framework.py 64.1 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

Y
Yu Yang 已提交
15
import collections
Q
qiaolongfei 已提交
16
import contextlib
F
fengjiayi 已提交
17
import re
18

Y
Yu Yang 已提交
19
import numpy as np
Q
qiaolongfei 已提交
20

21
import proto.framework_pb2 as framework_pb2
22 23 24 25 26 27 28
try:
    from . import core
except ImportError, e:
    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
Y
yuyang18 已提交
29
    directory. The original error is: \n""" + e.message)
30 31
except Exception, e:
    raise e
Y
Yu Yang 已提交
32
import unique_name
Y
Yu Yang 已提交
33

34
__all__ = [
35 36
    'Program',
    'Operator',
F
fengjiayi 已提交
37
    'Parameter',
38 39 40
    'default_startup_program',
    'default_main_program',
    'program_guard',
X
xuwei06 已提交
41
    'get_var',
42
]
Y
Yu Yang 已提交
43

Q
qiaolongfei 已提交
44 45 46 47 48 49 50 51
EMPTY_VAR_NAME = core.kEmptyVarName()
TEMP_VAR_NAME = core.kTempVarName()
GRAD_VAR_SUFFIX = core.kGradVarSuffix()
ZERO_VAR_SUFFIX = core.kZeroVarSuffix()


def grad_var_name(var_name):
    """
52 53
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
54 55 56
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
57

58
def convert_np_dtype_to_dtype_(np_dtype):
59 60
    """
    Convert the data type in numpy to the data type in Paddle
61

62
    Args:
63
        np_dtype(np.dtype): the data type in numpy.
64

65 66
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
67 68

    """
69 70
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
71
        return core.VarDesc.VarType.FP32
72
    elif dtype == np.float64:
73
        return core.VarDesc.VarType.FP64
74
    elif dtype == np.float16:
75
        return core.VarDesc.VarType.FP16
76
    elif dtype == np.int32:
77
        return core.VarDesc.VarType.INT32
78
    elif dtype == np.int16:
79
        return core.VarDesc.VarType.INT16
80
    elif dtype == np.int64:
81
        return core.VarDesc.VarType.INT64
82
    elif dtype == np.bool:
83
        return core.VarDesc.VarType.BOOL
84 85
    elif dtype == np.uint16:
        return core.VarDesc.VarType.INT16
86 87
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
88 89 90 91 92
    else:
        raise ValueError("Not supported numpy dtype " + str(dtype))


def dtype_is_floating(dtype):
93 94 95
    """
    Check the data type is floating or not.
    Args:
96
        dtype(np.dtype|core.VarDesc.VarType): data type.
97 98 99 100 101
            Could be numpy format or Paddle format

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

    """
102
    if not isinstance(dtype, core.VarDesc.VarType):
103 104
        dtype = convert_np_dtype_to_dtype_(dtype)

105 106 107 108
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
109 110


Y
Yang Yang(Tony) 已提交
111
def _debug_string_(proto, throw_on_error=True):
112 113 114 115 116 117 118 119 120 121 122
    """
    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
123
    error_fields = list()
Y
Yang Yang(Tony) 已提交
124
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
125 126
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
127 128 129
    return proto.__str__()


Y
Yu Yang 已提交
130
class Variable(object):
131
    """
132 133 134 135
    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 
    two variables in different blocks could have the same name.
136

137 138
    There are many kinds of variables. Each kind of them has its own attributes 
    and usages. Please reference the framework.proto for details. 
139

140
    Most of a Variable's member variables can be setted to be None. It mean 
141
    it is not available or will be specified later.
142 143

    Args:
144
        block(Block): The block that the variable belongs to.
145 146
        type(core.VarDesc.VarType): Variable type. Please reference the
            framework.proto for details.
147 148
        name(str|None): The name of the variable. If setted None, it will be
            generated automatically. Default: None
149
        shape(tuple|list|None): The shape of the variable. -1 means the batch size.
150
            Some kinds of variable do not contain shape, just set it to None.
151 152 153
            Default: None
        dtype(np.dtype|core.VarDesc.VarType|str|None): The data type of variable.
            Default: None
154
        lod_level (int|None): The level of lod tensor. 0 means it is not a time
155
            series data.
156
            Default: None
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
        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')
179 180
    """

Y
Yu Yang 已提交
181 182
    def __init__(self,
                 block,
Y
Yu Yang 已提交
183
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
184 185 186 187
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
188
                 capacity=None,
Q
QI JUN 已提交
189
                 persistable=None,
F
fengjiayi 已提交
190
                 error_clip=None,
Y
Yu Yang 已提交
191
                 stop_gradient=False,
F
fengjiayi 已提交
192
                 is_data=False,
Y
Yu Yang 已提交
193
                 **kwargs):
Y
Yu Yang 已提交
194
        self.block = block
F
fengjiayi 已提交
195
        self.error_clip = error_clip
Y
Yu Yang 已提交
196 197

        if name is None:
Y
Yu Yang 已提交
198
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
199 200 201 202
        is_new_var = False
        self.desc = self.block.desc.find_var(name)

        if self.desc is None:
D
dongzhihong 已提交
203
            self.desc = self.block.desc.var(name)
Y
Yu Yang 已提交
204
            is_new_var = True
Y
Yu Yang 已提交
205

