framework.py 73.6 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
X
Xin Pan 已提交
21
import sys
22

Y
Yu Yang 已提交
23
import numpy as np
Q
qiaolongfei 已提交
24

M
minqiyang 已提交
25
from .. import compat as cpt
26
from .proto import framework_pb2
27 28
try:
    from . import core
29
except ImportError as e:
30 31 32 33
    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
34
    directory. The original error is: \n""" + cpt.get_exception_message(e))
35
except Exception as e:
36
    raise e
37
from . import unique_name
Y
Yu Yang 已提交
38

39
__all__ = [
40 41 42 43
    'Program',
    'default_startup_program',
    'default_main_program',
    'program_guard',
44
    'name_scope',
45
]
Y
Yu Yang 已提交
46

Q
qiaolongfei 已提交
47 48 49 50
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
51 52
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()

X
Xin Pan 已提交
53 54 55 56 57 58 59 60 61 62
_imperative_tracer_ = None


def _in_imperative_mode():
    return _imperative_tracer_ is not None


def _imperative_tracer():
    return _imperative_tracer_

W
Wu Yi 已提交
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
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
T
Tink_Y 已提交
103

104 105 106 107
          with name_scope("encoder"):
             ...
          with name_scope("decoder"):
             ...
T
Tink_Y 已提交
108 109
          with name_scope("attention"):
             ...
110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128
    """
    # 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
129 130 131
def generate_control_dev_var_name():
    import random
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
132 133 134 135


def grad_var_name(var_name):
    """
136 137
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
138 139 140
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
141

142
def convert_np_dtype_to_dtype_(np_dtype):
143 144
    """
    Convert the data type in numpy to the data type in Paddle
145

146
    Args:
147
        np_dtype(np.dtype): the data type in numpy.
148

149 150
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
151 152

    """
153 154
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
155
        return core.VarDesc.VarType.FP32
156
    elif dtype == np.float64:
157
        return core.VarDesc.VarType.FP64
158
    elif dtype == np.float16:
159
        return core.VarDesc.VarType.FP16
160
    elif dtype == np.int32:
161
        return core.VarDesc.VarType.INT32
162
    elif dtype == np.int16:
163
        return core.VarDesc.VarType.INT16
164
    elif dtype == np.int64:
165
        return core.VarDesc.VarType.INT64
166
    elif dtype == np.bool:
167
        return core.VarDesc.VarType.BOOL
168 169
    elif dtype == np.uint16:
        return core.VarDesc.VarType.INT16
170 171
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
172 173
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
174
    else:
M
minqiyang 已提交
175
        raise ValueError("Not supported numpy dtype %s" % dtype)
176 177 178


def dtype_is_floating(dtype):
179 180 181
    """
    Check the data type is floating or not.
    Args:
182
        dtype(np.dtype|core.VarDesc.VarType): data type.
183 184 185 186 187
            Could be numpy format or Paddle format

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

    """
188
    if not isinstance(dtype, core.VarDesc.VarType):
189 190
        dtype = convert_np_dtype_to_dtype_(dtype)

191 192 193 194
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
195 196


Y
Yang Yang(Tony) 已提交
197
def _debug_string_(proto, throw_on_error=True):
198 199 200 201 202 203 204 205 206 207 208
    """
    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
209
    error_fields = list()
Y
Yang Yang(Tony) 已提交
210
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
211 212
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
213 214 215
    return proto.__str__()


X
Xin Pan 已提交
216
class Variable(core.VarBase):
217
    """
218 219 220
    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
221
    two variables in different blocks could have the same name.
222

223 224
    There are many kinds of variables. Each kind of them has its own attributes
    and usages. Please reference the framework.proto for details.
225

226
    Most of a Variable's member variables can be setted to be None. It mean
227
    it is not available or will be specified later.
228 229

    Args:
230
        block(Block): The block that the variable belongs to.
231 232
        type(core.VarDesc.VarType): Variable type. Please reference the
            framework.proto for details.
233 234
        name(str|None): The name of the variable. If setted None, it will be
            generated automatically. Default: None
235
        shape(tuple|list|None): The shape of the variable. -1 means the batch size.
236
            Some kinds of variable do not contain shape, just set it to None.
237 238 239
            Default: None
        dtype(np.dtype|core.VarDesc.VarType|str|None): The data type of variable.
            Default: None
240
        lod_level (int|None): The level of lod tensor. 0 means it is not a time
241
            series data.
242
            Default: None
243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264
        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')
265 266
    """

Y
Yu Yang 已提交
267 268
    def __init__(self,
                 block,
Y
Yu Yang 已提交
269
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
270 271 272 273
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
274
                 capacity=None,
Q
QI JUN 已提交
275
                 persistable=None,
F
fengjiayi 已提交
276
                 error_clip=None,
Y
Yu Yang 已提交
277
                 stop_gradient=False,
F
fengjiayi 已提交
278
                 is_data=False,
Y
Yu Yang 已提交
279
                 **kwargs):
X
Xin Pan 已提交
280
        core.VarBase.__init__(self)
Y
Yu Yang 已提交
281
        self.block = block
F
fengjiayi 已提交
282
        self.error_clip = error_clip
Y
Yu Yang 已提交
283 284

        if name is None:
Y
Yu Yang 已提交
285
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
286
        is_new_var = False
M
minqiyang 已提交
287
        name = cpt.to_text(name)
M
minqiyang 已提交
288
        self.desc = self.block.desc.find_var(cpt.to_bytes(name))
D
Dong Zhihong 已提交
289 290

        if self.desc is None:
M
minqiyang 已提交
291
            self.desc = self.block.desc.var(cpt.to_bytes(name))
Y
Yu Yang 已提交
292
            is_new_var = True
Y
Yu Yang 已提交
293

