 1250 1251
        """
        Return the Operator between start and end.

        Args:
            start(int): the start position.
            end(int): the end position.

        Returns:
            list: the Operators between start and end.
        """
Q
qiaolongfei 已提交
1252
        return self.ops[start:end]
Y
Yancey1989 已提交
1253

W
Wu Yi 已提交
1254 1255
    def _prepend_op(self, *args, **kwargs):
        op_desc = self.desc._prepend_op()
Y
Yu Yang 已提交
1256
        op = Operator(self, op_desc, *args, **kwargs)
Q
qiaolongfei 已提交
1257
        self.ops.insert(0, op)
Y
Yu Yang 已提交
1258 1259
        return op

W
Wu Yi 已提交
1260
    def _sync_with_cpp(self):
1261
        """
1262 1263
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
1264
        """
Q
Qiao Longfei 已提交
1265 1266 1267 1268 1269
        # sync variables from cpp
        for var in self.desc.all_vars():
            if not self.has_var(var.name()):
                self.create_var(name=var.name(), desc=var, type=var.type())

1270
        # sync variables removed from c++ end
1271
        for var in list(self.vars.keys()):
M
minqiyang 已提交
1272
            if not self.desc.find_var(cpt.to_bytes(var)):
1273 1274
                self.vars.pop(var)

Q
Qiao Longfei 已提交
1275
        # sync operators from cpp
1276 1277 1278 1279
        ops_in_cpp = []
        for op_idx in range(0, self.desc.op_size()):
            ops_in_cpp.append(self.desc.op(op_idx))

Y
Yu Yang 已提交
1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295
        if len(self.ops) != 0:
            first_op_in_python = self.ops[0].desc
            last_op_in_python = self.ops[len(self.ops) - 1].desc
            start_index = None
            end_index = None
            for index in range(len(ops_in_cpp)):
                if first_op_in_python == ops_in_cpp[index]:
                    start_index = index
                if last_op_in_python == ops_in_cpp[index]:
                    end_index = index
            assert start_index is not None
            assert end_index is not None
            assert start_index <= end_index
        else:
            start_index = 0
            end_index = -1
Q
Qiao Longfei 已提交
1296 1297 1298 1299 1300

        # sync ops append to the head of cpp_ops
        for index in range((start_index - 1 - 1), -1, -1):
            op_desc = ops_in_cpp[index]
            op = Operator(self, op_desc)
Q
qiaolongfei 已提交
1301
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
1302 1303 1304 1305 1306 1307 1308

        # sync ops append to the end of cpp_ops
        for index in range((end_index + 1), len(ops_in_cpp)):
            op_desc = ops_in_cpp[index]
            op = Operator(self, op_desc)
            self.ops.append(op)

1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321
        # sync ops removed from c++ end
        if end_index != -1 and end_index < len(self.ops):
            ops_in_cpp_index = 0
            ops_in_python_index = 0
            while ops_in_python_index < len(
                    self.ops) and ops_in_cpp_index < len(ops_in_cpp):
                if self.ops[ops_in_python_index].desc != ops_in_cpp[
                        ops_in_cpp_index]:
                    del self.ops[ops_in_python_index]
                else:
                    ops_in_cpp_index += 1
                    ops_in_python_index += 1

Q
Qiao Longfei 已提交
1322 1323 1324 1325
        assert len(self.ops) == len(ops_in_cpp)
        for index in range(len(self.ops)):
            assert self.ops[index].desc == ops_in_cpp[index]

W
Wu Yi 已提交
1326
    def _copy_param_info_from(self, other):
1327
        """
1328 1329
        Copy the information of parameters from the other block.

1330
        Args:
1331 1332 1333 1334 1335
            other(Block): the other block.

        Raises:
            ValueError: If type of input is not Block, or the `other` and this
                block is not in the same topology.
1336 1337 1338 1339 1340

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
1341 1342
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
1343
        for p in other.iter_parameters():
1344 1345 1346
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
W
Wu Yi 已提交
1347
                raise ValueError("_copy_param_info_from should be invoked with "
1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359
                                 "same topology")
            assert isinstance(v, Variable)
            new_p = Parameter(
                block=self,
                shape=v.shape,
                dtype=v.dtype,
                type=v.type,
                lod_level=v.lod_level,
                stop_gradient=p.stop_gradient,
                trainable=p.trainable,
                optimize_attr=p.optimize_attr,
                regularizer=p.regularizer,
F
fengjiayi 已提交
1360
                gradient_clip_attr=p.gradient_clip_attr,
F
fengjiayi 已提交
1361
                error_clip=p.error_clip,
1362 1363 1364
                name=v.name)
            self.vars[new_p.name] = new_p

W
Wu Yi 已提交
1365
    def _clone_variable(self, var):
1366 1367
        """
        Clone a variable into current block.
1368

1369 1370 1371 1372
        Args:
            var: the variable to be cloned.

