framework.py 115.3 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
X
Xin Pan 已提交
18
from collections import defaultdict
W
WangZhen 已提交
19
from collections import Iterable
Q
qiaolongfei 已提交
20
import contextlib
S
rename  
sneaxiy 已提交
21
from .wrapped_decorator import signature_safe_contextmanager
P
peizhilin 已提交
22
import os
F
fengjiayi 已提交
23
import re
24
import traceback
25
import six
26

Y
Yu Yang 已提交
27
import numpy as np
28
import subprocess
S
sneaxiy 已提交
29
import multiprocessing
Q
qiaolongfei 已提交
30

M
minqiyang 已提交
31
from .. import compat as cpt
32
from .proto import framework_pb2
33
try:
P
peizhilin 已提交
34
    if os.name == 'nt':
P
peizhilin 已提交
35
        import sys
P
peizhilin 已提交
36 37 38 39 40
        third_lib_path = os.path.abspath(os.path.dirname(
            __file__)) + os.sep + '..' + os.sep + 'libs'
        os.environ['path'] += ';' + third_lib_path
        sys.path.append(third_lib_path)

41
    from . import core
42
except ImportError as e:
P
peizhilin 已提交
43
    if os.name == 'nt':
44
        executable_path = os.path.abspath(os.path.dirname(sys.executable))
P
peizhilin 已提交
45
        raise ImportError(
46 47 48 49 50
            """NOTE: You may need to run \"set PATH=%s;%%PATH%%\"
        if you encounters \"DLL load failed\" errors. If you have python
        installed in other directory, replace \"%s\" with your own
        directory. The original error is: \n %s""" %
            (executable_path, executable_path, cpt.get_exception_message(e)))
P
peizhilin 已提交
51 52 53 54 55 56
    else:
        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
        directory. The original error is: \n""" + cpt.get_exception_message(e))
57
except Exception as e:
58
    raise e
59
from . import unique_name
Y
Yu Yang 已提交
60

61
__all__ = [
62 63 64 65
    'Program',
    'default_startup_program',
    'default_main_program',
    'program_guard',
66
    'name_scope',
S
sneaxiy 已提交
67 68 69
    'cuda_places',
    'cpu_places',
    'cuda_pinned_places',
L
lujun 已提交
70
    'in_dygraph_mode',
71
]
Y
Yu Yang 已提交
72

Q
qiaolongfei 已提交
73 74 75 76
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
77 78
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()

L
lujun 已提交
79 80
_dygraph_tracer_ = None
_dygraph_current_expected_place_ = None
81 82


L
lujun 已提交
83 84 85 86
def in_dygraph_mode():
    '''
    Returns(bool): True if the program is running in dynamic graph mode
    '''
L
lujun 已提交
87
    return _dygraph_tracer_ is not None
88 89


L
lujun 已提交
90 91
def _dygraph_tracer():
    return _dygraph_tracer_
92

W
Wu Yi 已提交
93

M
minqiyang 已提交
94
def _current_expected_place():
L
lujun 已提交
95
    return _dygraph_current_expected_place_
M
minqiyang 已提交
96 97


S
sneaxiy 已提交
98 99 100 101 102
def _cpu_num():
    return int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))


def cuda_places(device_ids=None):
S
add doc  
sneaxiy 已提交
103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123
    '''
    Create a list of :code:`fluid.CUDAPlace` objects.

    If :code:`device_ids` is None, environment variable of
    :code:`FLAGS_selected_gpus` would be checked first. If
    :code:`FLAGS_selected_gpus=0,1,2`, the returned list would
    be [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
    If :code:`FLAGS_selected_gpus` is not set, all visible
    gpu places would be returned.  

    If :code:`device_ids` is not None, it should be the device
    ids of gpus. For example, if :code:`device_ids=[0,1,2]`, 
    the returned list would be 
    [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
    
    Args: 
        device_ids (None|list(int)|tuple(int)): gpu device id list.

    Returns:
        out (list(fluid.CUDAPlace)): gpu place list.
    '''
S
sneaxiy 已提交
124 125 126 127 128 129 130 131 132 133 134 135 136 137
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_ids is None:
        gpus_env = os.getenv("FLAGS_selected_gpus")
        if gpus_env:
            device_ids = [int(s) for s in gpus_env.split(",")]
        else:
            device_ids = six.moves.range(core.get_cuda_device_count())
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.CUDAPlace(dev_id) for dev_id in device_ids]


def cpu_places(device_count=None):
S
add doc  
sneaxiy 已提交
138 139 140 141 142 143 144 145 146 147 148 149 150 151
    '''
    Create a list of :code:`fluid.CPUPlace` objects.
    
    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
    If :code:`CPU_NUM` is not set, the device count would
    be determined by :code:`multiprocessing.cpu_count()`. 

    Args:
        device_count (None|int): device number.

    Returns:
        out (list(fluid.CPUPlace)): cpu place list.
    '''
S
sneaxiy 已提交
152 153 154 155 156 157
    if device_count is None:
        device_count = _cpu_num()
    return [core.CPUPlace()] * device_count


def cuda_pinned_places(device_count=None):
S
add doc  
sneaxiy 已提交
158 159 160 161 162 163 164 165 166 167 168 169 170 171
    '''
    Create a list of :code:`fluid.CUDAPinnedPlace` objects.

    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
    If :code:`CPU_NUM` is not set, the device count would
    be determined by :code:`multiprocessing.cpu_count()`. 

    Args:
        device_count (None|int): device number.

    Returns:
        out (list(fluid.CUDAPinnedPlace)): cuda pinned place list.
    '''
S
sneaxiy 已提交
172 173 174 175 176 177 178
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_count is None:
        device_count = _cpu_num()
    return [core.cuda_pinned_places()] * device_count


179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204
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()


S
rename  
sneaxiy 已提交
205
@signature_safe_contextmanager
206 207 208 209 210 211 212 213 214 215 216 217
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
218

219 220 221 222
          with name_scope("encoder"):
             ...
          with name_scope("decoder"):
             ...
T
Tink_Y 已提交
223 224
          with name_scope("attention"):
             ...
225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243
    """
    # 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
244 245 246
def generate_control_dev_var_name():
    import random
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
247 248 249 250


def grad_var_name(var_name):
    """
251 252
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
253 254 255
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
256

257
def convert_np_dtype_to_dtype_(np_dtype):
258 259
    """
    Convert the data type in numpy to the data type in Paddle
260

261
    Args:
262
        np_dtype(np.dtype): the data type in numpy.
263

264 265
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
266 267

    """
268 269
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
270
        return core.VarDesc.VarType.FP32
271
    elif dtype == np.float64:
272
        return core.VarDesc.VarType.FP64
273
    elif dtype == np.float16:
274
        return core.VarDesc.VarType.FP16
275
    elif dtype == np.int32:
276
        return core.VarDesc.VarType.INT32
277
    elif dtype == np.int16:
278
        return core.VarDesc.VarType.INT16
279
    elif dtype == np.int64:
280
        return core.VarDesc.VarType.INT64
281
    elif dtype == np.bool:
282
        return core.VarDesc.VarType.BOOL
283 284
    elif dtype == np.uint16:
        return core.VarDesc.VarType.INT16
285 286
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
287 288
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
289
    else:
M
minqiyang 已提交
290
        raise ValueError("Not supported numpy dtype %s" % dtype)
291 292 293


def dtype_is_floating(dtype):
294 295 296
    """
    Check the data type is floating or not.
    Args:
297
        dtype(np.dtype|core.VarDesc.VarType): data type.
298 299 300 301 302
            Could be numpy format or Paddle format

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

    """
303
    if not isinstance(dtype, core.VarDesc.VarType):
304 305
        dtype = convert_np_dtype_to_dtype_(dtype)

306 307 308 309
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
310 311


Y
Yang Yang(Tony) 已提交
312
def _debug_string_(proto, throw_on_error=True):
313 314 315 316 317 318 319 320 321 322 323
    """
    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
324
    error_fields = list()
Y
Yang Yang(Tony) 已提交
325
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
326 327
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
328 329 330
    return proto.__str__()


X
Xin Pan 已提交
331
class Variable(object):
332
    """
333 334 335
    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
336
    two variables in different blocks could have the same name.
337

338 339
    There are many kinds of variables. Each kind of them has its own attributes
    and usages. Please reference the framework.proto for details.
340

341
    Most of a Variable's member variables can be setted to be None. It mean
342
    it is not available or will be specified later.
343 344

    Args:
345
        block(Block): The block that the variable belongs to.
346 347
        type(core.VarDesc.VarType): Variable type. Please reference the
            framework.proto for details.
348 349
        name(str|None): The name of the variable. If setted None, it will be
            generated automatically. Default: None
350
        shape(tuple|list|None): The shape of the variable. -1 means the batch size.
351
            Some kinds of variable do not contain shape, just set it to None.
352 353 354
            Default: None
        dtype(np.dtype|core.VarDesc.VarType|str|None): The data type of variable.
            Default: None
355
        lod_level (int|None): The level of lod tensor. 0 means it is not a time
356
            series data.
357
            Default: None
358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379
        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')
380 381
    """

Y
Yu Yang 已提交
382 383
    def __init__(self,
                 block,
Y
Yu Yang 已提交
384
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
385 386 387 388
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
389
                 capacity=None,
Q
QI JUN 已提交
390
                 persistable=None,
F
fengjiayi 已提交
391
                 error_clip=None,
Y
Yu Yang 已提交
392
                 stop_gradient=False,
F
fengjiayi 已提交
393
                 is_data=False,
Y
Yu Yang 已提交
394
                 **kwargs):
Y
Yu Yang 已提交
395 396
        self.block = block
        if name is None:
Y
Yu Yang 已提交
397
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
398

Y
Yu Yang 已提交
399
        if dtype is not None:
400
            if not isinstance(dtype, core.VarDesc.VarType):
401
                dtype = convert_np_dtype_to_dtype_(dtype)
402

L
lujun 已提交
403
        if in_dygraph_mode():
M
minqiyang 已提交
404
            # record vars in tracer rather than blocks
M
minqiyang 已提交
405 406
            self._ivar = kwargs.get("ivar", None)
            if not self._ivar:
407 408 409
                self._ivar = core.VarBase(
                    name, dtype if dtype else core.VarDesc.VarType.FP32,
                    list(shape) if shape else [],
X
fix  
Xin Pan 已提交
410 411
                    _current_expected_place(), stop_gradient, True
                    if persistable else False)
M
minqiyang 已提交
412
            if persistable:
L
lujun 已提交
413
                _dygraph_tracer().trace_var(name, self)
M
minqiyang 已提交
414
            self.op = None
M
minqiyang 已提交
415
        else:
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487
            self.error_clip = error_clip

            is_new_var = False
            name = cpt.to_text(name)
            self.desc = self.block.desc.find_var(cpt.to_bytes(name))

            if self.desc is None:
                self.desc = self.block.desc.var(cpt.to_bytes(name))
                is_new_var = True