Y
Yu Yang 已提交
294 295 296 297 298 299 300 301
        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
302
        if shape is not None:
Y
Yu Yang 已提交
303
            if is_new_var:
304
                self.desc.set_shape(shape)
Y
Yu Yang 已提交
305 306 307 308 309 310 311 312
            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
313
        if dtype is not None:
314
            if not isinstance(dtype, core.VarDesc.VarType):
315
                dtype = convert_np_dtype_to_dtype_(dtype)
Y
Yu Yang 已提交
316
            if is_new_var:
F
fengjiayi 已提交
317
                self.desc.set_dtype(dtype)
Y
Yu Yang 已提交
318
            else:
F
fengjiayi 已提交
319
                old_dtype = self.dtype
Q
QI JUN 已提交
320
                if dtype != old_dtype:
Y
Yu Yang 已提交
321 322 323 324 325
                    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
326 327

        if lod_level is not None:
Y
Yu Yang 已提交
328
            if is_new_var:
329
                self.desc.set_lod_level(lod_level)
Y
Yu Yang 已提交
330 331 332 333 334 335 336
            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))
337 338 339 340 341 342 343 344 345 346 347
        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))

348 349 350 351 352 353 354 355
        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
356
        self.block.vars[name] = self
Y
Yu Yang 已提交
357
        self.op = None
Y
Yu Yang 已提交
358
        self.stop_gradient = stop_gradient
F
fengjiayi 已提交
359
        self.is_data = is_data
Y
Yu Yang 已提交
360

X
Xin Pan 已提交
361 362 363 364 365 366 367 368 369 370 371 372
    def _numpy(self):
        scope = _imperative_tracer().get_scope(self.block.desc)
        tensor = core.get_variable_tensor(scope, self.desc.name())
        return np.array(tensor)

    def _backward(self):
        scope = _imperative_tracer().get_scope(self.block.desc)
        self._run_backward(scope)

    def _gradient(self):
        return np.array(self._grad())

373
    def __str__(self):
Y
Yang Yang(Tony) 已提交
374 375
        return self.to_string(True)

F
update  
fengjiayi 已提交
376
    def to_string(self, throw_on_error, with_details=False):
377 378 379 380
        """
        Get debug string.

        Args:
381 382
            throw_on_error(bool): True if raise an exception when self is
                not initialized.
F
update  
fengjiayi 已提交
383
            with_details(bool): more details about variables and parameters
384 385
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False;
386

387 388
        Returns:
            str: The debug string.
389
        """
F
update  
fengjiayi 已提交
390 391
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
392
        protostr = self.desc.serialize_to_string()
393
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
394 395 396 397
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
398 399
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
400
        return res_str
401 402 403

    __repr__ = __str__

W
Wu Yi 已提交
404
    def _set_desc(self, input):
405 406 407 408 409 410 411 412 413
        """
        Set the variable description.

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

        Returns:
            None
        """
414 415
        self.desc = input

416 417 418 419
    @property
    def persistable(self):
        return self.desc.persistable()

Y
Yu Yang 已提交
420 421 422 423
    @persistable.setter
    def persistable(self, p):
        self.desc.set_persistable(p)

Y
Yu Yang 已提交
424 425
    @property
    def name(self):
M
minqiyang 已提交
426
        return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
427

T
typhoonzero 已提交
428 429 430 431
    @name.setter
    def name(self, new_name):
        self.desc.set_name(new_name)

Y
Yu Yang 已提交
432 433 434
    @property
    def shape(self):
        # convert to tuple, make it as same as numpy API.
435
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
436 437

    @property
F
fengjiayi 已提交
438 439
    def dtype(self):
        return self.desc.dtype()
Y
Yu Yang 已提交
440 441 442

    @property
    def lod_level(self):
443
        return self.desc.lod_level()
Y
Yu Yang 已提交
444

Y
Yu Yang 已提交
445 446 447 448
    @property
    def type(self):
        return self.desc.type()

W
Wu Yi 已提交
449
    def _set_error_clip(self, error_clip):
450 451 452 453 454 455 456 457 458
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
459 460
        self.error_clip = error_clip

Y
Yu Yang 已提交
461

F
fengjiayi 已提交
462 463 464
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
465

466 467
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
468 469 470 471
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
472
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
473 474 475 476 477
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
478 479 480 481
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
482 483 484 485 486 487 488 489 490
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
491
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
492 493 494 495 496 497
        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):
498 499 500 501 502 503 504 505
        """
        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
506 507
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
508 509
        return self.op_proto_map[type]

510 511 512 513
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
514
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
515
            core.op_proto_and_checker_maker.kOpNameScopeAttrName()
516 517
        }

F
fengjiayi 已提交
518

X
Xin Pan 已提交
519
class Operator(core.OpBase):
520
    """
521 522 523 524 525 526 527
    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
528
        type(str): The type of operator. Default None.
529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548
        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
549
        Block.append_op or Block._prepend_op instead.
550 551 552 553 554 555 556 557 558 559

    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]})
560
    """
561 562 563 564
    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
565
        'ncclInit', 'select', 'checkpoint_notify', 'gen_nccl_id'
566
    }
567