        Returns:
1373
            Variable: the new  variable cloned from 'var' in current block.
1374 1375
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
1376 1377 1378 1379 1380
        ret_var = None
        # make STEP_SCOPES var can be safely cloned.
        if var.type == core.VarDesc.VarType.STEP_SCOPES:
            ret_var = self.create_var(
                name=var.name, persistable=var.persistable, type=var.type)
T
tangwei12 已提交
1381 1382
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
1383
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
1384 1385 1386 1387 1388 1389
        elif var.type == core.VarDesc.VarType.SELECTED_ROWS:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
F
fengjiayi 已提交
1390 1391
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1392 1393 1394 1395 1396 1397 1398
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
F
fengjiayi 已提交
1399 1400
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1401
        return ret_var
1402

Y
Yu Yang 已提交
1403 1404

class Program(object):
D
dzhwinter 已提交
1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415
    """
    Python Program. Beneath it is a ProgramDesc, which is used for
    create c++ Program. A program is a self-contained programing
    language like container. It has at least one Block, when the
    control flow op like conditional_block, while_op is included,
    it will contains nested block.
    Please reference the framework.proto for details.

    Notes: we have default_startup_program and default_main_program
    by default, a pair of them will shared the parameters.
    The default_startup_program only run once to initialize parameters,
Y
yuyang18 已提交
1416
    default_main_program run in every mini batch and adjust the weights.
D
dzhwinter 已提交
1417 1418

    Returns:
Y
yuyang18 已提交
1419
        A empty program.
D
dzhwinter 已提交
1420 1421

    Examples:
Y
yuyang18 已提交
1422 1423 1424 1425 1426 1427
        >>> main_program = fluid.Program()
        >>> startup_program = fluid.Program()
        >>> with fluid.program_guard(main_program=main_program, startup_program=startup_program):
        >>>     fluid.layers.data(name="x", shape=[-1, 784], dtype='float32')
        >>>     fluid.layers.data(name="y", shape=[-1, 1], dtype='int32')
        >>>     fluid.layers.fc(name="fc", shape=[10], dtype='float32', act="relu")
D
dzhwinter 已提交
1428 1429 1430

    """

1431 1432
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
1433 1434
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
D
dzhwinter 已提交
1435
        self._seed = 0
Y
yuyang18 已提交
1436
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
Y
yuyang18 已提交
1437
        self._op_role_var = []
T
tangwei12 已提交
1438 1439 1440 1441

        # for distribute
        self._is_distributed = False
        self._is_chief = False
T
tangwei12 已提交
1442
        self._slice_vars_and_attrs = []
T
tangwei12 已提交
1443 1444
        self._endpoints = []
        self._distributed_lookup_table = None
Y
yuyang18 已提交
1445 1446 1447

    @property
    def op_role(self):
Y
yuyang18 已提交
1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460
        """
        The operator role. In a enum {Forward, Backward, Optimize}.

        Notes: this is a low level API. It is used only for ParallelExecutor to
        duplicate or schedule operator to devices.

        For example, the forward operator should be executed on every device.
        The backward operator should be executed on every device and the
        parameter gradient of backward (use :code:`op_role_var` to get this
        variable) operator should be merged to one device. The optimization
        operators should be executed on only one device and broadcast the
        optimization result, i.e., the new parameter, to every other device.
        """
Y
yuyang18 已提交
1461 1462 1463 1464 1465 1466 1467 1468
        return self._current_role

    @op_role.setter
    def set_op_role(self, role):
        self._current_role = role

    @property
    def op_role_var(self):
Y
yuyang18 已提交
1469 1470 1471 1472 1473 1474 1475
        """
        The auxiliary variables for :code:`op_role` property.

