framework.py 201.7 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
21
from .wrapped_decorator import signature_safe_contextmanager, wrap_decorator
P
peizhilin 已提交
22
import os
F
fengjiayi 已提交
23
import re
24
import traceback
25
import six
26
import copy
27

Y
Yu Yang 已提交
28
import numpy as np
29
import subprocess
S
sneaxiy 已提交
30
import multiprocessing
31
import sys
32
import logging
M
minqiyang 已提交
33
from .. import compat as cpt
34
from .proto import framework_pb2
35 36

from . import core
37
from . import unique_name
38 39
import paddle.version as fluid_version
import warnings
40
import functools
Y
Yu Yang 已提交
41

42
__all__ = [
43 44 45 46
    'Program',
    'default_startup_program',
    'default_main_program',
    'program_guard',
47
    'name_scope',
S
sneaxiy 已提交
48 49
    'cuda_places',
    'cpu_places',
50
    'xpu_places',
S
sneaxiy 已提交
51
    'cuda_pinned_places',
L
lujun 已提交
52
    'in_dygraph_mode',
C
chengduo 已提交
53
    'is_compiled_with_cuda',
54
    'is_compiled_with_xpu',
55
    'Variable',
56
    'load_op_library',
57
    'require_version',
58
    'device_guard',
G
guofei 已提交
59 60
    'set_flags',
    'get_flags',
61
]
Y
Yu Yang 已提交
62

Q
qiaolongfei 已提交
63 64 65 66
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
67 68
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()

L
lujun 已提交
69
_dygraph_tracer_ = None
70
_global_expected_place_ = None
71
_current_device = None
72 73
global_prog_seed = 0

74

75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
def require_version(min_version, max_version=None):
    """
        Check if the installed version of PaddlePaddle is in [min_version, max_version],
        if the installed version is lower than ``min_version`` or higher than ``max_version``,
        an exception will be thrown, NO returns if the installed version is satisfied.

        Args:
            min_version (str): the minimum version required (like '1.4.0').
            max_version (str, optional): the max version required (like '1.6.0'), default is None,
                meaning any version equal or higher than ``min_version`` is acceptable.

        Returns:
            None.

        Raises:
            TypeError: if the type of ``min_version`` is not str.
            TypeError: if the type of ``max_version`` is not str or type(None).
            ValueError: if the value of ``min_version`` is not in version format.
            ValueError: if the value of ``max_version`` is not in version format or None.
            Exception: if the installed version is lower than ``min_version`` or higher than ``max_version``.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                # any version >= 0.1.0 is acceptable.
                fluid.require_version('0.1.0')

                # if 0.1.0 <= version <= 10.0.0, it is acceptable.
                fluid.require_version(min_version='0.1.0', max_version='10.0.0')
        """
    if not isinstance(min_version, str):
        raise TypeError(
            "The type of 'min_version' in require_version must be str, but received %s."
            % (type(min_version)))

    if not isinstance(max_version, (str, type(None))):
        raise TypeError(
            "The type of 'max_version' in require_version must be str or type(None), but received %s."
            % (type(max_version)))

    check_format = re.match(r'\d+(\.\d+){0,3}', min_version)
    if check_format is None or check_format.group() != min_version:
        raise ValueError(
            "The value of 'min_version' in require_version must be in format '\\d+(\\.\\d+){0,3}', "
            "like '1.5.2.0', but received %s" % min_version)

    if max_version is not None:
        check_format = re.match(r'\d+(\.\d+){0,3}', max_version)
        if check_format is None or check_format.group() != max_version:
            raise ValueError(
                "The value of 'max_version' in require_version must be in format '\\d+(\\.\\d+){0,3}', "
                "like '1.5.2.0', but received %s" % max_version)

    version_installed = [
        fluid_version.major, fluid_version.minor, fluid_version.patch,
        fluid_version.rc
    ]
    zero_version = ['0', '0', '0', '0']

    def version_cmp(ver_a, ver_b):
        for i in six.moves.range(len(ver_a)):
            if int(ver_a[i]) > int(ver_b[i]):
                return 1
            elif int(ver_a[i]) < int(ver_b[i]):
                return -1
        return 0

    if version_cmp(version_installed, zero_version) == 0:
        if max_version is not None:
            warnings.warn(
                "PaddlePaddle version in [%s, %s] required, but %s installed. "
                "Maybe you are using a develop version, "
                "please make sure the version is good with your code." %
                (min_version, max_version, fluid_version.full_version))
        else:
            warnings.warn(
                "PaddlePaddle version %s or higher is required, but %s installed, "
                "Maybe you are using a develop version, "
                "please make sure the version is good with your code." %
                (min_version, fluid_version.full_version))
        return

    min_version_split = min_version.split('.')
    min_version_to_check = min_version_split + zero_version[len(
        min_version_split):]

    if max_version is not None:
        max_version_split = max_version.split('.')
        max_version_to_check = max_version_split + zero_version[len(
            max_version_split):]

        if version_cmp(version_installed,
                       max_version_to_check) > 0 or version_cmp(
                           version_installed, min_version_to_check) < 0:
            raise Exception(
                "VersionError: PaddlePaddle version in [%s, %s] required, but %s installed."
                % (min_version, max_version, fluid_version.full_version))
    else:
        if version_cmp(version_installed, min_version_to_check) < 0:
            raise Exception(
                "VersionError: PaddlePaddle version %s or higher is required, but %s installed, "
                "please upgrade your PaddlePaddle to %s or other higher version."
                % (min_version, fluid_version.full_version, min_version))


L
lujun 已提交
182
def in_dygraph_mode():
L
lujun 已提交
183
    """
184

185 186 187 188 189 190 191
    .. note::
        Dynamic graph mode is turn ON by default since paddle 2.0.0

    This API checks whether paddle runs in dynamic graph mode.

    You can turn ON static graph mode by `enable_static <../dygraph/base/disable_dygraph_en.html>`_ ,
    and turn OFF static graph mode by `disable_static <../dygraph/base/enable_dygraph_en.html>`_  .
L
lujun 已提交
192 193

    Returns:
194
        bool: Whether paddle runs in dynamic graph mode.
L
lujun 已提交
195 196 197 198

    Examples:
        .. code-block:: python

199 200 201 202 203 204 205 206
            import paddle
            print(paddle.in_dynamic_mode())  # True, dynamic mode is turn ON by default since paddle 2.0.0

            paddle.enable_static()
            print(paddle.in_dynamic_mode())  # False, Now we are in static mode

            paddle.disable_static()
            print(paddle.in_dynamic_mode())  # True, Now we are in dynamic mode
L
lujun 已提交
207 208

    """
L
lujun 已提交
209
    return _dygraph_tracer_ is not None
210 211


212 213 214
def _dygraph_not_support_(func):
    def __impl__(*args, **kwargs):
        assert not in_dygraph_mode(
215
        ), "We don't support %s in imperative mode" % func.__name__
216 217 218 219 220 221 222 223
        return func(*args, **kwargs)

    return __impl__


def _dygraph_only_(func):
    def __impl__(*args, **kwargs):
        assert in_dygraph_mode(
224 225 226 227 228 229 230 231 232
        ), "We only support '%s()' in dynamic graph mode, please call 'paddle.disable_static()' to enter dynamic graph mode." % func.__name__
        return func(*args, **kwargs)

    return __impl__


def _static_only_(func):
    def __impl__(*args, **kwargs):
        assert not in_dygraph_mode(
233
        ), "In PaddlePaddle 2.x, we turn on dynamic graph mode by default, and '%s()' is only supported in static graph mode. So if you want to use this api, please call 'paddle.enable_static()' before this api to enter static graph mode." % func.__name__
234 235 236 237 238
        return func(*args, **kwargs)

    return __impl__


239 240 241 242 243 244
# NOTE(zhiqiu): This decorator is used for the APIs of Variable which is only
# used to make Variable and VarBase has same interfaces, like numpy. Since VarBase is not exposed in our
# official docments, logically, we want to keep VarBase and logically consistent. While, actually,
# in our implementation, there some APIs not supported, like numpy, because Variable contains the desc.
# So, those APIs are listed under class Variable to generate docs only.
# TODO(zhiqiu): We should make VarBase consistent with Variable in future, for example, by inheritting
T
tangwei12 已提交
245
# same base class.
246 247 248
def _fake_interface_only_(func):
    def __impl__(*args, **kwargs):
        raise AssertionError(
249 250 251 252
            "'%s' should be called by imperative Varible in imperative mode, please run it in dygraph "
            "mode. You can turn off paddle.enable_static() if you are in static mode, or turn off "
            "ProgramTranslator if you are using @paddle.jit.to_static" %
            func.__name__)
253 254 255 256

    return __impl__


T
tangwei12 已提交
257 258
# NOTE(chenweihang): There is argument name typo (stat_dict, correct name is state_dict)
# in fluid api Layer.set_dict, Optimizer.load, in order to correct the argument without
259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275
# introducing compatibility issues, add this decorator
# NOTE(chenweihang): not using `wrap_decorator` here is because `wrap_decorator` will
# move kwargs to args, which doesn't work in this decorate case
def deprecate_stat_dict(func):
    @functools.wraps(func)
    def wrapper(*args, **kwargs):
        if 'stat_dict' in kwargs:
            warnings.warn(
                "The argument `stat_dict` has deprecated, please change it to `state_dict`.",
                DeprecationWarning)
            kwargs['state_dict'] = kwargs['stat_dict']
            kwargs.pop('stat_dict')
        return func(*args, **kwargs)

    return wrapper


276 277
dygraph_not_support = wrap_decorator(_dygraph_not_support_)
dygraph_only = wrap_decorator(_dygraph_only_)
278
static_only = wrap_decorator(_static_only_)
279
fake_interface_only = wrap_decorator(_fake_interface_only_)
280 281


L
lujun 已提交
282 283
def _dygraph_tracer():
    return _dygraph_tracer_
284

W
Wu Yi 已提交
285

M
minqiyang 已提交
286
def _current_expected_place():
287 288 289
    global _global_expected_place_
    if _global_expected_place_ is None:
        if core.is_compiled_with_cuda():
290 291 292 293 294 295 296 297 298 299 300
            try:
                device_count = core.get_cuda_device_count()
            except Exception as e:
                device_count = 0
            if device_count > 0:
                _global_expected_place_ = core.CUDAPlace(0)
            else:
                warnings.warn(
                    "You are using GPU version Paddle, but your CUDA device is not set properly. CPU device will be used by default."
                )
                _global_expected_place_ = core.CPUPlace()
301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316
        else:
            _global_expected_place_ = core.CPUPlace()

    return _global_expected_place_


def _set_dygraph_tracer_expected_place(place):
    global _dygraph_tracer_
    if _dygraph_tracer_ is not None:
        _dygraph_tracer_._expected_place = place


def _set_expected_place(place):
    global _global_expected_place_
    _global_expected_place_ = place
    _set_dygraph_tracer_expected_place(place)
M
minqiyang 已提交
317 318


L
Leo Chen 已提交
319 320 321 322
# TODO(zhiqiu): remove this function.
def _var_base_to_np(var_base):
    """	
    convert VarBase tp numpy	
T
tangwei12 已提交
323

L
Leo Chen 已提交
324 325 326 327 328 329 330 331 332 333 334 335
    Args:	
        var_base(VarBase) : the VarBase to convert	
    Returns (np.ndarray): the np.ndarray contain the value of VarBase	
    """

    warnings.warn(
        "paddle.fluid.framework._var_base_to_np is deprecated, please use var_base.numpy() instead of _var_base_to_np(var_base)."
    )

    return var_base.numpy()


S
sneaxiy 已提交
336
def _cpu_num():
337
    if "CPU_NUM" not in os.environ.keys():
C
chengduo 已提交
338 339 340 341 342 343 344 345
        if multiprocessing.cpu_count() > 1:
            sys.stderr.write(
                '!!! The CPU_NUM is not specified, you should set CPU_NUM in the environment variable list.\n'
                'CPU_NUM indicates that how many CPUPlace are used in the current task.\n'
                'And if this parameter are set as N (equal to the number of physical CPU core) the program may be faster.\n\n'
                'export CPU_NUM={} # for example, set CPU_NUM as number of physical CPU core which is {}.\n\n'
                '!!! The default number of CPU_NUM=1.\n'.format(
                    multiprocessing.cpu_count(), multiprocessing.cpu_count()))
C
chengduo 已提交
346
        os.environ['CPU_NUM'] = str(1)
347
    cpu_num = os.environ.get('CPU_NUM')
C
chengduo 已提交
348 349 350 351 352 353 354 355 356 357
    return int(cpu_num)


def _cuda_ids():
    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())
    return device_ids
S
sneaxiy 已提交
358 359


360 361 362 363 364 365 366 367 368
def _xpu_ids():
    xpus_env = os.getenv("FLAGS_selected_xpus")
    if xpus_env:
        device_ids = [int(s) for s in xpus_env.split(",")]
    else:
        device_ids = six.moves.range(core.get_xpu_device_count())
    return device_ids


369 370 371 372 373 374 375 376 377 378 379 380 381 382 383
def is_compiled_with_xpu():
    """
    Whether this whl package can be used to run the model on XPU.

    Returns (bool): support xpu or not.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            support_xpu = fluid.is_compiled_with_xpu()
    """
    return core.is_compiled_with_xpu()


C
chengduo 已提交
384 385 386 387
def is_compiled_with_cuda():
    """
    Whether this whl package can be used to run the model on GPU.

388
    Returns (bool): `True` if CUDA is currently available, otherwise `False`.
C
chengduo 已提交
389 390 391 392

    Examples:
        .. code-block:: python

393 394
            import paddle
            support_gpu = paddle.is_compiled_with_cuda()
C
chengduo 已提交
395 396 397 398
    """
    return core.is_compiled_with_cuda()


S
sneaxiy 已提交
399
def cuda_places(device_ids=None):
L
lujun 已提交
400
    """
401 402 403 404
    **Note**:
        For multi-card tasks, please use `FLAGS_selected_gpus` environment variable to set the visible GPU device.
        The next version will fix the problem with `CUDA_VISIBLE_DEVICES` environment variable.

C
Chen Weihang 已提交
405
    This function creates a list of :code:`paddle.CUDAPlace` objects.
S
add doc  
sneaxiy 已提交
406 407

    If :code:`device_ids` is None, environment variable of
408
    :code:`FLAGS_selected_gpus` would be checked first. For example, if
S
add doc  
sneaxiy 已提交
409
    :code:`FLAGS_selected_gpus=0,1,2`, the returned list would
C
Chen Weihang 已提交
410
    be [paddle.CUDAPlace(0), paddle.CUDAPlace(1), paddle.CUDAPlace(2)].
S
add doc  
sneaxiy 已提交
411
    If :code:`FLAGS_selected_gpus` is not set, all visible
412
    gpu places would be returned according to the :code:`CUDA_VISIBLE_DEVICES` environment variable.
S
add doc  
sneaxiy 已提交
413 414

    If :code:`device_ids` is not None, it should be the device
415
    ids of GPUs. For example, if :code:`device_ids=[0,1,2]`,
S
add doc  
sneaxiy 已提交
416
    the returned list would be 
C
Chen Weihang 已提交
417
    [paddle.CUDAPlace(0), paddle.CUDAPlace(1), paddle.CUDAPlace(2)].
T
tangwei12 已提交
418

419 420
    Parameters:
        device_ids (list or tuple of int, optional): list of GPU device ids.
S
add doc  
sneaxiy 已提交
421 422

    Returns:
C
Chen Weihang 已提交
423
        list of paddle.CUDAPlace: Created GPU place list.
L
lujun 已提交
424 425 426 427

    Examples:
        .. code-block:: python

C
Chen Weihang 已提交
428 429
            import paddle
            import paddle.static as static
T
tangwei12 已提交
430

C
Chen Weihang 已提交
431 432 433
            paddle.enable_static()

            cuda_places = static.cuda_places()
L
lujun 已提交
434 435

    """
S
sneaxiy 已提交
436 437 438
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_ids is None:
C
chengduo 已提交
439
        device_ids = _cuda_ids()
S
sneaxiy 已提交
440 441 442 443 444
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.CUDAPlace(dev_id) for dev_id in device_ids]


445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466
def xpu_places(device_ids=None):
    """
    **Note**:
        For multi-card tasks, please use `FLAGS_selected_xpus` environment variable to set the visible XPU device.
    This function creates a list of :code:`paddle.XPUPlace` objects.
    If :code:`device_ids` is None, environment variable of
    :code:`FLAGS_selected_xpus` would be checked first. For example, if
    :code:`FLAGS_selected_xpus=0,1,2`, the returned list would
    be [paddle.XPUPlace(0), paddle.XPUPlace(1), paddle.XPUPlace(2)].
    If :code:`FLAGS_selected_xpus` is not set, all visible
    xpu places would be returned.
    If :code:`device_ids` is not None, it should be the device
    ids of XPUs. For example, if :code:`device_ids=[0,1,2]`,
    the returned list would be 
    [paddle.XPUPlace(0), paddle.XPUPlace(1), paddle.XPUPlace(2)].
    