            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))

            if shape is not None:
                if is_new_var:
                    self.desc.set_shape(shape)
                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))
            if dtype is not None:
                if is_new_var:
                    self.desc.set_dtype(dtype)
                else:
                    old_dtype = self.dtype
                    if dtype != old_dtype:
                        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))

            if lod_level is not None:
                if is_new_var:
                    self.desc.set_lod_level(lod_level)
                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))
            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))

            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

M
minqiyang 已提交
488
            self.block.vars[name] = self
489
            self.op = None
490
            self._stop_gradient = stop_gradient
491
            self.is_data = is_data
Y
Yu Yang 已提交
492

493
    def numpy(self):
M
minqiyang 已提交
494
        new_ivar = self._ivar._copy_to(core.CPUPlace(), True)
P
Paddle CI 已提交
495
        return np.array(new_ivar.value().get_tensor())
496

497
    def backward(self):
X
Xin Pan 已提交
498
        self._ivar._run_backward()
499

500
    def gradient(self):
501 502
        new_ivar = self._ivar._grad_ivar()._copy_to(core.CPUPlace(), True)
        return np.array(new_ivar.value().get_tensor())
503

504
    def clear_gradient(self):
X
Xin Pan 已提交
505
        self._ivar._clear_gradient()
X
Xin Pan 已提交
506

507
    def __str__(self):
Y
Yang Yang(Tony) 已提交
508 509
        return self.to_string(True)

F
update  
fengjiayi 已提交
510
    def to_string(self, throw_on_error, with_details=False):
511 512 513 514
        """
        Get debug string.

        Args:
515 516
            throw_on_error(bool): True if raise an exception when self is
                not initialized.
F
update  
fengjiayi 已提交
517
            with_details(bool): more details about variables and parameters
518 519
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False;
520

521 522
        Returns:
            str: The debug string.
523
        """
L
lujun 已提交
524
        if in_dygraph_mode():
L
lujun 已提交
525
            # TODO(panyx0718): add more dygraph debug info.
526 527 528
            return 'name %s, dtype: %s shape: %s' % (self.name, self.dtype,
                                                     self.shape)

F
update  
fengjiayi 已提交
529 530
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
531
        protostr = self.desc.serialize_to_string()
532
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
533 534 535 536
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
537 538
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
539
        return res_str
540 541 542

    __repr__ = __str__

543
    def set_desc(self, input):
544 545 546 547 548 549 550 551 552
        """
        Set the variable description.

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

        Returns:
            None
        """
553 554
        self.desc = input

555
    @property
556
    def stop_gradient(self):
L
lujun 已提交
557
        if in_dygraph_mode():
M
minqiyang 已提交
558 559
            return self._ivar.stop_gradient
        else:
560
            return self._stop_gradient
561

562 563
    @stop_gradient.setter
    def stop_gradient(self, s):
L
lujun 已提交
564
        if in_dygraph_mode():
M
minqiyang 已提交
565
            self._ivar.stop_gradient = s
566
        else:
567
            self._stop_gradient = s
568

569 570
    @property
    def persistable(self):
L
lujun 已提交
571
        if in_dygraph_mode():
572 573 574
            return self._ivar.persistable
        else:
            return self.desc.persistable()
575

Y
Yu Yang 已提交
576 577
    @persistable.setter
    def persistable(self, p):
L
lujun 已提交
578
        if in_dygraph_mode():
579 580 581
            return self._ivar.persistable
        else:
            self.desc.set_persistable(p)
Y
Yu Yang 已提交
582

Y
Yu Yang 已提交
583 584
    @property
    def name(self):
L
lujun 已提交
585
        if in_dygraph_mode():
586 587 588
            return self._ivar.name
        else:
            return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
589

T
typhoonzero 已提交
590 591
    @name.setter
    def name(self, new_name):
L
lujun 已提交
592
        if in_dygraph_mode():
593 594 595
            self._ivar.name = new_name
        else:
            self.desc.set_name(new_name)
T
typhoonzero 已提交
596

Y
Yu Yang 已提交
597 598 599
    @property
    def shape(self):
        # convert to tuple, make it as same as numpy API.
L
lujun 已提交
600
        if in_dygraph_mode():
601 602 603
            return self._ivar.shape
        else:
            return tuple(self.desc.shape())
Y
Yu Yang 已提交
604 605

    @property
F
fengjiayi 已提交
606
    def dtype(self):
L
lujun 已提交
607
        if in_dygraph_mode():
608 609 610
            return self._ivar.dtype
        else:
            return self.desc.dtype()
Y
Yu Yang 已提交
611 612 613

    @property
    def lod_level(self):
L
lujun 已提交
614
        # TODO(minqiyang): Support lod_level in dygraph mode
615
        return self.desc.lod_level()
Y
Yu Yang 已提交
616

Y
Yu Yang 已提交
617 618
    @property
    def type(self):
L
lujun 已提交
619
        if in_dygraph_mode():
620 621 622
            return self._ivar.dtype
        else:
            return self.desc.type()
Y
Yu Yang 已提交
623

W
Wu Yi 已提交
624
    def _set_error_clip(self, error_clip):
625 626 627 628 629 630 631 632 633
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
634 635
        self.error_clip = error_clip

636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722
    def _slice_indices(self, slice, length):
        """
        Reference implementation for the slice.indices method.
        """
        # Compute step and length as integers.
        step = 1 if slice.step is None else slice.step

        # Raise ValueError for negative length or zero step.
        if length < 0:
            raise ValueError("length should not be negative")
        if step == 0:
            raise ValueError("slice step cannot be zero")

        # Find lower and upper bounds for start and stop.
        lower = -1 if step < 0 else 0
        upper = length - 1 if step < 0 else length

        # Compute start.
        if slice.start is None:
            start = upper if step < 0 else lower
        else:
            start = slice.start
            start = max(start + length, lower) if start < 0 else min(start,
                                                                     upper)

        # Compute stop.
        if slice.stop is None:
            stop = lower if step < 0 else upper
        else:
            stop = slice.stop
            stop = max(stop + length, lower) if stop < 0 else min(stop, upper)

        return start, stop, step

    def _detectEllipsis(self, item):
        has_ellipsis = False
        start = 0
        end = len(self.shape)
        for index, o in enumerate(item):
            if o is Ellipsis:
                if has_ellipsis:
                    raise ValueError("Index can have one ellipsis only.")
                has_ellipsis = True
                start = index
            else:
                if has_ellipsis:
                    end = index
        return has_ellipsis, start, end

    def _reconstructSliceinfo(self, item):
        has_ellipsis, start, end = self._detectEllipsis(item)
        if has_ellipsis:
            newitem = []
            for i in range(start):
                newitem.append(item[i])
            for i in range(start, end):
                newitem.append(slice(None, None, None))
            for i in range(end, len(item)):
                newitem.append(item[i])
            return newitem
        else:
            return None

    def _detectContinuesSlice(self, item):
        starts = []
        ends = []
        for index, o in enumerate(item):
            if isinstance(o, int):
                start = int(o)
                if (index > 0 and index >= self.shape[index]) \
                        or (index < 0 and (index + self.shape[index]) < 0):
                    raise IndexError("invalid index")
                start = max(start + self.shape[index], 0) if start < 0 else min(
                    start, self.shape[index])
                starts.append(start)
                ends.append(start + 1)
            elif isinstance(o, slice):
                start, stop, step = self._slice_indices(o, self.shape[index])
                if step == 1 or step == -1:
                    starts.append(start)
                    ends.append(stop)
                else:
                    return False, None
            else:
                raise IndexError("Valid index accept int or slice or ellipsis")
        return True, [starts, ends]

L
lujun 已提交
723
    def _cloneVar(self, copy=False):
724 725 726 727 728
        if not copy:
            return self.block.create_var(
                name=unique_name.generate(".".join(self.name)),
                dtype=self.dtype,
                persistable=self.persistable,
729
                stop_gradient=self.stop_gradient, )
730 731 732 733
        else:
            return self

    def _sliceVar(self, axes, starts, ends):
L
lujun 已提交
734
        new_var = self._cloneVar()
735 736 737 738 739 740 741 742 743 744
        self.block.append_op(
            type="slice",
            inputs={'Input': [self]},
            outputs={'Out': [new_var]},
            attrs={'axes': axes,
                   'starts': starts,
                   'ends': ends})
        return new_var

    def _concatVar(self, inputs, axis):
L
lujun 已提交
745
        new_var = self._cloneVar()
746 747 748 749 750 751 752 753 754 755
        self.block.append_op(
            type="concat",
            inputs={'X': inputs},
            outputs={'Out': [new_var]},
            attrs={'axis': axis, })
        return new_var

    def _sliceAndConcatVar(self, item, axis):
        if isinstance(item, slice):
            if self.shape[axis] < 0:
L
lujun 已提交
756
                return self._cloneVar(True)
757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774
            start, stop, step = self._slice_indices(item, self.shape[axis])
            if step == 1:
                return self._sliceVar([axis], [start], [stop])
            else:
                vars = []
                if step > 0:
                    while start < stop:
                        vars.append(
                            self._sliceVar([axis], [start], [start + 1]))
                        start += step
                else:
                    while start > stop:
                        vars.append(
                            self._sliceVar([axis], [start], [start + 1]))
                        start += step
                return self._concatVar(vars, axis)
        elif isinstance(item, int):
            if self.shape[axis] < 0:
L
lujun 已提交
775
                return self._cloneVar(True)
776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797
            index = int(item)
            if (index > 0 and index >= self.shape[axis])\
                    or (index < 0 and (index + self.shape[axis]) < 0):
                raise IndexError("invalid index")
            return self._sliceVar([axis], [index], [index + 1])
        else:
            raise IndexError("Valid index accept int or slice or tuple")

    def __getitem__(self, item):
        """
        Slice the variable.

        Args:
            item(int/slice/tuple) : the index.

