framework.py 74.4 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15 16
from __future__ import print_function

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

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(object):
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):
Y
Yu Yang 已提交
280
        self.block = block
F
fengjiayi 已提交
281
        self.error_clip = error_clip
Y
Yu Yang 已提交
282 283

        if name is None:
Y
Yu Yang 已提交
284
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
285
        is_new_var = False
M
minqiyang 已提交
286
        name = cpt.to_text(name)
M
minqiyang 已提交
287
        self.desc = self.block.desc.find_var(cpt.to_bytes(name))
D
Dong Zhihong 已提交
288 289

        if self.desc is None:
M
minqiyang 已提交
290
            self.desc = self.block.desc.var(cpt.to_bytes(name))
Y
Yu Yang 已提交
291
            is_new_var = True
Y
Yu Yang 已提交
292

Y
Yu Yang 已提交
293 294 295 296 297 298 299 300
        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
301
        if shape is not None:
Y
Yu Yang 已提交
302
            if is_new_var:
303
                self.desc.set_shape(shape)
Y
Yu Yang 已提交
304 305 306 307 308 309 310 311
            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
312
        if dtype is not None:
313
            if not isinstance(dtype, core.VarDesc.VarType):
314
                dtype = convert_np_dtype_to_dtype_(dtype)
Y
Yu Yang 已提交
315
            if is_new_var:
F
fengjiayi 已提交
316
                self.desc.set_dtype(dtype)
Y
Yu Yang 已提交
317
            else:
F
fengjiayi 已提交
318
                old_dtype = self.dtype
Q
QI JUN 已提交
319
                if dtype != old_dtype:
Y
Yu Yang 已提交
320 321 322 323 324
                    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
325 326

        if lod_level is not None:
Y
Yu Yang 已提交
327
            if is_new_var:
328
                self.desc.set_lod_level(lod_level)
Y
Yu Yang 已提交
329 330 331 332 333 334 335
            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))
336 337 338 339 340 341 342 343 344 345 346
        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))

347 348 349 350 351 352 353 354
        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
355
        self.block.vars[name] = self
Y
Yu Yang 已提交
356
        self.op = None
Y
Yu Yang 已提交
357
        self.stop_gradient = stop_gradient
F
fengjiayi 已提交
358
        self.is_data = is_data
X
Xin Pan 已提交
359 360 361
        if _in_imperative_mode():
            self._ivar = core.VarBase()
            self._ivar.desc = self.desc
Y
Yu Yang 已提交
362

363 364 365 366 367 368 369
    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)
X
Xin Pan 已提交
370
        self._ivar._run_backward(scope)
371 372

    def _gradient(self):
X
Xin Pan 已提交
373
        return np.array(self._ivar._grad())
374

375
    def __str__(self):
Y
Yang Yang(Tony) 已提交
376 377
        return self.to_string(True)

F
update  
fengjiayi 已提交
378
    def to_string(self, throw_on_error, with_details=False):
379 380 381 382
        """
        Get debug string.

        Args:
383 384
            throw_on_error(bool): True if raise an exception when self is
                not initialized.
F
update  
fengjiayi 已提交
385
            with_details(bool): more details about variables and parameters
386 387
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False;
388

389 390
        Returns:
            str: The debug string.
391
        """
F
update  
fengjiayi 已提交
392 393
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
394
        protostr = self.desc.serialize_to_string()
395
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
396 397 398 399
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
400 401
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
402
        return res_str
403 404 405

    __repr__ = __str__

W
Wu Yi 已提交
406
    def _set_desc(self, input):
407 408 409 410 411 412 413 414 415
        """
        Set the variable description.

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

        Returns:
            None
        """
416 417
        self.desc = input

418 419 420 421
    @property
    def persistable(self):
        return self.desc.persistable()

Y
Yu Yang 已提交
422 423 424 425
    @persistable.setter
    def persistable(self, p):
        self.desc.set_persistable(p)

Y
Yu Yang 已提交
426 427
    @property
    def name(self):
M
minqiyang 已提交
428
        return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
429

T
typhoonzero 已提交
430 431 432 433
    @name.setter
    def name(self, new_name):
        self.desc.set_name(new_name)

Y
Yu Yang 已提交
434 435 436
    @property
    def shape(self):
        # convert to tuple, make it as same as numpy API.
437
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
438 439

    @property
F
fengjiayi 已提交
440 441
    def dtype(self):
        return self.desc.dtype()
Y
Yu Yang 已提交
442 443 444

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

Y
Yu Yang 已提交
447 448 449 450
    @property
    def type(self):
        return self.desc.type()

W
Wu Yi 已提交
451
    def _set_error_clip(self, error_clip):
452 453 454 455 456 457 458 459 460
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
461 462
        self.error_clip = error_clip

Y
Yu Yang 已提交
463

F
fengjiayi 已提交
464 465 466
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
467