    Parameters:
        device_ids (list or tuple of int, optional): list of XPU device ids.
    Returns:
        list of paddle.XPUPlace: Created XPU place list.
    Examples:
        .. code-block:: python
467
        
468 469 470 471 472 473 474 475 476 477 478 479 480 481 482
            import paddle
            import paddle.static as static
            
            paddle.enable_static()
            xpu_places = static.xpu_places()
    """
    assert core.is_compiled_with_xpu(), \
        "Not compiled with XPU"
    if device_ids is None:
        device_ids = _xpu_ids()
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.XPUPlace(dev_id) for dev_id in device_ids]


S
sneaxiy 已提交
483
def cpu_places(device_count=None):
L
lujun 已提交
484
    """
C
Chen Weihang 已提交
485
    This function creates a list of :code:`paddle.CPUPlace` objects, and returns the created list.
T
tangwei12 已提交
486

S
add doc  
sneaxiy 已提交
487 488
    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
C
chengduo 已提交
489 490
    If :code:`CPU_NUM` is not set, the default value is 1,
    i.e. CPU_NUM=1.
491 492
    :code:`CPU_NUM` indicates the number of devices used in the current task.
    The running of the program can be accelerated if :code:`CPU_NUM` is the same as the number of physical cores.
S
add doc  
sneaxiy 已提交
493

494 495
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
496 497

    Returns:
C
Chen Weihang 已提交
498
        list of paddle.CPUPlace: Created list of CPU places.
L
lujun 已提交
499 500 501 502

    Examples:
        .. code-block:: python

C
Chen Weihang 已提交
503 504
            import paddle
            import paddle.static as static
T
tangwei12 已提交
505

C
Chen Weihang 已提交
506 507 508
            paddle.enable_static()

            cpu_places = static.cpu_places()
L
lujun 已提交
509 510
    """

S
sneaxiy 已提交
511 512 513 514 515 516
    if device_count is None:
        device_count = _cpu_num()
    return [core.CPUPlace()] * device_count


def cuda_pinned_places(device_count=None):
L
lujun 已提交
517
    """
518
    This function creates a list of :code:`fluid.CUDAPinnedPlace` objects.
S
add doc  
sneaxiy 已提交
519 520 521

    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
522 523 524 525
    If :code:`CPU_NUM` is not set, the default value is 1,
    i.e. CPU_NUM=1.
    :code:`CPU_NUM` indicates the number of devices used in the current task.
    The running of the program can be accelerated if :code:`CPU_NUM` is the same as the number of physical cores.
S
add doc  
sneaxiy 已提交
526

527 528
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
529 530

    Returns:
531
        list of fluid.CUDAPinnedPlace: Created list of CUDA pinned places.
L
lujun 已提交
532 533 534 535

    Examples:
        .. code-block:: python

536
            import paddle.fluid as fluid
L
lujun 已提交
537 538 539 540 541
            cuda_pinned_places_cpu_num = fluid.cuda_pinned_places()
            # or
            cuda_pinned_places = fluid.cuda_pinned_places(1)

    """
S
sneaxiy 已提交
542 543 544
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_count is None:
545 546
        device_count = len(_cuda_ids())
    return [core.CUDAPinnedPlace()] * device_count
S
sneaxiy 已提交
547 548


549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574
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
575
@signature_safe_contextmanager
576 577
def name_scope(prefix=None):
    """
578 579
    :api_attr: Static Graph

580
    Generate hierarchical name prefix for the operators in Static Graph.
581

T
Tao Luo 已提交
582 583 584
    Note: 
        This should only used for debugging and visualization purpose.
        Don't use it for serious analysis such as graph/program transformations.
585
        Don't use it in dygraph, since it will cause memory leak.
586 587

    Args:
T
Tao Luo 已提交
588
        prefix(str, optional): prefix. Default is none.
589 590 591

    Examples:
        .. code-block:: python
T
Tink_Y 已提交
592

593 594 595
          import paddle
          paddle.enable_static()
          with paddle.static.name_scope("s1"):
596
             a = paddle.static.data(name='data', shape=[None, 1], dtype='int32')
T
Tao Luo 已提交
597
             b = a + 1
598
             with paddle.static.name_scope("s2"):
T
Tao Luo 已提交
599
                c = b * 1
600
             with paddle.static.name_scope("s3"):
T
Tao Luo 已提交
601
                d = c / 1
602 603 604
          with paddle.static.name_scope("s1"):
                f = paddle.tensor.pow(d, 2.0)
          with paddle.static.name_scope("s4"):
T
Tao Luo 已提交
605 606 607
                g = f - 1

          # Op are created in the default main program.  
608
          for op in paddle.static.default_main_program().block(0).ops:
T
Tao Luo 已提交
609 610 611 612 613 614 615 616 617 618 619 620 621 622 623
              # elementwise_add is created in /s1/
              if op.type == 'elementwise_add':
                  assert op.desc.attr("op_namescope") == '/s1/'
              # elementwise_mul is created in '/s1/s2'
              elif op.type == 'elementwise_mul':
                  assert op.desc.attr("op_namescope") == '/s1/s2/'
              # elementwise_div is created in '/s1/s3'
              elif op.type == 'elementwise_div':
                  assert op.desc.attr("op_namescope") == '/s1/s3/'
              # elementwise_sum is created in '/s4'
              elif op.type == 'elementwise_sub':
                  assert op.desc.attr("op_namescope") == '/s4/'
              # pow is created in /s1_1/
              elif op.type == 'pow':
                  assert op.desc.attr("op_namescope") == '/s1_1/'
624 625
    """
    # TODO(panyx0718): Only [0-9a-z].
626
    # in dygraph we don't need namescope since it will cause mem leak
L
Leo Chen 已提交
627 628 629
    if in_dygraph_mode():
        yield
    else:
T
tianshuo78520a 已提交
630
        assert prefix, "namescope prefix can not be empty."
631 632
        global _name_scope
        _name_scope = _name_scope.child(prefix)
633 634 635 636
        try:
            yield
        finally:
            _name_scope = _name_scope.parent()
637 638 639 640 641 642 643 644 645 646 647 648


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
649 650 651
def generate_control_dev_var_name():
    import random
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
652 653 654 655


def grad_var_name(var_name):
    """
656 657
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
658 659 660
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
661

662
def convert_np_dtype_to_dtype_(np_dtype):
663 664
    """
    Convert the data type in numpy to the data type in Paddle
665

666
    Args:
667
        np_dtype(np.dtype): the data type in numpy.
668

669 670
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
671 672

    """
673 674
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
675
        return core.VarDesc.VarType.FP32
676
    elif dtype == np.float64:
677
        return core.VarDesc.VarType.FP64
678
    elif dtype == np.float16:
679
        return core.VarDesc.VarType.FP16
680
    elif dtype == np.int32:
681
        return core.VarDesc.VarType.INT32
682
    elif dtype == np.int16:
683
        return core.VarDesc.VarType.INT16
684
    elif dtype == np.int64:
685
        return core.VarDesc.VarType.INT64
686
    elif dtype == np.bool:
687
        return core.VarDesc.VarType.BOOL
688
    elif dtype == np.uint16:
689 690 691
        # since there is still no support for bfloat16 in NumPy,
        # uint16 is used for casting bfloat16
        return core.VarDesc.VarType.BF16
692 693
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
694 695
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
696 697 698 699
    elif dtype == np.complex64:
        return core.VarDesc.VarType.COMPLEX64
    elif dtype == np.complex128:
        return core.VarDesc.VarType.COMPLEX128
700
    else:
M
minqiyang 已提交
701
        raise ValueError("Not supported numpy dtype %s" % dtype)
702 703 704


def dtype_is_floating(dtype):
705 706 707
    """
    Check the data type is floating or not.
    Args:
708
        dtype(np.dtype|core.VarDesc.VarType): data type.
709 710 711 712 713
            Could be numpy format or Paddle format

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

    """
714
    if not isinstance(dtype, core.VarDesc.VarType):
715 716
        dtype = convert_np_dtype_to_dtype_(dtype)

717 718 719 720
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
721 722


Y
Yang Yang(Tony) 已提交
723
def _debug_string_(proto, throw_on_error=True):
724 725 726 727 728 729 730 731 732 733 734
    """
    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
735
    error_fields = list()
Y
Yang Yang(Tony) 已提交
736
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
737 738
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
739 740 741
    return proto.__str__()


742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798
def _varbase_creator(type=core.VarDesc.VarType.LOD_TENSOR,
                     name=None,
                     shape=None,
                     dtype=None,
                     persistable=None,
                     **kwargs):
    if dtype is not None:
        if not isinstance(dtype, core.VarDesc.VarType):
            dtype = convert_np_dtype_to_dtype_(dtype)

    return core.VarBase(dtype if dtype else core.VarDesc.VarType.FP32,
                        list(shape) if shape else [], name, type
                        if type else core.VarDesc.VarType.LOD_TENSOR, True
                        if persistable else False)


class VariableMetaClass(type):
    @classmethod
    def __instancecheck__(cls, instance):
        t = type(instance)
        if in_dygraph_mode():
            return issubclass(t, core.VarBase)
        else:
            return issubclass(t, Variable)


class ParameterMetaClass(VariableMetaClass):
    @classmethod
    def __instancecheck__(cls, instance):
        t = type(instance)
        if in_dygraph_mode():
            return issubclass(t, ParamBase)
        else:
            return issubclass(t, Parameter)


def _getitem_impl_(var, item):
    """
    Slice the variable.

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

    Returns:
        Sliced variable
    """

    if not isinstance(item, tuple):
        item = [item]

    decrease_axis = []
    slice_axis = []
    slice_start = []
    slice_end = []
    slice_step = []
    use_strided_slice = False
    reverse_axis = []
799
    target_block = default_main_program().current_block()
800 801 802 803 804 805 806 807 808 809 810

    def fill_constant(shape, value, force_cpu=False, out=None):
        var.block.append_op(
            type='fill_constant',
            inputs={},
            outputs={'Out': [out]},
            attrs={
                'shape': shape,
                'dtype': out.dtype,
                'value': float(value),
                'force_cpu': force_cpu
811
            })
812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850
        out.stop_gradient = True
        return out

    for dim, slice_item in enumerate(item):
        if isinstance(slice_item, slice):
            start = slice_item.start
            end = slice_item.stop
            step = slice_item.step

            if start is None and end is None and step is None:
                continue

            if step is None:
                step = 1

            if start is None and end is None:
                assert (step == -1)
                reverse_axis.append(dim)
                continue

            if start is None:
                start = 0

            if end is None:
                end = 10000000

            if step != 1:
                use_strided_slice = True

            slice_axis.append(dim)
            slice_start.append(start)
            slice_end.append(end)
            slice_step.append(step)
        else:
            decrease_axis.append(dim)
            slice_axis.append(dim)
            slice_start.append(slice_item)
            slice_step.append(1)
            if isinstance(slice_item, Variable):
851
                temp_1 = var.block.create_var(dtype=slice_item.dtype)
852
                fill_constant([1], 1, force_cpu=True, out=temp_1)
853
                temp_end = target_block.create_var(dtype=slice_item.dtype)
854
                target_block.append_op(
855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893
                    type='elementwise_add',
                    inputs={'X': slice_item,
                            'Y': temp_1},
                    outputs={'Out': temp_end},
                    attrs={'axis': -1})
                slice_end.append(temp_end)
            else:
                slice_end.append(slice_item + 1
                                 if slice_item != -1 else 10000000)

    def contain_var(one_list):
        for ele in one_list:
            if isinstance(ele, Variable):
                return True
        return False

    def get_new_list_tensor(old_list):
        new_list_tensor = []
        for dim in old_list:
            if isinstance(dim, Variable):
                dim.stop_gradient = True
                new_list_tensor.append(dim)
            else:
                assert (isinstance(dim, int))
                temp_out = var.block.create_var(dtype='int32')
                fill_constant([1], dim, force_cpu=True, out=temp_out)
                new_list_tensor.append(temp_out)
        return new_list_tensor

    inputs = {'Input': [var]}
    attrs = {
        'axes': slice_axis,
        'starts': [],
        'ends': [],
        'decrease_axis': decrease_axis
    }
    if (use_strided_slice == True):
        attrs['strides'] = []
    infer_flags = list(1 for i in range(len(slice_axis)))
L
Leo Chen 已提交
894

895
    # starts
L
Leo Chen 已提交
896
    if contain_var(slice_start):
897 898 899 900 901 902 903 904
        inputs['StartsTensorList'] = get_new_list_tensor(slice_start)
        for i, dim in enumerate(slice_start):
            if isinstance(dim, Variable):
                attrs['starts'].append(-1)
                infer_flags[i] = -1
            else:
                attrs['starts'].append(dim)
    else:
L
Leo Chen 已提交
905 906 907 908
        attrs['starts'] = slice_start

    # ends
    if contain_var(slice_end):
909 910 911 912 913 914 915
        inputs['EndsTensorList'] = get_new_list_tensor(slice_end)
        for i, dim in enumerate(slice_end):
            if isinstance(dim, Variable):
                attrs['ends'].append(-1)
                infer_flags[i] = -1
            else:
                attrs['ends'].append(dim)
L
Leo Chen 已提交
916 917 918
    else:
        attrs['ends'] = slice_end

919 920
    # strides
    if use_strided_slice == True:
L
Leo Chen 已提交
921
        if contain_var(slice_step):
922 923 924 925 926 927 928
            inputs['StridesTensorList'] = get_new_list_tensor(slice_step)
            for i, dim in enumerate(slice_step):
                if isinstance(dim, Variable):
                    attrs['strides'].append(-1)
                    infer_flags[i] = -1
                else:
                    attrs['strides'].append(dim)
L
Leo Chen 已提交
929 930
        else:
            attrs['strides'] = slice_step
931 932 933 934 935 936
    # infer_flags
    attrs['infer_flags'] = infer_flags

    out = var
    if use_strided_slice == False and len(slice_axis) > 0:
        # append slice_op here
937
        slice_out_var = target_block.create_var(
938 939 940
            name=unique_name.generate_with_ignorable_key(var.name + "_slice"),
            dtype=var.dtype)

941
        target_block.append_op(
942 943 944 945 946 947 948
            type="slice",
            inputs=inputs,
            outputs={'Out': [slice_out_var]},
            attrs=attrs)

        out = slice_out_var
    elif use_strided_slice == True and len(slice_axis) > 0:
949
        strided_slice_out_var = target_block.create_var(
950 951 952
            name=unique_name.generate_with_ignorable_key(var.name +
                                                         "_strided_slice"),
            dtype=var.dtype)
953
        target_block.append_op(
954 955 956 957 958 959 960 961
            type="strided_slice",
            inputs=inputs,
            outputs={'Out': [strided_slice_out_var]},
            attrs=attrs)

        out = strided_slice_out_var

    if len(reverse_axis) > 0:
962
        reverse_out_var = target_block.create_var(
963 964 965
            name=unique_name.generate_with_ignorable_key(var.name +
                                                         "_slice_reverse"),
            dtype=var.dtype)
966
        target_block.append_op(
967 968 969 970 971 972 973 974 975 976 977
            type="reverse",
            inputs={'X': out},
            outputs={'Out': [reverse_out_var]},
            attrs={'axis': reverse_axis})

        out = reverse_out_var

    return out


@six.add_metaclass(VariableMetaClass)
X
Xin Pan 已提交
978
class Variable(object):
979
    """
J
Jiabin Yang 已提交
980
    **Notes**:
981
        **The constructor of Variable should not be invoked directly.**
J
Jiabin Yang 已提交
982

983 984
        **In Static Graph Mode: Please use** `Block.create_var` **to create a Static variable which has no data until being feed.**

J
Jiabin Yang 已提交
985 986 987
        **In Dygraph Mode: Please use** :ref:`api_fluid_dygraph_to_variable` **to create a dygraph variable with real data**

    In Fluid, every input and output of an OP is a variable. In most
988
    cases, variables are used for holding different kinds of data or training
J
Jiabin Yang 已提交
989 990
    labels. A variable belongs to a :ref:`api_guide_Block_en` . All variable has its own name and
    two variables in different :ref:`api_guide_Block_en` could have the same name.
991

992
    There are many kinds of variables. Each kind of them has its own attributes
J
Jiabin Yang 已提交
993
    and usages. Please refer to the `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_ for details.
994

T
tianshuo78520a 已提交
995
    Most of a Variable's member variables can be set to be None. It mean
996
    it is not available or will be specified later.
997

998
    Examples:
999 1000
        In Static Graph Mode:

1001 1002
        .. code-block:: python

1003
            import paddle.fluid as fluid
1004
            cur_program = fluid.Program()
1005 1006 1007 1008
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
J
Jiabin Yang 已提交
1009
        In `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_  Mode:
1010 1011 1012 1013 1014 1015 1016 1017 1018

        .. code-block:: python

            import paddle.fluid as fluid
            import numpy as np

            with fluid.dygraph.guard():
                new_variable = fluid.dygraph.to_variable(np.arange(10))

1019 1020
    """

Y
Yu Yang 已提交
1021 1022
    def __init__(self,
                 block,
Y
Yu Yang 已提交
1023
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
1024 1025 1026 1027
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
1028
                 capacity=None,
Q
QI JUN 已提交
1029
                 persistable=None,
F
fengjiayi 已提交
1030
                 error_clip=None,
Y
Yu Yang 已提交
1031
                 stop_gradient=False,
F
fengjiayi 已提交
1032
                 is_data=False,
H
Huihuang Zheng 已提交
1033
                 need_check_feed=False,
H
hong 已提交
1034
                 belong_to_optimizer=False,
Y
Yu Yang 已提交
1035
                 **kwargs):
Y
Yu Yang 已提交
1036 1037
        self.block = block
        if name is None:
Y
Yu Yang 已提交
1038
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
1039

Y
Yu Yang 已提交
1040
        if dtype is not None:
1041
            if not isinstance(dtype, core.VarDesc.VarType):
1042
                dtype = convert_np_dtype_to_dtype_(dtype)
1043

H
hong 已提交
1044 1045
        self.belong_to_optimizer = belong_to_optimizer

1046 1047 1048 1049 1050
        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))
1051

1052 1053 1054
        if self.desc is None:
            self.desc = self.block.desc.var(cpt.to_bytes(name))
            is_new_var = True
1055

1056 1057 1058
        if is_new_var:
            self.desc.set_type(type)
        elif self.desc.type() != type:
L
Leo Chen 已提交
1059 1060
            raise ValueError("Variable '{0}' has been created before. The "
                             "previous type is {1}, the new type is {2}. They"
1061 1062
                             " are not matched".format(self.name,
                                                       self.desc.type(), type))
1063

1064
        if shape is not None:
1065
            if is_new_var:
1066 1067 1068 1069 1070 1071
                self.desc.set_shape(shape)
            else:
                old_shape = self.shape
                shape = tuple(shape)
                if shape != old_shape:
                    raise ValueError(
L
Leo Chen 已提交
1072 1073
                        "Variable '{0}' has been created before. The previous "
                        "shape is {1}, the new shape is {2}. They are not "
1074 1075 1076 1077 1078 1079 1080
                        "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:
L
Leo Chen 已提交
1081 1082
                    raise ValueError("Variable '{0}' has been created before. "
                                     "The previous data type is {1}, the new "
1083 1084 1085 1086 1087 1088 1089 1090 1091
                                     "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:
L
Leo Chen 已提交
1092 1093
                    raise ValueError("Variable '{0}' has been created before. "
                                     "The previous lod_level is {1}, the new "
1094 1095 1096 1097 1098 1099 1100 1101 1102
                                     "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(
L
Leo Chen 已提交
1103 1104
                        "Variable '{0}' has been created before."
                        "The previous persistable is {1}, the new "
1105 1106
                        "persistable is {2}. They are not matched".format(
                            self.name, self.persistable, persistable))
1107

1108 1109
        if need_check_feed and is_new_var:
            self.desc.set_need_check_feed(need_check_feed)
H
Huihuang Zheng 已提交
1110

1111 1112 1113 1114 1115 1116 1117
        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
1118

1119 1120 1121 1122
        self.block.vars[name] = self
        self.op = None
        self._stop_gradient = stop_gradient
        self.is_data = is_data
Y
Yu Yang 已提交
1123

1124
    @fake_interface_only
1125 1126
    def detach(self):
        """
J
Jiabin Yang 已提交
1127
        **Notes**:
T
tianshuo78520a 已提交
1128
            **This API is ONLY available in Dygraph mode**
1129

1130
        Returns a new Variable, detached from the current graph.
1131

1132
        Returns:
J
Jiabin Yang 已提交
1133
             ( :ref:`api_guide_Variable_en` | dtype is same as current Variable): The detached Variable.
1134

1135

1136 1137 1138 1139 1140
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1141
                from paddle.fluid.dygraph import Linear
1142 1143 1144 1145
                import numpy as np

                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
                with fluid.dygraph.guard():
1146
                    linear = Linear(32, 64)
1147
                    data = to_variable(data)
1148
                    x = linear(data)
1149 1150 1151
                    y = x.detach()

        """
1152
        pass
1153

1154
    @fake_interface_only
1155
    def numpy(self):
1156
        """
J
Jiabin Yang 已提交
1157
        **Notes**:
T
tianshuo78520a 已提交
1158
            **This API is ONLY available in Dygraph mode**
1159

J
Jiabin Yang 已提交
1160
        Returns a numpy array shows the value of current :ref:`api_guide_Variable_en`
1161 1162 1163 1164 1165

        Returns:
            ndarray: The numpy value of current Variable.