        Returns:
            Sliced variable
        """
        new_var = None
        if isinstance(item, tuple):
            if len(item) > len(self.shape):
                raise IndexError("Too many indexes")
W
wopeizl 已提交
798 799 800 801 802 803
            fixedSize = True
            for i in range(len(self.shape)):
                if self.shape[i] == -1:
                    fixedSize = False
                    break

804
            newitem = self._reconstructSliceinfo(item) or item
W
wopeizl 已提交
805 806
            if fixedSize:
                check, info = self._detectContinuesSlice(newitem)
807
                if check:
W
wopeizl 已提交
808 809 810 811 812 813 814 815
                    starts = info[0]
                    ends = info[1]
                    axes = [i for i in range(len(starts))]
                    return self._sliceVar(axes, starts, ends)
                else:
                    new_var = self
                    for index, o in enumerate(newitem):
                        new_var = new_var._sliceAndConcatVar(o, index)
816 817 818 819 820 821 822 823
            else:
                new_var = self
                for index, o in enumerate(newitem):
                    new_var = new_var._sliceAndConcatVar(o, index)
        else:
            new_var = self._sliceAndConcatVar(item, 0)
        return new_var

Y
Yu Yang 已提交
824

F
fengjiayi 已提交
825 826 827
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
828

829 830
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
831 832 833 834
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
835
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
836 837 838 839 840
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
841 842 843 844
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
845 846 847 848 849 850 851 852 853
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
854
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
855 856 857 858 859 860
        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):
861 862 863 864 865 866 867 868
        """
        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
869 870
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
871 872
        return self.op_proto_map[type]

873 874 875 876
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
877
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
878 879
            core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
            core.op_proto_and_checker_maker.kOpCreationCallstackAttrName()
880 881
        }

F
fengjiayi 已提交
882

X
Xin Pan 已提交
883
class Operator(object):
884
    """
885 886 887 888 889 890 891
    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
892
        type(str): The type of operator. Default None.
893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912
        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
913
        Block.append_op or Block._prepend_op instead.
914 915 916 917 918 919 920 921 922 923

    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]})
924
    """
925
    OP_WITHOUT_KERNEL_SET = {
926 927 928
        'feed', 'fetch', 'recurrent', 'go', 'rnn_memory_helper_grad',
        'conditional_block', 'while', 'send', 'recv', 'listen_and_serv',
        'ncclInit', 'select', 'checkpoint_notify', 'gen_nccl_id'
929
    }
930

Y
Yu Yang 已提交
931 932
    def __init__(self,
                 block,
Y
Yu Yang 已提交
933
                 desc,
Y
Yu Yang 已提交
934 935 936
                 type=None,
                 inputs=None,
                 outputs=None,
M
minqiyang 已提交
937
                 attrs=None):
L
lujun 已提交
938
        if in_dygraph_mode():
939 940
            if type is None:
                raise ValueError(
941
                    "`type` to initialized an Operator can not be None.")
942
            self.iop = core.OpBase(type)
M
minqiyang 已提交
943
            self.previous_ops = []
M
minqiyang 已提交
944

M
minqiyang 已提交
945
            self.attrs = attrs if attrs else {}
946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039
        else:
            self.block = block
            self.desc = desc
            # 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()
            del attrs

            op_maker = core.op_proto_and_checker_maker

            if op_maker.kOpRoleAttrName() not in op_attrs:
                op_attrs[op_maker.kOpRoleAttrName(
                )] = self.block.program.op_role

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

            if role_var_name in op_attrs and len(op_attrs[role_var_name]) == 0:
                del op_attrs[role_var_name]

            if len(self.desc.type()) != 0:
                return
            if type is None:
                raise ValueError(
                    "`type` to initilized an Operator can not be None.")
            else:
                callstack_var_name = op_maker.kOpCreationCallstackAttrName()
                op_attrs[callstack_var_name] = list(
                    reversed(traceback.format_stack()))[1:]

            self.desc.set_type(type)
            proto = OpProtoHolder.instance().get_op_proto(type)

            namescope_var_name = op_maker.kOpNameScopeAttrName()
            op_attrs[namescope_var_name] = _full_name_scope()

            def find_name(var_list, name):
                for var_name in var_list:
                    if var_list[var_name] is not None and var_name == name:
                        return True
                return False

            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:
                        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:
                            raise ValueError(
                                "Input %s expects only one input, but %d are given."
                                % (in_proto.name, len(in_args)))
                        in_arg_names = []
                        for arg in in_args:
                            if isinstance(arg, six.string_types):
                                in_arg_names.append(arg)
                            elif isinstance(arg, six.binary_type):
                                in_arg_names.append(arg.decode())
                            else:
                                in_arg_names.append(cpt.to_text(arg.name))
                        self.desc.set_input(in_proto.name, in_arg_names)
                    else:
                        self.desc.set_input(in_proto.name, [])

            if outputs is not None:
                for m in proto.outputs:
                    if (m.name not in outputs) and m.dispensable:
                        continue
                    if not ((m.name in outputs) or m.dispensable):
                        raise ValueError(("Incorrect setting for output(s) of "
                                          "operator \"%s\", should set: [%s].")
                                         % (type, m.name))
                for out_proto in proto.outputs:
                    if out_proto.name not in outputs:
                        continue
                    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:
                        raise ValueError(
                            "Output %s expects only one output, but %d are given."
                            % (out_proto.name, len(out_args)))
                    out_arg_names = []
                    for arg in out_args:
                        out_arg_names.append(cpt.to_text(arg.name))
                        # TODO(minqiyang): could we remove variable's op in static mode?
L
lujun 已提交
1040
                        if not in_dygraph_mode():
1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059
                            arg.op = self
                    self.desc.set_output(out_proto.name, out_arg_names)

            if op_attrs is not None:
                if not isinstance(op_attrs, dict):
                    raise TypeError("'attrs' should be a dict.")
                for attr in proto.attrs:
                    attr_name = attr.name
                    if (attr_name not in op_attrs) or (
                            op_attrs[attr_name] is None):
                        continue
                    attr_val = op_attrs[attr_name]
                    self._update_desc_attr(attr_name, attr_val)

            self.desc.check_attrs()
            if self._has_kernel(type):
                self.desc.infer_var_type(self.block.desc)
                self.desc.infer_shape(self.block.desc)

W
Wu Yi 已提交
1060
    def _has_kernel(self, op_type):
1061 1062
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
1063
    def to_string(self, throw_on_error):
1064
        """
1065 1066
        Get debug string.

1067
        Args:
1068 1069
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
1070

1071 1072
        Returns:
            str: The debug string.
1073 1074

        """
1075
        protostr = self.desc.serialize_to_string()
1076
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
1077 1078 1079 1080
        return _debug_string_(proto, throw_on_error)

    def __str__(self):
        return self.to_string(True)
1081 1082 1083

    __repr__ = __str__

F
fengjiayi 已提交
1084 1085
    @property
    def type(self):
L
lujun 已提交
1086
        if in_dygraph_mode():
1087 1088 1089
            return self.iop.type
        else:
            return self.desc.type()
F
fengjiayi 已提交
1090 1091

    def input(self, name):
1092
        """
1093
        Get the input arguments according to the input parameter name.
1094

1095 1096
        Args:
            name(str): The input parameter name.
1097

1098 1099 1100
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
1101
        """
F
fengjiayi 已提交
1102 1103
        return self.desc.input(name)

W
Wu Yi 已提交
1104
    def _rename_input(self, old_name, new_name):
1105 1106 1107 1108 1109 1110 1111 1112 1113 1114
        """
        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
1115
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
1116

W
Wu Yi 已提交
1117
    def _rename_output(self, old_name, new_name):
1118 1119 1120 1121 1122 1123 1124 1125 1126 1127
        """
        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
1128
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
1129

F
fengjiayi 已提交
1130 1131 1132 1133
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
1134 1135 1136 1137 1138 1139 1140 1141
    @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
1142
    def output(self, name):
1143
        """
1144
        Get output arguments by the output parameter name.
1145

1146 1147
        Args:
            name(str): The output parameter name.
1148

1149 1150 1151
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
1152
        """
F
fengjiayi 已提交
1153 1154 1155 1156 1157 1158
        return self.desc.output(name)

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

1159 1160 1161 1162 1163 1164 1165 1166
    @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
1167
    def has_attr(self, name):
1168
        """
1169 1170
        Whether this Operator has the attribute with name or not.

1171
        Args:
1172
            name(str): the attribute name.
1173

1174 1175
        Returns:
            bool: True if has this attribute.
1176 1177

        """
F
fengjiayi 已提交
1178 1179 1180
        return self.desc.has_attr(name)

    def attr_type(self, name):
1181
        """
1182
        Get the type of attribute by attribute's name.
1183

1184 1185
        Args:
            name(str): the attribute name.
1186

1187 1188
        Returns:
            core.AttrType: the attribute type.
1189
        """
F
fengjiayi 已提交
1190 1191
        return self.desc.attr_type(name)

W
Wu Yi 已提交
1192
    def _set_attr(self, name, val):
1193 1194 1195 1196 1197 1198 1199 1200 1201 1202
        """
        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
1203 1204
        self._update_desc_attr(name, val)

1205 1206 1207
    def _remove_attr(self, name):
        self.desc.remove_attr(name)

G
gongweibao 已提交
1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218
    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
1219 1220
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
1221 1222
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
1223
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
1224 1225 1226 1227
        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
1228
            self.desc._set_attr(name, val)
Y
yuyang18 已提交
1229

F
fengjiayi 已提交
1230 1231 1232 1233 1234
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
1235
        """
1236 1237
        Get the attribute by name.

1238
        Args:
1239
            name(str): the attribute name.
1240

1241 1242
        Returns:
            bool|int|str|float|list: The attribute value. The return value
1243 1244
            can be any valid attribute type.
        """
F
fengjiayi 已提交
1245
        return self.desc.attr(name)
Y
Yu Yang 已提交
1246

W
Wu Yi 已提交
1247
    def _block_attr_id(self, name):
1248
        """
G
gongweibao 已提交
1249
        Get the block attribute's id by name.
1250

1251 1252
        Args:
            name(str): the attribute name.
1253

1254 1255
        Returns:
            int: the block index.
1256
        """
W
Wu Yi 已提交
1257
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
1258

W
Wu Yi 已提交
1259
    def _block_attr(self, name):
G
gongweibao 已提交
1260 1261 1262 1263 1264 1265 1266 1267 1268 1269
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
1270
        id = self._block_attr_id(name)
G
gongweibao 已提交
1271 1272 1273
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
1274
    def _blocks_attr(self, name):
G
gongweibao 已提交
1275 1276 1277 1278 1279 1280 1281 1282 1283 1284
        """
        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
1285
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
1286 1287 1288 1289 1290
            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
1291
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
1292 1293 1294 1295 1296 1297 1298 1299 1300 1301
        """
        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
1302
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
1303

J
JiayiFeng 已提交
1304
    def all_attrs(self):
F
fengjiayi 已提交
1305
        """
1306 1307 1308
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
1309
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
1310 1311 1312 1313
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
G
gongweibao 已提交
1314 1315
            attr_type = self.desc.attr_type(n)
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
1316
                attr_map[n] = self._block_attr(n)
G
gongweibao 已提交
1317 1318 1319
                continue

            if attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
1320
                attr_map[n] = self._blocks_attr(n)
G
gongweibao 已提交
1321 1322 1323 1324
                continue

            attr_map[n] = self.attr(n)

F
fengjiayi 已提交
1325 1326
        return attr_map

Y
Yu Yang 已提交
1327

Y
Yu Yang 已提交
1328
class Block(object):
1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342
    """
    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
1343
        use `Program._create_block()` to create a block.
1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357

    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
1358
    def __init__(self, program, idx):
Y
Yu Yang 已提交
1359
        self.desc = program.desc.block(idx)
1360
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
1361
        self.ops = list()  # operator list
Y
Yu Yang 已提交
1362
        self.program = program
1363
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
1364

1365
    def __str__(self):
Y
Yang Yang(Tony) 已提交
1366 1367
        return self.to_string(True)

F
fengjiayi 已提交
1368 1369
    def to_string(self, throw_on_error, with_details=False):
        """
1370 1371
        Get debug string.