Y
Yu Yang 已提交
206 207 208 209 210 211 212 213
        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
214
        if shape is not None:
Y
Yu Yang 已提交
215
            if is_new_var:
216
                self.desc.set_shape(shape)
Y
Yu Yang 已提交
217 218 219 220 221 222 223 224
            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
225
        if dtype is not None:
226
            if not isinstance(dtype, core.VarDesc.VarType):
227
                dtype = convert_np_dtype_to_dtype_(dtype)
Y
Yu Yang 已提交
228
            if is_new_var:
F
fengjiayi 已提交
229
                self.desc.set_dtype(dtype)
Y
Yu Yang 已提交
230
            else:
F
fengjiayi 已提交
231
                old_dtype = self.dtype
Q
QI JUN 已提交
232
                if dtype != old_dtype:
Y
Yu Yang 已提交
233 234 235 236 237
                    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
238 239

        if lod_level is not None:
Y
Yu Yang 已提交
240
            if is_new_var:
241
                self.desc.set_lod_level(lod_level)
Y
Yu Yang 已提交
242 243 244 245 246 247 248
            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))
249 250 251 252 253 254 255 256 257 258 259
        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))

260 261 262 263 264 265 266 267
        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
268
        self.block.vars[name] = self
Y
Yu Yang 已提交
269
        self.op = None
Y
Yu Yang 已提交
270
        self.stop_gradient = stop_gradient
F
fengjiayi 已提交
271
        self.is_data = is_data
Y
Yu Yang 已提交
272

273
    def __str__(self):
Y
Yang Yang(Tony) 已提交
274 275
        return self.to_string(True)

F
update  
fengjiayi 已提交
276
    def to_string(self, throw_on_error, with_details=False):
277 278 279 280
        """
        Get debug string.

        Args:
281 282
            throw_on_error(bool): True if raise an exception when self is
                not initialized.
F
update  
fengjiayi 已提交
283
            with_details(bool): more details about variables and parameters
284 285
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False;
286

287 288
        Returns:
            str: The debug string.
289
        """
F
update  
fengjiayi 已提交
290 291
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
292 293
        protostr = self.desc.serialize_to_string()
        proto = framework_pb2.VarDesc.FromString(str(protostr))
F
update  
fengjiayi 已提交
294 295 296 297 298 299 300
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
                res_str += "%s: %s\n" % (attr_name,
                                         str(getattr(self, attr_name)))
        return res_str
301 302 303

    __repr__ = __str__

W
Wu Yi 已提交
304
    def _set_desc(self, input):
305 306 307 308 309 310 311 312 313
        """
        Set the variable description.

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

        Returns:
            None
        """
314 315
        self.desc = input

316 317 318 319
    @property
    def persistable(self):
        return self.desc.persistable()

Y
Yu Yang 已提交
320 321 322 323
    @persistable.setter
    def persistable(self, p):
        self.desc.set_persistable(p)

Y
Yu Yang 已提交
324 325
    @property
    def name(self):
326
        return self.desc.name()
Y
Yu Yang 已提交
327

T
typhoonzero 已提交
328 329 330 331
    @name.setter
    def name(self, new_name):
        self.desc.set_name(new_name)

Y
Yu Yang 已提交
332 333 334
    @property
    def shape(self):
        # convert to tuple, make it as same as numpy API.
335
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
336 337

    @property
F
fengjiayi 已提交
338 339
    def dtype(self):
        return self.desc.dtype()
Y
Yu Yang 已提交
340 341 342

    @property
    def lod_level(self):
343
        return self.desc.lod_level()
Y
Yu Yang 已提交
344

Y
Yu Yang 已提交
345 346 347 348
    @property
    def type(self):
        return self.desc.type()

W
Wu Yi 已提交
349
    def _set_error_clip(self, error_clip):
350 351 352 353 354 355 356 357 358
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
359 360
        self.error_clip = error_clip

Y
Yu Yang 已提交
361

F
fengjiayi 已提交
362 363 364
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
365

366 367
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
368 369 370 371 372 373 374 375 376 377
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
        op_proto = framework_pb2.OpProto.FromString(str(pbstr))
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
378 379 380 381
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
382 383 384 385 386 387 388 389 390
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
391
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
392 393 394 395 396 397
        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):
398 399 400 401 402 403 404 405
        """
        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
406 407
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
408 409
        return self.op_proto_map[type]

410 411 412 413 414 415 416
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
            core.op_proto_and_checker_maker.kOpRoleVarAttrName()
        }

F
fengjiayi 已提交
417

Y
Yu Yang 已提交
418
class Operator(object):
419
    """
420 421 422 423 424 425 426
    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
427
        type(str): The type of operator. Default None.
428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447
        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
448
        Block.append_op or Block._prepend_op instead.
449 450 451 452 453 454 455 456 457 458

    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]})
459
    """
460 461 462 463 464
    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',
        'ncclInit', 'channel_create', 'channel_close', 'channel_send',
T
tangwei12 已提交
465
        'channel_recv', 'select', 'checkpoint_notify', 'gen_nccl_id'
466
    }
467