        See Also: :code:`Program.op_role`'s documentation for details.

        Notes: This is a very low-level API. Users should not use it directly.
        """
Y
yuyang18 已提交
1476 1477 1478 1479
        return self._op_role_var

    @op_role_var.setter
    def set_op_role_var(self, var_name):
Y
yuyang18 已提交
1480
        self._op_role_var = [var_name]
Y
yuyang18 已提交
1481 1482

    @contextlib.contextmanager
W
Wu Yi 已提交
1483
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
1484 1485 1486 1487 1488 1489 1490
        """
        A with guard to set :code:`Optimization` :code:`OpRole` and
        :code:`OpRoleVar` automatically.

        Notes: This is a very low level API. Users should not use it directly.

        Args:
1491
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
1492 1493 1494 1495

        Examples:

            >>> p, g = backward(...)
W
Wu Yi 已提交
1496
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
1497 1498
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
1499 1500 1501
        tmp_role = self._current_role
        tmp_var = self._op_role_var

Y
yuyang18 已提交
1502 1503
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
1504 1505 1506 1507
        self._op_role_var = [
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
Y
yuyang18 已提交
1508
        yield
X
Xin Pan 已提交
1509 1510
        self._op_role_var = tmp_var
        self._current_role = tmp_role
Y
Yu Yang 已提交
1511

1512
    @contextlib.contextmanager
X
Xin Pan 已提交
1513
    def _lr_schedule_guard(self, is_with_opt=False):
1514 1515 1516 1517 1518 1519 1520
        """
        A with guard to set :code:`LRSched` :code:`OpRole` and
        :code:`OpRoleVar` automatically. The :code:`OpRoleVar` is
        set to the target learning rate.

        Notes: This is a very low level API. Users should not use it directly.

X
Xin Pan 已提交
1521 1522 1523 1524
        Args:
            is_with_opt: Only set to true if these ops a in the middle
                 of a bunch of optimize ops so that it can be treated
                 correctly. For example, sgd->lr_op->sgd->lr_op->sgd.
1525 1526 1527 1528 1529 1530 1531

        Examples:

            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
1532 1533 1534 1535

        tmp_role = self._current_role
        tmp_var = self._op_role_var

1536 1537
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
1538 1539
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
1540 1541 1542
        # TODO(typhoonzero): how to set target learning rate var
        self._op_role_var = []
        yield
1543 1544
        self._op_role_var = tmp_var
        self._current_role = tmp_role
1545

1546
    def __str__(self):
Y
yuyang18 已提交
1547 1548 1549 1550 1551 1552 1553 1554 1555
        """
        Get the protobuf debug string of this Program.

        Returns:
            (str): The protobuf debug string.

        Raises:
            ValueError: If any of required fields is not set.
        """
Y
Yang Yang(Tony) 已提交
1556 1557
        return self.to_string(True)

F
fengjiayi 已提交
1558 1559 1560
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
1561

F
fengjiayi 已提交
1562
        Args:
Y
yuyang18 已提交
1563 1564
            throw_on_error(bool): raise Value error when any of required fields
                is not set.
F
fengjiayi 已提交
1565

Y
yuyang18 已提交
1566 1567 1568 1569 1570 1571 1572 1573 1574 1575
            with_details(bool): True if more details about variables and
                parameters, e.g., :code:`trainable`, :code:`optimize_attr`, need
                to print.

        Returns
            (str): The debug string.

        Raises:
            ValueError: If any of required fields is not set and throw_on_error is
                True.
F
fengjiayi 已提交
1576 1577 1578 1579 1580 1581 1582 1583 1584 1585

        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
            res_str = ""
            for block in self.blocks:
                res_str += block.to_string(throw_on_error, with_details)
        else:
            protostr = self.desc.serialize_to_string()
1586 1587
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
1588 1589
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1590

W
Wu Yi 已提交
1591
    def _get_desc(self):
Y
yuyang18 已提交
1592 1593 1594 1595 1596 1597 1598
        """
        Get the C++ side of `ProgramDesc` object pointer. The C++ object is
        exposed by :code:`pybind`.

        Notes: This is a very low level API. Users should not use this API
        directly.
        """
1599 1600
        return self.desc

X
version  
Xin Pan 已提交
1601 1602 1603
    def _version(self):
        return self.desc._version()

1604
    def clone(self, for_test=False):
Y
yuyang18 已提交
1605 1606 1607
        """
        Create a new, duplicated program.