468 469
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
470 471 472 473
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
474
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
475 476 477 478 479
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
480 481 482 483
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
484 485 486 487 488 489 490 491 492
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
493
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
494 495 496 497 498 499
        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):
500 501 502 503 504 505 506 507
        """
        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
508 509
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
510 511
        return self.op_proto_map[type]

512 513 514 515
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
516
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
517
            core.op_proto_and_checker_maker.kOpNameScopeAttrName()
518 519
        }

F
fengjiayi 已提交
520

X
Xin Pan 已提交
521
class Operator(object):
522
    """
523 524 525 526 527 528 529
    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
530
        type(str): The type of operator. Default None.
531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550
        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
551
        Block.append_op or Block._prepend_op instead.
552 553 554 555 556 557 558 559 560 561

    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]})
562
    """
563 564 565 566
    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
567
        'ncclInit', 'select', 'checkpoint_notify', 'gen_nccl_id'
568
    }
569

Y
Yu Yang 已提交
570 571
    def __init__(self,
                 block,
Y
Yu Yang 已提交
572
                 desc,
Y
Yu Yang 已提交
573 574 575 576 577
                 type=None,
                 inputs=None,
                 outputs=None,
                 attrs=None):
        self.block = block
Y
Yu Yang 已提交
578
        self.desc = desc
G
gongweibao 已提交
579 580 581 582 583
        # 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
584 585 586 587
        del attrs

        op_maker = core.op_proto_and_checker_maker

G
gongweibao 已提交
588 589
        if op_maker.kOpRoleAttrName() not in op_attrs:
            op_attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role
Y
yuyang18 已提交
590 591 592

        role_var_name = op_maker.kOpRoleVarAttrName()
        if len(self.block.program.
G
gongweibao 已提交
593 594
               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
595

G
gongweibao 已提交
596 597
        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
598

F
fengjiayi 已提交
599 600 601 602 603
        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
604
        self.desc.set_type(type)
F
fengjiayi 已提交
605
        proto = OpProtoHolder.instance().get_op_proto(type)
606

607 608 609
        namescope_var_name = op_maker.kOpNameScopeAttrName()
        op_attrs[namescope_var_name] = _full_name_scope()

Y
Yang Yang(Tony) 已提交
610 611
        def find_name(var_list, name):
            for var_name in var_list:
Q
Qiao Longfei 已提交
612
                if var_list[var_name] is not None and var_name == name:
Y
Yang Yang(Tony) 已提交
613 614
                    return True
            return False
Q
QI JUN 已提交
615

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

Y
Yu Yang 已提交
642
        if outputs is not None:
643 644 645 646 647 648 649
            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
650 651
                raise ValueError(("Incorrect setting for output(s) of "
                                  "operator \"%s\". Need: [%s] Given: [%s]") %
652 653 654
                                 (type,
                                  ", ".join(six.binary_type(e) for e in need),
                                  ", ".join(six.binary_type(e) for e in given)))
655

F
fengjiayi 已提交
656
            for out_proto in proto.outputs:
657 658 659 660
                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
661 662
                    raise ValueError(
                        "Output %s expects only one output, but %d are given." %
663 664 665
                        (out_proto.name, len(out_args)))
                out_arg_names = []
                for arg in out_args:
M
minqiyang 已提交
666
                    out_arg_names.append(cpt.to_text(arg.name))
667 668
                    arg.op = self
                self.desc.set_output(out_proto.name, out_arg_names)
F
Update  
fengjiayi 已提交
669

G
gongweibao 已提交
670 671
        if op_attrs is not None:
            if not isinstance(op_attrs, dict):
672
                raise TypeError("'attrs' should be a dict.")
F
fengjiayi 已提交
673
            for attr in proto.attrs:
F
Update  
fengjiayi 已提交
674
                attr_name = attr.name
G
gongweibao 已提交
675
                if (attr_name not in op_attrs) or (op_attrs[attr_name] is None):
F
Update  
fengjiayi 已提交
676
                    continue
G
gongweibao 已提交
677
                attr_val = op_attrs[attr_name]
G
gongweibao 已提交
678 679
                self._update_desc_attr(attr_name, attr_val)

680
        self.desc.check_attrs()
W
Wu Yi 已提交
681
        if self._has_kernel(type):
Q
QI JUN 已提交
682
            self.desc.infer_var_type(self.block.desc)
Y
Yu Yang 已提交
683
            self.desc.infer_shape(self.block.desc)
X
Xin Pan 已提交
684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700
        if _in_imperative_mode():
            self.iop = core.OpBase()
            self.iop.desc = self.desc
            self.inputs = []
            if inputs is not None:
                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 = []
            if outputs is not None:
                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[:])
F
fengjiayi 已提交
701

W
Wu Yi 已提交
702
    def _has_kernel(self, op_type):
703 704
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
705
    def to_string(self, throw_on_error):
706
        """
707 708
        Get debug string.