Y
Yu Yang 已提交
468 469
    def __init__(self,
                 block,
Y
Yu Yang 已提交
470
                 desc,
Y
Yu Yang 已提交
471 472 473 474
                 type=None,
                 inputs=None,
                 outputs=None,
                 attrs=None):
475

Y
Yu Yang 已提交
476
        self.block = block
Y
Yu Yang 已提交
477
        self.desc = desc
T
typhoonzero 已提交
478
        self.attrs = attrs
Y
yuyang18 已提交
479 480 481 482 483 484 485 486
        if self.attrs is None:
            self.attrs = dict()
        del attrs

        op_maker = core.op_proto_and_checker_maker

        if op_maker.kOpRoleAttrName() not in self.attrs:
            self.attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role
Y
yuyang18 已提交
487 488 489 490 491 492 493 494

        role_var_name = op_maker.kOpRoleVarAttrName()
        if len(self.block.program.
               op_role_var) != 0 and role_var_name not in self.attrs:
            self.attrs[role_var_name] = self.block.program.op_role_var

        if role_var_name in self.attrs and len(self.attrs[role_var_name]) == 0:
            del self.attrs[role_var_name]
Y
yuyang18 已提交
495

F
fengjiayi 已提交
496 497 498 499 500
        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
501
        self.desc.set_type(type)
F
fengjiayi 已提交
502
        proto = OpProtoHolder.instance().get_op_proto(type)
503

Y
Yang Yang(Tony) 已提交
504 505
        def find_name(var_list, name):
            for var_name in var_list:
Q
Qiao Longfei 已提交
506
                if var_list[var_name] is not None and var_name == name:
Y
Yang Yang(Tony) 已提交
507 508
                    return True
            return False
Q
QI JUN 已提交
509

Y
Yang Yang(Tony) 已提交
510 511 512 513 514 515 516
        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:
517 518 519 520
                    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
521 522
                        raise ValueError(
                            "Input %s expects only one input, but %d are given."
523 524 525
                            % (in_proto.name, len(in_args)))
                    in_arg_names = []
                    for arg in in_args:
Y
Yang Yu 已提交
526 527 528 529
                        if isinstance(arg, basestring):
                            in_arg_names.append(arg)
                        else:
                            in_arg_names.append(arg.name)
530
                    self.desc.set_input(in_proto.name, in_arg_names)
Y
Yang Yang(Tony) 已提交
531 532
                else:
                    self.desc.set_input(in_proto.name, [])
F
Update  
fengjiayi 已提交
533

Y
Yu Yang 已提交
534
        if outputs is not None:
535 536 537 538 539 540 541
            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
542 543
                raise ValueError(("Incorrect setting for output(s) of "
                                  "operator \"%s\". Need: [%s] Given: [%s]") %
544 545
                                 (type, ", ".join(str(e) for e in need),
                                  ", ".join(str(e) for e in given)))
546

F
fengjiayi 已提交
547
            for out_proto in proto.outputs:
548 549 550 551
                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
552 553
                    raise ValueError(
                        "Output %s expects only one output, but %d are given." %
554 555 556 557 558 559
                        (out_proto.name, len(out_args)))
                out_arg_names = []
                for arg in out_args:
                    out_arg_names.append(arg.name)
                    arg.op = self
                self.desc.set_output(out_proto.name, out_arg_names)
F
Update  
fengjiayi 已提交
560

Y
yuyang18 已提交
561 562
        if self.attrs is not None:
            if not isinstance(self.attrs, dict):
563
                raise TypeError("'attrs' should be a dict.")
F
fengjiayi 已提交
564
            for attr in proto.attrs:
F
Update  
fengjiayi 已提交
565
                attr_name = attr.name
Y
yuyang18 已提交
566 567
                if (attr_name not in self.attrs) or (
                        self.attrs[attr_name] is None):
F
Update  
fengjiayi 已提交
568
                    continue
Y
Yancey1989 已提交
569
                attr_val = self.attrs[attr_name]
G
gongweibao 已提交
570 571
                self._update_desc_attr(attr_name, attr_val)

572
        self.desc.check_attrs()
573
        if self.has_kernel(type):
Q
QI JUN 已提交
574
            self.desc.infer_var_type(self.block.desc)
Y
Yu Yang 已提交
575
            self.desc.infer_shape(self.block.desc)
F
fengjiayi 已提交
576

577 578 579
    def has_kernel(self, op_type):
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
580
    def to_string(self, throw_on_error):
581
        """
582 583
        Get debug string.