1608

Y
yuyang18 已提交
1609 1610 1611 1612
        Some operators, e.g., :code:`batch_norm`, behave differently between
        training and testing. They have an attribute, :code:`is_test`, to
        control this behaviour. This method will change the :code:`is_test`
        attribute of them to :code:`True` when :code:`for_test=True`.
1613

Y
yuyang18 已提交
1614 1615 1616 1617
        * Set for_test to False when we want to clone the program for training.
        * Set for_test to True when we want to clone the program for testing.

        Notes: This API DOES NOT prune any operator. Use
L
Luo Tao 已提交
1618 1619 1620 1621 1622
        :code:`clone(for_test=True)` before backward and optimization please. e.g.

            >>> test_program = fluid.default_main_program().clone(for_test=True)
            >>> optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
            >>> optimizer.minimize()
1623 1624

        Args:
Y
yuyang18 已提交
1625 1626
            for_test(bool): True if change the :code:`is_test` attribute of
                operators to :code:`True`.
1627

D
dzhwinter 已提交
1628
        Returns:
Y
yuyang18 已提交
1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681
            Program: The new, duplicated Program object.

        Examples:

            1. To clone a test program, the sample code is:

            >>> import paddle.fluid as fluid
            >>> train_program = fluid.Program()
            >>> startup_program = fluid.Program()
            >>> with fluid.program_guard(train_program, startup_program):
            >>>     img = fluid.layers.data(name='image', shape=[784])
            >>>     hidden = fluid.layers.fc(input=img, size=200, act='relu')
            >>>     hidden = fluid.layers.dropout(hidden, dropout_prob=0.5)
            >>>     loss = fluid.layers.cross_entropy(
            >>>                 input=fluid.layers.fc(hidden, size=10, act='softmax'),
            >>>                 label=fluid.layers.data(name='label', shape=[1], dtype='int64'))
            >>>
            >>> test_program = train_program.clone(for_test=True)
            >>>
            >>> sgd = fluid.optimizer.SGD(learning_rate=1e-3)
            >>> with fluid.program_guard(train_program, startup_program):
            >>>     sgd.minimize(loss)

            2. The :code:`clone` method can be avoid if you create program for
            training and program for testing individually.

            >>> import paddle.fluid as fluid
            >>>
            >>> def network(is_test):
            >>>     img = fluid.layers.data(name='image', shape=[784])
            >>>     hidden = fluid.layers.fc(input=img, size=200, act='relu')
            >>>     hidden = fluid.layers.dropout(hidden, dropout_prob=0.5, is_test=is_test)
            >>>     loss = fluid.layers.cross_entropy(
            >>>                 input=fluid.layers.fc(hidden, size=10, act='softmax'),
            >>>                 label=fluid.layers.data(name='label', shape=[1], dtype='int64'))
            >>>     return loss
            >>>
            >>> train_program = fluid.Program()
            >>> startup_program = fluid.Program()
            >>> test_program = fluid.Program()
            >>>
            >>> with fluid.program_guard(train_program, startup_program):
            >>>     with fluid.unique_name.guard():
            >>>         loss = network(is_test=False)
            >>>         sgd = fluid.optimizer.SGD(learning_rate=1e-3)
            >>>         sgd.minimize(loss)
            >>>
            >>> # the test startup program is not used.
            >>> with fluid.program_guard(test_program, fluid.Program()):
            >>>     with fluid.unique_name.guard():
            >>>         loss = network(is_test=True)

            The two code snippets above will generate same programs.
1682 1683
        """
        if for_test:
X
Xin Pan 已提交
1684
            p = self._inference_optimize(prune_read_op=False)
1685
        else:
1686
            p = Program()
G
gongweibao 已提交
1687 1688
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
1689
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
1690 1691 1692
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
1693 1694 1695 1696

            p._current_role = self._current_role
            p._op_role_var = self._op_role_var

W
Wu Yi 已提交
1697
            p._sync_with_cpp()
1698

W
Wu Yi 已提交
1699
        p._copy_param_info_from(self)
W
Wu Yi 已提交
1700
        p._copy_data_info_from(self)
1701
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
1702
        return p
1703

W
Wu Yi 已提交
1704
    def _prune(self, targets):
Y
yuyang18 已提交
1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719
        """
        Prune operators and variables which are not needed to generate
        :code:`targets`.