F
fengjiayi 已提交
1372 1373
        Args:
            throw_on_error(bool): raise exception when self is not initialized
1374
                when throw_on_error is True.
F
update  
fengjiayi 已提交
1375
            with_details(bool): more details about variables and parameters
1376 1377
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
1378

1379 1380
        Returns:
            str: The debug string.
F
fengjiayi 已提交
1381 1382 1383 1384
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
1385
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
1386 1387
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
1388
            for var in list(self.vars.values()):
F
fengjiayi 已提交
1389
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
1390
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
1391
            for op in self.ops:
F
fengjiayi 已提交
1392 1393
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
1394 1395 1396
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
1397 1398
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
1399 1400
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1401 1402 1403

    __repr__ = __str__

Y
Yu Yang 已提交
1404 1405
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
1406
        return self.desc.parent
Y
Yu Yang 已提交
1407

Y
Yu Yang 已提交
1408 1409 1410 1411
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
1412
    def _set_forward_block_idx(self, idx):
1413 1414 1415 1416 1417 1418 1419 1420 1421
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
1422
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
1423

Y
Yu Yang 已提交
1424 1425
    @property
    def idx(self):
Y
Yu Yang 已提交
1426
        return self.desc.id
Y
Yu Yang 已提交
1427

Q
Qiao Longfei 已提交
1428
    def var(self, name):
1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441
        """
        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.
        """
1442
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
1443 1444 1445
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
1446 1447
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
1448
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
1449
        return v
Q
Qiao Longfei 已提交
1450

X
Xin Pan 已提交
1451
    def _find_var_recursive(self, name):
1452 1453 1454 1455 1456 1457 1458
        """
        Get a Variable by name from this block recursively.

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

        Returns:
X
Xin Pan 已提交
1459
            Variable: the Variable with the giving name. Or None if not found.
1460
        """
Y
Yu Yang 已提交
1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484
        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
1485
        return None
Y
Yu Yang 已提交
1486

X
Xin Pan 已提交
1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505
    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
1506

Q
Qiao Longfei 已提交
1507
    def all_parameters(self):
1508
        return list(self.iter_parameters())
1509

1510
    def iter_parameters(self):
M
minqiyang 已提交
1511
        return (item[1] for item in six.iteritems(self.vars)
1512
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
1513

Y
Yu Yang 已提交
1514
    def create_var(self, *args, **kwargs):
1515
        var = Variable(block=self, *args, **kwargs)
1516 1517
        if 'initializer' in kwargs:
            kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
1518
        return var
Y
Yu Yang 已提交
1519

Q
Qiao Longfei 已提交
1520 1521 1522
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
1523
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
1524 1525
        """
        Rename variable in vars and ops' inputs and outputs
1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537

        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
1538
        """
M
minqiyang 已提交
1539 1540
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
1541

T
typhoonzero 已提交
1542
        if not self.has_var(name):
1543
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
1544 1545
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
1546
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
1547 1548 1549 1550 1551 1552 1553
            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
1554
            var_type = "Variable"
T
wip  
typhoonzero 已提交
1555 1556 1557 1558
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
1559
        orig_var_type = v.type
M
minqiyang 已提交
1560
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
1561
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
1562
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
1563
        if var_type == "Parameter":
T
wip  
typhoonzero 已提交
1564 1565 1566 1567
            var = Parameter(
                self,
                d.shape(),
                d.dtype(),
T
typhoonzero 已提交
1568
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1569 1570 1571 1572 1573 1574 1575
                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
1576
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
1577 1578
            var = Variable(
                self,
T
typhoonzero 已提交
1579
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1580 1581 1582 1583
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
1584
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
1585 1586 1587
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
1588
        self._sync_with_cpp()
1589
        return var
T
typhoonzero 已提交
1590

W
Wu Yi 已提交
1591 1592
    def _remove_var(self, name):
        self._sync_with_cpp()
M
minqiyang 已提交
1593
        self.desc._remove_var(cpt.to_bytes(name))
1594 1595
        del self.vars[name]

Y
Yu Yang 已提交
1596 1597
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
Q
Qiao Longfei 已提交
1598
        param = Parameter(global_block, *args, **kwargs)
1599
        if 'initializer' in kwargs:
1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619

            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
1620
        return param
Y
Yu Yang 已提交
1621

Y
Yu Yang 已提交
1622
    def append_op(self, *args, **kwargs):
1623 1624 1625 1626 1627 1628
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
L
lujun 已提交
1629
        if in_dygraph_mode():
1630 1631 1632 1633
            op = Operator(
                block=self,
                desc=None,
                type=kwargs.get("type", None),
M
minqiyang 已提交
1634 1635 1636
                inputs=None,
                outputs=None,
                attrs=kwargs.get("attrs", {}))
1637

M
minqiyang 已提交
1638 1639 1640
            # record ops in tracer rather than blocks
            #
            # TODO(minqiyang): add op stop_gradient support in static mode too.
L
lujun 已提交
1641
            # currently, we only support stop_gradient in dygraph mode.
M
minqiyang 已提交
1642 1643 1644 1645
            _dygraph_tracer().trace_op(op,
                                       kwargs.get("inputs", {}),
                                       kwargs.get("outputs", {}),
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
1646
        else:
1647 1648 1649 1650 1651 1652 1653 1654 1655
            op_desc = self.desc.append_op()
            op = Operator(
                block=self,
                desc=op_desc,
                type=kwargs.get("type", None),
                inputs=kwargs.get("inputs", None),
                outputs=kwargs.get("outputs", None),
                attrs=kwargs.get("attrs", None))

M
minqiyang 已提交
1656
            self.ops.append(op)
M
minqiyang 已提交
1657

1658 1659
        return op

W
Wu Yi 已提交
1660
    def _insert_op(self, index, *args, **kwargs):
1661 1662 1663 1664 1665 1666 1667 1668 1669
        """
        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
1670 1671
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
1672 1673 1674 1675
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

W
Wu Yi 已提交
1676
    def _remove_op(self, index):
1677 1678 1679 1680 1681 1682 1683 1684 1685
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
W
Wu Yi 已提交
1686 1687
        self._sync_with_cpp()
        self.desc._remove_op(index, index + 1)
1688 1689
        del self.ops[index]

W
Wu Yi 已提交
1690
    def _slice_ops(self, start, end):
1691 1692 1693 1694 1695 1696 1697 1698 1699 1700
        """
        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
1701
        return self.ops[start:end]
Y
Yancey1989 已提交
1702

W
Wu Yi 已提交
1703
    def _prepend_op(self, *args, **kwargs):
L
lujun 已提交
1704
        if in_dygraph_mode():
1705 1706 1707 1708
            op = Operator(
                self,
                None,
                type=kwargs.get("type", None),
M
minqiyang 已提交
1709 1710 1711 1712 1713 1714 1715 1716
                inputs=None,
                outputs=None,
                attrs=kwargs.get("attrs", {}))

            _dygraph_tracer().trace_op(op,
                                       kwargs.get("inputs", {}),
                                       kwargs.get("outputs", {}),
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
1717
        else:
1718 1719 1720 1721 1722 1723 1724 1725
            op_desc = self.desc._prepend_op()
            op = Operator(
                self,
                op_desc,
                type=kwargs.get("type", None),
                inputs=kwargs.get("inputs", None),
                outputs=kwargs.get("outputs", None),
                attrs=kwargs.get("attrs", None))
M
minqiyang 已提交
1726
            self.ops.insert(0, op)
1727

Y
Yu Yang 已提交
1728 1729
        return op

W
Wu Yi 已提交
1730
    def _sync_with_cpp(self):
1731
        """
1732 1733
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
1734
        """
Q
Qiao Longfei 已提交
1735 1736 1737 1738 1739
        # 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())

1740
        # sync variables removed from c++ end
1741
        for var in list(self.vars.keys()):
M
minqiyang 已提交
1742
            if not self.desc.find_var(cpt.to_bytes(var)):
1743 1744
                self.vars.pop(var)

Q
Qiao Longfei 已提交
1745
        # sync operators from cpp
1746 1747 1748 1749
        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
1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765
        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
1766 1767 1768 1769 1770

        # 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
1771
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
1772 1773 1774 1775 1776 1777 1778

        # 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)

1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791
        # 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
1792 1793 1794 1795
        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
1796
    def _copy_param_info_from(self, other):
1797
        """
1798 1799
        Copy the information of parameters from the other block.

1800
        Args:
1801 1802 1803 1804 1805
            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.
1806 1807 1808 1809 1810

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
1811 1812
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
1813
        for p in other.iter_parameters():
1814 1815 1816
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
W
Wu Yi 已提交
1817
                raise ValueError("_copy_param_info_from should be invoked with "
1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829
                                 "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
1830
                gradient_clip_attr=p.gradient_clip_attr,
F
fengjiayi 已提交
1831
                error_clip=p.error_clip,
1832 1833 1834
                name=v.name)
            self.vars[new_p.name] = new_p

1835
    def _clone_variable(self, var, force_persistable=True):
1836 1837
        """
        Clone a variable into current block.
1838

1839 1840
        Args:
            var: the variable to be cloned.
1841 1842 1843
            force_persistable(bool): True means setting the result variable to being persistable.
                                     False means setting the persistable the same with that of input var.
                                     default: True.
1844 1845

        Returns:
1846
            Variable: the new  variable cloned from 'var' in current block.
1847 1848
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
1849 1850 1851 1852 1853
        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
1854 1855
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
1856
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
1857 1858 1859 1860 1861 1862
        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,
1863
                persistable=True if force_persistable else var.persistable,
F
fengjiayi 已提交
1864
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1865 1866 1867 1868 1869 1870 1871
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
1872
                persistable=True if force_persistable else var.persistable,
F
fengjiayi 已提交
1873
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1874
        return ret_var
1875

Y
Yu Yang 已提交
1876

1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971
class IrNode(object):
    """
    Python IrNode. Beneath it is a core.Node, which is used for Ir Pass.
    """

    def __init__(self, node):
        """
        Construct an IrNode using core.Node.