584
        Args:
585 586
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
587

588 589
        Returns:
            str: The debug string.
590 591

        """
592 593
        protostr = self.desc.serialize_to_string()
        proto = framework_pb2.OpDesc.FromString(str(protostr))
Y
Yang Yang(Tony) 已提交
594 595 596 597
        return _debug_string_(proto, throw_on_error)

    def __str__(self):
        return self.to_string(True)
598 599 600

    __repr__ = __str__

F
fengjiayi 已提交
601 602 603 604 605
    @property
    def type(self):
        return self.desc.type()

    def input(self, name):
606
        """
607
        Get the input arguments according to the input parameter name.
608

609 610
        Args:
            name(str): The input parameter name.
611

612 613 614
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
615
        """
F
fengjiayi 已提交
616 617
        return self.desc.input(name)

T
typhoonzero 已提交
618
    def rename_input(self, old_name, new_name):
619 620 621 622 623 624 625 626 627 628
        """
        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
        """
T
typhoonzero 已提交
629 630 631
        self.desc.rename_input(old_name, new_name)

    def rename_output(self, old_name, new_name):
632 633 634 635 636 637 638 639 640 641
        """
        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
        """
T
typhoonzero 已提交
642 643
        self.desc.rename_output(old_name, new_name)

F
fengjiayi 已提交
644 645 646 647
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
648 649 650 651 652 653 654 655
    @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
656
    def output(self, name):
657
        """
658
        Get output arguments by the output parameter name.
659

660 661
        Args:
            name(str): The output parameter name.
662

663 664 665
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
666
        """
F
fengjiayi 已提交
667 668 669 670 671 672
        return self.desc.output(name)

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

673 674 675 676 677 678 679 680
    @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
681
    def has_attr(self, name):
682
        """
683 684
        Whether this Operator has the attribute with name or not.

685
        Args:
686
            name(str): the attribute name.
687

688 689
        Returns:
            bool: True if has this attribute.
690 691

        """
F
fengjiayi 已提交
692 693 694
        return self.desc.has_attr(name)

    def attr_type(self, name):
695
        """
696
        Get the type of attribute by attribute's name.
697

698 699
        Args:
            name(str): the attribute name.
700

701 702
        Returns:
            core.AttrType: the attribute type.
703
        """
F
fengjiayi 已提交
704 705
        return self.desc.attr_type(name)

Y
yuyang18 已提交
706
    def set_attr(self, name, val):
707 708 709 710 711 712 713 714 715 716
        """
        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).
        """
Y
yuyang18 已提交
717
        self.attrs[name] = val
G
gongweibao 已提交
718 719 720 721 722 723 724 725 726 727 728 729 730
        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
731 732
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
733 734
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
735
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
736 737 738 739 740
        elif isinstance(val, core.BlockDesc) or \
                isinstance(val, core.ProgramDesc):
            self.desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            self.desc.set_attr(name, val)
Y
yuyang18 已提交
741

F
fengjiayi 已提交
742 743 744 745 746
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
747
        """
748 749
        Get the attribute by name.

750
        Args:
751
            name(str): the attribute name.
752

753 754
        Returns:
            bool|int|str|float|list: The attribute value. The return value
755 756
            can be any valid attribute type.
        """
F
fengjiayi 已提交
757
        return self.desc.attr(name)
Y
Yu Yang 已提交
758

F
fengjiayi 已提交
759
    def block_attr(self, name):
760
        """
761
        Get the block attribute by name.
762

763 764
        Args:
            name(str): the attribute name.
765

766 767
        Returns:
            int: the block index.
768
        """
F
fengjiayi 已提交
769
        return self.desc.block_attr(name)
Y
Yu Yang 已提交
770

J
JiayiFeng 已提交
771
    def all_attrs(self):
F
fengjiayi 已提交
772
        """
773 774 775 776
        Get the attribute dict.

        Returns:
            dict: The Operator's attribute dict.
F
fengjiayi 已提交
777 778 779 780 781 782 783 784 785 786
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
            if n == 'sub_block':
                attr_map[n] = self.block_attr(n)
            else:
                attr_map[n] = self.attr(n)
        return attr_map

Y
Yu Yang 已提交
787

Y
Yu Yang 已提交
788
class Block(object):
789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817
    """
    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
        use `Program.create_block()` to create a block.

    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
818
    def __init__(self, program, idx):
Y
Yu Yang 已提交
819
        self.desc = program.desc.block(idx)
820
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
821
        self.ops = list()  # operator list
Y
Yu Yang 已提交
822
        self.program = program
823
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
824

825
    def __str__(self):
Y
Yang Yang(Tony) 已提交
826 827
        return self.to_string(True)

F
fengjiayi 已提交
828 829
    def to_string(self, throw_on_error, with_details=False):
        """
830 831
        Get debug string.

F
fengjiayi 已提交
832 833
        Args:
            throw_on_error(bool): raise exception when self is not initialized
834
                when throw_on_error is True.
F
update  
fengjiayi 已提交
835
            with_details(bool): more details about variables and parameters
836 837
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
838

839 840
        Returns:
            str: The debug string.
F
fengjiayi 已提交
841 842 843 844
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
845
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
846 847 848
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
            for var in self.vars.itervalues():
F
fengjiayi 已提交
849
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
850
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
851
            for op in self.ops:
F
fengjiayi 已提交
852 853
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
854 855 856 857 858 859
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
            proto = framework_pb2.BlockDesc.FromString(str(protostr))
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
860 861 862

    __repr__ = __str__

Y
Yu Yang 已提交
863 864
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
865
        return self.desc.parent
Y
Yu Yang 已提交
866

Y
Yu Yang 已提交
867 868 869 870
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
871
    def _set_forward_block_idx(self, idx):
872 873 874 875 876 877 878 879 880
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
881
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
882