        Returns type:
J
Jiabin Yang 已提交
1166
            ndarray: dtype is same as current Variable
1167 1168 1169 1170 1171 1172

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1173
                from paddle.fluid.dygraph import Linear
1174 1175 1176 1177
                import numpy as np

                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
                with fluid.dygraph.guard():
1178
                    linear = Linear(32, 64)
1179
                    data = to_variable(data)
1180
                    x = linear(data)
1181 1182 1183
                    print(x.numpy())

        """
1184
        pass
1185

1186
    @fake_interface_only
1187 1188
    def set_value(self, value):
        """
J
Jiabin Yang 已提交
1189
        **Notes**:
T
tianshuo78520a 已提交
1190
            **This API is ONLY available in Dygraph mode**
J
Jiabin Yang 已提交
1191

1192 1193 1194 1195 1196 1197 1198 1199 1200 1201
        Set a new value for this Variable.

        Args:
            value (Variable|np.ndarray): the new value.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1202
                from paddle.fluid.dygraph import Linear
1203 1204
                import numpy as np

1205
                data = np.ones([3, 1024], dtype='float32')
1206
                with fluid.dygraph.guard():
1207
                    linear = fluid.dygraph.Linear(1024, 4)
1208
                    t = to_variable(data)
1209
                    linear(t)  # call with default weight
1210
                    custom_weight = np.random.randn(1024, 4).astype("float32")
1211 1212
                    linear.weight.set_value(custom_weight)  # change existing weight
                    out = linear(t)  # call with different weight
1213 1214

        """
1215
        pass
1216

1217
    @fake_interface_only
1218
    def backward(self, retain_graph=False):
1219
        """
J
Jiabin Yang 已提交
1220
        **Notes**:
T
tianshuo78520a 已提交
1221
            **This API is ONLY available in Dygraph mode**
1222

1223
        Run backward of current Graph which starts from current Tensor.
1224

J
Jiabin Yang 已提交
1225
        Args:
1226 1227 1228 1229
            retain_graph(bool, optional): If False, the graph used to compute grads will be freed. If you would
                like to add more ops to the built graph after calling this method( :code:`backward` ), set the parameter
                :code:`retain_graph` to True, then the grads will be retained. Thus, seting it to False is much more memory-efficient.
                Defaults to False.
1230

J
Jiabin Yang 已提交
1231 1232
        Returns:
            NoneType: None
1233 1234 1235 1236 1237

        Examples:
            .. code-block:: python

                import numpy as np
1238 1239
                import paddle
                paddle.disable_static()
1240 1241

                x = np.ones([2, 2], np.float32)
1242 1243 1244 1245 1246 1247 1248
                inputs = []
                for _ in range(10):
                    tmp = paddle.to_tensor(x)
                    # if we don't set tmp's stop_gradient as False then, all path to loss will has no gradient since
                    # there is no one need gradient on it.
                    tmp.stop_gradient=False
                    inputs.append(tmp)
1249 1250
                ret = paddle.add_n(inputs)
                loss = paddle.sum(ret)
1251
                loss.backward()
1252 1253

        """
1254
        pass
1255

1256
    @fake_interface_only
1257
    def gradient(self):
1258
        """
J
Jiabin Yang 已提交
1259
        **Notes**:
T
tianshuo78520a 已提交
1260
            **This API is ONLY available in Dygraph mode**
1261 1262 1263

        Get the Gradient of Current Variable

J
Jiabin Yang 已提交
1264
        Returns:
1265
            ndarray or tuple of ndarray: if Variable's type is LoDTensor, return numpy value of the gradient of current Variable, if Variable's type is SelectedRows, return tuple of ndarray, first element of tuple is numpy value of the gradient of current Variable, second element of tuple is numpy value of the rows of current Variable.
1266 1267 1268 1269 1270 1271 1272

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

1273
                # example1: return ndarray
1274 1275 1276 1277 1278 1279 1280 1281 1282
                x = np.ones([2, 2], np.float32)
                with fluid.dygraph.guard():
                    inputs2 = []
                    for _ in range(10):
                        tmp = fluid.dygraph.base.to_variable(x)
                        tmp.stop_gradient=False
                        inputs2.append(tmp)
                    ret2 = fluid.layers.sums(inputs2)
                    loss2 = fluid.layers.reduce_sum(ret2)
1283
                    loss2.backward()
1284 1285
                    print(loss2.gradient())

1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
                # example2: return tuple of ndarray
                with fluid.dygraph.guard():
                    embedding = fluid.dygraph.Embedding(
                        size=[20, 32],
                        param_attr='emb.w',
                        is_sparse=True)
                    x_data = np.arange(12).reshape(4, 3).astype('int64')
                    x_data = x_data.reshape((-1, 3, 1))
                    x = fluid.dygraph.base.to_variable(x_data)
                    out = embedding(x)
                    out.backward()
                    print(embedding.weight.gradient())

1299
        """
1300
        pass
1301

1302
    @fake_interface_only
1303
    def clear_gradient(self):
1304
        """
J
Jiabin Yang 已提交
1305
        **Notes**:
T
tianshuo78520a 已提交
1306
            **1. This API is ONLY available in Dygraph mode**
J
Jiabin Yang 已提交
1307 1308

            **2. Use it only Variable has gradient, normally we use this for Parameters since other temporal Variable will be deleted by Python's GC**
1309

J
Jiabin Yang 已提交
1310
        Clear  (set to ``0`` ) the Gradient of Current Variable
1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328

        Returns:  None

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

                x = np.ones([2, 2], np.float32)
                with fluid.dygraph.guard():
                    inputs2 = []
                    for _ in range(10):
                        tmp = fluid.dygraph.base.to_variable(x)
                        tmp.stop_gradient=False
                        inputs2.append(tmp)
                    ret2 = fluid.layers.sums(inputs2)
                    loss2 = fluid.layers.reduce_sum(ret2)
1329
                    loss2.backward()
1330 1331 1332 1333 1334
                    print(loss2.gradient())
                    loss2.clear_gradient()
                    print("After clear {}".format(loss2.gradient()))

        """
1335
        pass
X
Xin Pan 已提交
1336

1337
    def __str__(self):
1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353
        return self._to_readable_code()

    def _to_readable_code(self):
        """
        Get readable debug string of Variable.

        .. note::
            If you want to get the debug string in protobuf format,
            please use :code:`to_string` method.

        Returns:
            string: The formatted Variable string.

        Examples:
            .. code-block:: python

1354 1355
                import paddle
                import paddle.static as static
1356

1357 1358 1359
                paddle.enable_static()

                cur_program = static.Program()
1360 1361 1362 1363 1364 1365
                cur_block = cur_program.current_block()
                new_variable = cur_block.create_var(name="X",
                                                    shape=[-1, 23, 48],
                                                    dtype='float32')
                print(new_variable._to_readable_code())
        """
1366 1367
        # VarType.LOD_TENSOR -> LOD_TENSOR
        type_str = str(self.type).split('.')[1]
1368
        if self.type == core.VarDesc.VarType.SELECTED_ROWS or self.type == core.VarDesc.VarType.LOD_TENSOR:
1369 1370
            dtype_str = str(self.dtype).split('.')[1]
            var_str = "{name} : {type}.shape{shape}.dtype({dtype}).stop_gradient({stop_gradient})".\
T
tangwei12 已提交
1371 1372
                format(name=self.name, type=type_str, shape=self.shape,
                       dtype=dtype_str, stop_gradient=self.stop_gradient)
1373
        else:
1374 1375
            var_str = "{name} : {type})".\
                format(name=self.name, type=type_str)
1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388

        if type(self) == Parameter:
            if self.trainable:
                var_str = "trainable param " + var_str
            else:
                var_str = "param " + var_str
        else:
            var_str = "var " + var_str

        if self.persistable:
            var_str = "persist " + var_str

        return var_str
Y
Yang Yang(Tony) 已提交
1389

F
update  
fengjiayi 已提交
1390
    def to_string(self, throw_on_error, with_details=False):
1391 1392 1393
        """
        Get debug string.

J
Jiabin Yang 已提交
1394 1395 1396 1397 1398
        Args:

            throw_on_error (bool): True if raise an exception when self is not initialized.

            with_details (bool): more details about variables and parameters (e.g. trainable, optimize_attr, ...) will be printed when with_details is True. Default value is False;
1399

1400 1401
        Returns:
            str: The debug string.
1402 1403 1404 1405 1406

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
1407
                import paddle
1408

1409
                paddle.enable_static()
1410 1411 1412 1413 1414
                cur_program = fluid.Program()
                cur_block = cur_program.current_block()
                new_variable = cur_block.create_var(name="X",
                                                    shape=[-1, 23, 48],
                                                    dtype='float32')
1415
                print(new_variable.to_string(True))
J
Jiabin Yang 已提交
1416
                print("=============with detail===============")
1417
                print(new_variable.to_string(True, True))
1418
        """
F
update  
fengjiayi 已提交
1419 1420
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
1421
        protostr = self.desc.serialize_to_string()
1422
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
1423 1424 1425 1426
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
1427 1428 1429
                res_str += "%s: %s\n" % (attr_name,
                                         cpt.to_text(getattr(self, attr_name)))

F
update  
fengjiayi 已提交
1430
        return res_str
1431 1432 1433

    __repr__ = __str__

1434
    @property
1435
    def stop_gradient(self):
J
Jiabin Yang 已提交
1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450
        """
        Indicating if we stop gradient from current Variable

        **Notes: This Property has default value as** ``True`` **in** `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_ **mode, while Parameter's default value is False. However, in Static Graph Mode all Variable's default stop_gradient value is** ``False``

        Examples:
          .. code-block:: python

            import paddle.fluid as fluid
            import numpy as np

            with fluid.dygraph.guard():
                value0 = np.arange(26).reshape(2, 13).astype("float32")
                value1 = np.arange(6).reshape(2, 3).astype("float32")
                value2 = np.arange(10).reshape(2, 5).astype("float32")
1451 1452
                linear = fluid.Linear(13, 5, dtype="float32")
                linear2 = fluid.Linear(3, 3, dtype="float32")
J
Jiabin Yang 已提交
1453 1454 1455
                a = fluid.dygraph.to_variable(value0)
                b = fluid.dygraph.to_variable(value1)
                c = fluid.dygraph.to_variable(value2)
1456 1457
                out1 = linear(a)
                out2 = linear2(b)
J
Jiabin Yang 已提交
1458 1459 1460 1461
                out1.stop_gradient = True
                out = fluid.layers.concat(input=[out1, out2, c], axis=1)
                out.backward()

1462
                assert linear.weight.gradient() is None
J
Jiabin Yang 已提交
1463 1464
                assert (out1.gradient() == 0).all()
        """
1465
        return self._stop_gradient
1466

1467 1468
    @stop_gradient.setter
    def stop_gradient(self, s):
1469
        self._stop_gradient = s
1470

1471 1472
    @property
    def persistable(self):
J
Jiabin Yang 已提交
1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493
        """
        Indicating if we current Variable should be long-term alive


        **Notes: This Property will be deprecated and this API is just to help user understand concept**

            **1. All Variable's persistable is** ``False`` **except Parameters.**

            **2. In** `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_ **mode, this property should not be changed**

        Examples:
          .. code-block:: python

            import paddle.fluid as fluid
            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
            print("persistable of current Var is: {}".format(new_variable.persistable))
        """
1494
        return self.desc.persistable()
1495

Y
Yu Yang 已提交
1496 1497
    @persistable.setter
    def persistable(self, p):
1498
        self.desc.set_persistable(p)
Y
Yu Yang 已提交
1499

Y
Yu Yang 已提交
1500 1501
    @property
    def name(self):
J
Jiabin Yang 已提交
1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517
        """
        Indicating name of current Variable

        **Notes: If it has two or more Varaible share the same name in the same** :ref:`api_guide_Block_en` **, it means these Variable will share content in no-** `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_ **mode. This is how we achieve Parameter sharing**

        Examples:
          .. code-block:: python

            import paddle.fluid as fluid
            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
            print("name of current Var is: {}".format(new_variable.name))
        """
1518
        return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
1519

1520 1521 1522 1523 1524 1525 1526 1527
    @property
    def grad_name(self):
        """
        Indicating name of the gradient Variable of current Variable.

        **Notes: This is a read-only property. It simply returns name of
          gradient Variable from a naming convention but doesn't guarantee
          the gradient exists.**
T
tangwei12 已提交
1528

1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539
        Examples:
          .. code-block:: python

          import paddle.fluid as fluid

          x = fluid.data(name="x", shape=[-1, 23, 48], dtype='float32')
          print(x.grad_name) # output is "x@GRAD"

        """
        return self.name + "@GRAD"

T
typhoonzero 已提交
1540 1541
    @name.setter
    def name(self, new_name):
1542
        self.desc.set_name(new_name)
T
typhoonzero 已提交
1543

Y
Yu Yang 已提交
1544 1545
    @property
    def shape(self):
J
Jiabin Yang 已提交
1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562
        """
        Indicating shape of current Variable

        **Notes: This is a read-only property**

        Examples:
          .. code-block:: python

            import paddle.fluid as fluid
            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
            print("shape of current Var is: {}".format(new_variable.shape))

        """
Y
Yu Yang 已提交
1563
        # convert to tuple, make it as same as numpy API.
1564
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
1565 1566

    @property
F
fengjiayi 已提交
1567
    def dtype(self):
J
Jiabin Yang 已提交
1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583
        """
        Indicating data type of current Variable

        **Notes: This is a read-only property**

        Examples:
          .. code-block:: python

            import paddle.fluid as fluid
            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
            print("Dtype of current Var is: {}".format(new_variable.dtype))
        """
1584
        return self.desc.dtype()
Y
Yu Yang 已提交
1585 1586 1587

    @property
    def lod_level(self):
J
Jiabin Yang 已提交
1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608
        """
        Indicating ``LoD`` info of current Variable, please refer to  :ref:`api_fluid_LoDTensor_en` to check the meaning
        of ``LoD``

        **Notes**:

            **1. This is a read-only property**

            **2. Don't support this property in** `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_ **mode, it's value should be** ``0(int)``

        Examples:
          .. code-block:: python

            import paddle.fluid as fluid
            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
            print("LoD Level of current Var is: {}".format(new_variable.lod_level))
        """
1609 1610 1611
        if self.type == core.VarDesc.VarType.SELECTED_ROWS:
            raise Exception("SelectedRows DO NOT supprt lod")

1612
        return self.desc.lod_level()
Y
Yu Yang 已提交
1613

Y
Yu Yang 已提交
1614 1615
    @property
    def type(self):
J
Jiabin Yang 已提交
1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631
        """
        Indicating Type of current Variable

        **Notes: This is a read-only property**

        Examples:
          .. code-block:: python

            import paddle.fluid as fluid
            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
            print("Type of current Var is: {}".format(new_variable.type))
        """
1632
        return self.desc.type()
Y
Yu Yang 已提交
1633

1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667
    def clone(self):
        """
        Returns a new static Variable, which is the clone of the original static
        Variable. It remains in the current graph, that is, the cloned Variable 
        provides gradient propagation. Calling ``out = tensor.clone()`` is same
        as ``out = assign(tensor)`` .

        Returns:
            Variable: The cloned Variable.

        Examples:
            .. code-block:: python

                import paddle

                paddle.enable_static()

                # create a static Variable
                x = paddle.static.data(name='x', shape=[3, 2, 1])
                # create a cloned Variable
                y = x.clone()

        """
        output = self.block.create_var(
            name=unique_name.generate_with_ignorable_key(self.name + "_clone"),
            dtype=self.dtype,
            type=self.type,
            persistable=self.persistable,
            stop_gradient=self.stop_gradient)

        self.block.append_op(
            type='assign', inputs={'X': [self]}, outputs={'Out': [output]})
        return output

W
Wu Yi 已提交
1668
    def _set_error_clip(self, error_clip):
1669 1670 1671 1672 1673 1674 1675 1676 1677
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
1678 1679
        self.error_clip = error_clip

1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708
    def _set_info(self, key, value):
        """
        Set key-value information for this variable.

        Args:
            key(str): Key for this information.
            value(object): The value associated to the key.

        Returns: 
            None
        """
        if not hasattr(self, "_info"):
            self._info = {}
        self._info[key] = value

    def _get_info(self, key):
        """
        Get the information of this variable corresponding to key.

        Args:
            key(str): Key for this information.