709
        Args:
710 711
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
712

713 714
        Returns:
            str: The debug string.
715 716

        """
717
        protostr = self.desc.serialize_to_string()
718
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
719 720 721 722
        return _debug_string_(proto, throw_on_error)

    def __str__(self):
        return self.to_string(True)
723 724 725

    __repr__ = __str__

F
fengjiayi 已提交
726 727 728 729 730
    @property
    def type(self):
        return self.desc.type()

    def input(self, name):
731
        """
732
        Get the input arguments according to the input parameter name.
733

734 735
        Args:
            name(str): The input parameter name.
736

737 738 739
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
740
        """
F
fengjiayi 已提交
741 742
        return self.desc.input(name)

W
Wu Yi 已提交
743
    def _rename_input(self, old_name, new_name):
744 745 746 747 748 749 750 751 752 753
        """
        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
754
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
755

W
Wu Yi 已提交
756
    def _rename_output(self, old_name, new_name):
757 758 759 760 761 762 763 764 765 766
        """
        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
767
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
768

F
fengjiayi 已提交
769 770 771 772
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
773 774 775 776 777 778 779 780
    @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
781
    def output(self, name):
782
        """
783
        Get output arguments by the output parameter name.
784

785 786
        Args:
            name(str): The output parameter name.
787

788 789 790
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
791
        """
F
fengjiayi 已提交
792 793 794 795 796 797
        return self.desc.output(name)

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

798 799 800 801 802 803 804 805
    @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
806
    def has_attr(self, name):
807
        """
808 809
        Whether this Operator has the attribute with name or not.

810
        Args:
811
            name(str): the attribute name.
812

813 814
        Returns:
            bool: True if has this attribute.
815 816

        """
F
fengjiayi 已提交
817 818 819
        return self.desc.has_attr(name)

    def attr_type(self, name):
820
        """
821
        Get the type of attribute by attribute's name.
822

823 824
        Args:
            name(str): the attribute name.
825

826 827
        Returns:
            core.AttrType: the attribute type.
828
        """
F
fengjiayi 已提交
829 830
        return self.desc.attr_type(name)

W
Wu Yi 已提交
831
    def _set_attr(self, name, val):
832 833 834 835 836 837 838 839 840 841
        """
        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
842 843 844 845 846 847 848 849 850 851 852 853 854
        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
855 856
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
857 858
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
859
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
860 861 862 863
        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
864
            self.desc._set_attr(name, val)
Y
yuyang18 已提交
865

F
fengjiayi 已提交
866 867 868 869 870
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
871
        """
872 873
        Get the attribute by name.

874
        Args:
875
            name(str): the attribute name.
876

877 878
        Returns:
            bool|int|str|float|list: The attribute value. The return value
879 880
            can be any valid attribute type.
        """
F
fengjiayi 已提交
881
        return self.desc.attr(name)
Y
Yu Yang 已提交
882

W
Wu Yi 已提交
883
    def _block_attr_id(self, name):
884
        """
G
gongweibao 已提交
885
        Get the block attribute's id by name.
886

887 888
        Args:
            name(str): the attribute name.
889

890 891
        Returns:
            int: the block index.
892
        """
W
Wu Yi 已提交
893
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
894

W
Wu Yi 已提交
895
    def _block_attr(self, name):
G
gongweibao 已提交
896 897 898 899 900 901 902 903 904 905
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
906
        id = self._block_attr_id(name)
G
gongweibao 已提交
907 908 909
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
910
    def _blocks_attr(self, name):
G
gongweibao 已提交
911 912 913 914 915 916 917 918 919 920
        """
        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
921
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
922 923 924 925 926
            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
927
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
928 929 930 931 932 933 934 935 936 937
        """
        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
938
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
939

J
JiayiFeng 已提交
940
    def all_attrs(self):
F
fengjiayi 已提交
941
        """
942 943 944
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
945
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
946 947 948 949
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
G
gongweibao 已提交
950 951
            attr_type = self.desc.attr_type(n)
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
952
                attr_map[n] = self._block_attr(n)
G
gongweibao 已提交
953 954 955
                continue

            if attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
956
                attr_map[n] = self._blocks_attr(n)
G
gongweibao 已提交
957 958 959 960
                continue

            attr_map[n] = self.attr(n)

F
fengjiayi 已提交
961 962
        return attr_map

Y
Yu Yang 已提交
963

Y
Yu Yang 已提交
964
class Block(object):
965 966 967 968 969 970 971 972 973 974 975 976 977 978
    """
    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
979
        use `Program._create_block()` to create a block.
980 981 982 983 984 985 986 987 988 989 990 991 992 993

    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
994
    def __init__(self, program, idx):
Y
Yu Yang 已提交
995
        self.desc = program.desc.block(idx)
996
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
997
        self.ops = list()  # operator list
Y
Yu Yang 已提交
998
        self.program = program
999
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
1000

1001
    def __str__(self):
Y
Yang Yang(Tony) 已提交
1002 1003
        return self.to_string(True)

F
fengjiayi 已提交
1004 1005
    def to_string(self, throw_on_error, with_details=False):
        """
1006 1007
        Get debug string.