Y
Yu Yang 已提交
568 569
    def __init__(self,
                 block,
Y
Yu Yang 已提交
570
                 desc,
Y
Yu Yang 已提交
571 572 573 574
                 type=None,
                 inputs=None,
                 outputs=None,
                 attrs=None):
X
Xin Pan 已提交
575
        core.OpBase.__init__(self)
Y
Yu Yang 已提交
576
        self.block = block
Y
Yu Yang 已提交
577
        self.desc = desc
G
gongweibao 已提交
578 579 580 581 582
        # 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
583 584 585 586
        del attrs

        op_maker = core.op_proto_and_checker_maker

G
gongweibao 已提交
587 588
        if op_maker.kOpRoleAttrName() not in op_attrs:
            op_attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role
Y
yuyang18 已提交
589 590 591

        role_var_name = op_maker.kOpRoleVarAttrName()
        if len(self.block.program.
G
gongweibao 已提交
592 593
               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
594

G
gongweibao 已提交
595 596
        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
597

F
fengjiayi 已提交
598 599 600 601 602
        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
603
        self.desc.set_type(type)
F
fengjiayi 已提交
604
        proto = OpProtoHolder.instance().get_op_proto(type)
605

606 607 608
        namescope_var_name = op_maker.kOpNameScopeAttrName()
        op_attrs[namescope_var_name] = _full_name_scope()

Y
Yang Yang(Tony) 已提交
609 610
        def find_name(var_list, name):
            for var_name in var_list:
Q
Qiao Longfei 已提交
611
                if var_list[var_name] is not None and var_name == name:
Y
Yang Yang(Tony) 已提交
612 613
                    return True
            return False
Q
QI JUN 已提交
614

X
Xin Pan 已提交
615
        self.inputs = []
Y
Yang Yang(Tony) 已提交
616 617 618 619 620 621 622
        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:
623 624 625 626
                    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
627 628
                        raise ValueError(
                            "Input %s expects only one input, but %d are given."
629 630 631
                            % (in_proto.name, len(in_args)))
                    in_arg_names = []
                    for arg in in_args:
632
                        if isinstance(arg, six.string_types):
Y
Yang Yu 已提交
633
                            in_arg_names.append(arg)
634 635
                        elif isinstance(arg, six.binary_type):
                            in_arg_names.append(arg.decode())
Y
Yang Yu 已提交
636
                        else:
M
minqiyang 已提交
637
                            in_arg_names.append(cpt.to_text(arg.name))
638
                    self.desc.set_input(in_proto.name, in_arg_names)
Y
Yang Yang(Tony) 已提交
639 640
                else:
                    self.desc.set_input(in_proto.name, [])
F
Update  
fengjiayi 已提交
641

X
Xin Pan 已提交
642 643 644 645 646 647 648
            for inp in inputs.values():
                if isinstance(inp, Variable):
                    self.inputs.append(inp)
                elif isinstance(inp, list) or isinstance(inp, tuple):
                    self.inputs.extend(inp[:])

        self.outputs = []
Y
Yu Yang 已提交
649
        if outputs is not None:
650 651 652 653 654 655 656
            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
657 658
                raise ValueError(("Incorrect setting for output(s) of "
                                  "operator \"%s\". Need: [%s] Given: [%s]") %
659 660 661
                                 (type,
                                  ", ".join(six.binary_type(e) for e in need),
                                  ", ".join(six.binary_type(e) for e in given)))
662

F
fengjiayi 已提交
663
            for out_proto in proto.outputs:
664 665 666 667
                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
668 669
                    raise ValueError(
                        "Output %s expects only one output, but %d are given." %
670 671 672
                        (out_proto.name, len(out_args)))
                out_arg_names = []
                for arg in out_args:
M
minqiyang 已提交
673
                    out_arg_names.append(cpt.to_text(arg.name))
674 675
                    arg.op = self
                self.desc.set_output(out_proto.name, out_arg_names)
F
Update  
fengjiayi 已提交
676

X
Xin Pan 已提交
677 678 679 680 681 682
            for out in outputs.values():
                if isinstance(out, Variable):
                    self.outputs.append(out)
                elif isinstance(out, list) or isinstance(out, tuple):
                    self.outputs.extend(out[:])

G
gongweibao 已提交
683 684
        if op_attrs is not None:
            if not isinstance(op_attrs, dict):
685
                raise TypeError("'attrs' should be a dict.")
F
fengjiayi 已提交
686
            for attr in proto.attrs:
F
Update  
fengjiayi 已提交
687
                attr_name = attr.name
G
gongweibao 已提交
688
                if (attr_name not in op_attrs) or (op_attrs[attr_name] is None):
F
Update  
fengjiayi 已提交
689
                    continue
G
gongweibao 已提交
690
                attr_val = op_attrs[attr_name]
G
gongweibao 已提交
691 692
                self._update_desc_attr(attr_name, attr_val)

693
        self.desc.check_attrs()
W
Wu Yi 已提交
694
        if self._has_kernel(type):
Q
QI JUN 已提交
695
            self.desc.infer_var_type(self.block.desc)
Y
Yu Yang 已提交
696
            self.desc.infer_shape(self.block.desc)
F
fengjiayi 已提交
697

W
Wu Yi 已提交
698
    def _has_kernel(self, op_type):
699 700
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
701
    def to_string(self, throw_on_error):
702
        """
703 704
        Get debug string.