        Returns: 
            object
        """
        if hasattr(self, "_info") and key in self._info:
            return self._info[key]
        return None

1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719
    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:
T
tianshuo78520a 已提交
1720
            raise ValueError("slice step can not be zero")
1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795

        # 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
1796
    def _cloneVar(self, copy=False):
1797 1798
        if not copy:
            return self.block.create_var(
H
Hongyu Liu 已提交
1799 1800
                name=unique_name.generate_with_ignorable_key(self.name),
                dtype=self.dtype)
1801 1802 1803 1804
        else:
            return self

    def _sliceVar(self, axes, starts, ends):
L
lujun 已提交
1805
        new_var = self._cloneVar()
1806 1807 1808 1809 1810 1811 1812 1813 1814 1815
        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
1816
        new_var = self._cloneVar()
1817 1818 1819 1820 1821 1822 1823 1824 1825 1826
        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
1827
                return self._cloneVar(True)
1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845
            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
1846
                return self._cloneVar(True)
1847
            index = int(item)
1848
            if (index > 0 and index >= self.shape[axis]) \
1849 1850 1851 1852 1853 1854 1855
                    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):
1856
        return _getitem_impl_(self, item)
1857

1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899
    def __setitem__(self, item, value):
        inputs = {'Input': self}

        # 1. Parse item
        if not isinstance(item, tuple):
            item = [item]

        axes = []
        starts = []
        ends = []
        max_integer = sys.maxsize
        for dim, slice_item in enumerate(item):
            if isinstance(slice_item, slice):
                start = slice_item.start
                end = slice_item.stop
                step = slice_item.step

                if start is None and end is None and step is None:
                    continue

                start = 0 if start is None else start
                step = 1 if step is None else step

                # TODO: support cases when step != 1
                if step != 1:
                    raise ValueError(
                        "When assign a value to a paddle.Tensor, only support step is 1, "
                        "but received step is {}.".format(step))
                end = max_integer if end is None else end
            else:
                start = slice_item
                end = slice_item + 1 if slice_item != -1 else max_integer
            axes.append(dim)
            starts.append(start)
            ends.append(end)

        attrs = {'axes': axes, 'starts': starts, 'ends': ends}

        # 2. Parse value
        dtype = self.dtype
        attrs['dtype'] = dtype

1900
        from .data_feeder import convert_dtype
1901 1902
        #  2.1 value is an integer of float
        if isinstance(value, (int, float)):
1903
            value = np.array([value]).astype(convert_dtype(dtype))
1904 1905 1906 1907 1908 1909 1910 1911 1912 1913

        #  2.2 value is a np.ndarray
        if isinstance(value, np.ndarray):
            shape = list(value.shape)
            if dtype == core.VarDesc.VarType.BOOL:
                value_name = "bool_values"
                values = [bool(v) for v in value.flat]
            elif dtype == core.VarDesc.VarType.FP32:
                value_name = "fp32_values"
                values = [float(v) for v in value.flat]
1914 1915 1916
            elif dtype == core.VarDesc.VarType.FP64:
                value_name = "fp64_values"
                values = [float(v) for v in value.flat]
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
            elif dtype == core.VarDesc.VarType.INT32:
                value_name = "int32_values"
                values = [int(v) for v in value.flat]
            elif dtype == core.VarDesc.VarType.INT64:
                value_name = "int64_values"
                values = [int(v) for v in value.flat]
            else:
                raise TypeError(
                    "When assign a numpy.ndarray, integer or float to a paddle.Tensor, "
                    "the data type of the paddle.Tensor must be bool, float32, int32 or int64, but "
                    "received %s." % convert_dtype(dtype))
            attrs[value_name] = values
            attrs["shape"] = shape

        elif isinstance(value, Variable):
            inputs["ValueTensor"] = value
        else:
            raise TypeError(
                "Only support to assign an integer, float, numpy.ndarray or "
                "paddle.Tensor to a paddle.Tensor, but received {}".format(
                    type(value)))

        self.block.append_op(
            type="set_value", inputs=inputs, outputs={'Out': self}, attrs=attrs)
        return self

Y
Yu Yang 已提交
1943

F
fengjiayi 已提交
1944 1945 1946
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
1947

1948 1949
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
1950 1951 1952 1953
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
1954
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
1955 1956 1957 1958 1959
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
1960 1961 1962 1963
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
1964 1965 1966 1967 1968 1969 1970 1971 1972
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
1973
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
1974 1975 1976 1977 1978 1979
        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):
1980 1981 1982 1983 1984 1985 1986 1987
        """
        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
1988 1989
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
1990 1991
        return self.op_proto_map[type]

1992 1993 1994 1995 1996 1997
    def update_op_proto(self):
        op_protos = get_all_op_protos()
        for proto in op_protos:
            if proto.type not in self.op_proto_map:
                self.op_proto_map[proto.type] = proto

1998 1999 2000 2001
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
2002
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
2003
            core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
2004 2005
            core.op_proto_and_checker_maker.kOpCreationCallstackAttrName(),
            core.op_proto_and_checker_maker.kOpDeviceAttrName()
2006 2007
        }

F
fengjiayi 已提交
2008

X
Xin Pan 已提交
2009
class Operator(object):
2010
    """
2011 2012 2013 2014 2015 2016 2017
    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
2018
        type(str): The type of operator. Default None.
2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038
        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
2039
        Block.append_op or Block._prepend_op instead.
2040 2041 2042 2043

    Examples:
        .. code-block:: python

2044
            import paddle.fluid as fluid
2045
            cur_program = fluid.Program()
2046 2047 2048 2049 2050
            cur_block = cur_program.current_block()
            # var1 += var2 + var3
            cur_block.append_op(type="sum",
                                inputs={"X": [var1, var2, var3]},
                                outputs={"Out": [var1]})
2051
    """
2052
    OP_WITHOUT_KERNEL_SET = {
2053 2054
        'feed', 'fetch', 'recurrent', 'go', 'rnn_memory_helper_grad',
        'conditional_block', 'while', 'send', 'recv', 'listen_and_serv',
2055 2056
        'fl_listen_and_serv', 'ncclInit', 'select', 'checkpoint_notify',
        'gen_nccl_id', 'c_gen_nccl_id', 'c_comm_init', 'c_sync_calc_stream',
T
tangwei12 已提交
2057 2058
        'c_sync_comm_stream', 'queue_generator', 'dequeue', 'enqueue',
        'heter_listen_and_serv'
2059
    }
2060

Y
Yu Yang 已提交
2061 2062
    def __init__(self,
                 block,
Y
Yu Yang 已提交
2063
                 desc,
Y
Yu Yang 已提交
2064 2065 2066
                 type=None,
                 inputs=None,
                 outputs=None,
M
minqiyang 已提交
2067
                 attrs=None):
L
lujun 已提交
2068
        if in_dygraph_mode():
2069 2070
            if type is None:
                raise ValueError(
2071
                    "`type` to initialized an Operator can not be None.")
J
Jiabin Yang 已提交
2072
            self._type = type
M
minqiyang 已提交
2073
            self.attrs = attrs if attrs else {}
2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087
        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(
2088
                )] = self.block.program._op_role
2089 2090 2091

            role_var_name = op_maker.kOpRoleVarAttrName()
            if len(self.block.program.
2092 2093
                   _op_role_var) != 0 and role_var_name not in op_attrs:
                op_attrs[role_var_name] = self.block.program._op_role_var
2094 2095 2096 2097 2098 2099 2100 2101

            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(
2102
                    "`type` to initialized an Operator can not be None.")
2103 2104
            else:
                callstack_var_name = op_maker.kOpCreationCallstackAttrName()
2105 2106 2107 2108 2109 2110 2111
                op_attrs[callstack_var_name] = []
                for frame in traceback.extract_stack():
                    op_attrs[callstack_var_name].append(
                        '  File "{}", line {}, in {}'.format(frame[0], frame[1],
                                                             frame[2]))
                    op_attrs[callstack_var_name].append('    {}'.format(frame[
                        3]))
2112 2113 2114 2115 2116 2117 2118

            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()

2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136
            # set device for op with kernels, give warning for op without kernels
            # when force_cpu and device_guard are used at the same time, a warning will be given.
            # TODO(zhangting2020): when force_cpu is removed, clear warning below.
            if _current_device is not None:
                if self._has_kernel(type):
                    op_device = op_maker.kOpDeviceAttrName()
                    op_attrs[op_device] = _current_device
                else:
                    warnings.warn("The Op(%s) is not support to set device." %
                                  type)
                if 'force_cpu' in op_attrs:
                    if (type is 'less_than' and op_attrs['force_cpu'] != None
                        ) or op_attrs['force_cpu'] != False:
                        warnings.warn(
                            "The Attr(force_cpu) of Op(%s) will be deprecated in the future, "
                            "please use 'device_guard' instead. 'device_guard' has higher priority when they are "
                            "used at the same time." % type)

2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149
            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]
2150
                        if not isinstance(in_args, (list, tuple)):
2151 2152 2153 2154 2155 2156
                            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 = []
2157
                        for index, arg in enumerate(in_args):
2158 2159 2160 2161
                            if isinstance(arg, six.string_types):
                                in_arg_names.append(arg)
                            elif isinstance(arg, six.binary_type):
                                in_arg_names.append(arg.decode())
2162
                            elif isinstance(arg, (Variable, core.VarBase)):
2163
                                in_arg_names.append(cpt.to_text(arg.name))
2164
                            else:
2165 2166 2167 2168
                                raise TypeError(
                                    "The type of '%s' in operator %s should be "
                                    "one of [basestring(), str, Varibale] in python2, "
                                    "or one of [str, bytes, Variable] in python3."
2169 2170
                                    "but received : %s" %
                                    (in_proto.name, type, arg))
2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194
                        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:
2195 2196 2197 2198
                        if isinstance(arg, six.string_types):
                            out_arg_names.append(arg)
                        else:
                            out_arg_names.append(cpt.to_text(arg.name))
2199
                        # TODO(minqiyang): could we remove variable's op in static mode?
L
lujun 已提交
2200
                        if not in_dygraph_mode():
2201 2202 2203 2204
                            if isinstance(arg, six.string_types):
                                block.var(arg).op = self
                            else:
                                arg.op = self
2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222
                    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
2223
    def _has_kernel(self, op_type):
2224 2225
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
2226
    def to_string(self, throw_on_error):
2227
        """
2228 2229
        Get debug string.

2230
        Args:
2231 2232
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
2233

2234 2235
        Returns:
            str: The debug string.
2236 2237

        """
2238
        protostr = self.desc.serialize_to_string()
2239
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
2240 2241
        return _debug_string_(proto, throw_on_error)

2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328
    def _to_readable_code(self, skip_op_callstack=True):
        """
        Get readable debug string of Operator.

        .. note::
            If you want to get the debug string in protobuf format,
            please use :code:`to_string` method.

        Args:
            skip_op_callstack(bool): whether to skip parsing Operator's attribute
                op_callstack, default value is True

        Returns:
            string: The formatted Operator string.

        Examples:
            .. code-block:: python

            import paddle.fluid as fluid

            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            var = cur_block.create_var(name="X",
                                       shape=[-1, 23, 48],
                                       dtype='float32')
            new_op = cur_block.append_op(type="abs",
                                inputs={"X": [var]},
                                outputs={"Out": [var]})
            print(new_op._to_readable_code())
        """
        assert isinstance(
            skip_op_callstack, bool
        ), "skip_op_callstack parameter's type is error, expect bool, received %s".format(
            type(skip_op_callstack))
        outputs_str = "{"
        for i in range(0, len(self.output_names)):
            outputs_str += "{name}=".format(name=self.output_names[i])
            o = self.output(self.output_names[i])
            outputs_str += "{value}".format(value=o)
            if i != len(self.output_names) - 1:
                outputs_str += ", "
        outputs_str += "}"

        inputs_str = "{"
        for i in range(0, len(self.input_names)):
            inputs_str += "{name}=".format(name=self.input_names[i])
            o = self.input(self.input_names[i])
            inputs_str += "{value}".format(value=o)

            if i != len(self.input_names) - 1:
                inputs_str += ", "
        inputs_str += "}"

        attr_names = sorted(self.attr_names)
        attrs_str = ""
        for i in range(0, len(attr_names)):
            name = attr_names[i]
            if skip_op_callstack and name == "op_callstack":
                continue

            attr_type = self.desc.attr_type(name)
            if attr_type == core.AttrType.BLOCK:
                a = "{name} = block[{value}]".format(
                    name=name, type=attr_type, value=self._block_attr_id(name))
                attrs_str += a
                if i != len(attr_names) - 1:
                    attrs_str += ", "
                continue

            if attr_type == core.AttrType.BLOCKS:
                a = "{name} = blocks{value}".format(
                    name=name,
                    type=attr_type,
                    value=self._blocks_attr_ids(name))
                attrs_str += a
                if i != len(attr_names) - 1:
                    attrs_str += ", "
                continue

            a = "{name} = {value}".format(
                name=name, type=attr_type, value=self.desc.attr(name))
            attrs_str += a
            if i != len(attr_names) - 1:
                attrs_str += ", "

        if outputs_str != "{}":
            op_str = "{outputs} = {op_type}(inputs={inputs}, {attrs})".\
T
tangwei12 已提交
2329 2330
                format(outputs=outputs_str, op_type=self.type,
                       inputs=inputs_str, attrs=attrs_str)
2331 2332 2333 2334 2335
        else:
            op_str = "{op_type}(inputs={inputs}, {attrs})".\
                format(op_type=self.type, inputs=inputs_str, attrs=attrs_str)
        return op_str

Y
Yang Yang(Tony) 已提交
2336
    def __str__(self):
2337
        return self._to_readable_code()
2338 2339 2340

    __repr__ = __str__

F
fengjiayi 已提交
2341 2342
    @property
    def type(self):
2343
        return self.desc.type()
F
fengjiayi 已提交
2344 2345

    def input(self, name):
2346
        r"""
2347
        Get the input arguments according to the input parameter name.
2348

2349 2350
        Args:
            name(str): The input parameter name.
2351

2352 2353 2354
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
2355
        """
F
fengjiayi 已提交
2356 2357
        return self.desc.input(name)

W
Wu Yi 已提交
2358
    def _rename_input(self, old_name, new_name):
2359 2360 2361 2362 2363 2364 2365 2366 2367 2368
        """
        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
2369
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
2370

W
Wu Yi 已提交
2371
    def _rename_output(self, old_name, new_name):
2372 2373 2374 2375 2376 2377 2378 2379 2380 2381
        """
        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
2382
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
2383

F
fengjiayi 已提交
2384 2385 2386 2387
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
2388 2389 2390 2391 2392 2393 2394 2395
    @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
2396
    def output(self, name):
2397
        r"""
2398
        Get output arguments by the output parameter name.
2399

2400 2401
        Args:
            name(str): The output parameter name.
2402

2403 2404 2405
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
2406
        """
F
fengjiayi 已提交
2407 2408 2409 2410 2411 2412
        return self.desc.output(name)

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

2413 2414 2415 2416 2417 2418 2419 2420
    @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
2421
    def has_attr(self, name):
2422
        """
2423 2424
        Whether this Operator has the attribute with name or not.

2425
        Args:
2426
            name(str): the attribute name.
2427

2428 2429
        Returns:
            bool: True if has this attribute.
2430 2431

        """
F
fengjiayi 已提交
2432 2433 2434
        return self.desc.has_attr(name)

    def attr_type(self, name):
2435
        """
2436
        Get the type of attribute by attribute's name.
2437

2438 2439
        Args:
            name(str): the attribute name.
2440

2441 2442
        Returns:
            core.AttrType: the attribute type.
2443
        """
F
fengjiayi 已提交
2444 2445
        return self.desc.attr_type(name)

W
Wu Yi 已提交
2446
    def _set_attr(self, name, val):
2447 2448 2449 2450 2451 2452 2453 2454 2455 2456
        """
        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
2457 2458
        self._update_desc_attr(name, val)

2459 2460 2461
    def _remove_attr(self, name):
        self.desc.remove_attr(name)

G
gongweibao 已提交
2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472
    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
2473 2474
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
2475 2476
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
2477
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
2478 2479 2480 2481
        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
2482
            self.desc._set_attr(name, val)
Y
yuyang18 已提交
2483

F
fengjiayi 已提交
2484 2485 2486 2487 2488
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
2489
        """
2490 2491
        Get the attribute by name.

2492
        Args:
2493
            name(str): the attribute name.
2494

2495 2496
        Returns:
            bool|int|str|float|list: The attribute value. The return value
2497 2498
            can be any valid attribute type.
        """
F
fengjiayi 已提交
2499
        return self.desc.attr(name)
Y
Yu Yang 已提交
2500

W
Wu Yi 已提交
2501
    def _block_attr_id(self, name):
2502
        """
G
gongweibao 已提交
2503
        Get the block attribute's id by name.
2504

2505 2506
        Args:
            name(str): the attribute name.
2507

2508 2509
        Returns:
            int: the block index.
2510
        """
W
Wu Yi 已提交
2511
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
2512

W
Wu Yi 已提交
2513
    def _block_attr(self, name):
G
gongweibao 已提交
2514 2515 2516 2517 2518 2519 2520 2521 2522 2523
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
2524
        id = self._block_attr_id(name)
G
gongweibao 已提交
2525 2526 2527
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
2528
    def _blocks_attr(self, name):
G
gongweibao 已提交
2529 2530 2531 2532 2533 2534 2535 2536 2537 2538
        """
        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
2539
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
2540 2541 2542 2543 2544
            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
2545
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
2546 2547 2548 2549 2550 2551 2552 2553 2554 2555
        """
        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
2556
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
2557

J
JiayiFeng 已提交
2558
    def all_attrs(self):
F
fengjiayi 已提交
2559
        """
2560 2561 2562
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
2563
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
2564 2565 2566 2567
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
G
gongweibao 已提交
2568 2569
            attr_type = self.desc.attr_type(n)
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
2570
                attr_map[n] = self._block_attr(n)
G
gongweibao 已提交
2571 2572 2573
                continue

            if attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
2574
                attr_map[n] = self._blocks_attr(n)
G
gongweibao 已提交
2575 2576 2577 2578
                continue

            attr_map[n] = self.attr(n)

F
fengjiayi 已提交
2579 2580
        return attr_map

2581 2582 2583
    def _is_optimize_op(self):
        op_maker = core.op_proto_and_checker_maker
        OPTIMIZE = core.op_proto_and_checker_maker.OpRole.Optimize
2584 2585 2586 2587

        if not self.desc.has_attr(op_maker.kOpRoleAttrName()):
            return False

2588 2589 2590
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(OPTIMIZE):
            return True
2591 2592 2593 2594 2595 2596 2597 2598

        return False

    def _is_backward_op(self):
        op_maker = core.op_proto_and_checker_maker
        BACKWARD = core.op_proto_and_checker_maker.OpRole.Backward

        if not self.desc.has_attr(op_maker.kOpRoleAttrName()):
2599 2600
            return False

2601 2602 2603 2604 2605 2606
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(BACKWARD):
            return True

        return False

Y
Yu Yang 已提交
2607

Y
Yu Yang 已提交
2608
class Block(object):
2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622
    """
    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
2623
        use `Program._create_block()` to create a block.
2624 2625 2626 2627

    Examples:
        .. code-block:: python

2628 2629 2630
            import paddle.fluid as fluid

            cur_program = fluid.Program()
2631 2632 2633 2634 2635 2636 2637 2638 2639
            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
2640
    def __init__(self, program, idx):
Y
Yu Yang 已提交
2641
        self.desc = program.desc.block(idx)
2642
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
2643
        self.ops = list()  # operator list
Y
Yu Yang 已提交
2644
        self.program = program
2645
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
2646

2647
    def __str__(self):
2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693
        return self._to_readable_code()

    def _to_readable_code(self, skip_op_callstack=True):
        """
        Get readable debug string of Block.

        .. note::
            If you want to get the debug string in protobuf format,
            please use :code:`to_string` method.

        Args:
            skip_op_callstack(bool): whether to skip parsing Operator's attribute
                op_callstack, default value is True

        Returns:
            string: The formatted Block string.