F
fengjiayi 已提交
1008 1009
        Args:
            throw_on_error(bool): raise exception when self is not initialized
1010
                when throw_on_error is True.
F
update  
fengjiayi 已提交
1011
            with_details(bool): more details about variables and parameters
1012 1013
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
1014

1015 1016
        Returns:
            str: The debug string.
F
fengjiayi 已提交
1017 1018 1019 1020
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
1021
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
1022 1023
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
1024
            for var in list(self.vars.values()):
F
fengjiayi 已提交
1025
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
1026
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
1027
            for op in self.ops:
F
fengjiayi 已提交
1028 1029
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
1030 1031 1032
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
1033 1034
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
1035 1036
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1037 1038 1039

    __repr__ = __str__

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

Y
Yu Yang 已提交
1044 1045 1046 1047
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
1048
    def _set_forward_block_idx(self, idx):
1049 1050 1051 1052 1053 1054 1055 1056 1057
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
1058
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
1059

Y
Yu Yang 已提交
1060 1061
    @property
    def idx(self):
Y
Yu Yang 已提交
1062
        return self.desc.id
Y
Yu Yang 已提交
1063

Q
Qiao Longfei 已提交
1064
    def var(self, name):
1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077
        """
        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.
        """
1078
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
1079 1080 1081
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
1082 1083
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
1084
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
1085
        return v
Q
Qiao Longfei 已提交
1086

X
Xin Pan 已提交
1087
    def _find_var_recursive(self, name):
1088 1089 1090 1091 1092 1093 1094
        """
        Get a Variable by name from this block recursively.

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

        Returns:
X
Xin Pan 已提交
1095
            Variable: the Variable with the giving name. Or None if not found.
1096
        """
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
        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))
X
Xin Pan 已提交
1121
        return None
Y
Yu Yang 已提交
1122

X
Xin Pan 已提交
1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141
    def _var_recursive(self, name):
        """
        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.
        """
        var = self._find_var_recursive(name)
        if var:
            return var
        else:
            raise ValueError("Var {0} is not found recursively".format(name))
F
fengjiayi 已提交
1142

Q
Qiao Longfei 已提交
1143
    def all_parameters(self):
1144
        return list(self.iter_parameters())
1145

1146
    def iter_parameters(self):
M
minqiyang 已提交
1147
        return (item[1] for item in six.iteritems(self.vars)
1148
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
1149

Y
Yu Yang 已提交
1150
    def create_var(self, *args, **kwargs):
1151
        var = Variable(block=self, *args, **kwargs)
1152 1153
        if 'initializer' in kwargs:
            kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
1154
        return var
Y
Yu Yang 已提交
1155

Q
Qiao Longfei 已提交
1156 1157 1158
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
1159
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
1160 1161
        """
        Rename variable in vars and ops' inputs and outputs
1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173

        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
1174
        """
M
minqiyang 已提交
1175 1176
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
1177

T
typhoonzero 已提交
1178
        if not self.has_var(name):
1179
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
1180 1181
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
1182
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
1183 1184 1185 1186 1187 1188 1189
            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
1190
            var_type = "Variable"
T
wip  
typhoonzero 已提交
1191 1192 1193 1194
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
1195
        orig_var_type = v.type
M
minqiyang 已提交
1196
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
1197
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
1198
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
1199
        if var_type == "Parameter":
T
wip  
typhoonzero 已提交
1200 1201 1202 1203
            var = Parameter(
                self,
                d.shape(),
                d.dtype(),
T
typhoonzero 已提交
1204
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1205 1206 1207 1208 1209 1210 1211
                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
1212
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
1213 1214
            var = Variable(
                self,
T
typhoonzero 已提交
1215
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1216 1217 1218 1219
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
1220
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
1221 1222 1223
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
1224
        self._sync_with_cpp()
1225
        return var
T
typhoonzero 已提交
1226

W
Wu Yi 已提交
1227 1228
    def _remove_var(self, name):
        self._sync_with_cpp()
M
minqiyang 已提交
1229
        self.desc._remove_var(cpt.to_bytes(name))
1230 1231
        del self.vars[name]

Y
Yu Yang 已提交
1232 1233
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
Q
Qiao Longfei 已提交
1234
        param = Parameter(global_block, *args, **kwargs)
1235
        if 'initializer' in kwargs:
1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255

            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
1256
        return param
Y
Yu Yang 已提交
1257