Y
Yu Yang 已提交
883 884
    @property
    def idx(self):
Y
Yu Yang 已提交
885
        return self.desc.id
Y
Yu Yang 已提交
886

Q
Qiao Longfei 已提交
887
    def var(self, name):
888 889 890 891 892 893 894 895 896 897 898 899 900
        """
        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.
        """
Y
Yu Yang 已提交
901
        if not isinstance(name, basestring):
902 903 904
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
905 906
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
907
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
908
        return v
Q
Qiao Longfei 已提交
909

W
Wu Yi 已提交
910
    def _var_recursive(self, name):
911 912 913 914 915 916 917 918 919 920 921 922 923
        """
        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
924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949
        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
950

Q
Qiao Longfei 已提交
951
    def all_parameters(self):
W
Wu Yi 已提交
952
        return list(self._iter_parameters())
953

W
Wu Yi 已提交
954
    def _iter_parameters(self):
955 956
        return (item[1] for item in self.vars.iteritems()
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
957

Y
Yu Yang 已提交
958
    def create_var(self, *args, **kwargs):
959
        var = Variable(block=self, *args, **kwargs)
960 961
        if 'initializer' in kwargs:
            kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
962
        return var
Y
Yu Yang 已提交
963

Q
Qiao Longfei 已提交
964 965 966
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
967
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
968 969
        """
        Rename variable in vars and ops' inputs and outputs
970 971 972 973 974 975 976 977 978 979 980 981

        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
982 983
        """
        if not self.has_var(name):
984
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
985 986
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
987
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
988 989 990 991 992 993 994
            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
995
            var_type = "Variable"
T
wip  
typhoonzero 已提交
996 997 998 999
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
1000
        orig_var_type = v.type
W
Wu Yi 已提交
1001 1002
        self.desc._rename_var(name, new_name)
        # NOTE: v is destroyed by C++ after calling _rename_var.
T
wip  
typhoonzero 已提交
1003
        d = self.desc.find_var(new_name)
T
typhoonzero 已提交
1004
        if var_type == "Parameter":
T
wip  
typhoonzero 已提交
1005 1006 1007 1008
            var = Parameter(
                self,
                d.shape(),
                d.dtype(),
T
typhoonzero 已提交
1009
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1010 1011 1012 1013 1014 1015 1016
                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
1017
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
1018 1019
            var = Variable(
                self,
T
typhoonzero 已提交
1020
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1021 1022 1023 1024
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
1025
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
1026 1027 1028
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
1029
        self._sync_with_cpp()
1030
        return var
T
typhoonzero 已提交
1031

W
Wu Yi 已提交
1032 1033 1034
    def _remove_var(self, name):
        self._sync_with_cpp()
        self.desc._remove_var(name)
1035 1036
        del self.vars[name]

Y
Yu Yang 已提交
1037 1038
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
Q
Qiao Longfei 已提交
1039
        param = Parameter(global_block, *args, **kwargs)
1040 1041
        if 'initializer' in kwargs:
            kwargs['initializer'](param, self)
Q
Qiao Longfei 已提交
1042
        return param
Y
Yu Yang 已提交
1043

Y
Yu Yang 已提交
1044
    def append_op(self, *args, **kwargs):
1045 1046 1047 1048 1049 1050
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
Y
Yu Yang 已提交
1051
        op_desc = self.desc.append_op()
1052
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
Y
Yu Yang 已提交
1053 1054 1055
        self.ops.append(op)
        return op

W
Wu Yi 已提交
1056
    def _insert_op(self, index, *args, **kwargs):
1057 1058 1059 1060 1061 1062 1063 1064 1065
        """
        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
1066 1067
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
1068 1069 1070 1071
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

W
Wu Yi 已提交
1072
    def _remove_op(self, index):
1073 1074 1075 1076 1077 1078 1079 1080 1081
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
W
Wu Yi 已提交
1082 1083
        self._sync_with_cpp()
        self.desc._remove_op(index, index + 1)
1084 1085
        del self.ops[index]

W
Wu Yi 已提交
1086
    def _slice_ops(self, start, end):
1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
        """
        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
1097
        return self.ops[start:end]
Y
Yancey1989 已提交
1098

W
Wu Yi 已提交
1099 1100
    def _prepend_op(self, *args, **kwargs):
        op_desc = self.desc._prepend_op()
Y
Yu Yang 已提交
1101
        op = Operator(self, op_desc, *args, **kwargs)
Q
qiaolongfei 已提交
1102
        self.ops.insert(0, op)
Y
Yu Yang 已提交
1103 1104
        return op

W
Wu Yi 已提交
1105
    def _sync_with_cpp(self):
1106
        """
1107 1108
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
1109
        """
Q
Qiao Longfei 已提交
1110 1111 1112 1113 1114
        # 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())

1115 1116 1117 1118 1119
        # sync variables removed from c++ end
        for var in self.vars.keys():
            if not self.desc.find_var(var):
                self.vars.pop(var)

Q
Qiao Longfei 已提交
1120
        # sync operators from cpp
1121 1122 1123 1124
        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
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140
        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
1141 1142 1143 1144 1145

        # 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
1146
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
1147 1148 1149 1150 1151 1152 1153

        # 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)

1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166
        # 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
1167 1168 1169 1170
        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
1171
    def _copy_param_info_from(self, other):
1172
        """
1173 1174
        Copy the information of parameters from the other block.