        Args:
            node(core.Node): C++ Node.
        """
        assert isinstance(node,
                          core.Node), 'node must be the instance of core.Node.'
        self.node = node

    def name(self):
        """
        Return the node name.

        Returns:
            str: node name.
        """
        return self.node.name()

    def node_type(self):
        """
        Return the node type.

        Returns:
            core.Node.Type: node type(core.Node.Type.Operation or core.Node.Type.Variable).
        """
        return self.node.node_type()

    def var(self):
        """
        Return the node variable description.

        Returns:
            core.VarDesc: node variable description.
        """
        return self.node.var()

    def op(self):
        """
        Return the node operator description.

        Returns:
            core.OpDesc: node operator description.
        """
        return self.node.op()

    def id(self):
        """
        Return the node id.

        Returns:
            int: node id.
        """
        return self.node.id()

    def is_op(self):
        """
        If the node is an operator, then return true.

        Returns:
            bool: indicate whether the node is an operator.
        """
        return self.node.is_op()

    def is_var(self):
        """
        If the node is a variable, then return true.

        Returns:
            bool: indicate whether the node is a variable.
        """
        return self.node.is_var()

    def is_ctrl_var(self):
        """
        If the node is a control dependence variable, then return true.

        Returns:
            bool: indicate whether the node is a control dependence variable.
        """
        return self.node.is_ctrl_var()

    def clear_inputs(self):
        """
        Clear the node inputs. After executing the `clear_inputs` function,
        the node inputs will be empty.
        """
        self.node.clear_inputs()

1972
    def remove_input_by_id(self, node_id):
1973 1974 1975 1976 1977 1978
        """
        Remove a node from inputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
1979
        self.node.remove_input(node_id)
1980

1981
    def remove_input(self, node):
1982 1983 1984 1985
        """
        Remove a node from inputs.

        Args:
1986
            node(IrNode): the node being removed.
1987
        """
1988
        self.node.remove_input(node.node)
1989

1990
    def append_input(self, node):
1991 1992 1993 1994
        """
        Append a node in inputs.

        Args:
1995
            node(IrNode): the node being appended.
1996
        """
1997
        self.node.append_input(node.node)
1998 1999 2000 2001 2002 2003 2004 2005

    def clear_outputs(self):
        """
        Clear the node outputs. After executing the `clear_outputs` function,
        the node outputs will be empty.
        """
        self.node.clear_outputs()

2006
    def remove_output_by_id(self, node_id):
2007 2008 2009 2010 2011 2012
        """
        Remove a node from outputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
2013
        self.node.remove_output(node_id)
2014

2015
    def remove_output(self, node):
2016 2017 2018 2019
        """
        Remove a node from outputs.

        Args:
2020
            node(IrNode): the node being removed.
2021
        """
2022
        self.node.remove_output(node.node)
2023

2024
    def append_output(self, node):
2025 2026 2027 2028
        """
        Append a node in outputs.

        Args:
2029
            node(IrNode): the node being appended.
2030
        """
2031
        self.node.append_output(node.node)
2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092

    @property
    def inputs(self):
        """
        Return the node inputs.

        Returns:
            list(IrNode): node inputs wrapped by IrNode.
        """
        return [IrNode(n) for n in self.node.inputs]

    @property
    def outputs(self):
        """
        Return the node outputs.

        Returns:
            list(IrNode): node outputs wrapped by IrNode.
        """
        return [IrNode(n) for n in self.node.outputs]


class IrVarNode(IrNode):
    """
    Python IrVarNode. Beneath it is a core.Node, it inherits from IrNode.
    """

    def __init__(self, node):
        """
        Construct an IrVarNode using core.Node.

        Args:
            node(core.Node): C++ Node.
        """
        assert isinstance(node, core.Node) and node.is_var(), \
            'node must be the instance of core.Node and it must be a variable node.'
        super(IrVarNode, self).__init__(node)
        self.node = node

    def set_shape(self, shape):
        """
        Set the node variable shape.

        Args:
            shape(list): shape to be set.
        """
        assert self.node.var() is not None, \
            "The node variable description cannot be None."
        self.node.var().set_shape(shape)

    def persistable(self):
        """
        If the variable node is a persistable variable, then return true.

        Returns:
            bool: indicate whether the variable is persistable.
        """
        assert self.node.var() is not None, \
            "The node variable description cannot be None."
        return self.node.var().persistable()

2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125
    def type(self):
        """
        Return the variable type.

        Returns:
            core.VarDesc.VarType: the variable type.
        """
        assert self.node.var() is not None, \
            "The node variable description cannot be None."
        return self.node.var().type()

    def dtype(self):
        """
        Return the variable data type.

        Returns:
            core.VarDesc.VarType: the variable data type.
        """
        assert self.node.var() is not None, \
            "The node variable description cannot be None."
        return self.node.var().dtype()

    def shape(self):
        """
        Return the variable shape.

        Returns:
            list: the variable shape.
        """
        assert self.node.var() is not None, \
            "The node variable description cannot be None."
        return self.node.var().shape()

2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175
    @property
    def inputs(self):
        """
        Return the node inputs.

        Returns:
            list(IrOpNode): node inputs wrapped by IrOpNode.
        """
        return [IrOpNode(n) for n in self.node.inputs]

    @property
    def outputs(self):
        """
        Return the node outputs.

        Returns:
            list(IrOpNode): node outputs wrapped by IrOpNode.
        """
        return [IrOpNode(n) for n in self.node.outputs]


class IrOpNode(IrNode):
    """
    Python IrOpNode. Beneath it is a core.Node, it inherits from IrNode.
    """

    def __init__(self, node):
        """
        Construct an IrOpNode using core.Node.

        Args:
            node(core.Node): C++ Node.
        """
        assert isinstance(node, core.Node) and node.is_op(), \
            'node must be the instance of core.Node and it must be a operator node.'
        super(IrOpNode, self).__init__(node)
        self.node = node

    def rename_input(self, old_input_name, new_input_name):
        """
        Rename the input of this node.

        Args:
            old_input_name(str): the old input name.
            new_input_name(str): the new input name.
        """
        assert self.node.op() is not None, \
            "The node operator description cannot be None."
        self.node.op()._rename_input(old_input_name, new_input_name)

2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214
    def input(self, name):
        """
        Get the argument name list by the parameter name for input.

        Args:
            name(str): the parameter name.

        Returns:
            list(str): the argument name list.
        """
        assert self.node.op() is not None, \
            "The node operator description cannot be None."
        return self.node.op().input(name)

    def output(self, name):
        """
        Get the argument name list by the parameter name for output.

        Args:
            name(str): the parameter name.

        Returns:
            list(str): the argument name list.
        """
        assert self.node.op() is not None, \
            "The node operator description cannot be None."
        return self.node.op().output(name)

    def set_type(self, new_type):
        """
        Change the operator type into new type.

        Args:
            new_type(str): new operator type to be set.
        """
        assert self.node.op() is not None, \
            "The node operator description cannot be None."
        return self.node.op().set_type(new_type)

2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242
    def set_attr(self, name, val):
        """
        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.
        """
        self._update_desc_attr(name, val)

    def _update_desc_attr(self, name, val):
        """
        Update the value of the op desc's attribute by attribute's name.
        """
        assert self.node.op() is not None, \
            "The node operator description cannot be None."
        desc = self.node.op()
        if isinstance(val, Block):
            desc.set_block_attr(name, val.desc)
        elif isinstance(val, list) and val and \
            all(isinstance(v, Block) for v in val):
            desc.set_blocks_attr(name, [v.desc for v in val])
        elif isinstance(val, core.BlockDesc) or \
            isinstance(val, core.ProgramDesc):
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)

2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264
    def input_arg_names(self):
        """
        Return input arguments' names of this op node.

        Returns:
            list(str): input arguments' names of this op node.
        """
        assert self.node.op() is not None, \
            "The node operator description cannot be None."
        return self.node.op().input_arg_names()

    def output_arg_names(self):
        """
        Return output arguments' names of this op node.

        Returns:
            list(str): output arguments' names of this op node.
        """
        assert self.node.op() is not None, \
            "The node operator description cannot be None."
        return self.node.op().output_arg_names()

2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285
    @property
    def inputs(self):
        """
        Return the node inputs.

        Returns:
            list(IrVarNode): node inputs wrapped by IrVarNode.
        """
        return [IrVarNode(n) for n in self.node.inputs]

    @property
    def outputs(self):
        """
        Return the node outputs.

        Returns:
            list(IrVarNode): node outputs wrapped by IrVarNode.
        """
        return [IrVarNode(n) for n in self.node.outputs]


2286 2287
class IrGraph(object):
    """
2288
    Python IrGraph. Beneath it is a core.Graph, which is used for
2289
    creating a c++ Ir Pass Graph. An IrGraph is just a graph view of
2290 2291
    a Program. In an IrGraph, both Variables and Operators are graph
    nodes.
2292 2293 2294 2295
    """

    def __init__(self, graph, for_test=False):
        """
2296 2297
        Construct an IrGraph using core.Graph.

2298 2299 2300 2301 2302 2303 2304 2305 2306
        Args:
            graph(core.Graph): C++ Graph.
            for_test(bool): True for the test graph and false for the train graph.
        """
        assert isinstance(
            graph, core.Graph), 'graph must be the instance of core.Graph.'
        self.graph = graph
        self._for_test = for_test

2307 2308 2309 2310
    def clone(self):
        """
        Create a new and duplicated IrGraph.

2311 2312 2313
        Warns:
            The method only clones the graph structure, not its attributes.

2314 2315 2316
        Returns:
            IrGraph: A new and duplicated graph.
        """
2317
        g = self.graph.clone()
2318 2319
        return IrGraph(g, self._for_test)

2320
    def is_test(self):
2321 2322 2323
        """
        If the graph is used for testing, the function returns true. Otherwise, returns false.
        """
2324 2325
        return self._for_test

W
WangZhen 已提交
2326
    def all_nodes(self):
2327 2328 2329
        """
        Return all nodes included in the graph as a set.
        """
2330
        return {IrNode(node) for node in self.graph.nodes()}
2331

2332
    def all_var_nodes(self):
2333 2334 2335
        """
        Return all variable nodes included in the graph as a set.
        """
2336
        return {IrVarNode(node) for node in self.graph.nodes() if node.is_var()}
2337

2338
    def all_persistable_nodes(self):
2339 2340 2341
        """
        Return all persistable variable nodes included in the graph as a set.
        """
W
WangZhen 已提交
2342 2343 2344 2345 2346
        persistable_nodes = set()
        for node in self.graph.nodes():
            if node.is_var() and node.var() is not None and node.var(
            ).persistable():
                persistable_nodes.add(node)
2347
        return {IrVarNode(p) for p in persistable_nodes}
W
WangZhen 已提交
2348

2349
    def all_op_nodes(self):
2350 2351 2352
        """
        Return all operator nodes included in the graph as a set.
        """
2353
        return {IrOpNode(node) for node in self.graph.nodes() if node.is_op()}
2354

2355
    def create_persistable_node(self, name, var_type, shape, var_dtype):
2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366
        """
        Create a persistable variable node in the graph. In IrGraph,
        it can not distinguish between persistable variables and parameters.