Y
Yu Yang 已提交
1258
    def append_op(self, *args, **kwargs):
1259 1260 1261 1262 1263 1264
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
Y
Yu Yang 已提交
1265
        op_desc = self.desc.append_op()
1266
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
1267
        if _in_imperative_mode():
X
Xin Pan 已提交
1268 1269
            _imperative_tracer().trace(op.iop, [v._ivar for v in op.inputs],
                                       [v._ivar for v in op.outputs], self.desc)
Y
Yu Yang 已提交
1270 1271 1272
        self.ops.append(op)
        return op

W
Wu Yi 已提交
1273
    def _insert_op(self, index, *args, **kwargs):
1274 1275 1276 1277 1278 1279 1280 1281 1282
        """
        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
1283 1284
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
1285 1286 1287 1288
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

W
Wu Yi 已提交
1289
    def _remove_op(self, index):
1290 1291 1292 1293 1294 1295 1296 1297 1298
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
W
Wu Yi 已提交
1299 1300
        self._sync_with_cpp()
        self.desc._remove_op(index, index + 1)
1301 1302
        del self.ops[index]

W
Wu Yi 已提交
1303
    def _slice_ops(self, start, end):
1304 1305 1306 1307 1308 1309 1310 1311 1312 1313
        """
        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
1314
        return self.ops[start:end]
Y
Yancey1989 已提交
1315

W
Wu Yi 已提交
1316 1317
    def _prepend_op(self, *args, **kwargs):
        op_desc = self.desc._prepend_op()
Y
Yu Yang 已提交
1318
        op = Operator(self, op_desc, *args, **kwargs)
Q
qiaolongfei 已提交
1319
        self.ops.insert(0, op)
Y
Yu Yang 已提交
1320 1321
        return op

W
Wu Yi 已提交
1322
    def _sync_with_cpp(self):
1323
        """
1324 1325
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
1326
        """
Q
Qiao Longfei 已提交
1327 1328 1329 1330 1331
        # 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())

1332
        # sync variables removed from c++ end
1333
        for var in list(self.vars.keys()):
M
minqiyang 已提交
1334
            if not self.desc.find_var(cpt.to_bytes(var)):
1335 1336
                self.vars.pop(var)

Q
Qiao Longfei 已提交
1337
        # sync operators from cpp
1338 1339 1340 1341
        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
1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357
        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
1358 1359 1360 1361 1362

        # 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
1363
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
1364 1365 1366 1367 1368 1369 1370

        # 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)

1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383
        # 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
1384 1385 1386 1387
        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
1388
    def _copy_param_info_from(self, other):
1389
        """
1390 1391
        Copy the information of parameters from the other block.

1392
        Args:
1393 1394 1395 1396 1397
            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.
1398 1399 1400 1401 1402

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
1403 1404
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
1405
        for p in other.iter_parameters():
1406 1407 1408
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
W
Wu Yi 已提交
1409
                raise ValueError("_copy_param_info_from should be invoked with "
1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421
                                 "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
1422
                gradient_clip_attr=p.gradient_clip_attr,
F
fengjiayi 已提交
1423
                error_clip=p.error_clip,
1424 1425 1426
                name=v.name)
            self.vars[new_p.name] = new_p

W
Wu Yi 已提交
1427
    def _clone_variable(self, var):
1428 1429
        """
        Clone a variable into current block.
1430

1431 1432 1433 1434
        Args:
            var: the variable to be cloned.

        Returns:
1435
            Variable: the new  variable cloned from 'var' in current block.
1436 1437
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
1438 1439 1440 1441 1442
        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
1443 1444
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
1445
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
1446 1447 1448 1449 1450 1451
        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
1452 1453
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1454 1455 1456 1457 1458 1459 1460
        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
1461 1462
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1463
        return ret_var
1464

Y
Yu Yang 已提交
1465 1466

class Program(object):
D
dzhwinter 已提交
1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477
    """
    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
1478
    default_main_program run in every mini batch and adjust the weights.
D
dzhwinter 已提交
1479 1480

    Returns:
Y
yuyang18 已提交
1481
        A empty program.
D
dzhwinter 已提交
1482 1483

    Examples:
Y
yuyang18 已提交
1484 1485 1486 1487 1488 1489
        >>> 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
1490 1491 1492

    """

1493 1494
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
1495 1496
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
D
dzhwinter 已提交
1497
        self._seed = 0
Y
yuyang18 已提交
1498
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
Y
yuyang18 已提交
1499
        self._op_role_var = []
T
tangwei12 已提交
1500 1501 1502 1503

        # for distribute
        self._is_distributed = False
        self._is_chief = False
T
tangwei12 已提交
1504
        self._slice_vars_and_attrs = []
T
tangwei12 已提交
1505
        self._endpoints = []
1506
        self._trainers_endpoints = []
T
tangwei12 已提交
1507
        self._distributed_lookup_table = None
Y
yuyang18 已提交
1508 1509 1510