705
        Args:
706 707
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
708

709 710
        Returns:
            str: The debug string.
711 712

        """
713
        protostr = self.desc.serialize_to_string()
714
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
715 716 717 718
        return _debug_string_(proto, throw_on_error)

    def __str__(self):
        return self.to_string(True)
719 720 721

    __repr__ = __str__

F
fengjiayi 已提交
722 723 724 725 726
    @property
    def type(self):
        return self.desc.type()

    def input(self, name):
727
        """
728
        Get the input arguments according to the input parameter name.
729

730 731
        Args:
            name(str): The input parameter name.
732

733 734 735
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
736
        """
F
fengjiayi 已提交
737 738
        return self.desc.input(name)

W
Wu Yi 已提交
739
    def _rename_input(self, old_name, new_name):
740 741 742 743 744 745 746 747 748 749
        """
        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
750
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
751

W
Wu Yi 已提交
752
    def _rename_output(self, old_name, new_name):
753 754 755 756 757 758 759 760 761 762
        """
        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
763
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
764

F
fengjiayi 已提交
765 766 767 768
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
769 770 771 772 773 774 775 776
    @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
777
    def output(self, name):
778
        """
779
        Get output arguments by the output parameter name.
780

781 782
        Args:
            name(str): The output parameter name.
783

784 785 786
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
787
        """
F
fengjiayi 已提交
788 789 790 791 792 793
        return self.desc.output(name)

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

794 795 796 797 798 799 800 801
    @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
802
    def has_attr(self, name):
803
        """
804 805
        Whether this Operator has the attribute with name or not.

806
        Args:
807
            name(str): the attribute name.
808

809 810
        Returns:
            bool: True if has this attribute.
811 812

        """
F
fengjiayi 已提交
813 814 815
        return self.desc.has_attr(name)

    def attr_type(self, name):
816
        """
817
        Get the type of attribute by attribute's name.
818

819 820
        Args:
            name(str): the attribute name.
821

822 823
        Returns:
            core.AttrType: the attribute type.
824
        """
F
fengjiayi 已提交
825 826
        return self.desc.attr_type(name)

W
Wu Yi 已提交
827
    def _set_attr(self, name, val):
828 829 830 831 832 833 834 835 836 837
        """
        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
838 839 840 841 842 843 844 845 846 847 848 849 850
        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
851 852
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
853 854
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
855
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
856 857 858 859
        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
860
            self.desc._set_attr(name, val)
Y
yuyang18 已提交
861

F
fengjiayi 已提交
862 863 864 865 866
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
867
        """
868 869
        Get the attribute by name.

870
        Args:
871
            name(str): the attribute name.
872

873 874
        Returns:
            bool|int|str|float|list: The attribute value. The return value
875 876
            can be any valid attribute type.
        """
F
fengjiayi 已提交
877
        return self.desc.attr(name)
Y
Yu Yang 已提交
878

W
Wu Yi 已提交
879
    def _block_attr_id(self, name):
880
        """
G
gongweibao 已提交
881
        Get the block attribute's id by name.
882

883 884
        Args:
            name(str): the attribute name.
885

886 887
        Returns:
            int: the block index.
888
        """
W
Wu Yi 已提交
889
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
890

W
Wu Yi 已提交
891
    def _block_attr(self, name):
G
gongweibao 已提交
892 893 894 895 896 897 898 899 900 901
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
902
        id = self._block_attr_id(name)
G
gongweibao 已提交
903 904 905
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
906
    def _blocks_attr(self, name):
G
gongweibao 已提交
907 908 909 910 911 912 913 914 915 916
        """
        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
917
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
918 919 920 921 922
            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
923
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
924 925 926 927 928 929 930 931 932 933
        """
        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
934
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
935

J
JiayiFeng 已提交
936
    def all_attrs(self):
F
fengjiayi 已提交
937
        """
938 939 940
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
941
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
942 943 944 945
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
G
gongweibao 已提交
946 947
            attr_type = self.desc.attr_type(n)
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
948
                attr_map[n] = self._block_attr(n)
G
gongweibao 已提交
949 950 951
                continue

            if attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
952
                attr_map[n] = self._blocks_attr(n)
G
gongweibao 已提交
953 954 955 956
                continue

            attr_map[n] = self.attr(n)

F
fengjiayi 已提交
957 958
        return attr_map

Y
Yu Yang 已提交
959

Y
Yu Yang 已提交
960
class Block(object):
961 962 963 964 965 966 967 968 969 970 971 972 973 974
    """
    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
975
        use `Program._create_block()` to create a block.
976 977 978 979 980 981 982 983 984 985 986 987 988 989

    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
990
    def __init__(self, program, idx):
Y
Yu Yang 已提交
991
        self.desc = program.desc.block(idx)
992
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
993
        self.ops = list()  # operator list
Y
Yu Yang 已提交
994
        self.program = program
995
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
996

997
    def __str__(self):
Y
Yang Yang(Tony) 已提交
998 999
        return self.to_string(True)

F
fengjiayi 已提交
1000 1001
    def to_string(self, throw_on_error, with_details=False):
        """
1002 1003
        Get debug string.

F
fengjiayi 已提交
1004 1005
        Args:
            throw_on_error(bool): raise exception when self is not initialized
1006
                when throw_on_error is True.
F
update  
fengjiayi 已提交
1007
            with_details(bool): more details about variables and parameters
1008 1009
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
1010

1011 1012
        Returns:
            str: The debug string.
F
fengjiayi 已提交
1013 1014 1015 1016
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
1017
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
1018 1019
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
1020
            for var in list(self.vars.values()):
F
fengjiayi 已提交
1021
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
1022
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
1023
            for op in self.ops:
F
fengjiayi 已提交
1024 1025
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
1026 1027 1028
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
1029 1030
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
1031 1032
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1033 1034 1035

    __repr__ = __str__

Y
Yu Yang 已提交
1036 1037
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
1038
        return self.desc.parent
Y
Yu Yang 已提交
1039

Y
Yu Yang 已提交
1040 1041 1042 1043
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
1044
    def _set_forward_block_idx(self, idx):
1045 1046 1047 1048 1049 1050 1051 1052 1053
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
1054
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
1055