        Examples:
            .. code-block:: python

            import paddle.fluid as fluid

            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_var = cur_block.create_var(name="X",
                                           shape=[-1, 23, 48],
                                           dtype='float32')
            new_op = cur_block.append_op(type="abs",
                                inputs={"X": [new_var]},
                                outputs={"Out": [new_var]})
            print(cur_block._to_readable_code())
        """
        assert isinstance(
            skip_op_callstack, bool
        ), "skip_op_callstack parameter's type is error, expect bool, received %s".format(
            type(skip_op_callstack))
        block_str = "{ // block "
        block_str += "{}\n".format(self.idx)
        for var in list(self.vars.values()):
            block_str += "    {}\n".format(var._to_readable_code())
        block_str += "\n"
        for op in self.ops:
            block_str += "    {}\n".format(
                op._to_readable_code(skip_op_callstack))
        block_str += "}"
        return block_str
Y
Yang Yang(Tony) 已提交
2694

F
fengjiayi 已提交
2695 2696
    def to_string(self, throw_on_error, with_details=False):
        """
2697 2698
        Get debug string.

F
fengjiayi 已提交
2699 2700
        Args:
            throw_on_error(bool): raise exception when self is not initialized
2701
                when throw_on_error is True.
F
update  
fengjiayi 已提交
2702
            with_details(bool): more details about variables and parameters
2703 2704
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
2705

2706 2707
        Returns:
            str: The debug string.
F
fengjiayi 已提交
2708 2709 2710 2711
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
2712
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
2713 2714
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
2715
            for var in list(self.vars.values()):
F
fengjiayi 已提交
2716
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
2717
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
2718
            for op in self.ops:
F
fengjiayi 已提交
2719 2720
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
2721 2722 2723
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
2724 2725
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
2726 2727
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
2728 2729 2730

    __repr__ = __str__

Y
Yu Yang 已提交
2731 2732
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
2733
        return self.desc.parent
Y
Yu Yang 已提交
2734

Y
Yu Yang 已提交
2735 2736 2737 2738
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
2739
    def _set_forward_block_idx(self, idx):
2740 2741 2742 2743 2744 2745 2746 2747 2748
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
2749
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
2750

2751 2752 2753 2754 2755 2756 2757 2758
    @property
    def backward_block_idx(self):
        cur_block_idx = self.idx
        for block in self.program.blocks:
            if block.forward_block_idx == cur_block_idx:
                return block.idx
        return -1

Y
Yu Yang 已提交
2759 2760
    @property
    def idx(self):
Y
Yu Yang 已提交
2761
        return self.desc.id
Y
Yu Yang 已提交
2762

Q
Qiao Longfei 已提交
2763
    def var(self, name):
2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776
        """
        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.
        """
2777
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
2778 2779 2780
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
2781 2782
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
2783
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
2784
        return v
Q
Qiao Longfei 已提交
2785

X
Xin Pan 已提交
2786
    def _find_var_recursive(self, name):
2787 2788 2789 2790 2791 2792 2793
        """
        Get a Variable by name from this block recursively.

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

        Returns:
X
Xin Pan 已提交
2794
            Variable: the Variable with the giving name. Or None if not found.
2795
        """
Y
Yu Yang 已提交
2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819
        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
2820
        return None
Y
Yu Yang 已提交
2821

X
Xin Pan 已提交
2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840
    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
2841

Q
Qiao Longfei 已提交
2842
    def all_parameters(self):
2843
        return list(self.iter_parameters())
2844

2845
    def iter_parameters(self):
M
minqiyang 已提交
2846
        return (item[1] for item in six.iteritems(self.vars)
2847
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
2848

Y
Yu Yang 已提交
2849
    def create_var(self, *args, **kwargs):
L
Leo Chen 已提交
2850 2851 2852
        if in_dygraph_mode():
            var = _varbase_creator(*args, **kwargs)
        else:
2853 2854 2855
            var = Variable(block=self, *args, **kwargs)
            if 'initializer' in kwargs:
                kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
2856
        return var
Y
Yu Yang 已提交
2857

Q
Qiao Longfei 已提交
2858 2859 2860
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
2861
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
2862 2863
        """
        Rename variable in vars and ops' inputs and outputs
2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875

        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
2876
        """
M
minqiyang 已提交
2877 2878
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
2879

T
typhoonzero 已提交
2880
        if not self.has_var(name):
2881
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
2882 2883
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
2884
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
2885 2886 2887 2888 2889 2890
            stop_gradient = v.stop_gradient
            trainable = v.trainable
            optimize_attr = v.optimize_attr
            regularizer = v.regularizer
            error_clip = v.error_clip
        elif type(v) == Variable:
T
typhoonzero 已提交
2891
            var_type = "Variable"
T
wip  
typhoonzero 已提交
2892 2893 2894 2895
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
2896
        orig_var_type = v.type
M
minqiyang 已提交
2897
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
2898
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
2899
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
2900
        if var_type == "Parameter":
L
Leo Chen 已提交
2901 2902
            if in_dygraph_mode():
                var = ParamBase(
2903 2904 2905 2906 2907 2908 2909 2910 2911 2912
                    d.shape(),
                    d.dtype(),
                    type=orig_var_type,
                    name=new_name,
                    stop_gradient=stop_gradient,
                    trainable=trainable,
                    optimize_attr=optimize_attr,
                    regularizer=regularizer,
                    error_clip=error_clip)
            else:
L
Leo Chen 已提交
2913 2914
                var = Parameter(
                    self,
2915 2916 2917 2918 2919 2920 2921 2922 2923
                    d.shape(),
                    d.dtype(),
                    type=orig_var_type,
                    name=new_name,
                    stop_gradient=stop_gradient,
                    trainable=trainable,
                    optimize_attr=optimize_attr,
                    regularizer=regularizer,
                    error_clip=error_clip)
T
typhoonzero 已提交
2924
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
2925 2926
            var = Variable(
                self,
T
typhoonzero 已提交
2927
                type=orig_var_type,
T
wip  
typhoonzero 已提交
2928 2929 2930 2931
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
2932
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
2933 2934 2935
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
2936
        self._sync_with_cpp()
2937
        return var
T
typhoonzero 已提交
2938

2939 2940 2941
    def _remove_var(self, name, sync=True):
        if sync == True:
            self._sync_with_cpp()
M
minqiyang 已提交
2942
        self.desc._remove_var(cpt.to_bytes(name))
2943 2944
        del self.vars[name]

Y
Yu Yang 已提交
2945 2946
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
2947
        param = None
L
Leo Chen 已提交
2948
        if in_dygraph_mode():
2949
            param = ParamBase(*args, **kwargs)
L
Leo Chen 已提交
2950 2951
        else:
            param = Parameter(global_block, *args, **kwargs)
2952 2953 2954 2955 2956 2957
            # NOTE: Why only set stop_gradient=False in static mode
            # Because in dygraph mode, the `stop_gradient` and `trainable`
            # are related, and `trainable` default vallue is `True` or
            # it is specified by users, there is no need to set
            # `stop_gradient` for ParamBase here.
            param.stop_gradient = False
2958
        if 'initializer' in kwargs:
2959 2960 2961 2962 2963

            def _is_inited_by(block, var):
                init_ops = []
                for op in block.ops:
                    if var.name in op.output_arg_names:
2964
                        # In startup_program, "c_broadcast" and "c_sync_comm_stream"
T
tangwei12 已提交
2965
                        # are treated as initialization ops that cause error.
2966 2967 2968
                        # Think of "c_broadcast" and "c_sync_comm_stream" as a special case here.
                        if op.type in ["c_broadcast", "c_sync_comm_stream"]:
                            continue
2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979
                        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:
2980
                # TODO already inited, do nothing, should log a warning
2981 2982 2983
                pass
            else:
                initializer(param, self)
Q
Qiao Longfei 已提交
2984
        return param
Y
Yu Yang 已提交
2985

Y
Yu Yang 已提交
2986
    def append_op(self, *args, **kwargs):
2987 2988 2989 2990 2991 2992
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
L
lujun 已提交
2993
        if in_dygraph_mode():
2994
            attrs = kwargs.get("attrs", {})
J
Jiabin Yang 已提交
2995
            type = kwargs.get("type", None)
2996 2997 2998
            op = Operator(
                block=self,
                desc=None,
J
Jiabin Yang 已提交
2999
                type=type,
M
minqiyang 已提交
3000 3001
                inputs=None,
                outputs=None,
3002
                attrs=attrs)
3003

M
minqiyang 已提交
3004 3005 3006
            # record ops in tracer rather than blocks
            #
            # TODO(minqiyang): add op stop_gradient support in static mode too.
L
lujun 已提交
3007
            # currently, we only support stop_gradient in dygraph mode.
J
Jiabin Yang 已提交
3008 3009

            _dygraph_tracer().trace_op(type,
M
minqiyang 已提交
3010
                                       kwargs.get("inputs", {}),
J
Jiabin Yang 已提交
3011 3012
                                       kwargs.get("outputs", {}), attrs
                                       if attrs else {},
M
minqiyang 已提交
3013
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
3014
        else:
3015 3016 3017 3018 3019 3020 3021 3022 3023
            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
3024
            self.ops.append(op)
M
minqiyang 已提交
3025

3026 3027
        return op

W
Wu Yi 已提交
3028
    def _insert_op(self, index, *args, **kwargs):
3029 3030 3031 3032 3033 3034 3035 3036 3037
        """
        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
3038 3039
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
3040 3041 3042 3043
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060
    def _insert_op_without_sync(self, index, *args, **kwargs):
        """
        Insert an Operator according to the giving arguments, 
        without sync_with_cpp to meke the compilation faster.

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

        Returns:
            Operator: the insert Operator.
        """
        op_desc = self.desc._insert_op(index)
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

    def _remove_op(self, index, sync=True):
3061 3062 3063 3064 3065 3066 3067 3068 3069
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
3070 3071
        if sync == True:
            self._sync_with_cpp()
W
Wu Yi 已提交
3072
        self.desc._remove_op(index, index + 1)
3073 3074
        del self.ops[index]

W
Wu Yi 已提交
3075
    def _slice_ops(self, start, end):
3076 3077 3078 3079 3080 3081 3082 3083 3084 3085
        """
        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
3086
        return self.ops[start:end]
Y
Yancey1989 已提交
3087

W
Wu Yi 已提交
3088
    def _prepend_op(self, *args, **kwargs):
L
lujun 已提交
3089
        if in_dygraph_mode():
J
Jiabin Yang 已提交
3090 3091
            type = kwargs.get("type", None)
            attrs = kwargs.get("attrs", {})
3092
            op = Operator(
J
Jiabin Yang 已提交
3093
                self, None, type=type, inputs=None, outputs=None, attrs=attrs)
M
minqiyang 已提交
3094

J
Jiabin Yang 已提交
3095
            _dygraph_tracer().trace_op(type,
M
minqiyang 已提交
3096
                                       kwargs.get("inputs", {}),
J
Jiabin Yang 已提交
3097 3098
                                       kwargs.get("outputs", {}), attrs
                                       if attrs else {},
M
minqiyang 已提交
3099
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
3100
        else:
3101 3102 3103 3104 3105 3106 3107 3108
            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
3109
            self.ops.insert(0, op)
3110

Y
Yu Yang 已提交
3111 3112
        return op

W
Wu Yi 已提交
3113
    def _sync_with_cpp(self):
3114
        """
3115 3116
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
3117
        """
Q
Qiao Longfei 已提交
3118 3119 3120 3121 3122
        # 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())

3123
        # sync variables removed from c++ end
3124
        for var in list(self.vars.keys()):
M
minqiyang 已提交
3125
            if not self.desc.find_var(cpt.to_bytes(var)):
3126 3127
                self.vars.pop(var)

Q
Qiao Longfei 已提交
3128
        # sync operators from cpp
3129 3130 3131 3132
        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
3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148
        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
3149 3150 3151 3152 3153

        # 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
3154
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
3155 3156 3157 3158 3159 3160 3161

        # 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)

3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174
        # 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
3175 3176 3177 3178
        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
3179
    def _copy_param_info_from(self, other):
3180
        """
3181 3182
        Copy the information of parameters from the other block.

3183
        Args:
3184 3185 3186 3187 3188
            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.
3189 3190 3191 3192 3193

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
3194 3195
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
3196
        for p in other.iter_parameters():
3197 3198 3199
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
3200 3201
                # if the Parameter is pruned, v may be None
                continue
3202
            assert isinstance(v, Variable)
3203
            new_p = None
L
Leo Chen 已提交
3204 3205
            if in_dygraph_mode():
                new_p = ParamBase(
3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216
                    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,
                    error_clip=p.error_clip,
                    name=v.name)
            else:
L
Leo Chen 已提交
3217 3218
                new_p = Parameter(
                    block=self,
3219 3220 3221
                    shape=v.shape,
                    dtype=v.dtype,
                    type=v.type,
3222 3223
                    lod_level=v.lod_level
                    if v.type == core.VarDesc.VarType.LOD_TENSOR else None,
3224 3225 3226 3227 3228 3229
                    stop_gradient=p.stop_gradient,
                    trainable=p.trainable,
                    optimize_attr=p.optimize_attr,
                    regularizer=p.regularizer,
                    error_clip=p.error_clip,
                    name=v.name)
3230 3231
            self.vars[new_p.name] = new_p

3232
    def _clone_variable(self, var, force_persistable=True):
3233 3234
        """
        Clone a variable into current block.
3235

3236 3237
        Args:
            var: the variable to be cloned.
3238 3239 3240
            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.
3241 3242

        Returns:
3243
            Variable: the new  variable cloned from 'var' in current block.
3244 3245
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
3246 3247 3248 3249 3250
        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
3251 3252
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
3253
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
3254 3255 3256 3257 3258 3259
        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,
3260
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
3261 3262
                is_data=var.is_data,
                need_check_feed=var.desc.need_check_feed())
T
update  
typhoonzero 已提交
3263 3264 3265 3266 3267 3268 3269
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
3270
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
3271 3272
                is_data=var.is_data,
                need_check_feed=var.desc.need_check_feed())
T
update  
typhoonzero 已提交
3273
        return ret_var
3274

Y
Yu Yang 已提交
3275

3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370
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()

3371
    def remove_input_by_id(self, node_id):
3372 3373 3374 3375 3376 3377
        """
        Remove a node from inputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
3378
        self.node.remove_input(node_id)
3379

3380
    def remove_input(self, node):
3381 3382 3383 3384
        """
        Remove a node from inputs.

        Args:
3385
            node(IrNode): the node being removed.
3386
        """
3387
        self.node.remove_input(node.node)
3388

3389
    def append_input(self, node):
3390 3391 3392 3393
        """
        Append a node in inputs.

        Args:
3394
            node(IrNode): the node being appended.
3395
        """
3396
        self.node.append_input(node.node)
3397 3398 3399 3400 3401 3402 3403 3404

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

3405
    def remove_output_by_id(self, node_id):
3406 3407 3408 3409 3410 3411
        """
        Remove a node from outputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
3412
        self.node.remove_output(node_id)
3413

3414
    def remove_output(self, node):
3415 3416 3417 3418
        """
        Remove a node from outputs.

        Args:
3419
            node(IrNode): the node being removed.
3420
        """
3421
        self.node.remove_output(node.node)
3422

3423
    def append_output(self, node):
3424 3425 3426 3427
        """
        Append a node in outputs.

        Args:
3428
            node(IrNode): the node being appended.
3429
        """
3430
        self.node.append_output(node.node)
3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477

    @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, \
T
tianshuo78520a 已提交
3478
            "The node variable description can not be None."
3479 3480 3481 3482 3483 3484 3485 3486 3487 3488
        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, \
T
tianshuo78520a 已提交
3489
            "The node variable description can not be None."
3490 3491
        return self.node.var().persistable()

3492 3493 3494 3495 3496 3497 3498 3499
    def type(self):
        """
        Return the variable type.

        Returns:
            core.VarDesc.VarType: the variable type.
        """
        assert self.node.var() is not None, \
T
tianshuo78520a 已提交
3500
            "The node variable description can not be None."
3501 3502 3503 3504 3505 3506 3507 3508 3509 3510
        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, \
T
tianshuo78520a 已提交
3511
            "The node variable description can not be None."
3512 3513 3514 3515 3516 3517 3518 3519 3520 3521
        return self.node.var().dtype()

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

        Returns:
            list: the variable shape.
        """
        assert self.node.var() is not None, \
T
tianshuo78520a 已提交
3522
            "The node variable description can not be None."
3523 3524
        return self.node.var().shape()

3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571
    @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, \
T
tianshuo78520a 已提交
3572
            "The node operator description can not be None."
3573 3574
        self.node.op()._rename_input(old_input_name, new_input_name)

3575 3576 3577 3578 3579 3580 3581 3582 3583
    def rename_output(self, old_output_name, new_output_name):
        """
        Rename the output of this node.

        Args:
            old_output_name(str): the old output name.
            new_output_name(str): the new output name.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
3584
            "The node operator description can not be None."
3585 3586
        self.node.op()._rename_output(old_output_name, new_output_name)

3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597
    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, \
T
tianshuo78520a 已提交
3598
            "The node operator description can not be None."
3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611
        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, \
T
tianshuo78520a 已提交
3612
            "The node operator description can not be None."
3613 3614 3615 3616 3617 3618 3619 3620 3621 3622
        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, \
T
tianshuo78520a 已提交
3623
            "The node operator description can not be None."
3624 3625
        return self.node.op().set_type(new_type)

3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640
    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, \
T
tianshuo78520a 已提交
3641
            "The node operator description can not be None."
3642 3643 3644 3645
        desc = self.node.op()
        if isinstance(val, Block):
            desc.set_block_attr(name, val.desc)
        elif isinstance(val, list) and val and \
3646
                all(isinstance(v, Block) for v in val):
3647 3648
            desc.set_blocks_attr(name, [v.desc for v in val])
        elif isinstance(val, core.BlockDesc) or \
3649
                isinstance(val, core.ProgramDesc):
3650 3651 3652 3653
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)

3654 3655 3656 3657 3658 3659 3660 3661
    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, \
T
tianshuo78520a 已提交
3662
            "The node operator description can not be None."
3663 3664 3665 3666 3667 3668 3669 3670 3671 3672
        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, \
T
tianshuo78520a 已提交
3673
            "The node operator description can not be None."
3674 3675
        return self.node.op().output_arg_names()

3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696
    @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]


3697 3698
class IrGraph(object):
    """
3699
    Python IrGraph. Beneath it is a core.Graph, which is used for
3700
    creating a c++ Ir Pass Graph. An IrGraph is just a graph view of
3701 3702
    a Program. In an IrGraph, both Variables and Operators are graph
    nodes.
3703 3704 3705 3706
    """

    def __init__(self, graph, for_test=False):
        """
3707 3708
        Construct an IrGraph using core.Graph.

3709 3710 3711 3712 3713 3714 3715 3716 3717
        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

3718 3719 3720 3721
    def clone(self):
        """
        Create a new and duplicated IrGraph.

3722 3723 3724
        Warns:
            The method only clones the graph structure, not its attributes.