1175
        Args:
1176 1177 1178 1179 1180
            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.
1181 1182 1183 1184 1185

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
1186 1187 1188
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
        for p in other._iter_parameters():
1189 1190 1191
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
W
Wu Yi 已提交
1192
                raise ValueError("_copy_param_info_from should be invoked with "
1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204
                                 "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
1205
                gradient_clip_attr=p.gradient_clip_attr,
F
fengjiayi 已提交
1206
                error_clip=p.error_clip,
1207 1208 1209
                name=v.name)
            self.vars[new_p.name] = new_p

W
Wu Yi 已提交
1210
    def _clone_variable(self, var):
1211 1212
        """
        Clone a variable into current block.
1213

1214 1215 1216 1217
        Args:
            var: the variable to be cloned.

        Returns:
1218
            Variable: the new  variable cloned from 'var' in current block.
1219 1220
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
1221 1222 1223 1224 1225
        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
1226 1227
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
1228
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
1229 1230 1231 1232 1233 1234
        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
1235 1236
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1237 1238 1239 1240 1241 1242 1243
        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
1244 1245
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1246
        return ret_var
1247

Y
Yu Yang 已提交
1248 1249

class Program(object):
D
dzhwinter 已提交
1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260
    """
    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
1261
    default_main_program run in every mini batch and adjust the weights.
D
dzhwinter 已提交
1262 1263

    Returns:
Y
yuyang18 已提交
1264
        A empty program.
D
dzhwinter 已提交
1265 1266

    Examples:
Y
yuyang18 已提交
1267 1268 1269 1270 1271 1272
        >>> 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
1273 1274 1275

    """

1276 1277
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
1278 1279
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
D
dzhwinter 已提交
1280
        self._seed = 0
Y
yuyang18 已提交
1281
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
Y
yuyang18 已提交
1282
        self._op_role_var = []
Y
yuyang18 已提交
1283 1284 1285

    @property
    def op_role(self):
Y
yuyang18 已提交
1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
        """
        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
1299 1300 1301 1302 1303 1304 1305 1306
        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
1307 1308 1309 1310 1311 1312 1313
        """
        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
1314 1315 1316 1317
        return self._op_role_var

    @op_role_var.setter
    def set_op_role_var(self, var_name):
Y
yuyang18 已提交
1318
        self._op_role_var = [var_name]
Y
yuyang18 已提交
1319 1320

    @contextlib.contextmanager
1321
    def optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
1322 1323 1324 1325 1326 1327 1328
        """
        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:
1329
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
1330 1331 1332 1333

        Examples:

            >>> p, g = backward(...)
1334
            >>> with program.optimized_guard([p,g]):
Y
yuyang18 已提交
1335 1336
            >>>     p = p - 0.001 * g
        """
Y
yuyang18 已提交
1337 1338
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
1339 1340 1341 1342
        self._op_role_var = [
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
Y
yuyang18 已提交
1343
        yield
Y
yuyang18 已提交
1344
        self._op_role_var = []
Y
yuyang18 已提交
1345
        self._current_role = OpRole.Forward
Y
Yu Yang 已提交
1346

1347
    def __str__(self):
Y
yuyang18 已提交
1348 1349 1350 1351 1352 1353 1354 1355 1356
        """
        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
1357 1358
        return self.to_string(True)

F
fengjiayi 已提交
1359 1360 1361
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
1362

F
fengjiayi 已提交
1363
        Args:
Y
yuyang18 已提交
1364 1365
            throw_on_error(bool): raise Value error when any of required fields
                is not set.
F
fengjiayi 已提交
1366

Y
yuyang18 已提交
1367 1368 1369 1370 1371 1372 1373 1374 1375 1376
            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
1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389

        """
        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()
            proto = framework_pb2.ProgramDesc.FromString(str(protostr))
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1390

1391
    def get_desc(self):
Y
yuyang18 已提交
1392 1393 1394 1395 1396 1397 1398
        """
        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.
        """
1399 1400
        return self.desc

1401
    def clone(self, for_test=False):
Y
yuyang18 已提交
1402 1403 1404
        """
        Create a new, duplicated program.

1405

Y
yuyang18 已提交
1406 1407 1408 1409
        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`.
1410

Y
yuyang18 已提交
1411 1412 1413 1414
        * 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
1415 1416 1417 1418 1419
        :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()
1420 1421

        Args:
Y
yuyang18 已提交
1422 1423
            for_test(bool): True if change the :code:`is_test` attribute of
                operators to :code:`True`.
1424

D
dzhwinter 已提交
1425
        Returns:
Y
yuyang18 已提交
1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478
            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.
1479 1480
        """
        if for_test:
1481
            p = self.inference_optimize()
1482
        else:
1483
            p = Program()
1484
            p.desc = core.ProgramDesc(self.desc)
1485
            p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())]
W
Wu Yi 已提交
1486
            p._sync_with_cpp()
1487

W
Wu Yi 已提交
1488
        p._copy_param_info_from(self)
F
fengjiayi 已提交
1489
        p.copy_data_info_from(self)
Y
Yu Yang 已提交
1490
        return p
1491

1492
    def prune(self, targets):
Y
yuyang18 已提交
1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507
        """
        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.