        Args:
            name(str): the name of the persistable variable node.
            vart_type(core.VarDesc.VarType): the type of the persistable variable node.
            shape(list): the shape of the persistable variable node.
            var_dtype(core.VarDesc.VarType): the data type of the persistable variable node.

        Returns:
2367
            IrVarNode: the created persistable variable node.
2368
        """
2369 2370 2371 2372 2373
        var_desc = core.VarDesc(name)
        var_desc.set_type(var_type)
        var_desc.set_shape(shape)
        var_desc.set_dtype(var_dtype)
        var_desc.set_persistable(True)
2374
        return IrVarNode(self.graph.create_var_node(var_desc))
2375 2376

    def create_var_node(self, name, var_type, shape, var_dtype):
2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387
        """
        Create a variable node in the graph. The created variable node is
        not persistable.

        Args:
            name(str): the name of the variable node.
            vart_type(core.VarDesc.VarType): the type of the variable node.
            shape(list): the shape of the variable node.
            var_dtype(core.VarDesc.VarType): the data type of the variable node.

        Returns:
2388
            IrVarNode: the created variable node.
2389 2390
        """

2391 2392 2393 2394
        var_desc = core.VarDesc(name)
        var_desc.set_type(var_type)
        var_desc.set_shape(shape)
        var_desc.set_dtype(var_dtype)
2395
        return IrVarNode(self.graph.create_var_node(var_desc))
2396 2397

    def create_var_node_from_desc(self, var_desc):
2398 2399 2400 2401 2402 2403 2404 2405
        """
        Create a variable node by using an existing VarDesc in the graph.
        Depend on the giving VarDesc, the created variable node may be persistable.

        Args:
            var_desc(core.VarDesc): the giving variable description.

        Returns:
2406
            IrVarNode: the created variable node.
2407
        """
2408
        return IrVarNode(self.graph.create_var_node(var_desc))
2409 2410

    def create_op_node(self, op_type, attrs, inputs, outputs):
2411 2412 2413 2414 2415 2416 2417 2418 2419 2420
        """
        Create a operator node in the graph.

        Args:
            op_type(str): the type of the operator node.
            attrs(dict): the attributes of the operator node.
            inputs(dict): the inputs of the operator node.
            outputs(dict): the outpus of the operator node.

        Returns:
2421
            IrOpNode: the created operator node.
2422
        """
2423 2424
        op_desc = core.OpDesc()
        op_desc.set_type(op_type)
2425
        for attr, value in six.iteritems(attrs):
2426
            self._update_desc_attr(op_desc, attr, value)
2427
        for input_name, var_nodes in six.iteritems(inputs):
2428 2429 2430 2431
            if not isinstance(var_nodes, list):
                var_nodes = [var_nodes]
            op_desc.set_input(input_name,
                              [var_node.name() for var_node in var_nodes])
2432
        for output_name, var_nodes in six.iteritems(outputs):
2433 2434 2435 2436
            if not isinstance(var_nodes, list):
                var_nodes = [var_nodes]
            op_desc.set_output(output_name,
                               [var_node.name() for var_node in var_nodes])
2437
        return IrOpNode(self.graph.create_op_node(op_desc))
2438 2439

    def create_op_node_from_desc(self, op_desc):
2440 2441 2442 2443 2444 2445 2446
        """
        Create a operator node by using an existing OpDesc in the graph.

        Args:
            op_desc(core.VarDesc): the giving operator description.

        Returns:
2447
            IrOpNode: the created operator node.
2448
        """
2449
        return IrOpNode(self.graph.create_op_node(op_desc))
2450 2451

    def update_input_link(self, old_input_node, new_input_node, op_node):
2452 2453 2454 2455
        """
        Update the input's link of a operator node.

        Args:
2456 2457 2458
            old_input_node(IrNode): the old input node of the giving op_node.
            new_input_node(IrNode): the new input node of the giving op_node.
            op_node(IrOpNode): the operator node that is needed to update input's link.
2459
        """
2460 2461
        assert old_input_node.node in self.graph.nodes() and new_input_node.node in \
        self.graph.nodes() and op_node.node in self.graph.nodes(), \
W
WangZhen 已提交
2462
        'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.'
2463 2464 2465 2466
        old_input_node.remove_output(op_node)
        op_node.remove_input(old_input_node)
        new_input_node.append_output(op_node)
        op_node.append_input(new_input_node)
2467
        op_node.rename_input(old_input_node.name(), new_input_node.name())
2468 2469

    def link_to(self, node_in, node_out):
2470 2471 2472 2473
        """
        Connect two nodes.

        Args:
2474 2475
            node_in(IrNode): the input node.
            node_out(IrNode): the output node.
2476
        """
2477
        assert node_in.node in self.graph.nodes() and node_out.node in self.graph.nodes(), \
W
WangZhen 已提交
2478
            'The two arguments(node_in&node_out) must be in the graph nodes.'
2479 2480
        node_in.append_output(node_out)
        node_out.append_input(node_in)
2481 2482

    def safe_remove_nodes(self, remove_nodes):
2483 2484 2485 2486 2487 2488 2489
        """
        Remove nodes safely since links connected to these removed nodes are
        also removed.

        Args:
            remove_nodes(set): the nodes prepared to be removed.
        """
2490
        if not isinstance(remove_nodes, set):
W
WangZhen 已提交
2491 2492 2493 2494
            if isinstance(remove_nodes, Iterable):
                remove_nodes = set(remove_nodes)
            else:
                remove_nodes = {remove_nodes}
2495 2496
        original_nodes = {n.node for n in remove_nodes}
        core.graph_safe_remove_nodes(self.graph, original_nodes)
2497

Z
Zhen Wang 已提交
2498 2499 2500 2501 2502 2503 2504 2505
    def resolve_hazard(self):
        ordered_nodes = core.topology_sort(self.graph)
        var_nodes = dict()
        for node in ordered_nodes:
            if node.is_op() and node.op() is not None:
                for each_var_name in node.op().input_arg_names():
                    if each_var_name not in var_nodes:
                        var_nodes[each_var_name] = [
2506
                            self._find_node_by_name(node.inputs, each_var_name)
Z
Zhen Wang 已提交
2507 2508 2509 2510
                        ]
                for each_var_name in node.op().output_arg_names():
                    if each_var_name not in var_nodes:
                        var_nodes[each_var_name] = [
2511
                            self._find_node_by_name(node.outputs, each_var_name)
Z
Zhen Wang 已提交
2512 2513 2514
                        ]
                    else:
                        var_nodes[each_var_name].append(
2515 2516
                            self._find_node_by_name(node.outputs,
                                                    each_var_name))
Z
Zhen Wang 已提交
2517 2518
        self.graph.resolve_hazard(var_nodes)

W
WangZhen 已提交
2519
    def has_circle(self):
2520 2521 2522 2523 2524 2525
        """
        Check if the graph has a circle.

        Returns:
            bool: True if the graph has a circle else False.
        """
W
WangZhen 已提交
2526 2527 2528
        return core.has_circle(self.graph)

    def graph_num(self):
2529 2530 2531 2532 2533 2534
        """
        Count the number of unconnected graphs in this graph.

        Returns:
            int: the number of unconnected graphs.
        """
W
WangZhen 已提交
2535 2536 2537
        return core.graph_num(self.graph)

    def topology_sort(self):
2538 2539 2540 2541 2542 2543
        """
        Perform the topology sort operation on the graph.

        Notes: the `graph` cannot contain a circle.

        Returns:
Z
Zhen Wang 已提交
2544
            list(IrNode): nodes in topology order.
2545
        """
2546
        ordered_nodes = core.topology_sort(self.graph)
Z
Zhen Wang 已提交
2547
        return [IrNode(n) for n in ordered_nodes]
W
WangZhen 已提交
2548 2549

    def build_adjacency_list(self):
2550 2551 2552 2553
        """
        Build an adjacency list of operations for the `graph`.

        Returns:
2554
            dict{IrNode: set(IrNode)}: the adjacency list.
2555
        """
2556 2557 2558 2559 2560
        adj_list = core.build_adjacency_list(self.graph)
        wrapped_adj_list = dict()
        for k, v in six.iteritems(adj_list):
            wrapped_adj_list[IrNode(k)] = {IrNode(n) for n in v}
        return wrapped_adj_list
W
WangZhen 已提交
2561

2562 2563 2564 2565 2566 2567 2568 2569
    def draw(self, save_path, name, marked_nodes=None, remove_ctr_var=True):
        """
        Draw the graph. If `dot` command is installed, the drawn graph
        will be saved as pdf file type, otherwise dot file type is used.

        Args:
            save_path(str): the save path of drawn graph.
            name(str): the name of drawn graph.
2570
            marked_nodes(set(IrNode)): nodes that are needed to be marked.
2571 2572 2573 2574 2575
            Default value is None.
            remove_ctr_var(bool): If it is set True, all control variable nodes
            in the graph will be removed. Default value is True.
        """

2576 2577 2578 2579 2580 2581 2582 2583 2584
        def _convert_to_pdf(dot_file_path):
            pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf'
            exited_code = subprocess.call('dot -Tpdf ' + dot_file_path \
                            + ' -o ' + pdf_save_path, shell=True)
            if exited_code != 0:
                print('The dot command is needed for creating pdf files.')
                print('The {} is saved as the dot filetype.'.format(
                    dot_file_path))

2585
        remove_ctr_vars = set()
2586
        if remove_ctr_var:
2587
            for node in self.all_var_nodes():
2588 2589 2590
                if node.is_ctrl_var():
                    remove_ctr_vars.add(node)
            self.safe_remove_nodes(remove_ctr_vars)
2591 2592
        print('Total ops num = {}.'.format(len(self.all_op_nodes())))

2593 2594
        if marked_nodes is not None:
            if not isinstance(marked_nodes, set):
2595 2596 2597 2598 2599 2600
                if isinstance(marked_nodes, Iterable):
                    marked_nodes = set(marked_nodes)
                else:
                    marked_nodes = {marked_nodes}
            marked_nodes = {n.node for n in marked_nodes}
            remove_ctr_vars = {n.node for n in remove_ctr_vars}
2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611
            marked_nodes = marked_nodes - remove_ctr_vars
            if self.graph.has('__graphviz__marked_node__'):
                self.graph.erase('__graphviz__marked_node__')
            self.graph.set('__graphviz__marked_node__', marked_nodes)
        viz_dot_path = os.path.join(save_path, name) + '.dot'
        viz_pass = core.get_pass('graph_viz_pass')
        viz_pass.set('graph_viz_path', viz_dot_path)
        viz_pass.apply(self.graph)
        _convert_to_pdf(viz_dot_path)

    def to_program(self):
2612 2613 2614
        """
        Convert the graph into a Program.