3725 3726 3727
        Returns:
            IrGraph: A new and duplicated graph.
        """
3728
        g = self.graph.clone()
3729 3730
        return IrGraph(g, self._for_test)

3731
    def is_test(self):
3732 3733 3734
        """
        If the graph is used for testing, the function returns true. Otherwise, returns false.
        """
3735 3736
        return self._for_test

W
WangZhen 已提交
3737
    def all_nodes(self):
3738 3739 3740
        """
        Return all nodes included in the graph as a set.
        """
3741
        return {IrNode(node) for node in self.graph.nodes()}
3742

3743
    def all_var_nodes(self):
3744 3745 3746
        """
        Return all variable nodes included in the graph as a set.
        """
3747
        return {IrVarNode(node) for node in self.graph.nodes() if node.is_var()}
3748

3749
    def all_persistable_nodes(self):
3750 3751 3752
        """
        Return all persistable variable nodes included in the graph as a set.
        """
W
WangZhen 已提交
3753 3754 3755 3756 3757
        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)
3758
        return {IrVarNode(p) for p in persistable_nodes}
W
WangZhen 已提交
3759

3760
    def all_op_nodes(self):
3761 3762 3763
        """
        Return all operator nodes included in the graph as a set.
        """
3764
        return {IrOpNode(node) for node in self.graph.nodes() if node.is_op()}
3765

3766
    def create_persistable_node(self, name, var_type, shape, var_dtype):
3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777
        """
        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:
3778
            IrVarNode: the created persistable variable node.
3779
        """
3780 3781 3782 3783 3784
        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)
3785
        return IrVarNode(self.graph.create_var_node(var_desc))
3786 3787

    def create_var_node(self, name, var_type, shape, var_dtype):
3788 3789 3790 3791 3792 3793 3794 3795 3796 3797 3798
        """
        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:
3799
            IrVarNode: the created variable node.
3800 3801
        """

3802 3803 3804 3805
        var_desc = core.VarDesc(name)
        var_desc.set_type(var_type)
        var_desc.set_shape(shape)
        var_desc.set_dtype(var_dtype)
3806
        return IrVarNode(self.graph.create_var_node(var_desc))
3807

3808 3809 3810 3811 3812 3813
    def create_control_dep_var(self):
        """
        create a control var
        """
        return IrVarNode(self.graph.create_control_dep_var())

3814
    def create_var_node_from_desc(self, var_desc):
3815 3816 3817 3818 3819 3820 3821 3822
        """
        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:
3823
            IrVarNode: the created variable node.
3824
        """
3825
        return IrVarNode(self.graph.create_var_node(var_desc))
3826 3827

    def create_op_node(self, op_type, attrs, inputs, outputs):
3828 3829 3830 3831 3832 3833 3834
        """
        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.
T
tianshuo78520a 已提交
3835
            outputs(dict): the outputs of the operator node.
3836 3837

        Returns:
3838
            IrOpNode: the created operator node.
3839
        """
3840 3841
        op_desc = core.OpDesc()
        op_desc.set_type(op_type)
3842
        for attr, value in six.iteritems(attrs):
3843
            self._update_desc_attr(op_desc, attr, value)
3844
        for input_name, var_nodes in six.iteritems(inputs):
3845 3846 3847 3848
            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])
3849
        for output_name, var_nodes in six.iteritems(outputs):
3850 3851 3852 3853
            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])
3854
        return IrOpNode(self.graph.create_op_node(op_desc))
3855 3856

    def create_op_node_from_desc(self, op_desc):
3857 3858 3859 3860 3861 3862 3863
        """
        Create a operator node by using an existing OpDesc in the graph.

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

        Returns:
3864
            IrOpNode: the created operator node.
3865
        """
3866
        return IrOpNode(self.graph.create_op_node(op_desc))
3867 3868

    def update_input_link(self, old_input_node, new_input_node, op_node):
3869 3870 3871 3872
        """
        Update the input's link of a operator node.

        Args:
3873 3874 3875
            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.
3876
        """
3877
        assert old_input_node.node in self.graph.nodes() and new_input_node.node in \
T
tangwei12 已提交
3878
            self.graph.nodes() and op_node.node in self.graph.nodes(), \
3879
            'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.'
3880 3881 3882 3883
        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)
3884
        op_node.rename_input(old_input_node.name(), new_input_node.name())
3885

3886 3887 3888 3889 3890 3891 3892 3893 3894 3895
    def update_output_link(self, old_output_node, new_output_node, op_node):
        """
        Update the output's link of an operator node.

        Args:
            old_output_node(IrNode): the old output node of the giving op_node.
            new_output_node(IrNode): the new output node of the giving op_node.
            op_node(IrOpNode): the operator node that is needed to update input's link.
        """
        assert old_output_node.node in self.graph.nodes() and new_output_node.node in \
T
tangwei12 已提交
3896
            self.graph.nodes() and op_node.node in self.graph.nodes(), \
3897
            'The three arguments(old_output_node &new_output_node &op_node) must be in the graph nodes.'
3898 3899 3900 3901 3902 3903
        old_output_node.remove_input(op_node)
        op_node.remove_output(old_output_node)
        new_output_node.append_input(op_node)
        op_node.append_output(new_output_node)
        op_node.rename_output(old_output_node.name(), new_output_node.name())

3904
    def link_to(self, node_in, node_out):
3905 3906 3907 3908
        """
        Connect two nodes.

        Args:
3909 3910
            node_in(IrNode): the input node.
            node_out(IrNode): the output node.
3911
        """
3912
        assert node_in.node in self.graph.nodes() and node_out.node in self.graph.nodes(), \
W
WangZhen 已提交
3913
            'The two arguments(node_in&node_out) must be in the graph nodes.'
3914 3915
        node_in.append_output(node_out)
        node_out.append_input(node_in)
3916 3917

    def safe_remove_nodes(self, remove_nodes):
3918 3919 3920 3921 3922 3923 3924
        """
        Remove nodes safely since links connected to these removed nodes are
        also removed.

        Args:
            remove_nodes(set): the nodes prepared to be removed.
        """
3925
        if not isinstance(remove_nodes, set):
W
WangZhen 已提交
3926 3927 3928 3929
            if isinstance(remove_nodes, Iterable):
                remove_nodes = set(remove_nodes)
            else:
                remove_nodes = {remove_nodes}
3930 3931
        original_nodes = {n.node for n in remove_nodes}
        core.graph_safe_remove_nodes(self.graph, original_nodes)
3932

Z
Zhen Wang 已提交
3933 3934 3935 3936 3937 3938 3939 3940
    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] = [
3941
                            self._find_node_by_name(node.inputs, each_var_name)
Z
Zhen Wang 已提交
3942 3943 3944 3945
                        ]
                for each_var_name in node.op().output_arg_names():
                    if each_var_name not in var_nodes:
                        var_nodes[each_var_name] = [
3946
                            self._find_node_by_name(node.outputs, each_var_name)
Z
Zhen Wang 已提交
3947 3948 3949
                        ]
                    else:
                        var_nodes[each_var_name].append(
3950 3951
                            self._find_node_by_name(node.outputs,
                                                    each_var_name))
Z
Zhen Wang 已提交
3952 3953
        self.graph.resolve_hazard(var_nodes)

W
WangZhen 已提交
3954
    def has_circle(self):
3955 3956 3957 3958 3959 3960
        """
        Check if the graph has a circle.

        Returns:
            bool: True if the graph has a circle else False.
        """
W
WangZhen 已提交
3961 3962 3963
        return core.has_circle(self.graph)

    def graph_num(self):
3964 3965 3966 3967 3968 3969
        """
        Count the number of unconnected graphs in this graph.

        Returns:
            int: the number of unconnected graphs.
        """
W
WangZhen 已提交
3970 3971 3972
        return core.graph_num(self.graph)

    def topology_sort(self):
3973 3974 3975
        """
        Perform the topology sort operation on the graph.

T
tianshuo78520a 已提交
3976
        Notes: the `graph` can not contain a circle.
3977 3978

        Returns:
Z
Zhen Wang 已提交
3979
            list(IrNode): nodes in topology order.
3980
        """
3981
        ordered_nodes = core.topology_sort(self.graph)
Z
Zhen Wang 已提交
3982
        return [IrNode(n) for n in ordered_nodes]
W
WangZhen 已提交
3983 3984

    def build_adjacency_list(self):
3985 3986 3987 3988
        """
        Build an adjacency list of operations for the `graph`.

        Returns:
3989
            dict{IrNode: set(IrNode)}: the adjacency list.
3990
        """
3991 3992 3993 3994 3995
        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
3996

3997 3998 3999 4000 4001 4002 4003 4004
    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.
4005
            marked_nodes(set(IrNode)): nodes that are needed to be marked.
4006 4007 4008 4009 4010
            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.
        """

4011 4012
        def _convert_to_pdf(dot_file_path):
            pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf'
T
tangwei12 已提交
4013 4014 4015
            exited_code = subprocess.call(
                'dot -Tpdf ' + dot_file_path + ' -o ' + pdf_save_path,
                shell=True)
4016 4017 4018 4019 4020
            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))

4021
        remove_ctr_vars = set()
4022
        if remove_ctr_var:
4023
            for node in self.all_var_nodes():
4024 4025 4026
                if node.is_ctrl_var():
                    remove_ctr_vars.add(node)
            self.safe_remove_nodes(remove_ctr_vars)
4027 4028
        print('Total ops num = {}.'.format(len(self.all_op_nodes())))

4029 4030
        if marked_nodes is not None:
            if not isinstance(marked_nodes, set):
4031 4032 4033 4034 4035 4036
                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}
4037 4038 4039 4040
            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)
4041 4042
        if not os.path.exists(save_path):
            os.makedirs(save_path)
4043 4044 4045 4046 4047 4048 4049
        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):
4050 4051 4052
        """
        Convert the graph into a Program.

Z
Zhen Wang 已提交
4053
        WARN: When the graph includes backward operator nodes, the
4054 4055 4056 4057 4058 4059
        conversion process may be failed. Usually, this function is
        only used to convert a test graph.

        Returns:
            Program: a program converted from the graph.
        """
4060
        convert_pass = core.get_pass('graph_to_program_pass')
4061 4062
        desc = core.ProgramDesc()
        convert_pass.set_not_owned('program', desc)
4063 4064 4065 4066
        convert_pass.apply(self.graph)
        program = Program._construct_from_desc(desc)
        return program

4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077
    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

4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093
    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
4094
class Program(object):
D
dzhwinter 已提交
4095
    """
4096
    Create Python Program.  It has at least one :ref:`api_guide_Block_en`, when the
4097
    control flow op like conditional_block, while :ref:`api_paddle_fluid_layers_While` is included,
J
Jiabin Yang 已提交
4098
    it will contain nested block.
4099

J
Jiabin Yang 已提交
4100 4101 4102
    Please reference the
    `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_
    for details.
D
dzhwinter 已提交
4103

J
Jiabin Yang 已提交
4104
    A set of Program usually contains startup program and main program.
J
Jiabin Yang 已提交
4105
    A startup program is set to contain some initial work, eg. initialize the ``Parameter``, and the main
J
Jiabin Yang 已提交
4106 4107 4108 4109 4110 4111 4112
    program will contain the network structure and vars for train.

    A set of Program can be used for test or train, in train program ,
    Paddle will contain all content to build a train network,  in test
    program Paddle will prune some content which is irrelevant to test, eg.
    backward ops and vars.

J
Jiabin Yang 已提交
4113
    **Notes**:
4114 4115 4116
        **we have** :ref:`api_paddle_fluid_framework_default_startup_program` **and** :ref:`api_paddle_fluid_framework_default_main_program`
        **by default, a pair of them will shared the parameters. The** :ref:`api_paddle_fluid_framework_default_startup_program` **only run once to initialize parameters,**
        :ref:`api_paddle_fluid_framework_default_main_program` **run in every mini batch and adjust the weights.**
D
dzhwinter 已提交
4117 4118

    Returns:
J
Jiabin Yang 已提交
4119
        Program: An empty Program.
D
dzhwinter 已提交
4120 4121

    Examples:
4122 4123
        .. code-block:: python

4124 4125 4126 4127
            import paddle
            import paddle.static as static

            paddle.enable_static()
4128

4129 4130 4131 4132 4133
            main_program = static.Program()
            startup_program = static.Program()
            with static.program_guard(main_program=main_program, startup_program=startup_program):
                x = static.data(name="x", shape=[-1, 784], dtype='float32')
                y = static.data(name="y", shape=[-1, 1], dtype='int32')
4134
                z = static.nn.fc(name="fc", x=x, size=10, activation="relu")
4135 4136 4137

            print("main program is: {}".format(main_program))
            print("start up program is: {}".format(startup_program))
D
dzhwinter 已提交
4138 4139 4140

    """

4141 4142
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
4143 4144
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
4145 4146
        global global_prog_seed
        self._seed = global_prog_seed
Y
yuyang18 已提交
4147
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
4148
        self.__op_role_var = []
T
tangwei12 已提交
4149

4150 4151
        # for distribute training
        # _is_distributed = True if under distributed training
T
tangwei12 已提交
4152
        self._is_distributed = False
4153
        # _is_chief = True if the trainer is the first one, usually No.0
T
tangwei12 已提交
4154
        self._is_chief = False
4155 4156 4157
        # _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
4158
        self._endpoints = []
4159 4160 4161
        # if current role is parameter server, the _ps_endpoint is its "ip:port"
        self._ps_endpoint = None
        # trainers_endpoints, it is used for distribution.
4162
        self._trainers_endpoints = []
4163
        # the distributed lookup table names
T
tangwei12 已提交
4164
        self._distributed_lookup_table = None
4165 4166 4167

        # use Deep gradient comrepssion or not
        self._enable_dgc = False
4168 4169
        self._use_lamb = False

4170 4171 4172
        self._nccl_comm_num = 1
        self._use_hierarchical_allreduce = False
        self._hierarchical_allreduce_inter_nranks = 0
4173

4174 4175 4176
        # 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
4177
        self._program_config = None
4178

H
hutuxian 已提交
4179 4180 4181
        # assigned if this program has been parsed by a pipeline optimizer
        self._pipeline_opt = None

4182 4183 4184
        # appending gradients times
        self._appending_grad_times = 0

4185 4186 4187 4188
        # identifier for auto checkpoint
        self._auto_checkpoint_name = unique_name.generate(
            "__auto_checkpoint_program__")

4189 4190 4191
        # compiled program, i.e. Graph
        self._graph = None

4192 4193 4194 4195 4196 4197 4198 4199 4200 4201
    def global_seed(self, seed=0):
        """
        Set global seed for Program

        Returns:
            None.

        Examples:
            .. code-block:: python

4202 4203
                import paddle
                import paddle.static as static
4204

4205 4206 4207
                paddle.enable_static()

                prog = static.default_main_program()
4208 4209 4210 4211 4212
                print(prog.random_seed)
                ## 0
                ## the default random seed is 0

                prog.global_seed(102)
4213
                prog1 = static.default_main_program()
4214 4215 4216 4217 4218 4219 4220 4221
                print(prog1.random_seed)
                ## 102
                ## the random seed is 102
        """
        global global_prog_seed
        global_prog_seed = seed
        self._seed = global_prog_seed

Y
yuyang18 已提交
4222
    @property
4223
    def _op_role(self):
Y
yuyang18 已提交
4224 4225 4226 4227 4228 4229 4230 4231
        """
        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
4232
        parameter gradient of backward (use :code:`_op_role_var` to get this
Y
yuyang18 已提交
4233 4234 4235 4236
        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
4237 4238
        return self._current_role

4239 4240
    @_op_role.setter
    def _op_role(self, role):
Y
yuyang18 已提交
4241 4242 4243
        self._current_role = role

    @property
4244
    def _op_role_var(self):
Y
yuyang18 已提交
4245
        """
4246
        The auxiliary variables for :code:`_op_role` property.
Y
yuyang18 已提交
4247

4248
        See Also: :code:`Program._op_role`'s documentation for details.
Y
yuyang18 已提交
4249 4250 4251

        Notes: This is a very low-level API. Users should not use it directly.
        """
4252
        return self.__op_role_var
Y
yuyang18 已提交
4253

4254
    @signature_safe_contextmanager
4255 4256 4257 4258 4259
    def _backward_role_guard(self):
        tmp_role = self._current_role

        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Backward
4260 4261 4262 4263
        try:
            yield
        finally:
            self._current_role = tmp_role
4264

S
rename  
sneaxiy 已提交
4265
    @signature_safe_contextmanager
W
Wu Yi 已提交
4266
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
4267 4268 4269 4270 4271 4272 4273
        """
        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:
4274
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
4275 4276 4277

        Examples:

4278
            >>> import paddle.fluid as fluid
Y
yuyang18 已提交
4279
            >>> p, g = backward(...)
W
Wu Yi 已提交
4280
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
4281 4282
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
4283
        tmp_role = self._current_role
4284
        tmp_var = self.__op_role_var
X
Xin Pan 已提交
4285

Y
yuyang18 已提交
4286 4287
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
4288
        self.__op_role_var = [
4289 4290 4291
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
4292 4293 4294 4295 4296
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
Y
Yu Yang 已提交
4297

S
rename  
sneaxiy 已提交
4298
    @signature_safe_contextmanager
X
Xin Pan 已提交
4299
    def _lr_schedule_guard(self, is_with_opt=False):
4300 4301 4302 4303 4304 4305 4306
        """
        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
4307 4308 4309 4310
        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.
4311 4312 4313

        Examples:

4314
            >>> import paddle.fluid as fluid
4315 4316 4317 4318
            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
4319 4320

        tmp_role = self._current_role
4321
        tmp_var = self.__op_role_var
4322

4323 4324
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
4325 4326
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
4327
        # TODO(typhoonzero): how to set target learning rate var
4328
        self.__op_role_var = []
4329 4330 4331 4332 4333
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
4334

4335
    def __str__(self):
Y
yuyang18 已提交
4336 4337 4338 4339 4340 4341 4342 4343 4344
        """
        Get the protobuf debug string of this Program.

        Returns:
            (str): The protobuf debug string.

        Raises:
            ValueError: If any of required fields is not set.
        """
4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364
        return self._to_readable_code()

    def _to_readable_code(self, skip_op_callstack=True):
        """
        Get readable debug string of Program.

        .. note::
            If you want to get the debug string in protobuf format,
            please use :code:`to_string` method.

        Args:
            skip_op_callstack(bool): whether to skip parsing Operator's attribute
                op_callstack, default value is True

        Returns:
            string: The formatted Program string.