Y
Yu Yang 已提交
1056 1057
    @property
    def idx(self):
Y
Yu Yang 已提交
1058
        return self.desc.id
Y
Yu Yang 已提交
1059

Q
Qiao Longfei 已提交
1060
    def var(self, name):
1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073
        """
        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.
        """
1074
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
1075 1076 1077
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
1078 1079
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
1080
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
1081
        return v
Q
Qiao Longfei 已提交
1082

W
Wu Yi 已提交
1083
    def _var_recursive(self, name):
1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
        """
        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
1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122
        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
1123

Q
Qiao Longfei 已提交
1124
    def all_parameters(self):
1125
        return list(self.iter_parameters())
1126

1127
    def iter_parameters(self):
M
minqiyang 已提交
1128
        return (item[1] for item in six.iteritems(self.vars)
1129
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
1130

Y
Yu Yang 已提交
1131
    def create_var(self, *args, **kwargs):
1132
        var = Variable(block=self, *args, **kwargs)
1133 1134
        if 'initializer' in kwargs:
            kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
1135
        return var
Y
Yu Yang 已提交
1136

Q
Qiao Longfei 已提交
1137 1138 1139
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
1140
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
1141 1142
        """
        Rename variable in vars and ops' inputs and outputs
1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154

        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
1155
        """
M
minqiyang 已提交
1156 1157
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
1158

T
typhoonzero 已提交
1159
        if not self.has_var(name):
1160
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
1161 1162
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
1163
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
1164 1165 1166 1167 1168 1169 1170
            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
1171
            var_type = "Variable"
T
wip  
typhoonzero 已提交
1172 1173 1174 1175
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
1176
        orig_var_type = v.type
M
minqiyang 已提交
1177
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
1178
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
1179
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
1180
        if var_type == "Parameter":
T
wip  
typhoonzero 已提交
1181 1182 1183 1184
            var = Parameter(
                self,
                d.shape(),
                d.dtype(),
T
typhoonzero 已提交
1185
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1186 1187 1188 1189 1190 1191 1192
                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
1193
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
1194 1195
            var = Variable(
                self,
T
typhoonzero 已提交
1196
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1197 1198 1199 1200
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
1201
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
1202 1203 1204
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
1205
        self._sync_with_cpp()
1206
        return var
T
typhoonzero 已提交
1207

W
Wu Yi 已提交
1208 1209
    def _remove_var(self, name):
        self._sync_with_cpp()
M
minqiyang 已提交
1210
        self.desc._remove_var(cpt.to_bytes(name))
1211 1212
        del self.vars[name]

Y
Yu Yang 已提交
1213 1214
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
Q
Qiao Longfei 已提交
1215
        param = Parameter(global_block, *args, **kwargs)
1216
        if 'initializer' in kwargs:
1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236

            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
1237
        return param
Y
Yu Yang 已提交
1238

Y
Yu Yang 已提交
1239
    def append_op(self, *args, **kwargs):
1240 1241 1242 1243 1244 1245
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
Y
Yu Yang 已提交
1246
        op_desc = self.desc.append_op()
1247
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
X
Xin Pan 已提交
1248 1249
        if _in_imperative_mode():
            _imperative_tracer().trace(op, op.inputs, op.outputs, self.desc)
Y
Yu Yang 已提交
1250 1251 1252
        self.ops.append(op)
        return op

W
Wu Yi 已提交
1253
    def _insert_op(self, index, *args, **kwargs):
1254 1255 1256 1257 1258 1259 1260 1261 1262
        """
        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
1263 1264
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
1265 1266 1267 1268
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

W
Wu Yi 已提交
1269
    def _remove_op(self, index):
1270 1271 1272 1273 1274 1275 1276 1277 1278
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
W
Wu Yi 已提交
1279 1280
        self._sync_with_cpp()
        self.desc._remove_op(index, index + 1)
1281 1282
        del self.ops[index]

W
Wu Yi 已提交
1283
    def _slice_ops(self, start, end):
1284 1285 1286 1287 1288 1289 1290 1291 1292 1293
        """
        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
1294
        return self.ops[start:end]
Y
Yancey1989 已提交
1295

W
Wu Yi 已提交
1296 1297
    def _prepend_op(self, *args, **kwargs):
        op_desc = self.desc._prepend_op()
Y
Yu Yang 已提交
1298
        op = Operator(self, op_desc, *args, **kwargs)
Q
qiaolongfei 已提交
1299
        self.ops.insert(0, op)
Y
Yu Yang 已提交
1300 1301
        return op

W
Wu Yi 已提交
1302
    def _sync_with_cpp(self):
1303
        """
1304 1305
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
1306
        """
Q
Qiao Longfei 已提交
1307 1308 1309 1310 1311
        # 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())

1312
        # sync variables removed from c++ end
1313
        for var in list(self.vars.keys()):
M
minqiyang 已提交
1314
            if not self.desc.find_var(cpt.to_bytes(var)):
1315 1316
                self.vars.pop(var)

Q
Qiao Longfei 已提交
1317
        # sync operators from cpp
1318 1319 1320 1321
        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
1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337
        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
1338 1339 1340 1341 1342

        # 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
1343
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
1344 1345 1346 1347 1348 1349 1350

        # 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)

1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363
        # 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
1364 1365 1366 1367
        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
1368
    def _copy_param_info_from(self, other):
1369
        """
1370 1371
        Copy the information of parameters from the other block.