Z
Zhen Wang 已提交
2615
        WARN: When the graph includes backward operator nodes, the
2616 2617 2618 2619 2620 2621
        conversion process may be failed. Usually, this function is
        only used to convert a test graph.

        Returns:
            Program: a program converted from the graph.
        """
2622
        convert_pass = core.get_pass('graph_to_program_pass')
2623 2624
        desc = core.ProgramDesc()
        convert_pass.set_not_owned('program', desc)
2625 2626 2627 2628
        convert_pass.apply(self.graph)
        program = Program._construct_from_desc(desc)
        return program

2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639
    def _find_node_by_name(self, nodes, node_name):
        """
        Find a node in the giving nodes set by the name.
        """
        target_node = None
        for n in nodes:
            if n.name() == node_name:
                target_node = n
        assert target_node is not None, "Cannot find the target node in the giving set."
        return target_node

2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655
    def _update_desc_attr(self, desc, name, val):
        """
        Update the value of desc's attribute by attribute's name.
        """
        if isinstance(val, Block):
            desc.set_block_attr(name, val.desc)
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
            desc.set_blocks_attr(name, [v.desc for v in val])
        elif isinstance(val, core.BlockDesc) or \
                isinstance(val, core.ProgramDesc):
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)


Y
Yu Yang 已提交
2656
class Program(object):
D
dzhwinter 已提交
2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667
    """
    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
2668
    default_main_program run in every mini batch and adjust the weights.
D
dzhwinter 已提交
2669 2670

    Returns:
Y
yuyang18 已提交
2671
        A empty program.
D
dzhwinter 已提交
2672 2673

    Examples:
Y
yuyang18 已提交
2674 2675 2676 2677 2678 2679
        >>> 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
2680 2681 2682

    """

2683 2684
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
2685 2686
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
D
dzhwinter 已提交
2687
        self._seed = 0
Y
yuyang18 已提交
2688
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
Y
yuyang18 已提交
2689
        self._op_role_var = []
T
tangwei12 已提交
2690

2691 2692
        # for distribute training
        # _is_distributed = True if under distributed training
T
tangwei12 已提交
2693
        self._is_distributed = False
2694
        # _is_chief = True if the trainer is the first one, usually No.0
T
tangwei12 已提交
2695
        self._is_chief = False
2696 2697 2698
        # _parameters_on_pservers records all the parameters distributed on parameter servers.
        self._parameters_on_pservers = None
        # _endpoints is a list about parameter servers ip:port, such as ["ip:port","ip:port"]
T
tangwei12 已提交
2699
        self._endpoints = []
2700 2701 2702
        # if current role is parameter server, the _ps_endpoint is its "ip:port"
        self._ps_endpoint = None
        # trainers_endpoints, it is used for distribution.
2703
        self._trainers_endpoints = []
2704
        # the distributed lookup table names
T
tangwei12 已提交
2705
        self._distributed_lookup_table = None
2706 2707 2708 2709

        # use Deep gradient comrepssion or not
        self._enable_dgc = False

D
dzhwinter 已提交
2710
        # @deprecated(the python memory optimize transpiler is deprecated)
D
dzhwinter 已提交
2711
        # whether the program is optimized by memory_optimize_transpiler
D
dzhwinter 已提交
2712
        self.__is_mem_optimized = False
D
dzhwinter 已提交
2713

2714 2715 2716
        # if this program has been optimized by distributed optimizer
        # fleet_opt will be given a value
        self._fleet_opt = None
D
dongdaxiang 已提交
2717
        self._program_config = None
2718

D
dzhwinter 已提交
2719
    @property
D
dzhwinter 已提交
2720
    def _is_mem_optimized(self):
D
dzhwinter 已提交
2721 2722
        # if the program is optimized, operator input/outputs
        # maybe same, which conflict with save_inference_model.
D
dzhwinter 已提交
2723
        return self.__is_mem_optimized
D
dzhwinter 已提交
2724

D
dzhwinter 已提交
2725 2726 2727
    @_is_mem_optimized.setter
    def _is_mem_optimized(self, target):
        self.__is_mem_optimized = target
Y
yuyang18 已提交
2728 2729 2730

    @property
    def op_role(self):
Y
yuyang18 已提交
2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743
        """
        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
2744 2745 2746
        return self._current_role

    @op_role.setter
D
dzhwinter 已提交
2747
    def op_role(self, role):
Y
yuyang18 已提交
2748 2749 2750 2751
        self._current_role = role

    @property
    def op_role_var(self):
Y
yuyang18 已提交
2752 2753 2754 2755 2756 2757 2758
        """
        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
2759 2760 2761 2762
        return self._op_role_var

    @op_role_var.setter
    def set_op_role_var(self, var_name):
Y
yuyang18 已提交
2763
        self._op_role_var = [var_name]
Y
yuyang18 已提交
2764

2765 2766 2767 2768 2769 2770 2771 2772 2773
    @contextlib.contextmanager
    def _backward_role_guard(self):
        tmp_role = self._current_role

        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Backward
        yield
        self._current_role = tmp_role

S
rename  
sneaxiy 已提交
2774
    @signature_safe_contextmanager
W
Wu Yi 已提交
2775
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
2776 2777 2778 2779 2780 2781 2782
        """
        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:
2783
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
2784 2785 2786 2787

        Examples:

            >>> p, g = backward(...)
W
Wu Yi 已提交
2788
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
2789 2790
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
2791 2792 2793
        tmp_role = self._current_role
        tmp_var = self._op_role_var

Y
yuyang18 已提交
2794 2795
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
2796 2797 2798 2799
        self._op_role_var = [
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
Y
yuyang18 已提交
2800
        yield
X
Xin Pan 已提交
2801 2802
        self._op_role_var = tmp_var
        self._current_role = tmp_role
Y
Yu Yang 已提交
2803

S
rename  
sneaxiy 已提交
2804
    @signature_safe_contextmanager
X
Xin Pan 已提交
2805
    def _lr_schedule_guard(self, is_with_opt=False):
2806 2807 2808 2809 2810 2811 2812
        """
        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
2813 2814 2815 2816
        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.
2817 2818 2819 2820 2821 2822 2823

        Examples:

            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
2824 2825 2826 2827

        tmp_role = self._current_role
        tmp_var = self._op_role_var

2828 2829
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
2830 2831
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
2832 2833 2834
        # TODO(typhoonzero): how to set target learning rate var
        self._op_role_var = []
        yield
2835 2836
        self._op_role_var = tmp_var
        self._current_role = tmp_role
2837

2838
    def __str__(self):
Y
yuyang18 已提交
2839 2840 2841 2842 2843 2844 2845 2846 2847
        """
        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
2848 2849
        return self.to_string(True)

F
fengjiayi 已提交
2850 2851 2852
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
2853

F
fengjiayi 已提交
2854
        Args:
Y
yuyang18 已提交
2855 2856
            throw_on_error(bool): raise Value error when any of required fields
                is not set.
F
fengjiayi 已提交
2857

Y
yuyang18 已提交
2858 2859 2860 2861
            with_details(bool): True if more details about variables and
                parameters, e.g., :code:`trainable`, :code:`optimize_attr`, need
                to print.

H
haowang101779990 已提交
2862 2863
        Returns:
            str : The debug string.
Y
yuyang18 已提交
2864 2865 2866 2867

        Raises:
            ValueError: If any of required fields is not set and throw_on_error is
                True.
F
fengjiayi 已提交
2868 2869 2870 2871 2872 2873 2874 2875 2876 2877

        """
        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()
2878 2879
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
2880 2881
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
2882

W
Wu Yi 已提交
2883
    def _get_desc(self):
Y
yuyang18 已提交
2884 2885 2886 2887 2888 2889 2890
        """
        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.
        """
2891 2892
        return self.desc

X
version  
Xin Pan 已提交
2893 2894 2895
    def _version(self):
        return self.desc._version()

2896
    def clone(self, for_test=False):
Y
yuyang18 已提交
2897 2898 2899
        """
        Create a new, duplicated program.

2900

Y
yuyang18 已提交
2901 2902 2903 2904
        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`.
2905

Y
yuyang18 已提交
2906 2907 2908 2909
        * 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
2910 2911 2912 2913 2914
        :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()
2915 2916

        Args:
Y
yuyang18 已提交
2917 2918
            for_test(bool): True if change the :code:`is_test` attribute of
                operators to :code:`True`.
2919

D
dzhwinter 已提交
2920
        Returns:
Y
yuyang18 已提交
2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973
            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.
2974 2975
        """
        if for_test:
X
Xin Pan 已提交
2976
            p = self._inference_optimize(prune_read_op=False)
2977
        else:
2978
            p = Program()
G
gongweibao 已提交
2979 2980
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
2981
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
2982 2983 2984
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
2985 2986 2987 2988

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

W
Wu Yi 已提交
2989
            p._sync_with_cpp()
2990

W
Wu Yi 已提交
2991
        p._copy_param_info_from(self)
W
Wu Yi 已提交
2992
        p._copy_data_info_from(self)
2993
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
2994
        return p
2995

W
Wu Yi 已提交
2996
    def _prune(self, targets):
Y
yuyang18 已提交
2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011
        """
        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.

        """
3012 3013 3014 3015 3016 3017
        if not isinstance(targets, list):
            targets = [targets]
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
3018 3019
                    # After transpiler processing, the op that output this
                    # variable maybe has been changed, so t.op is not reliable
3020
                    # and we need to find the current op that generate this
3021 3022 3023 3024 3025 3026 3027 3028
                    # 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

3029
                    t = t.op
3030 3031 3032 3033
                    if t is None:
                        raise ValueError(
                            "The target variable must have an "
                            "associated operator that generates it.")
3034
                else:
3035 3036
                    raise ValueError("All targets of prune() can only be "
                                     "Variable or Operator.")
3037 3038 3039 3040

            targets_idx.append([t.block.idx, t.idx])
        res = Program()
        res.desc = core.prune(self.desc, targets_idx)
M
minqiyang 已提交
3041 3042 3043
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
3044
        res._sync_with_cpp()
3045 3046
        return res

X
Xin Pan 已提交
3047
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
3048
        """
F
fengjiayi 已提交
3049 3050 3051 3052 3053
        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.