        Examples:
            .. code-block:: python

4365 4366
            import paddle
            import paddle.static as static
4367

4368 4369 4370
            paddle.enable_static()

            cur_program = static.Program()
4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386
            cur_block = cur_program.current_block()
            new_var = cur_block.create_var(name="X",
                                           shape=[-1, 23, 48],
                                           dtype='float32')
            new_op = cur_block.append_op(type="abs",
                                inputs={"X": [new_var]},
                                outputs={"Out": [new_var]})
            print(cur_program._to_readable_code())
        """
        assert isinstance(
            skip_op_callstack, bool
        ), "skip_op_callstack parameter's type is error, expect bool, received %s".format(
            type(skip_op_callstack))
        program_str = ""
        for block in self.blocks:
            program_str += block._to_readable_code(skip_op_callstack)
4387
            program_str += '\n'
4388
        return program_str
Y
Yang Yang(Tony) 已提交
4389

F
fengjiayi 已提交
4390 4391 4392
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
4393

J
Jiabin Yang 已提交
4394 4395 4396
        Args:

            throw_on_error (bool): raise Value error when any of required fields is not set.
F
fengjiayi 已提交
4397

J
Jiabin Yang 已提交
4398
            with_details (bool): True if more details about variables and parameters, e.g., :code:`trainable`, :code:`optimize_attr`, need to print.
Y
yuyang18 已提交
4399

H
haowang101779990 已提交
4400
        Returns:
J
Jiabin Yang 已提交
4401
            str: The debug string describe current Program.
Y
yuyang18 已提交
4402 4403

        Raises:
J
Jiabin Yang 已提交
4404
            ValueError: If any of required fields is not set and throw_on_error is True.
F
fengjiayi 已提交
4405

4406 4407 4408
        Examples:
            .. code-block:: python

4409 4410 4411 4412
                import paddle
                import paddle.static as static

                paddle.enable_static()
4413

4414 4415 4416
                prog = static.default_main_program()
                x = static.data(name="X", shape=[2,3], dtype="float32")
                pred = static.nn.fc(x, size=3)
4417
                prog_string = prog.to_string(throw_on_error=True, with_details=False)
4418
                prog_string_with_details = prog.to_string(throw_on_error=False, with_details=True)
T
tianshuo78520a 已提交
4419
                print("program string without detail: {}".format(prog_string))
4420
                print("program string with detail: {}".format(prog_string_with_details))
F
fengjiayi 已提交
4421
        """
4422 4423 4424 4425 4426 4427 4428 4429 4430
        assert isinstance(
            throw_on_error, bool
        ), "The type of throw_on_error parameter is wrong, expected bool, but received {}.".format(
            type(throw_on_error))
        assert isinstance(
            with_details, bool
        ), "The type of with_details parameter is wrong, expected bool, but received {}.".format(
            type(with_details))

F
fengjiayi 已提交
4431 4432 4433 4434 4435 4436
        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()
4437 4438
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
4439 4440
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
4441

W
Wu Yi 已提交
4442
    def _get_desc(self):
Y
yuyang18 已提交
4443 4444 4445 4446 4447 4448 4449
        """
        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.
        """
4450 4451
        return self.desc

X
version  
Xin Pan 已提交
4452 4453 4454
    def _version(self):
        return self.desc._version()

4455
    def clone(self, for_test=False):
Y
yuyang18 已提交
4456
        """
4457 4458 4459 4460
        .. note:::
            1. :code:`Program.clone()` method DOES NOT clone :ref:`api_paddle_io_DataLoader` . 
            2. Recommend you to use :code:`clone` before using :code:`Opimizer.minimize` . 
            3. This API has no effect in Dygraph Mode.
Y
yuyang18 已提交
4461

4462
        Create a new Program with forward content of original one when ``for_test=True``.
4463
        Create a new Program as same as the original one when ``for_test=False``.
4464

4465
        Some operators, e.g., :ref:`api_paddle_fluid_layers_batch_norm` , behave differently between
Y
yuyang18 已提交
4466 4467 4468
        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`.
4469

4470 4471
        * Set for_test to False when you want to clone the program for training.
        * Set for_test to True when you want to clone the program for testing.
4472 4473
          We will prune the backward and optimize part of the program when you
          use :code:`clone` after :code:`Opimizer.minimize`, but we still
J
Jiabin Yang 已提交
4474
          recommend you to use :code:`clone` before using :code:`Opimizer.minimize`.
Y
yuyang18 已提交
4475

J
Jiabin Yang 已提交
4476
        For Example:
4477
          ::
L
Luo Tao 已提交
4478

4479 4480 4481 4482 4483 4484
            import paddle
            import paddle.static as static

            paddle.enable_static()

            img = static.data(name='image', shape=[None, 784])
4485
            pred = static.nn.fc(x=img, size=10, actvation='relu')
4486
            loss = paddle.mean(pred)
4487
            # Here we use clone before Momentum
4488 4489
            test_program = static.default_main_program().clone(for_test=True)
            optimizer = paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
4490
            optimizer.minimize(loss)
4491

J
Jiabin Yang 已提交
4492
        Args:
4493

4494 4495
            for_test (bool): True if change the :code:`is_test` attribute of operators to :code:`True`
                and prune the backward and optimize part of the program. The default value is :code:`False` .
4496

J
Jiabin Yang 已提交
4497
        Returns:
4498
            Program: A new Program with forward content of original one when ``for_test=True``.  A new Program as same as the original one when ``for_test=False``
4499

Y
yuyang18 已提交
4500 4501 4502

        Examples:

4503 4504 4505 4506 4507 4508 4509
            .. note::
                The Program's order maybe different after :code:`clone` and
                this will not affect your training or testing progress. In the following
                example we give you an simple method :code:`print_prog(program)` to
                print Program Descs inorder to make sure you have same print result
                after :code:`clone`:

4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525
            .. code-block:: python

                import six

                def print_prog(prog):
                    for name, value in sorted(six.iteritems(prog.block(0).vars)):
                        print(value)
                    for op in prog.block(0).ops:
                        print("op type is {}".format(op.type))
                        print("op inputs are {}".format(op.input_arg_names))
                        print("op outputs are {}".format(op.output_arg_names))
                        for key, value in sorted(six.iteritems(op.all_attrs())):
                            if key not in ['op_callstack', 'op_role_var']:
                                print(" [ attrs: {}:   {} ]".format(key, value))


4526
            1. To clone a test program, the sample code is:
4527 4528 4529
                .. code-block:: python

                    import six
4530 4531 4532 4533 4534 4535
                    import paddle
                    import paddle.static as static
                    import paddle.utils as utils
                    import paddle.nn.functional as F

                    paddle.enable_static()
4536 4537 4538 4539 4540 4541 4542 4543 4544 4545 4546 4547

                    def print_prog(prog):
                        for name, value in sorted(six.iteritems(prog.block(0).vars)):
                            print(value)
                        for op in prog.block(0).ops:
                            print("op type is {}".format(op.type))
                            print("op inputs are {}".format(op.input_arg_names))
                            print("op outputs are {}".format(op.output_arg_names))
                            for key, value in sorted(six.iteritems(op.all_attrs())):
                                if key not in ['op_callstack', 'op_role_var']:
                                    print(" [ attrs: {}:   {} ]".format(key, value))

4548 4549
                    train_program = static.Program()
                    startup_program = static.Program()
J
Jiabin Yang 已提交
4550 4551 4552

                    # startup_program is used to do some parameter init work,
                    # and main program is used to hold the network
4553 4554 4555
                    with static.program_guard(train_program, startup_program):
                        with utils.unique_name.guard():
                            img = static.data(name='image', shape=[None, 784])
4556
                            hidden = static.nn.fc(x=img, size=200, activation='relu')
4557 4558
                            hidden = F.dropout(hidden, p=0.5)
                            loss = F.cross_entropy(
4559
                                input=static.nn.fc(x=hidden, size=10, activation='softmax'),
4560 4561
                                label=static.data(name='label', shape=[1], dtype='int64'))
                            avg_loss = paddle.mean(loss)
4562
                            test_program = train_program.clone(for_test=True)
4563
                    print_prog(test_program)
J
Jiabin Yang 已提交
4564 4565 4566 4567

                    # Due to parameter sharing usage for train and test, so we need to use startup program of train
                    # instead of using test startup program, while nothing is in test's startup program

4568
                    # In Paddle we will share weights by using the same Tensor name. In train and test program
J
Jiabin Yang 已提交
4569 4570 4571 4572
                    # all parameters will have the same name and this can make train and test program sharing parameters,
                    # that's why we need to use startup program of train. And for startup program of test, it has nothing,
                    # since it is a new program.

4573 4574 4575
                    with static.program_guard(train_program, startup_program):
                        with utils.unique_name.guard():
                            sgd = paddle.optimizer.SGD(learning_rate=1e-3)
4576 4577 4578
                            sgd.minimize(avg_loss)


4579
            2. The clone method can be avoid if you create program for training and program for testing individually.
4580 4581 4582
                .. code-block:: python

                    import six
4583 4584 4585 4586 4587 4588
                    import paddle
                    import paddle.static as static
                    import paddle.utils as utils
                    import paddle.nn.functional as F

                    paddle.enable_static()
4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599

                    def print_prog(prog):
                        for name, value in sorted(six.iteritems(prog.block(0).vars)):
                            print(value)
                        for op in prog.block(0).ops:
                            print("op type is {}".format(op.type))
                            print("op inputs are {}".format(op.input_arg_names))
                            print("op outputs are {}".format(op.output_arg_names))
                            for key, value in sorted(six.iteritems(op.all_attrs())):
                                if key not in ['op_callstack', 'op_role_var']:
                                    print(" [ attrs: {}:   {} ]".format(key, value))
4600

4601
                    def network():
4602
                        img = static.data(name='image', shape=[None, 784])
4603
                        hidden = static.nn.fc(x=img, size=200, activation='relu')
4604 4605
                        hidden = F.dropout(hidden, p=0.5)
                        loss = F.cross_entropy(
4606
                            input=static.nn.fc(x=hidden, size=10, activation='softmax'),
4607 4608
                            label=static.data(name='label', shape=[1], dtype='int64'))
                        avg_loss = paddle.mean(loss)
4609 4610
                        return avg_loss

4611 4612 4613 4614 4615
                    train_program_2 = static.Program()
                    startup_program_2 = static.Program()
                    test_program_2 = static.Program()
                    with static.program_guard(train_program_2, startup_program_2):
                        with utils.unique_name.guard():
4616
                            avg_loss = network()
4617
                            sgd = paddle.optimizer.SGD(learning_rate=1e-3)
4618
                            sgd.minimize(avg_loss)
4619
                    # the test startup program is not used.
4620 4621
                    with static.program_guard(test_program_2, startup_program_2):
                        with utils.unique_name.guard():
4622 4623
                            avg_loss = network()
                    print_prog(test_program_2)
4624

4625
            The two code snippets above will generate and print same programs.
4626
        """
4627

T
tangwei12 已提交
4628
        # NOTE(zhiqiu): we sync the original program first, since its program may diff with
4629 4630 4631
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

4632
        pruned_origin_block_id_map = None
4633
        if for_test:
4634 4635 4636 4637 4638 4639 4640 4641 4642
            forward_prog = Program()
            forward_prog.desc, pruned_origin_block_id_map = core.prune_backward(
                self.desc)
            forward_prog.blocks = [
                Block(forward_prog, i)
                for i in six.moves.range(forward_prog.desc.num_blocks())
            ]
            forward_prog._sync_with_cpp()
            p = forward_prog._inference_optimize(prune_read_op=False)
4643
        else:
4644
            p = Program()
G
gongweibao 已提交
4645 4646
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
4647
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
4648 4649 4650
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
4651 4652

            p._current_role = self._current_role
4653
            p.__op_role_var = self.__op_role_var
4654
            p._appending_grad_times = self._appending_grad_times
4655 4656
            if hasattr(self, 'lr_sheduler'):
                p.lr_sheduler = self.lr_sheduler
G
gongweibao 已提交
4657

T
tangwei12 已提交
4658
            # NOTE(zhiqiu): we sync the cloned program, to update its program by
4659
            # its desc.
W
Wu Yi 已提交
4660
            p._sync_with_cpp()
4661

W
Wu Yi 已提交
4662
        p._copy_param_info_from(self)
4663
        p._copy_data_info_from(self, pruned_origin_block_id_map)
4664
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
4665
        return p
4666

4667
    def _prune(self, targets):
Y
yuyang18 已提交
4668 4669 4670 4671 4672 4673 4674 4675
        """
        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:
4676
            targets(list|Variable|Operator): A list of variables, operators, or variable names
Y
yuyang18 已提交
4677 4678 4679 4680
                need to be pruned

        Returns:
            Program:  A new, pruned program.
4681
        """
4682
        return self._prune_with_input([], targets)
4683 4684

    def _prune_with_input(self, feeded_var_names, targets):
Y
yuyang18 已提交
4685
        """
4686 4687 4688 4689 4690 4691 4692 4693 4694 4695
        Prune operators and variables which are not needed to generate
        :code:`targets`. Prune operators and variables which are needed 
        to generate feeded_var 

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

        Args:
            feeded_var_names(list|str): A list of variable names from where
                pruning start. If it is set as [], this API works just like _prune()
4696
            targets(list|Variable|Operator): A list of variables, operators, or variable names
4697 4698 4699 4700 4701 4702
                need to be pruned

        Returns:
            Program:  A new, pruned program.
        """

T
tangwei12 已提交
4703
        # NOTE(zhiqiu): we sync the original program first, since its program may diff with
4704 4705 4706
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

4707 4708
        if not isinstance(feeded_var_names, list):
            feeded_var_names = [feeded_var_names]
4709 4710
        if not isinstance(targets, list):
            targets = [targets]
4711 4712 4713

        for var in feeded_var_names:
            if not isinstance(var, six.string_types):
4714 4715 4716
                raise ValueError(
                    "All feeded_var_names of Program._prune_with_input() can only be "
                    "str, but received %s." % type(var))
4717

4718 4719 4720 4721
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
4722 4723 4724
                    name = t.name
                elif isinstance(t, six.string_types):
                    name = str(t)
4725
                else:
4726 4727 4728
                    raise ValueError(
                        "All targets of Program._prune_with_input() can only be "
                        "Variable or Operator, but received %s." % type(t))
4729 4730 4731 4732 4733 4734 4735 4736

                # NOTEZ(zhiqiu): For variable to be fed in fetch_list, there two cases:
                # (1) the variable is leaf, it has no op that generates it;
                # (2) the variable is not leaf, and we need to prune the op that generates it.
                # In both cases, wo can just skip target_op of that it.
                if name in feeded_var_names:
                    continue

4737 4738 4739 4740 4741 4742 4743 4744 4745
                # After transpiler processing, the op that output this
                # variable maybe has been changed, so t.op is not reliable
                # and we need to find the current op that generate this
                # variable here.
                target_op = None
                global_block = self.global_block()
                for idx, op in enumerate(global_block.ops):
                    if name in op.output_arg_names:
                        # NOTE(zhiqiu): Find op that generate target name.
T
tangwei12 已提交
4746
                        # Skip optimize op except for optimize op in targets,
4747 4748 4749 4750 4751 4752
                        # since optimize op generates parameters.
                        if op._is_optimize_op() and op not in targets:
                            continue
                        else:
                            target_op = op
                            break
4753 4754 4755 4756 4757 4758 4759 4760
                if target_op is None:
                    raise ValueError(
                        "The target variable used for pruning should have an "
                        "associated operator that generates it.")
                else:
                    targets_idx.append([target_op.block.idx, target_op.idx])
            else:
                targets_idx.append([t.block.idx, t.idx])
4761

4762
        res = Program()
4763 4764 4765
        res.desc, pruned_origin_block_id_map = core.prune(self.desc,
                                                          set(feeded_var_names),
                                                          targets_idx)
M
minqiyang 已提交
4766 4767 4768
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
4769
        res._sync_with_cpp()
4770 4771 4772 4773 4774

        res._copy_param_info_from(self)
        res._copy_data_info_from(self, pruned_origin_block_id_map)
        res._copy_dist_param_info_from(self)

4775 4776
        return res

X
Xin Pan 已提交
4777
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
4778
        """
F
fengjiayi 已提交
4779 4780 4781 4782 4783
        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.

4784
        3. change the :code:`is_test`
Y
yuyang18 已提交
4785 4786 4787
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

4788
        Args:
X
Xin Pan 已提交
4789 4790
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
4791

Y
yuyang18 已提交
4792 4793 4794 4795 4796 4797
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
4798
        res = Program()
4799
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
4800 4801 4802 4803

        # 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
4804
        if prune_read_op:
4805 4806 4807 4808 4809 4810 4811 4812 4813
            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
4814
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
4815 4816

        # change all `is_test` attributes to True
M
minqiyang 已提交
4817
        for i in six.moves.range(res.desc.num_blocks()):
4818
            block = res.desc.block(i)
M
minqiyang 已提交
4819
            for j in six.moves.range(block.op_size()):
4820 4821
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
4822
                    op._set_attr('is_test', True)
M
minqiyang 已提交
4823 4824 4825
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
4826
        res._sync_with_cpp()
4827 4828
        return res

4829 4830
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
4831
        """
4832 4833 4834
        .. note::
            1. All information about parameters will be lost after serialization; 
            2. This API has no effect in Dygraph mode.
Y
yuyang18 已提交
4835

4836 4837
        Deserialize a Program from  `protobuf <https://en.wikipedia.org/wiki/Protocol_Buffers>`_  binary string.
        This method always use to save and load model
Y
yuyang18 已提交
4838

J
Jiabin Yang 已提交
4839
        Args:
Y
yuyang18 已提交
4840

J
Jiabin Yang 已提交
4841
            binary_str_type (str): the binary prootbuf string.
4842

J
Jiabin Yang 已提交
4843 4844
        Returns:
            Program: A deserialized Program.
4845 4846 4847 4848

        Examples:
            .. code-block:: python

4849 4850 4851 4852
                import paddle
                import paddle.static as static

                paddle.enable_static()
4853

4854 4855 4856 4857
                startup_prog = static.Program()
                main_prog = static.Program()
                with static.program_guard(startup_prog, main_prog):
                    x = static.data(name='X', shape=[1000, 784], dtype='float32')
4858

4859
                    y = static.data(name='Y', shape=[784, 100], dtype='float32')
4860

4861
                    z = paddle.matmul(x=x, y=y)
4862

4863 4864
                    binary_str = static.default_main_program().desc.serialize_to_string()
                    prog_restored = static.default_main_program().parse_from_string(binary_str)
4865

4866
                    print(static.default_main_program())
4867
                    print(prog_restored)
Y
yuyang18 已提交
4868
        """
4869 4870
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
4871
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
4872
        p._sync_with_cpp()
4873
        return p
Y
Yu Yang 已提交
4874

4875
    @staticmethod
4876
    def _construct_from_desc(desc):
4877 4878 4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891
        """
        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
4892 4893
    @property
    def random_seed(self):
Y
yuyang18 已提交
4894
        """
J
Jiabin Yang 已提交
4895
        The default random seed for random operators in Program. ``0`` means get
Y
yuyang18 已提交
4896 4897
        the random seed from random device.