1372
        Args:
1373 1374 1375 1376 1377
            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.
1378 1379 1380 1381 1382

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
1383 1384
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
1385
        for p in other.iter_parameters():
1386 1387 1388
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
W
Wu Yi 已提交
1389
                raise ValueError("_copy_param_info_from should be invoked with "
1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401
                                 "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
1402
                gradient_clip_attr=p.gradient_clip_attr,
F
fengjiayi 已提交
1403
                error_clip=p.error_clip,
1404 1405 1406
                name=v.name)
            self.vars[new_p.name] = new_p

W
Wu Yi 已提交
1407
    def _clone_variable(self, var):
1408 1409
        """
        Clone a variable into current block.
1410

1411 1412 1413 1414
        Args:
            var: the variable to be cloned.

        Returns:
1415
            Variable: the new  variable cloned from 'var' in current block.
1416 1417
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
1418 1419 1420 1421 1422
        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
1423 1424
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
1425
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
1426 1427 1428 1429 1430 1431
        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
1432 1433
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1434 1435 1436 1437 1438 1439 1440
        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
1441 1442
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1443
        return ret_var
1444

Y
Yu Yang 已提交
1445 1446

class Program(object):
D
dzhwinter 已提交
1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457
    """
    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
1458
    default_main_program run in every mini batch and adjust the weights.
D
dzhwinter 已提交
1459 1460

    Returns:
Y
yuyang18 已提交
1461
        A empty program.
D
dzhwinter 已提交
1462 1463

    Examples:
Y
yuyang18 已提交
1464 1465 1466 1467 1468 1469
        >>> 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
1470 1471 1472

    """

1473 1474
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
1475 1476
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
D
dzhwinter 已提交
1477
        self._seed = 0
Y
yuyang18 已提交
1478
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
Y
yuyang18 已提交
1479
        self._op_role_var = []
T
tangwei12 已提交
1480 1481 1482 1483

        # for distribute
        self._is_distributed = False
        self._is_chief = False
T
tangwei12 已提交
1484
        self._slice_vars_and_attrs = []
T
tangwei12 已提交
1485 1486
        self._endpoints = []
        self._distributed_lookup_table = None
Y
yuyang18 已提交
1487 1488 1489

    @property
    def op_role(self):
Y
yuyang18 已提交
1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502
        """
        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
1503 1504 1505 1506 1507 1508 1509 1510
        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
1511 1512 1513 1514 1515 1516 1517
        """
        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
1518 1519 1520 1521
        return self._op_role_var

    @op_role_var.setter
    def set_op_role_var(self, var_name):
Y
yuyang18 已提交
1522
        self._op_role_var = [var_name]
Y
yuyang18 已提交
1523 1524

    @contextlib.contextmanager
W
Wu Yi 已提交
1525
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
1526 1527 1528 1529 1530 1531 1532
        """
        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:
1533
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
1534 1535 1536 1537

        Examples:

            >>> p, g = backward(...)
W
Wu Yi 已提交
1538
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
1539 1540
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
1541 1542 1543
        tmp_role = self._current_role
        tmp_var = self._op_role_var

Y
yuyang18 已提交
1544 1545
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
1546 1547 1548 1549
        self._op_role_var = [
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
Y
yuyang18 已提交
1550
        yield
X
Xin Pan 已提交
1551 1552
        self._op_role_var = tmp_var
        self._current_role = tmp_role
Y
Yu Yang 已提交
1553

1554
    @contextlib.contextmanager
X
Xin Pan 已提交
1555
    def _lr_schedule_guard(self, is_with_opt=False):
1556 1557 1558 1559 1560 1561 1562
        """
        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
1563 1564 1565 1566
        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.
1567 1568 1569 1570 1571 1572 1573

        Examples:

            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
1574 1575 1576 1577

        tmp_role = self._current_role
        tmp_var = self._op_role_var

1578 1579
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
1580 1581
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
1582 1583 1584
        # TODO(typhoonzero): how to set target learning rate var
        self._op_role_var = []
        yield
1585 1586
        self._op_role_var = tmp_var
        self._current_role = tmp_role
1587

1588
    def __str__(self):
Y
yuyang18 已提交
1589 1590 1591 1592 1593 1594 1595 1596 1597
        """
        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
1598 1599
        return self.to_string(True)

F
fengjiayi 已提交
1600 1601 1602
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
1603

F
fengjiayi 已提交
1604
        Args:
Y
yuyang18 已提交
1605 1606
            throw_on_error(bool): raise Value error when any of required fields
                is not set.
F
fengjiayi 已提交
1607

Y
yuyang18 已提交
1608 1609 1610 1611 1612 1613 1614 1615 1616 1617
            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
1618 1619 1620 1621 1622 1623 1624 1625 1626 1627

        """
        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()
1628 1629
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
1630 1631
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1632

W
Wu Yi 已提交
1633
    def _get_desc(self):
Y
yuyang18 已提交
1634 1635 1636 1637 1638 1639 1640
        """
        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.
        """
1641 1642
        return self.desc

X
version  
Xin Pan 已提交
1643 1644 1645
    def _version(self):
        return self.desc._version()

1646
    def clone(self, for_test=False):
Y
yuyang18 已提交
1647 1648 1649
        """
        Create a new, duplicated program.