        Notes: This is a very low level API. Users should not use this API
        directly. This API is in flux and not stable.

        Args:
            targets(list|Variable|Operator): A list of variables or operators
                need to be pruned

        Returns:
            Program:  A new, pruned program.

        """
1720 1721 1722 1723 1724 1725
        if not isinstance(targets, list):
            targets = [targets]
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
1726 1727
                    # After transpiler processing, the op that output this
                    # variable maybe has been changed, so t.op is not reliable
1728
                    # and we need to find the current op that generate this
1729 1730 1731 1732 1733 1734 1735 1736
                    # variable here.
                    t.op = None
                    global_block = self.global_block()
                    for idx, op in enumerate(global_block.ops):
                        if t.name in op.output_arg_names:
                            t.op = op
                            break

1737
                    t = t.op
1738 1739 1740 1741
                    if t is None:
                        raise ValueError(
                            "The target variable must have an "
                            "associated operator that generates it.")
1742
                else:
1743 1744
                    raise ValueError("All targets of prune() can only be "
                                     "Variable or Operator.")
1745 1746 1747 1748

            targets_idx.append([t.block.idx, t.idx])
        res = Program()
        res.desc = core.prune(self.desc, targets_idx)
M
minqiyang 已提交
1749 1750 1751
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
1752
        res._sync_with_cpp()
1753 1754
        return res

X
Xin Pan 已提交
1755
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
1756
        """
F
fengjiayi 已提交
1757 1758 1759 1760 1761
        This method will create a new program and do following adjustments on it:
        1. Remove all reader variables and their creator ops if exist.

        2. Remove the :code:`read_op` if exists.

1762
        3. change the :code:`is_test`
Y
yuyang18 已提交
1763 1764 1765
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

1766
        Args:
X
Xin Pan 已提交
1767 1768
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
1769

Y
yuyang18 已提交
1770 1771 1772 1773 1774 1775
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
1776
        res = Program()
1777
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
1778 1779 1780 1781

        # remove all readers and the read_op if exist
        read_op_idx = 0
        root_block = res.desc.block(0)
X
Xin Pan 已提交
1782
        if prune_read_op:
1783 1784 1785 1786 1787 1788 1789 1790 1791
            while True:
                if read_op_idx >= root_block.op_size() or root_block.op(
                        read_op_idx).type() == 'read':
                    break
                read_op_idx += 1
            if read_op_idx < root_block.op_size():
                root_block._remove_op(0, read_op_idx + 1)
            for var in root_block.all_vars():
                if var.type() == core.VarDesc.VarType.READER:
M
minqiyang 已提交
1792
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
1793 1794

        # change all `is_test` attributes to True
M
minqiyang 已提交
1795
        for i in six.moves.range(res.desc.num_blocks()):
1796
            block = res.desc.block(i)
M
minqiyang 已提交
1797
            for j in six.moves.range(block.op_size()):
1798 1799
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
1800
                    op._set_attr('is_test', True)
M
minqiyang 已提交
1801 1802 1803
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
1804
        res._sync_with_cpp()
1805 1806
        return res

1807 1808
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
1809 1810 1811 1812 1813 1814 1815
        """
        Deserialize a program desc from protobuf binary string.

        Notes: All information about parameters will be lost after serialization
        and deserialization.

        Args:
1816
            binary_str_type(str): The binary prootbuf string.
Y
yuyang18 已提交
1817 1818 1819 1820

        Returns:
            Program: A deserialized program desc.
        """
1821 1822
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
1823
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
1824
        p._sync_with_cpp()
1825
        return p
Y
Yu Yang 已提交
1826

D
dzhwinter 已提交
1827 1828
    @property
    def random_seed(self):
Y
yuyang18 已提交
1829 1830 1831 1832 1833 1834
        """
        The default random seed for random operators in Program. Zero means get
        the random seed from random device.