    @property
    def op_role(self):
Y
yuyang18 已提交
1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523
        """
        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
1524 1525 1526 1527 1528 1529 1530 1531
        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
1532 1533 1534 1535 1536 1537 1538
        """
        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
1539 1540 1541 1542
        return self._op_role_var

    @op_role_var.setter
    def set_op_role_var(self, var_name):
Y
yuyang18 已提交
1543
        self._op_role_var = [var_name]
Y
yuyang18 已提交
1544 1545

    @contextlib.contextmanager
W
Wu Yi 已提交
1546
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
1547 1548 1549 1550 1551 1552 1553
        """
        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:
1554
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
1555 1556 1557 1558

        Examples:

            >>> p, g = backward(...)
W
Wu Yi 已提交
1559
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
1560 1561
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
1562 1563 1564
        tmp_role = self._current_role
        tmp_var = self._op_role_var

Y
yuyang18 已提交
1565 1566
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
1567 1568 1569 1570
        self._op_role_var = [
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
Y
yuyang18 已提交
1571
        yield
X
Xin Pan 已提交
1572 1573
        self._op_role_var = tmp_var
        self._current_role = tmp_role
Y
Yu Yang 已提交
1574

1575
    @contextlib.contextmanager
X
Xin Pan 已提交
1576
    def _lr_schedule_guard(self, is_with_opt=False):
1577 1578 1579 1580 1581 1582 1583
        """
        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
1584 1585 1586 1587
        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.
1588 1589 1590 1591 1592 1593 1594

        Examples:

            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
1595 1596 1597 1598

        tmp_role = self._current_role
        tmp_var = self._op_role_var

1599 1600
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
1601 1602
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
1603 1604 1605
        # TODO(typhoonzero): how to set target learning rate var
        self._op_role_var = []
        yield
1606 1607
        self._op_role_var = tmp_var
        self._current_role = tmp_role
1608

1609
    def __str__(self):
Y
yuyang18 已提交
1610 1611 1612 1613 1614 1615 1616 1617 1618
        """
        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
1619 1620
        return self.to_string(True)

F
fengjiayi 已提交
1621 1622 1623
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
1624

F
fengjiayi 已提交
1625
        Args:
Y
yuyang18 已提交
1626 1627
            throw_on_error(bool): raise Value error when any of required fields
                is not set.
F
fengjiayi 已提交
1628

Y
yuyang18 已提交
1629 1630 1631 1632 1633 1634 1635 1636 1637 1638
            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
1639 1640 1641 1642 1643 1644 1645 1646 1647 1648

        """
        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()
1649 1650
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
1651 1652
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1653

W
Wu Yi 已提交
1654
    def _get_desc(self):
Y
yuyang18 已提交
1655 1656 1657 1658 1659 1660 1661
        """
        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.
        """
1662 1663
        return self.desc

X
version  
Xin Pan 已提交
1664 1665 1666
    def _version(self):
        return self.desc._version()

1667
    def clone(self, for_test=False):
Y
yuyang18 已提交
1668 1669 1670
        """
        Create a new, duplicated program.

1671

Y
yuyang18 已提交
1672 1673 1674 1675
        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`.
1676

Y
yuyang18 已提交
1677 1678 1679 1680
        * 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
1681 1682 1683 1684 1685
        :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()
1686 1687

        Args:
Y
yuyang18 已提交
1688 1689
            for_test(bool): True if change the :code:`is_test` attribute of
                operators to :code:`True`.
1690

D
dzhwinter 已提交
1691
        Returns:
Y
yuyang18 已提交
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 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744
            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.
1745 1746
        """
        if for_test:
X
Xin Pan 已提交
1747
            p = self._inference_optimize(prune_read_op=False)
1748
        else:
1749
            p = Program()
G
gongweibao 已提交
1750 1751
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
1752
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
1753 1754 1755
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
1756 1757 1758 1759

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

W
Wu Yi 已提交
1760
            p._sync_with_cpp()
1761

W
Wu Yi 已提交
1762
        p._copy_param_info_from(self)
W
Wu Yi 已提交
1763
        p._copy_data_info_from(self)
1764
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
1765
        return p
1766

W
Wu Yi 已提交
1767
    def _prune(self, targets):
Y
yuyang18 已提交
1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782
        """
        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.