        """
1508 1509 1510 1511 1512 1513
        if not isinstance(targets, list):
            targets = [targets]
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
1514 1515
                    # After transpiler processing, the op that output this
                    # variable maybe has been changed, so t.op is not reliable
1516
                    # and we need to find the current op that generate this
1517 1518 1519 1520 1521 1522 1523 1524
                    # 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

1525
                    t = t.op
1526 1527 1528 1529
                    if t is None:
                        raise ValueError(
                            "The target variable must have an "
                            "associated operator that generates it.")
1530
                else:
1531 1532
                    raise ValueError("All targets of prune() can only be "
                                     "Variable or Operator.")
1533 1534 1535 1536 1537

            targets_idx.append([t.block.idx, t.idx])
        res = Program()
        res.desc = core.prune(self.desc, targets_idx)
        res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())]
W
Wu Yi 已提交
1538
        res._sync_with_cpp()
1539 1540
        return res

1541
    def inference_optimize(self):
Y
yuyang18 已提交
1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552
        """
        This method will create a new program and change the :code:`is_test`
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
1553 1554
        # this is an alternative implement before
        # core.inference_optimize being fixed.
1555
        res = Program()
1556 1557 1558 1559 1560 1561 1562
        res.desc = core.ProgramDesc(self.desc)
        for i in xrange(res.desc.num_blocks()):
            block = res.desc.block(i)
            for j in xrange(block.op_size()):
                op = block.op(j)
                if op.has_attr('is_test'):
                    op.set_attr('is_test', True)
1563
        res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())]
W
Wu Yi 已提交
1564
        res._sync_with_cpp()
1565 1566
        return res

1567 1568
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580
        """
        Deserialize a program desc from protobuf binary string.

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

        Args:
            binary_str(str): The binary prootbuf string.

        Returns:
            Program: A deserialized program desc.
        """
1581 1582
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
1583
        p.blocks = [Block(p, i) for i in xrange(p.desc.num_blocks())]
W
Wu Yi 已提交
1584
        p._sync_with_cpp()
1585
        return p
Y
Yu Yang 已提交
1586

D
dzhwinter 已提交
1587 1588
    @property
    def random_seed(self):
Y
yuyang18 已提交
1589 1590 1591 1592 1593 1594
        """
        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
1595 1596
        return self._seed

Q
qiaolongfei 已提交
1597 1598
    @property
    def num_blocks(self):
Y
yuyang18 已提交
1599 1600 1601
        """
        The number of blocks in this program.
        """
Q
qiaolongfei 已提交
1602 1603
        return self.desc.num_blocks()

D
dzhwinter 已提交
1604 1605 1606 1607 1608 1609
    @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
1610 1611
    def __repr__(self):
        return str(self)
1612

Y
Yu Yang 已提交
1613
    def global_block(self):
Y
yuyang18 已提交
1614 1615 1616
        """
        Get the first block of this program.
        """
Y
Yu Yang 已提交
1617 1618
        return self.blocks[0]

Q
Qiao Longfei 已提交
1619
    def block(self, index):
Y
yuyang18 已提交
1620 1621 1622 1623 1624 1625 1626 1627
        """
        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
1628 1629
        return self.blocks[index]

Y
Yu Yang 已提交
1630
    def current_block(self):
Y
yuyang18 已提交
1631 1632 1633 1634
        """
        Get the current block. The :code:`current` block is the block to append
        operators.
        """
Y
Yu Yang 已提交
1635 1636
        return self.blocks[self.current_block_idx]

F
update  
fengjiayi 已提交
1637
    def create_block(self, parent_idx=None):
Y
yuyang18 已提交
1638 1639 1640 1641 1642 1643 1644 1645 1646 1647
        """
        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
1648
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
1649 1650 1651
        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
1652 1653 1654 1655 1656
        self.current_block_idx = new_block_idx
        self.blocks.append(Block(self, self.current_block_idx))
        return self.current_block()

    def rollback(self):
Y
yuyang18 已提交
1657 1658 1659 1660 1661
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
1662 1663
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
1664
    def _sync_with_cpp(self):
Y
yuyang18 已提交
1665 1666 1667 1668 1669 1670 1671 1672 1673 1674
        """
        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
1675 1676 1677
        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
1678
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
1679

W
Wu Yi 已提交
1680
    def _copy_param_info_from(self, other):
1681
        """
1682
        Copy the information of parameters from other program.
D
dzhwinter 已提交
1683

Y
yuyang18 已提交
1684 1685 1686
        Notes: This is a very low level API. Users should not invoke it
        directly.