1650

Y
yuyang18 已提交
1651 1652 1653 1654
        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`.
1655

Y
yuyang18 已提交
1656 1657 1658 1659
        * 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
1660 1661 1662 1663 1664
        :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()
1665 1666

        Args:
Y
yuyang18 已提交
1667 1668
            for_test(bool): True if change the :code:`is_test` attribute of
                operators to :code:`True`.
1669

D
dzhwinter 已提交
1670
        Returns:
Y
yuyang18 已提交
1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723
            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.
1724 1725
        """
        if for_test:
X
Xin Pan 已提交
1726
            p = self._inference_optimize(prune_read_op=False)
1727
        else:
1728
            p = Program()
G
gongweibao 已提交
1729 1730
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
1731
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
1732 1733 1734
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
1735 1736 1737 1738

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

W
Wu Yi 已提交
1739
            p._sync_with_cpp()
1740

W
Wu Yi 已提交
1741
        p._copy_param_info_from(self)
W
Wu Yi 已提交
1742
        p._copy_data_info_from(self)
1743
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
1744
        return p
1745

W
Wu Yi 已提交
1746
    def _prune(self, targets):
Y
yuyang18 已提交
1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761
        """
        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.

        """
1762 1763 1764 1765 1766 1767
        if not isinstance(targets, list):
            targets = [targets]
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
1768 1769
                    # After transpiler processing, the op that output this
                    # variable maybe has been changed, so t.op is not reliable
1770
                    # and we need to find the current op that generate this
1771 1772 1773 1774 1775 1776 1777 1778
                    # 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

1779
                    t = t.op
1780 1781 1782 1783
                    if t is None:
                        raise ValueError(
                            "The target variable must have an "
                            "associated operator that generates it.")
1784
                else:
1785 1786
                    raise ValueError("All targets of prune() can only be "
                                     "Variable or Operator.")
1787 1788 1789 1790

            targets_idx.append([t.block.idx, t.idx])
        res = Program()
        res.desc = core.prune(self.desc, targets_idx)
M
minqiyang 已提交
1791 1792 1793
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
1794
        res._sync_with_cpp()
1795 1796
        return res

X
Xin Pan 已提交
1797
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
1798
        """
F
fengjiayi 已提交
1799 1800 1801 1802 1803
        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.

1804
        3. change the :code:`is_test`
Y
yuyang18 已提交
1805 1806 1807
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

1808
        Args:
X
Xin Pan 已提交
1809 1810
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
1811

Y
yuyang18 已提交
1812 1813 1814 1815 1816 1817
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
1818
        res = Program()
1819
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
1820 1821 1822 1823

        # 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
1824
        if prune_read_op:
1825 1826 1827 1828 1829 1830 1831 1832 1833
            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
1834
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
1835 1836

        # change all `is_test` attributes to True
M
minqiyang 已提交
1837
        for i in six.moves.range(res.desc.num_blocks()):
1838
            block = res.desc.block(i)
M
minqiyang 已提交
1839
            for j in six.moves.range(block.op_size()):
1840 1841
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
1842
                    op._set_attr('is_test', True)
M
minqiyang 已提交
1843 1844 1845
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
1846
        res._sync_with_cpp()
1847 1848
        return res

1849 1850
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
1851 1852 1853 1854 1855 1856 1857
        """
        Deserialize a program desc from protobuf binary string.

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

        Args:
1858
            binary_str_type(str): The binary prootbuf string.
Y
yuyang18 已提交
1859 1860 1861 1862

        Returns:
            Program: A deserialized program desc.
        """
1863 1864
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
1865
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
1866
        p._sync_with_cpp()
1867
        return p
Y
Yu Yang 已提交
1868

D
dzhwinter 已提交
1869 1870
    @property
    def random_seed(self):
Y
yuyang18 已提交
1871 1872 1873 1874 1875 1876
        """
        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
1877 1878
        return self._seed

Q
qiaolongfei 已提交
1879 1880
    @property
    def num_blocks(self):
Y
yuyang18 已提交
1881 1882 1883
        """
        The number of blocks in this program.
        """
Q
qiaolongfei 已提交
1884 1885
        return self.desc.num_blocks()

D
dzhwinter 已提交
1886 1887 1888 1889 1890 1891
    @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
1892
    def __repr__(self):
1893
        return self.__str__()
1894

Y
Yu Yang 已提交
1895
    def global_block(self):
Y
yuyang18 已提交
1896 1897 1898
        """
        Get the first block of this program.
        """
Y
Yu Yang 已提交
1899 1900
        return self.blocks[0]

Q
Qiao Longfei 已提交
1901
    def block(self, index):
Y
yuyang18 已提交
1902 1903 1904 1905 1906 1907 1908 1909
        """
        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
1910 1911
        return self.blocks[index]

Y
Yu Yang 已提交
1912
    def current_block(self):
Y
yuyang18 已提交
1913 1914 1915 1916
        """
        Get the current block. The :code:`current` block is the block to append
        operators.
        """
Y
Yu Yang 已提交
1917 1918
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
1919
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
1920 1921 1922 1923 1924 1925 1926 1927 1928 1929
        """
        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
1930
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
1931 1932 1933
        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
1934 1935 1936 1937
        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
1938
    def _rollback(self):
Y
yuyang18 已提交
1939 1940 1941 1942 1943
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
1944 1945
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
1946
    def _sync_with_cpp(self):
Y
yuyang18 已提交
1947 1948 1949 1950 1951 1952 1953 1954 1955 1956
        """
        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
1957 1958 1959
        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
1960
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
1961

W
Wu Yi 已提交
1962
    def _copy_param_info_from(self, other):
1963
        """
1964
        Copy the information of parameters from other program.
D
dzhwinter 已提交
1965

Y
yuyang18 已提交
1966 1967 1968
        Notes: This is a very low level API. Users should not invoke it
        directly.