        """
1783 1784 1785 1786 1787 1788
        if not isinstance(targets, list):
            targets = [targets]
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
1789 1790
                    # After transpiler processing, the op that output this
                    # variable maybe has been changed, so t.op is not reliable
1791
                    # and we need to find the current op that generate this
1792 1793 1794 1795 1796 1797 1798 1799
                    # 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

1800
                    t = t.op
1801 1802 1803 1804
                    if t is None:
                        raise ValueError(
                            "The target variable must have an "
                            "associated operator that generates it.")
1805
                else:
1806 1807
                    raise ValueError("All targets of prune() can only be "
                                     "Variable or Operator.")
1808 1809 1810 1811

            targets_idx.append([t.block.idx, t.idx])
        res = Program()
        res.desc = core.prune(self.desc, targets_idx)
M
minqiyang 已提交
1812 1813 1814
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
1815
        res._sync_with_cpp()
1816 1817
        return res

X
Xin Pan 已提交
1818
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
1819
        """
F
fengjiayi 已提交
1820 1821 1822 1823 1824
        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.

1825
        3. change the :code:`is_test`
Y
yuyang18 已提交
1826 1827 1828
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

1829
        Args:
X
Xin Pan 已提交
1830 1831
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
1832

Y
yuyang18 已提交
1833 1834 1835 1836 1837 1838
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
1839
        res = Program()
1840
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
1841 1842 1843 1844

        # 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
1845
        if prune_read_op:
1846 1847 1848 1849 1850 1851 1852 1853 1854
            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
1855
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
1856 1857

        # change all `is_test` attributes to True
M
minqiyang 已提交
1858
        for i in six.moves.range(res.desc.num_blocks()):
1859
            block = res.desc.block(i)
M
minqiyang 已提交
1860
            for j in six.moves.range(block.op_size()):
1861 1862
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
1863
                    op._set_attr('is_test', True)
M
minqiyang 已提交
1864 1865 1866
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
1867
        res._sync_with_cpp()
1868 1869
        return res

1870 1871
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
1872 1873 1874 1875 1876 1877 1878
        """
        Deserialize a program desc from protobuf binary string.

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

        Args:
1879
            binary_str_type(str): The binary prootbuf string.
Y
yuyang18 已提交
1880 1881 1882 1883

        Returns:
            Program: A deserialized program desc.
        """
1884 1885
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
1886
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
1887
        p._sync_with_cpp()
1888
        return p
Y
Yu Yang 已提交
1889

D
dzhwinter 已提交
1890 1891
    @property
    def random_seed(self):
Y
yuyang18 已提交
1892 1893 1894 1895 1896 1897
        """
        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
1898 1899
        return self._seed

Q
qiaolongfei 已提交
1900 1901
    @property
    def num_blocks(self):
Y
yuyang18 已提交
1902 1903 1904
        """
        The number of blocks in this program.
        """
Q
qiaolongfei 已提交
1905 1906
        return self.desc.num_blocks()

D
dzhwinter 已提交
1907 1908 1909 1910 1911 1912
    @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
1913
    def __repr__(self):
1914
        return self.__str__()
1915

Y
Yu Yang 已提交
1916
    def global_block(self):
Y
yuyang18 已提交
1917 1918 1919
        """
        Get the first block of this program.
        """
Y
Yu Yang 已提交
1920 1921
        return self.blocks[0]

Q
Qiao Longfei 已提交
1922
    def block(self, index):
Y
yuyang18 已提交
1923 1924 1925 1926 1927 1928 1929 1930
        """
        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
1931 1932
        return self.blocks[index]

Y
Yu Yang 已提交
1933
    def current_block(self):
Y
yuyang18 已提交
1934 1935 1936 1937
        """
        Get the current block. The :code:`current` block is the block to append
        operators.
        """
Y
Yu Yang 已提交
1938 1939
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
1940
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
1941 1942 1943 1944 1945 1946 1947 1948 1949 1950
        """
        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
1951
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
1952 1953 1954
        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
1955 1956 1957 1958
        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
1959
    def _rollback(self):
Y
yuyang18 已提交
1960 1961 1962 1963 1964
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
1965 1966
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
1967
    def _sync_with_cpp(self):
Y
yuyang18 已提交
1968 1969 1970 1971 1972 1973 1974 1975 1976 1977
        """
        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
1978 1979 1980
        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
1981
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
1982

W
Wu Yi 已提交
1983
    def _copy_param_info_from(self, other):
1984
        """
1985
        Copy the information of parameters from other program.
D
dzhwinter 已提交
1986

Y
yuyang18 已提交
1987 1988 1989
        Notes: This is a very low level API. Users should not invoke it
        directly.

1990 1991 1992 1993 1994 1995 1996
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
1997
            raise TypeError("_copy_param_info_from should be invoked with "
1998 1999 2000
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
2001
            raise ValueError("_copy_param_info_from should be invoked with two "
2002
                             "program, with represent the same topology")
W
Wu Yi 已提交
2003
        self.global_block()._copy_param_info_from(other.global_block())
2004

2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023
    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
2024
    def _copy_data_info_from(self, other):
F
fengjiayi 已提交
2025 2026
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
2027

Y
yuyang18 已提交
2028 2029 2030
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
2031 2032 2033 2034 2035 2036 2037
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
2038
            raise TypeError("_copy_param_info_from should be invoked with "
F
fengjiayi 已提交
2039 2040 2041
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
2042
            raise ValueError("_copy_param_info_from should be invoked with two "
F
fengjiayi 已提交
2043
                             "program, with represent the same topology")
2044
        for var in list(other.global_block().vars.values()):
F
fengjiayi 已提交
2045 2046 2047
            if var.is_data:
                self.global_block().var(var.name).is_data = True

2048
    def list_vars(self):
Y
yuyang18 已提交
2049 2050 2051 2052 2053 2054
        """
        Get all variables from this Program. A iterable object is returned.