3054
        3. change the :code:`is_test`
Y
yuyang18 已提交
3055 3056 3057
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

3058
        Args:
X
Xin Pan 已提交
3059 3060
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
3061

Y
yuyang18 已提交
3062 3063 3064 3065 3066 3067
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
3068
        res = Program()
3069
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
3070 3071 3072 3073

        # 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
3074
        if prune_read_op:
3075 3076 3077 3078 3079 3080 3081 3082 3083
            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
3084
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
3085 3086

        # change all `is_test` attributes to True
M
minqiyang 已提交
3087
        for i in six.moves.range(res.desc.num_blocks()):
3088
            block = res.desc.block(i)
M
minqiyang 已提交
3089
            for j in six.moves.range(block.op_size()):
3090 3091
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
3092
                    op._set_attr('is_test', True)
M
minqiyang 已提交
3093 3094 3095
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
3096
        res._sync_with_cpp()
3097 3098
        return res

3099 3100
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
3101 3102 3103 3104 3105 3106 3107
        """
        Deserialize a program desc from protobuf binary string.

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

        Args:
3108
            binary_str_type(str): The binary prootbuf string.
Y
yuyang18 已提交
3109 3110 3111 3112

        Returns:
            Program: A deserialized program desc.
        """
3113 3114
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
3115
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
3116
        p._sync_with_cpp()
3117
        return p
Y
Yu Yang 已提交
3118

3119
    @staticmethod
3120
    def _construct_from_desc(desc):
3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135
        """
        Construct a program from program desc.

        Args:
            desc(core.ProgramDesc): The program desc for constructing.

        Returns:
            Program: A program.
        """
        p = Program()
        p.desc = desc
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
        p._sync_with_cpp()
        return p

D
dzhwinter 已提交
3136 3137
    @property
    def random_seed(self):
Y
yuyang18 已提交
3138 3139 3140 3141 3142 3143
        """
        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
3144 3145
        return self._seed

Q
qiaolongfei 已提交
3146 3147
    @property
    def num_blocks(self):
Y
yuyang18 已提交
3148 3149 3150
        """
        The number of blocks in this program.
        """
Q
qiaolongfei 已提交
3151 3152
        return self.desc.num_blocks()

D
dzhwinter 已提交
3153 3154 3155 3156 3157 3158
    @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
3159
    def __repr__(self):
3160
        return self.__str__()
3161

Y
Yu Yang 已提交
3162
    def global_block(self):
Y
yuyang18 已提交
3163 3164 3165
        """
        Get the first block of this program.
        """
Y
Yu Yang 已提交
3166 3167
        return self.blocks[0]

Q
Qiao Longfei 已提交
3168
    def block(self, index):
Y
yuyang18 已提交
3169 3170 3171 3172 3173 3174 3175 3176
        """
        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
3177 3178
        return self.blocks[index]

Y
Yu Yang 已提交
3179
    def current_block(self):
Y
yuyang18 已提交
3180 3181 3182 3183
        """
        Get the current block. The :code:`current` block is the block to append
        operators.
        """
Y
Yu Yang 已提交
3184 3185
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
3186
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
3187 3188 3189 3190 3191 3192 3193 3194 3195 3196
        """
        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
3197
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
3198 3199 3200
        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
3201 3202 3203 3204
        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
3205
    def _rollback(self):
Y
yuyang18 已提交
3206 3207 3208 3209 3210
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
3211 3212
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
3213
    def _sync_with_cpp(self):
Y
yuyang18 已提交
3214 3215 3216 3217 3218 3219 3220 3221 3222 3223
        """
        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
3224 3225 3226
        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
3227
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
3228

W
Wu Yi 已提交
3229
    def _copy_param_info_from(self, other):
3230
        """
3231
        Copy the information of parameters from other program.
D
dzhwinter 已提交
3232

Y
yuyang18 已提交
3233 3234 3235
        Notes: This is a very low level API. Users should not invoke it
        directly.

3236 3237 3238 3239 3240 3241 3242
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
3243
            raise TypeError("_copy_param_info_from should be invoked with "
3244 3245 3246
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
3247
            raise ValueError("_copy_param_info_from should be invoked with two "
3248
                             "program, with represent the same topology")
W
Wu Yi 已提交
3249
        self.global_block()._copy_param_info_from(other.global_block())
3250

3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265
    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
3266
        self._parameters_on_pservers = other._parameters_on_pservers
3267
        self._endpoints = other._endpoints
3268
        self._ps_endpoint = other._ps_endpoint
3269 3270
        self._distributed_lookup_table = other._distributed_lookup_table

W
Wu Yi 已提交
3271
    def _copy_data_info_from(self, other):
F
fengjiayi 已提交
3272 3273
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
3274

Y
yuyang18 已提交
3275 3276 3277
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
3278 3279 3280 3281 3282 3283 3284
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
3285
            raise TypeError("_copy_param_info_from should be invoked with "
F
fengjiayi 已提交
3286 3287 3288
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
3289
            raise ValueError("_copy_param_info_from should be invoked with two "
F
fengjiayi 已提交
3290
                             "program, with represent the same topology")
3291
        for var in list(other.global_block().vars.values()):
F
fengjiayi 已提交
3292 3293 3294
            if var.is_data:
                self.global_block().var(var.name).is_data = True

3295
    def list_vars(self):
Y
yuyang18 已提交
3296 3297 3298 3299 3300 3301
        """
        Get all variables from this Program. A iterable object is returned.

        Returns:
            iterable: The generator will yield every variable in this program.
        """
3302
        for each_block in self.blocks:
3303
            for each_var in list(each_block.vars.values()):
3304 3305
                yield each_var

Y
Yu Yang 已提交
3306

Y
Yu Yang 已提交
3307
class Parameter(Variable):
3308
    """
3309
    Parameter is derived from Variable. A parameter is a persistable
3310
    Variable, and will be updated by optimizers after each iteration.
3311
    The training of a neural network is essentially the updating of
3312 3313
    its parameters.

3314
    Relative to a general Variable, a Parameter has several its own
3315 3316
    member variables:

3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328
    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.
3329 3330
    """

Y
Yu Yang 已提交
3331 3332 3333 3334 3335 3336 3337 3338 3339 3340
    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")
3341 3342 3343

        Variable.__init__(
            self, block, persistable=True, shape=shape, dtype=dtype, **kwargs)
Y
Yu Yang 已提交
3344 3345 3346 3347
        self.trainable = kwargs.get('trainable', True)

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

3348 3349
        self.regularizer = kwargs.get('regularizer', None)

F
fengjiayi 已提交
3350
        self.gradient_clip_attr = kwargs.get('gradient_clip_attr', None)
Y
Yu Yang 已提交
3351

W
wanghaoshuang 已提交
3352
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
3353

F
fengjiayi 已提交
3354 3355 3356
    def __str__(self):
        return self.to_string(True)

F
update  
fengjiayi 已提交
3357 3358 3359
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
3360

F
update  
fengjiayi 已提交
3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374
        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
3375
                               "gradient_clip_attr", "do_model_average")
F
update  
fengjiayi 已提交
3376
            for attr_name in additional_attr:
3377 3378
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
3379 3380
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
3381 3382 3383 3384
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
3385

Y
Yu Yang 已提交
3386
# program is a global instance.
Y
Yu Yang 已提交
3387 3388
_main_program_ = Program()
_startup_program_ = Program()
3389

3390

3391
def default_startup_program():
Y
Yu Yang 已提交
3392
    """
Y
yuyang18 已提交
3393 3394 3395 3396 3397 3398 3399 3400 3401
    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.
3402

Y
Yu Yang 已提交
3403 3404 3405
    Returns:
        Program: startup program
    """
Y
Yu Yang 已提交
3406
    return _startup_program_
3407

3408

3409
def default_main_program():
Y
Yu Yang 已提交
3410
    """
Y
yuyang18 已提交
3411 3412 3413 3414 3415 3416 3417 3418 3419
    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.
3420

Y
Yu Yang 已提交
3421 3422 3423
    Returns:
        Program: main program
    """
Y
Yu Yang 已提交
3424
    return _main_program_
Y
Yu Yang 已提交
3425 3426 3427 3428 3429


def switch_main_program(program):
    """
    Switch the main program to a new program.
3430

Y
Yu Yang 已提交
3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444
    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):
    """
3445
    Switch the startup program to a new program
Y
Yu Yang 已提交
3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457
    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


S
rename  
sneaxiy 已提交
3458
@signature_safe_contextmanager
Y
Yu Yang 已提交
3459 3460
def program_guard(main_program, startup_program=None):
    """
Y
yuyang18 已提交
3461 3462 3463
    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.
3464

Y
Yu Yang 已提交
3465
    Examples:
Y
yuyang18 已提交
3466 3467 3468 3469 3470 3471 3472 3473 3474 3475

        >>> 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.
3476

Y
Yu Yang 已提交
3477
    Examples:
Y
yuyang18 已提交
3478 3479 3480 3481 3482 3483

        >>> 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 = ...
3484

Y
Yu Yang 已提交
3485
    Args:
Y
yuyang18 已提交
3486
        main_program(Program): New main program inside `with` statement.
3487
        startup_program(Program): New startup program inside `with` statement.
Y
Yu Yang 已提交
3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500
            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
3501 3502


W
Wu Yi 已提交
3503
def _get_var(name, program=None):
X
xuwei06 已提交
3504
    """
Y
yuyang18 已提交
3505
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
3506

X
xuwei06 已提交
3507 3508 3509
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
3510
        If None, default_global_program() will be used.
X
xuwei06 已提交
3511 3512 3513 3514 3515 3516 3517

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
3518
    assert isinstance(program, Program)
X
xuwei06 已提交
3519 3520

    return program.global_block().var(name)
3521 3522


S
rename  
sneaxiy 已提交
3523
@signature_safe_contextmanager
L
lujun 已提交
3524 3525 3526 3527
def _dygraph_guard(tracer):
    global _dygraph_tracer_
    tmp_trace = _dygraph_tracer_
    _dygraph_tracer_ = tracer
M
minqiyang 已提交
3528

3529
    yield
P
Paddle CI 已提交
3530

L
lujun 已提交
3531
    _dygraph_tracer_ = tmp_trace
P
Paddle CI 已提交
3532 3533


S
rename  
sneaxiy 已提交
3534
@signature_safe_contextmanager
L
lujun 已提交
3535 3536 3537 3538
def _dygraph_place_guard(place):
    global _dygraph_current_expected_place_
    tmp_place = _dygraph_current_expected_place_
    _dygraph_current_expected_place_ = place
M
minqiyang 已提交
3539

3540
    yield
M
minqiyang 已提交
3541

L
lujun 已提交
3542
    _dygraph_current_expected_place_ = tmp_place