1687 1688 1689 1690 1691 1692 1693
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
1694
            raise TypeError("_copy_param_info_from should be invoked with "
1695 1696 1697
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
1698
            raise ValueError("_copy_param_info_from should be invoked with two "
1699
                             "program, with represent the same topology")
W
Wu Yi 已提交
1700
        self.global_block()._copy_param_info_from(other.global_block())
1701

F
fengjiayi 已提交
1702 1703 1704
    def copy_data_info_from(self, other):
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
1705

Y
yuyang18 已提交
1706 1707 1708
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
1709 1710 1711 1712 1713 1714 1715
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
1716
            raise TypeError("_copy_param_info_from should be invoked with "
F
fengjiayi 已提交
1717 1718 1719
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
1720
            raise ValueError("_copy_param_info_from should be invoked with two "
F
fengjiayi 已提交
1721 1722 1723 1724 1725
                             "program, with represent the same topology")
        for var in other.global_block().vars.itervalues():
            if var.is_data:
                self.global_block().var(var.name).is_data = True

1726
    def list_vars(self):
Y
yuyang18 已提交
1727 1728 1729 1730 1731 1732
        """
        Get all variables from this Program. A iterable object is returned.

        Returns:
            iterable: The generator will yield every variable in this program.
        """
1733 1734 1735 1736
        for each_block in self.blocks:
            for each_var in each_block.vars.itervalues():
                yield each_var

Y
Yu Yang 已提交
1737

Y
Yu Yang 已提交
1738
class Parameter(Variable):
1739 1740 1741 1742 1743 1744
    """
    Parameter is derived from Variable. A parameter is a persistable 
    Variable, and will be updated by optimizers after each iteration.
    The training of a neural network is essentially the updating of 
    its parameters.

1745
    Relative to a general Variable, a Parameter has several its own
1746 1747
    member variables:

1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759
    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.
1760 1761
    """

Y
Yu Yang 已提交
1762 1763 1764 1765 1766 1767 1768 1769 1770 1771
    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")
1772 1773 1774

        Variable.__init__(
            self, block, persistable=True, shape=shape, dtype=dtype, **kwargs)
Y
Yu Yang 已提交
1775 1776 1777 1778
        self.trainable = kwargs.get('trainable', True)

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

1779 1780
        self.regularizer = kwargs.get('regularizer', None)

F
fengjiayi 已提交
1781
        self.gradient_clip_attr = kwargs.get('gradient_clip_attr', None)
Y
Yu Yang 已提交
1782

W
wanghaoshuang 已提交
1783
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
1784

F
fengjiayi 已提交
1785 1786 1787
    def __str__(self):
        return self.to_string(True)

F
update  
fengjiayi 已提交
1788 1789 1790
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
1791

F
update  
fengjiayi 已提交
1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805
        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
1806
                               "gradient_clip_attr", "do_model_average")
F
update  
fengjiayi 已提交
1807 1808 1809 1810 1811
            for attr_name in additional_attr:
                res_str += "%s: %s\n" % (attr_name,
                                         str(getattr(self, attr_name)))
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
1812 1813 1814 1815
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
1816

Y
Yu Yang 已提交
1817
# program is a global instance.
Y
Yu Yang 已提交
1818 1819
_main_program_ = Program()
_startup_program_ = Program()
1820

1821

1822
def default_startup_program():
Y
Yu Yang 已提交
1823
    """
Y
yuyang18 已提交
1824 1825 1826 1827 1828 1829 1830 1831 1832
    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.
1833

Y
Yu Yang 已提交
1834 1835 1836
    Returns:
        Program: startup program
    """
Y
Yu Yang 已提交
1837
    return _startup_program_
1838

1839

1840
def default_main_program():
Y
Yu Yang 已提交
1841
    """
Y
yuyang18 已提交
1842 1843 1844 1845 1846 1847 1848 1849 1850
    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.
1851

Y
Yu Yang 已提交
1852 1853 1854
    Returns:
        Program: main program
    """
Y
Yu Yang 已提交
1855
    return _main_program_
Y
Yu Yang 已提交
1856 1857 1858 1859 1860


def switch_main_program(program):
    """
    Switch the main program to a new program.
1861

Y
Yu Yang 已提交
1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875
    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):
    """
1876
    Switch the startup program to a new program
Y
Yu Yang 已提交
1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891
    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
1892 1893 1894
    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.
1895

Y
Yu Yang 已提交
1896
    Examples:
Y
yuyang18 已提交
1897 1898 1899 1900 1901 1902 1903 1904 1905 1906

        >>> 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.
1907

Y
Yu Yang 已提交
1908
    Examples:
Y
yuyang18 已提交
1909 1910 1911 1912 1913 1914

        >>> 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 = ...
1915

Y
Yu Yang 已提交
1916
    Args:
Y
yuyang18 已提交
1917
        main_program(Program): New main program inside `with` statement.
1918
        startup_program(Program): New startup program inside `with` statement.
Y
Yu Yang 已提交
1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931
            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
1932 1933 1934 1935


def get_var(name, program=None):
    """
Y
yuyang18 已提交
1936
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
1937

X
xuwei06 已提交
1938 1939 1940
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
1941
        If None, default_global_program() will be used.
X
xuwei06 已提交
1942 1943 1944 1945 1946 1947 1948

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
1949
    assert isinstance(program, Program)
X
xuwei06 已提交
1950 1951

    return program.global_block().var(name)