        Notes: It must be set before the operators have been added.
        """
D
dzhwinter 已提交
1835 1836
        return self._seed

Q
qiaolongfei 已提交
1837 1838
    @property
    def num_blocks(self):
Y
yuyang18 已提交
1839 1840 1841
        """
        The number of blocks in this program.
        """
Q
qiaolongfei 已提交
1842 1843
        return self.desc.num_blocks()

D
dzhwinter 已提交
1844 1845 1846 1847 1848 1849
    @random_seed.setter
    def random_seed(self, seed):
        if not isinstance(seed, int):
            raise ValueError("Seed must be a integer.")
        self._seed = seed

Y
Yu Yang 已提交
1850
    def __repr__(self):
1851
        return self.__str__()
1852

Y
Yu Yang 已提交
1853
    def global_block(self):
Y
yuyang18 已提交
1854 1855 1856
        """
        Get the first block of this program.
        """
Y
Yu Yang 已提交
1857 1858
        return self.blocks[0]

Q
Qiao Longfei 已提交
1859
    def block(self, index):
Y
yuyang18 已提交
1860 1861 1862 1863 1864 1865 1866 1867
        """
        Get the :code:`index` block of this program
        Args:
            index(int): The index of block to get

        Returns:
            Block: The :code:`index` block
        """
Q
Qiao Longfei 已提交
1868 1869
        return self.blocks[index]

Y
Yu Yang 已提交
1870
    def current_block(self):
Y
yuyang18 已提交
1871 1872 1873 1874
        """
        Get the current block. The :code:`current` block is the block to append
        operators.
        """
Y
Yu Yang 已提交
1875 1876
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
1877
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
1878 1879 1880 1881 1882 1883 1884 1885 1886 1887
        """
        Create a new block with the :code:`parent_idx` and change the current block
        to new block.

        Args:
            parent_idx(int): The parent block index.

        Returns:
            Block: The new block.
        """
Y
Yu Yang 已提交
1888
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
1889 1890 1891
        parent = self.current_block() if parent_idx is None else self.block(
            parent_idx)
        self.desc.append_block(parent.desc)
Y
Yu Yang 已提交
1892 1893 1894 1895
        self.current_block_idx = new_block_idx
        self.blocks.append(Block(self, self.current_block_idx))
        return self.current_block()

W
Wu Yi 已提交
1896
    def _rollback(self):
Y
yuyang18 已提交
1897 1898 1899 1900 1901
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
1902 1903
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
1904
    def _sync_with_cpp(self):
Y
yuyang18 已提交
1905 1906 1907 1908 1909 1910 1911 1912 1913 1914
        """
        Synchronize Python instance to its binding C++ object instance.
        If the program is modified in C++ space, this method should be invoked.

        Notes: This is a very low level API. Users should not invoke it
        directly.

        Returns:
            None
        """
Q
Qiao Longfei 已提交
1915 1916 1917
        for block_idx in range(len(self.blocks), self.desc.num_blocks()):
            self.blocks.append(Block(self, block_idx))
        for block in self.blocks:
W
Wu Yi 已提交
1918
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
1919

W
Wu Yi 已提交
1920
    def _copy_param_info_from(self, other):
1921
        """
1922
        Copy the information of parameters from other program.
D
dzhwinter 已提交
1923

Y
yuyang18 已提交
1924 1925 1926
        Notes: This is a very low level API. Users should not invoke it
        directly.

1927 1928 1929 1930 1931 1932 1933
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
1934
            raise TypeError("_copy_param_info_from should be invoked with "
1935 1936 1937
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
1938
            raise ValueError("_copy_param_info_from should be invoked with two "
1939
                             "program, with represent the same topology")
W
Wu Yi 已提交
1940
        self.global_block()._copy_param_info_from(other.global_block())
1941

1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960
    def _copy_dist_param_info_from(self, other):
        """
        Copy the information of distributed information from other program.