1969 1970 1971 1972 1973 1974 1975
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
1976
            raise TypeError("_copy_param_info_from should be invoked with "
1977 1978 1979
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
1980
            raise ValueError("_copy_param_info_from should be invoked with two "
1981
                             "program, with represent the same topology")
W
Wu Yi 已提交
1982
        self.global_block()._copy_param_info_from(other.global_block())
1983

1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002
    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
2003
    def _copy_data_info_from(self, other):
F
fengjiayi 已提交
2004 2005
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
2006

Y
yuyang18 已提交
2007 2008 2009
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
2010 2011 2012 2013 2014 2015 2016
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
2017
            raise TypeError("_copy_param_info_from should be invoked with "
F
fengjiayi 已提交
2018 2019 2020
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
2021
            raise ValueError("_copy_param_info_from should be invoked with two "
F
fengjiayi 已提交
2022
                             "program, with represent the same topology")
2023
        for var in list(other.global_block().vars.values()):
F
fengjiayi 已提交
2024 2025 2026
            if var.is_data:
                self.global_block().var(var.name).is_data = True

2027
    def list_vars(self):
Y
yuyang18 已提交
2028 2029 2030 2031 2032 2033
        """
        Get all variables from this Program. A iterable object is returned.

        Returns:
            iterable: The generator will yield every variable in this program.
        """
2034
        for each_block in self.blocks:
2035
            for each_var in list(each_block.vars.values()):
2036 2037
                yield each_var

Y
Yu Yang 已提交
2038

Y
Yu Yang 已提交
2039
class Parameter(Variable):
2040
    """
2041
    Parameter is derived from Variable. A parameter is a persistable
2042
    Variable, and will be updated by optimizers after each iteration.
2043
    The training of a neural network is essentially the updating of
2044 2045
    its parameters.

2046
    Relative to a general Variable, a Parameter has several its own
2047 2048
    member variables:

2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060
    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.
2061 2062
    """

Y
Yu Yang 已提交
2063 2064 2065 2066 2067 2068 2069 2070 2071 2072
    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")
2073 2074 2075

        Variable.__init__(
            self, block, persistable=True, shape=shape, dtype=dtype, **kwargs)
Y
Yu Yang 已提交
2076 2077 2078 2079
        self.trainable = kwargs.get('trainable', True)

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

2080 2081
        self.regularizer = kwargs.get('regularizer', None)

F
fengjiayi 已提交
2082
        self.gradient_clip_attr = kwargs.get('gradient_clip_attr', None)
Y
Yu Yang 已提交
2083

W
wanghaoshuang 已提交
2084
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
2085

F
fengjiayi 已提交
2086 2087 2088
    def __str__(self):
        return self.to_string(True)

F
update  
fengjiayi 已提交
2089 2090 2091
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
2092

F
update  
fengjiayi 已提交
2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106
        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
2107
                               "gradient_clip_attr", "do_model_average")
F
update  
fengjiayi 已提交
2108
            for attr_name in additional_attr:
2109 2110
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
2111 2112
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
2113 2114 2115 2116
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
2117

Y
Yu Yang 已提交
2118
# program is a global instance.
Y
Yu Yang 已提交
2119 2120
_main_program_ = Program()
_startup_program_ = Program()
2121

2122

2123
def default_startup_program():
Y
Yu Yang 已提交
2124
    """
Y
yuyang18 已提交
2125 2126 2127 2128 2129 2130 2131 2132 2133
    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.
2134

Y
Yu Yang 已提交
2135 2136 2137
    Returns:
        Program: startup program
    """
Y
Yu Yang 已提交
2138
    return _startup_program_
2139

2140

2141
def default_main_program():
Y
Yu Yang 已提交
2142
    """
Y
yuyang18 已提交
2143 2144 2145 2146 2147 2148 2149 2150 2151
    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.
2152

Y
Yu Yang 已提交
2153 2154 2155
    Returns:
        Program: main program
    """
Y
Yu Yang 已提交
2156
    return _main_program_
Y
Yu Yang 已提交
2157 2158 2159 2160 2161


def switch_main_program(program):
    """
    Switch the main program to a new program.
2162

Y
Yu Yang 已提交
2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176
    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):
    """
2177
    Switch the startup program to a new program
Y
Yu Yang 已提交
2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192
    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
2193 2194 2195
    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.
2196

Y
Yu Yang 已提交
2197
    Examples:
Y
yuyang18 已提交
2198 2199 2200 2201 2202 2203 2204 2205 2206 2207

        >>> 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.
2208

Y
Yu Yang 已提交
2209
    Examples:
Y
yuyang18 已提交
2210 2211 2212 2213 2214 2215

        >>> 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 = ...
2216

Y
Yu Yang 已提交
2217
    Args:
Y
yuyang18 已提交
2218
        main_program(Program): New main program inside `with` statement.
2219
        startup_program(Program): New startup program inside `with` statement.
Y
Yu Yang 已提交
2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232
            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
2233 2234


W
Wu Yi 已提交
2235
def _get_var(name, program=None):
X
xuwei06 已提交
2236
    """
Y
yuyang18 已提交
2237
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
2238

X
xuwei06 已提交
2239 2240 2241
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
2242
        If None, default_global_program() will be used.
X
xuwei06 已提交
2243 2244 2245 2246 2247 2248 2249

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
2250
    assert isinstance(program, Program)
X
xuwei06 已提交
2251 2252

    return program.global_block().var(name)
X
Xin Pan 已提交
2253 2254 2255 2256 2257 2258 2259 2260 2261


@contextlib.contextmanager
def _imperative_guard(tracer):
    global _imperative_tracer_
    tmp_trace = _imperative_tracer_
    _imperative_tracer_ = tracer
    yield
    _imperative_tracer_ = tmp_trace