        Returns:
            iterable: The generator will yield every variable in this program.
        """
2055
        for each_block in self.blocks:
2056
            for each_var in list(each_block.vars.values()):
2057 2058
                yield each_var

Y
Yu Yang 已提交
2059

Y
Yu Yang 已提交
2060
class Parameter(Variable):
2061
    """
2062
    Parameter is derived from Variable. A parameter is a persistable
2063
    Variable, and will be updated by optimizers after each iteration.
2064
    The training of a neural network is essentially the updating of
2065 2066
    its parameters.

2067
    Relative to a general Variable, a Parameter has several its own
2068 2069
    member variables:

2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081
    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.
2082 2083
    """

Y
Yu Yang 已提交
2084 2085 2086 2087 2088 2089 2090 2091 2092 2093
    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")
2094 2095 2096

        Variable.__init__(
            self, block, persistable=True, shape=shape, dtype=dtype, **kwargs)
Y
Yu Yang 已提交
2097 2098 2099 2100
        self.trainable = kwargs.get('trainable', True)

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

2101 2102
        self.regularizer = kwargs.get('regularizer', None)

F
fengjiayi 已提交
2103
        self.gradient_clip_attr = kwargs.get('gradient_clip_attr', None)
Y
Yu Yang 已提交
2104

W
wanghaoshuang 已提交
2105
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
2106

F
fengjiayi 已提交
2107 2108 2109
    def __str__(self):
        return self.to_string(True)

F
update  
fengjiayi 已提交
2110 2111 2112
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
2113

F
update  
fengjiayi 已提交
2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127
        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
2128
                               "gradient_clip_attr", "do_model_average")
F
update  
fengjiayi 已提交
2129
            for attr_name in additional_attr:
2130 2131
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
2132 2133
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
2134 2135 2136 2137
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
2138

Y
Yu Yang 已提交
2139
# program is a global instance.
Y
Yu Yang 已提交
2140 2141
_main_program_ = Program()
_startup_program_ = Program()
2142

2143

2144
def default_startup_program():
Y
Yu Yang 已提交
2145
    """
Y
yuyang18 已提交
2146 2147 2148 2149 2150 2151 2152 2153 2154
    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.
2155

Y
Yu Yang 已提交
2156 2157 2158
    Returns:
        Program: startup program
    """
Y
Yu Yang 已提交
2159
    return _startup_program_
2160

2161

2162
def default_main_program():
Y
Yu Yang 已提交
2163
    """
Y
yuyang18 已提交
2164 2165 2166 2167 2168 2169 2170 2171 2172
    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.
2173

Y
Yu Yang 已提交
2174 2175 2176
    Returns:
        Program: main program
    """
Y
Yu Yang 已提交
2177
    return _main_program_
Y
Yu Yang 已提交
2178 2179 2180 2181 2182


def switch_main_program(program):
    """
    Switch the main program to a new program.
2183

Y
Yu Yang 已提交
2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197
    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):
    """
2198
    Switch the startup program to a new program
Y
Yu Yang 已提交
2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213
    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
2214 2215 2216
    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.
2217

Y
Yu Yang 已提交
2218
    Examples:
Y
yuyang18 已提交
2219 2220 2221 2222 2223 2224 2225 2226 2227 2228

        >>> 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.
2229

Y
Yu Yang 已提交
2230
    Examples:
Y
yuyang18 已提交
2231 2232 2233 2234 2235 2236

        >>> 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 = ...
2237

Y
Yu Yang 已提交
2238
    Args:
Y
yuyang18 已提交
2239
        main_program(Program): New main program inside `with` statement.
2240
        startup_program(Program): New startup program inside `with` statement.
Y
Yu Yang 已提交
2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253
            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
2254 2255


W
Wu Yi 已提交
2256
def _get_var(name, program=None):
X
xuwei06 已提交
2257
    """
Y
yuyang18 已提交
2258
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
2259

X
xuwei06 已提交
2260 2261 2262
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
2263
        If None, default_global_program() will be used.
X
xuwei06 已提交
2264 2265 2266 2267 2268 2269 2270

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
2271
    assert isinstance(program, Program)
X
xuwei06 已提交
2272 2273

    return program.global_block().var(name)
2274 2275 2276 2277 2278 2279 2280 2281 2282


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