        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
            raise TypeError("_copy_dist_param_info_from should be invoked with "
                            "Program")
        self._is_distributed = other._is_distributed
        self._is_chief = other._is_chief
        self._slice_vars_and_attrs = other._slice_vars_and_attrs
        self._endpoints = other._endpoints
        self._distributed_lookup_table = other._distributed_lookup_table

W
Wu Yi 已提交
1961
    def _copy_data_info_from(self, other):
F
fengjiayi 已提交
1962 1963
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
1964

Y
yuyang18 已提交
1965 1966 1967
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
1968 1969 1970 1971 1972 1973 1974
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
1975
            raise TypeError("_copy_param_info_from should be invoked with "
F
fengjiayi 已提交
1976 1977 1978
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
1979
            raise ValueError("_copy_param_info_from should be invoked with two "
F
fengjiayi 已提交
1980
                             "program, with represent the same topology")
1981
        for var in list(other.global_block().vars.values()):
F
fengjiayi 已提交
1982 1983 1984
            if var.is_data:
                self.global_block().var(var.name).is_data = True

1985
    def list_vars(self):
Y
yuyang18 已提交
1986 1987 1988 1989 1990 1991
        """
        Get all variables from this Program. A iterable object is returned.

        Returns:
            iterable: The generator will yield every variable in this program.
        """
1992
        for each_block in self.blocks:
1993
            for each_var in list(each_block.vars.values()):
1994 1995
                yield each_var

Y
Yu Yang 已提交
1996

Y
Yu Yang 已提交
1997
class Parameter(Variable):
1998
    """
1999
    Parameter is derived from Variable. A parameter is a persistable
2000
    Variable, and will be updated by optimizers after each iteration.
2001
    The training of a neural network is essentially the updating of
2002 2003
    its parameters.

2004
    Relative to a general Variable, a Parameter has several its own
2005 2006
    member variables:

2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018
    Args:
        trainable(bool): True if the parameter need to be updated after
            iterations.
        optimize_attr(map): Parameter attributes related with optimizing.
            Currently, it only contains 'learning_rate'.
            Default: {'learning_rate': 1.0}
        regularizer(WeightDecayRegularizer): The Regularizer which will
            be applied on the parameter. Default: None
        gradient_clip_attr(BaseGradientClipAttr): The gradint clip strategy
            which will be applied on the parameter. Default: None
        do_model_average(bool): True if the model average strategy will
            be applied on this parameter.
2019 2020
    """

Y
Yu Yang 已提交
2021 2022 2023 2024 2025 2026 2027 2028 2029 2030
    def __init__(self, block, shape, dtype, **kwargs):
        if shape is None or dtype is None:
            raise ValueError("Parameter must set shape and dtype")
        if len(shape) == 0:
            raise ValueError("Parameter shape cannot be empty")

        for each in shape:
            if each < 0:
                raise ValueError("Parameter shape should not be related with "
                                 "batch-size")
2031 2032 2033

        Variable.__init__(
            self, block, persistable=True, shape=shape, dtype=dtype, **kwargs)
Y
Yu Yang 已提交
2034 2035 2036 2037
        self.trainable = kwargs.get('trainable', True)

        self.optimize_attr = kwargs.get('optimize_attr', {'learning_rate': 1.0})

2038 2039
        self.regularizer = kwargs.get('regularizer', None)

F
fengjiayi 已提交
2040
        self.gradient_clip_attr = kwargs.get('gradient_clip_attr', None)
Y
Yu Yang 已提交
2041

W
wanghaoshuang 已提交
2042
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
2043

F
fengjiayi 已提交
2044 2045 2046
    def __str__(self):
        return self.to_string(True)

F
update  
fengjiayi 已提交
2047 2048 2049
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
2050

F
update  
fengjiayi 已提交
2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064
        Args:
            throw_on_error(bool): raise exception when self is not initialized
                when throw_on_error is True
            with_details(bool): more details about variables and parameters
                (e.g. trainable, optimize_attr, ...) will be printed when with_details is True

        Returns(str): The debug string.

        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
            res_str = Variable.to_string(self, throw_on_error, True)
            additional_attr = ("trainable", "optimize_attr", "regularizer",
W
wanghaoshuang 已提交
2065
                               "gradient_clip_attr", "do_model_average")
F
update  
fengjiayi 已提交
2066
            for attr_name in additional_attr:
2067 2068
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
2069 2070
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
2071 2072 2073 2074
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
2075

Y
Yu Yang 已提交
2076
# program is a global instance.
Y
Yu Yang 已提交
2077 2078
_main_program_ = Program()
_startup_program_ = Program()
2079

2080

2081
def default_startup_program():
Y
Yu Yang 已提交
2082
    """
Y
yuyang18 已提交
2083 2084 2085 2086 2087 2088 2089 2090 2091
    Get default/global startup program.

    The layer function in :code:`fluid.layers` will create parameters, readers,
    NCCL handles as global variables. The :code:`startup_program` will
    initialize them by the operators in startup program. The layer function will
    append these initialization operators into startup program.

    This method will return the :code:`default` or the :code:`current` startup
    program. Users can use :code:`fluid.program_guard` to switch program.
2092

Y
Yu Yang 已提交
2093 2094 2095
    Returns:
        Program: startup program
    """
Y
Yu Yang 已提交
2096
    return _startup_program_
2097

2098

2099
def default_main_program():
Y
Yu Yang 已提交
2100
    """
Y
yuyang18 已提交
2101 2102 2103 2104 2105 2106 2107 2108 2109
    Get default/global main program. The main program is used for training or
    testing.

    All layer function in :code:`fluid.layers` will append operators and
    variables to the :code:`default_main_program`.

    The :code:`default_main_program` is the default program in a lot of APIs.
    For example, the :code:`Executor.run()` will execute the
    :code:`default_main_program` when the program is not specified.
2110

Y
Yu Yang 已提交
2111 2112 2113
    Returns:
        Program: main program
    """
Y
Yu Yang 已提交
2114
    return _main_program_
Y
Yu Yang 已提交
2115 2116 2117 2118 2119


def switch_main_program(program):
    """
    Switch the main program to a new program.
2120

Y
Yu Yang 已提交
2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134
    Args:
        program(Program): The new main program

    Returns:
        Program: The previous main program
    """
    global _main_program_
    prev_program = _main_program_
    _main_program_ = program
    return prev_program


def switch_startup_program(program):
    """
2135
    Switch the startup program to a new program
Y
Yu Yang 已提交
2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150
    Args:
        program(Program): The new startup program

    Returns:
        Program: The previous startup program
    """
    global _startup_program_
    prev_program = _startup_program_
    _startup_program_ = program
    return prev_program


@contextlib.contextmanager
def program_guard(main_program, startup_program=None):
    """
Y
yuyang18 已提交
2151 2152 2153
    Change the global main program and startup program with `with` statement.
    Layer functions in the Python `with` block will append operators and
    variables to the new main programs.
2154

Y
Yu Yang 已提交
2155
    Examples:
Y
yuyang18 已提交
2156 2157 2158 2159 2160 2161 2162 2163 2164 2165

        >>> import paddle.fluid as fluid
        >>> main_program = fluid.Program()
        >>> startup_program = fluid.Program()
        >>> with fluid.program_guard(main_program, startup_program):
        >>>     data = fluid.layers.data(...)
        >>>     hidden = fluid.layers.fc(...)

    Notes: The temporary :code:`Program` can be used if the user does not need
    to construct either of startup program or main program.
2166

Y
Yu Yang 已提交
2167
    Examples:
Y
yuyang18 已提交
2168 2169 2170 2171 2172 2173

        >>> import paddle.fluid as fluid
        >>> main_program = fluid.Program()
        >>> # does not care about startup program. Just pass a temporary value.
        >>> with fluid.program_guard(main_program, fluid.Program()):
        >>>     data = ...
2174

Y
Yu Yang 已提交
2175
    Args:
Y
yuyang18 已提交
2176
        main_program(Program): New main program inside `with` statement.
2177
        startup_program(Program): New startup program inside `with` statement.
Y
Yu Yang 已提交
2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190
            None means do not change startup program.
    """
    if not isinstance(main_program, Program):
        raise TypeError("main_program should be Program")
    main_program = switch_main_program(main_program)
    if startup_program is not None:
        if not isinstance(startup_program, Program):
            raise TypeError("startup_program should be Program")
        startup_program = switch_startup_program(startup_program)
    yield
    switch_main_program(main_program)
    if startup_program is not None:
        switch_startup_program(startup_program)
X
xuwei06 已提交
2191 2192


W
Wu Yi 已提交
2193
def _get_var(name, program=None):
X
xuwei06 已提交
2194
    """
Y
yuyang18 已提交
2195
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
2196

X
xuwei06 已提交
2197 2198 2199
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
2200
        If None, default_global_program() will be used.
X
xuwei06 已提交
2201 2202 2203 2204 2205 2206 2207

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
2208
    assert isinstance(program, Program)
X
xuwei06 已提交
2209 2210

    return program.global_block().var(name)