4898 4899
        .. note:: 
            It must be set before the operators have been added.
J
Jiabin Yang 已提交
4900 4901 4902

        Returns:
            int64: Random seed in current Program
4903

4904 4905 4906 4907

        Examples:
            .. code-block:: python

4908 4909 4910
                import paddle
                import paddle.static as static
                import paddle.nn.functional as F
4911

4912 4913 4914
                paddle.enable_static()

                prog = static.default_main_program()
4915
                random_seed = prog.random_seed
4916
                x_var = static.data(name="X", shape=[3,3], dtype="float32")
4917 4918 4919
                print(random_seed)
                ## 0
                ## the default random seed is 0
4920

4921
                # Here we need to set random seed before we use paddle.nn.functional.dropout
4922
                prog.random_seed = 1
4923
                z_var = F.dropout(x_var, 0.7)
4924

4925
                print(prog.random_seed)
4926 4927
                ## 1
                ## the random seed is change to 1
Y
yuyang18 已提交
4928
        """
D
dzhwinter 已提交
4929 4930
        return self._seed

Q
qiaolongfei 已提交
4931 4932
    @property
    def num_blocks(self):
Y
yuyang18 已提交
4933
        """
4934 4935
        The number of :ref:`api_guide_Block_en`  in this Program.

4936 4937
        .. note:: 
            This API has no effect in Dygraph mode.
J
Jiabin Yang 已提交
4938 4939 4940

        Returns:
            int(Platform-dependent size): num of :ref:`api_guide_Block_en`  in current Program
4941

4942 4943 4944 4945

        Examples:
            .. code-block:: python

4946 4947 4948 4949
                import paddle
                import paddle.static as static

                paddle.enable_static()
4950

4951
                prog = static.default_main_program()
4952 4953
                num_blocks = prog.num_blocks
                print(num_blocks)
4954

4955 4956
                # print result:
                # 1
Y
yuyang18 已提交
4957
        """
Q
qiaolongfei 已提交
4958 4959
        return self.desc.num_blocks()

D
dzhwinter 已提交
4960 4961 4962
    @random_seed.setter
    def random_seed(self, seed):
        if not isinstance(seed, int):
4963 4964 4965
            raise ValueError(
                "Program.random_seed's input seed must be an integer, but received %s."
                % type(seed))
D
dzhwinter 已提交
4966 4967
        self._seed = seed

Y
Yu Yang 已提交
4968
    def __repr__(self):
4969
        return self.__str__()
4970

Y
Yu Yang 已提交
4971
    def global_block(self):
Y
yuyang18 已提交
4972
        """
4973 4974
        .. note::
            This API has no effect in Dygraph mode.
4975 4976 4977

        Get the first :ref:`api_guide_Block_en` of this Program.

J
Jiabin Yang 已提交
4978 4979
        Returns:
            :ref:`api_guide_Block_en`: The first :ref:`api_guide_Block_en`  of this Program.
4980

4981 4982 4983 4984

        Examples:
            .. code-block:: python

4985 4986 4987 4988
                import paddle
                import paddle.static as static

                paddle.enable_static()
4989

4990
                prog = static.default_main_program()
4991 4992
                gb_block = prog.global_block()
                print(gb_block)
4993

Y
yuyang18 已提交
4994
        """
Y
Yu Yang 已提交
4995 4996
        return self.blocks[0]

Q
Qiao Longfei 已提交
4997
    def block(self, index):
Y
yuyang18 已提交
4998
        """
4999 5000
        .. note::
            This API has no effect in Dygraph mode.
Y
yuyang18 已提交
5001

5002 5003
        Get the :code:`index`  :ref:`api_guide_Block_en`  of this Program

J
Jiabin Yang 已提交
5004 5005
        Args:
            index (int) - The index of  :ref:`api_guide_Block_en`  to get
5006

J
Jiabin Yang 已提交
5007 5008
        Returns:
            :ref:`api_guide_Block_en`: The :code:`index` block
5009 5010 5011 5012

        Examples:
            .. code-block:: python

5013 5014 5015 5016
                import paddle
                import paddle.static as static

                paddle.enable_static()
5017

5018
                prog = static.default_main_program()
5019 5020
                block_0 = prog.block(0)
                print(block_0)
Y
yuyang18 已提交
5021
        """
Q
Qiao Longfei 已提交
5022 5023
        return self.blocks[index]

Y
Yu Yang 已提交
5024
    def current_block(self):
Y
yuyang18 已提交
5025
        """
5026 5027
        .. note::
            This API has no effect in Dygraph mode.
5028

J
Jiabin Yang 已提交
5029 5030
        Get the current  :ref:`api_guide_Block_en` . The :code:`current`  :ref:`api_guide_Block_en`
        is the  :ref:`api_guide_Block_en`  to append operators.
5031

J
Jiabin Yang 已提交
5032 5033
        Returns:
             :ref:`api_guide_Block_en`: The :code:`index`  :ref:`api_guide_Block_en`
5034

5035 5036 5037
        Examples:
            .. code-block:: python

5038 5039 5040 5041
                import paddle
                import paddle.static as static

                paddle.enable_static()
5042

5043
                prog = static.default_main_program()
5044 5045
                current_blk = prog.current_block()
                print(current_blk)
Y
yuyang18 已提交
5046
        """
Y
Yu Yang 已提交
5047 5048
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
5049
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
5050 5051 5052 5053 5054
        """
        Create a new block with the :code:`parent_idx` and change the current block
        to new block.

        Args:
J
Jiabin Yang 已提交
5055

Y
yuyang18 已提交
5056 5057 5058 5059 5060
            parent_idx(int): The parent block index.

        Returns:
            Block: The new block.
        """
Y
Yu Yang 已提交
5061
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
5062 5063 5064
        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
5065 5066 5067 5068
        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
5069
    def _rollback(self):
Y
yuyang18 已提交
5070 5071 5072 5073 5074
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
5075 5076
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
5077
    def _sync_with_cpp(self):
Y
yuyang18 已提交
5078 5079 5080 5081 5082 5083 5084 5085 5086 5087
        """
        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
5088 5089 5090
        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
5091
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
5092

W
Wu Yi 已提交
5093
    def _copy_param_info_from(self, other):
5094
        """
5095
        Copy the information of parameters from other program.
D
dzhwinter 已提交
5096

Y
yuyang18 已提交
5097 5098 5099
        Notes: This is a very low level API. Users should not invoke it
        directly.

5100 5101 5102 5103 5104 5105 5106
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
5107 5108 5109
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
5110

W
Wu Yi 已提交
5111
        self.global_block()._copy_param_info_from(other.global_block())
5112

5113 5114 5115 5116 5117 5118 5119 5120 5121 5122 5123
    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):
5124 5125 5126
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
5127 5128
        self._is_distributed = other._is_distributed
        self._is_chief = other._is_chief
5129
        self._parameters_on_pservers = other._parameters_on_pservers
5130
        self._endpoints = other._endpoints
5131
        self._ps_endpoint = other._ps_endpoint
5132 5133
        self._distributed_lookup_table = other._distributed_lookup_table

5134
    def _copy_data_info_from(self, other, pruned_origin_block_id_map=None):
F
fengjiayi 已提交
5135 5136
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
5137

Y
yuyang18 已提交
5138 5139 5140
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
5141 5142
        Args:
            other(Program): Other program
5143 5144 5145 5146
            pruned_origin_block_id_map(dict{int:int}): A dict which maps the block id in program
            self to the block id in program other. For example, {0:0, 1:1, 2:3} means block 0 in self is 
            cloned from block 0 in other, etc. Default is None, which means default mapped, 
            {0:0, 1:1,..., n:n}.
F
fengjiayi 已提交
5147 5148 5149 5150 5151

        Returns:
            None
        """
        if not isinstance(other, Program):
5152 5153 5154
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
F
fengjiayi 已提交
5155

5156 5157 5158 5159 5160
        if not pruned_origin_block_id_map:
            pruned_origin_block_id_map = {
                i: i
                for i in six.moves.range(self.desc.num_blocks())
            }
5161 5162 5163

        # NOTE(zhiqiu): All vars in cloned program exist in original program.
        # The reverse is not true, due to backward pruning.
5164 5165
        for i, block in enumerate(self.blocks):
            other_block = other.blocks[pruned_origin_block_id_map[i]]
5166
            for var in list(block.vars.values()):
5167 5168 5169 5170 5171 5172 5173
                other_var = other_block.var(var.name)
                if other_var.is_data:
                    var.is_data = True
                if other_var.desc.need_check_feed():
                    var.desc.set_need_check_feed(True)
                if other_var.stop_gradient:
                    var.stop_gradient = True
F
fengjiayi 已提交
5174

5175
    def list_vars(self):
Y
yuyang18 已提交
5176
        """
5177
        Get all Tensors from this Program. A iterable object is returned.
Y
yuyang18 已提交
5178

J
Jiabin Yang 已提交
5179
        Returns:
5180
            iterable Tensors: The Generator will yield every Tensor in this program.
5181 5182 5183 5184

        Examples:
            .. code-block:: python

5185 5186
                import paddle
                import paddle.static as static
5187

5188 5189 5190 5191 5192
                paddle.enable_static()

                prog = static.default_main_program()
                img = static.data(name='img', shape=[None, 1,28,28], dtype='float32')
                label = static.data(name='label', shape=[None,1], dtype='int64')
5193 5194
                for var in prog.list_vars():
                    print(var)
T
tangwei12 已提交
5195

5196 5197
                # var img : paddle.VarType.LOD_TENSOR.shape(-1, 1, 28, 28).astype(VarType.FP32)
                # var label : paddle.VarType.LOD_TENSOR.shape(-1, 1).astype(VarType.INT64)
Y
yuyang18 已提交
5198
        """
5199
        for each_block in self.blocks:
5200
            for each_var in list(each_block.vars.values()):
5201 5202
                yield each_var

5203 5204 5205 5206 5207 5208 5209 5210 5211 5212
    def all_parameters(self):
        """
        Get all :ref:`api_guide_parameter_en` from this Program. A list object is returned.

        Returns:
            list[ :ref:`api_guide_parameter_en` ]: The list contians all parameters in this program.

        Examples:
            .. code-block:: python

5213 5214 5215 5216
                import paddle
                import paddle.static as static

                paddle.enable_static()
5217

5218 5219
                program = static.default_main_program()
                data = static.data(name='x', shape=[None, 13], dtype='float32')
5220
                hidden = static.nn.fc(x=data, size=10)
5221 5222
                loss = paddle.mean(hidden)
                paddle.optimizer.SGD(learning_rate=0.01).minimize(loss)
5223 5224 5225 5226 5227 5228 5229

                for param in program.all_parameters():
                    print(param)

                # Here will print all parameters in current program, in this example,
                # the result is like:
                #
5230 5231
                # persist trainable param fc_0.w_0 : paddle.VarType.LOD_TENSOR.shape(13, 10).astype(VarType.FP32)
                # persist trainable param fc_0.b_0 : paddle.VarType.LOD_TENSOR.shape(10,).astype(VarType.FP32)
5232 5233 5234 5235 5236 5237 5238 5239 5240 5241
                #
                # Here print(param) will print out all the properties of a parameter,
                # including name, type and persistable, you can access to specific
                # property of a parameter, such as param.name, param.type
        """
        parameters = []
        for each_block in self.blocks:
            parameters.extend(each_block.all_parameters())
        return parameters

Y
Yu Yang 已提交
5242

5243
@six.add_metaclass(ParameterMetaClass)
Y
Yu Yang 已提交
5244
class Parameter(Variable):
5245
    """
5246
    Parameter is derived from Variable. A parameter is a persistable
5247
    Variable, and will be updated by optimizers after each iteration.
5248
    The training of a neural network is essentially the updating of
5249 5250
    its parameters.

5251
    Relative to a general Variable, a Parameter has several its own
5252 5253
    member variables:

5254 5255 5256 5257 5258 5259 5260 5261 5262 5263
    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
        do_model_average(bool): True if the model average strategy will
            be applied on this parameter.
5264 5265
        need_clip (bool): Whether the parameter gradient need to be cliped 
            in optimizer. Default is True.
5266 5267
    """

5268 5269 5270 5271 5272 5273
    def __init__(self,
                 block,
                 shape,
                 dtype,
                 type=core.VarDesc.VarType.LOD_TENSOR,
                 **kwargs):
5274 5275 5276 5277 5278
        if shape is None:
            raise ValueError("The shape of Parameter should not be None")
        if dtype is None:
            raise ValueError("The dtype of Parameter should not be None")

Y
Yu Yang 已提交
5279
        if len(shape) == 0:
5280 5281
            raise ValueError(
                "The dimensions of shape for Parameter must be greater than 0")
Y
Yu Yang 已提交
5282 5283 5284

        for each in shape:
            if each < 0:
5285 5286 5287
                raise ValueError(
                    "Each dimension of shape for Parameter must be greater than 0, but received %s"
                    % list(shape))
5288 5289

        Variable.__init__(
5290 5291 5292 5293 5294 5295 5296
            self,
            block,
            persistable=True,
            shape=shape,
            dtype=dtype,
            type=type,
            **kwargs)
Y
Yu Yang 已提交
5297 5298 5299 5300
        self.trainable = kwargs.get('trainable', True)

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

5301 5302
        self.regularizer = kwargs.get('regularizer', None)

W
wanghaoshuang 已提交
5303
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
5304

5305 5306
        self.need_clip = kwargs.get('need_clip', True)

5307 5308
        self.is_distributed = False

F
fengjiayi 已提交
5309
    def __str__(self):
5310
        return self._to_readable_code()
F
fengjiayi 已提交
5311

F
update  
fengjiayi 已提交
5312 5313 5314
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
5315

F
update  
fengjiayi 已提交
5316 5317 5318 5319 5320 5321 5322 5323
        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.

5324 5325 5326 5327 5328 5329 5330 5331 5332
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
                rlt = fluid.layers.data("fake_data", shape=[1,1], dtype='float32')
                debug_str = prog.to_string(throw_on_error=True, with_details=False)
                print(debug_str)
F
update  
fengjiayi 已提交
5333 5334 5335 5336 5337 5338
        """
        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",
5339
                               "do_model_average", "need_clip")
F
update  
fengjiayi 已提交
5340
            for attr_name in additional_attr:
5341 5342
                res_str += "%s: %s\n" % (attr_name,
                                         cpt.to_text(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
5343 5344
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
5345 5346 5347 5348
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
5349

5350 5351
class ParamBase(core.VarBase):
    """
5352 5353 5354
    ParamBase is derived from Tensor( Which is the concept in Dygraph Mode). 
    A ParamBase is a persistable Tensor, and will be updated by optimizers 
    after each iteration.
5355 5356 5357
    The training of a neural network is essentially the updating of
    its ParamBase.

5358
    Relative to a general Tensor, a ParamBase has several its own
5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370
    member variables:

    Args:
        trainable(bool): True if the ParamBase need to be updated after
            iterations.
        optimize_attr(map): ParamBase 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 ParamBase. Default: None
        do_model_average(bool): True if the model average strategy will
            be applied on this ParamBase.
5371 5372
        need_clip (bool): Whether the parameter gradient need to be cliped 
            in optimizer. Default is True.
5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399 5400 5401 5402
    """

    @dygraph_only
    def __init__(self, shape, dtype, **kwargs):
        if shape is None:
            raise ValueError("The shape of Parameter should not be None")
        if dtype is None:
            raise ValueError("The dtype of Parameter should not be None")

        if len(shape) == 0:
            raise ValueError(
                "The dimensions of shape for Parameter must be greater than 0")

        for each in shape:
            if each < 0:
                raise ValueError(
                    "Each dimension of shape for Parameter must be greater than 0, but received %s"
                    % list(shape))

        if dtype is not None:
            if not isinstance(dtype, core.VarDesc.VarType):
                dtype = convert_np_dtype_to_dtype_(dtype)

        name = kwargs.get('name', unique_name.generate('_param_base'))

        super(ParamBase, self).__init__(dtype
                                        if dtype else core.VarDesc.VarType.FP32,
                                        list(shape) if shape else [], name,
                                        core.VarDesc.VarType.LOD_TENSOR, True)

5403 5404
        trainable = kwargs.get('trainable', True)
        self.stop_gradient = not trainable
5405 5406 5407 5408 5409 5410 5411

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

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

        self.do_model_average = kwargs.get('do_model_average', None)

5412 5413
        self.need_clip = kwargs.get('need_clip', True)

5414
        self.is_distributed = False
5415
        # self.block = default_main_program().global_block()
5416

5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429
    @property
    def trainable(self):
        return not self.stop_gradient

    @trainable.setter
    def trainable(self, trainable):
        if isinstance(trainable, bool):
            self.stop_gradient = not trainable
        else:
            raise ValueError(
                "The type of trainable MUST be bool, but the type is ",
                type(trainable))

5430
    def __str__(self):
5431
        """
5432
        Convert a ParamBase object to a readable string.
5433

5434
        Returns(str): A readable string.
5435 5436 5437 5438

        Examples:
            .. code-block:: python

5439
                import paddle
5440 5441 5442 5443 5444 5445 5446
                linear = paddle.nn.Linear(3, 3)
                print(linear.weight)
                # Parameter containing:
                # Tensor(shape=[3, 3], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
                #        [[ 0.48948765,  0.05829060, -0.25524026],
                #         [-0.70368278,  0.52986908, -0.68742192],
                #         [-0.54217887,  0.48439729,  0.34082305]])
5447
        """
5448 5449
        return "Parameter containing:\n{tensor}".format(
            tensor=super(ParamBase, self).__str__())
5450

5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461
    def __deepcopy__(self, memo):
        """
        Deep copy parameter, it will always performs Tensor copy.

        Examples:
            .. code-block:: python

                import paddle
                import copy
                linear = paddle.nn.Linear(1, 3)
                linear_copy = copy.deepcopy(linear)
T
tangwei12 已提交
5462

5463 5464 5465 5466 5467 5468 5469 5470 5471 5472 5473 5474 5475 5476 5477 5478 5479 5480
                print(linear.weight)
                # Parameter containing:
                # Tensor(shape=[1, 3], dtype=float32, place=CPUPlace, stop_gradient=False,
                #     [[-0.30929261, -0.90929240, -1.07851017]])

                print(linear_copy.weight)
                # Parameter containing:
                # Tensor(shape=[1, 3], dtype=float32, place=CPUPlace, stop_gradient=False,
                #     [[-0.30929261, -0.90929240, -1.07851017]])

        """
        state = copy.deepcopy(self.__dict__, memo)
        state["name"] = self.name + unique_name.generate("_deepcopy")
        new_param = ParamBase(self.shape, self.dtype, **state)
        memo[id(self)] = new_param
        new_param.copy_(self, True)
        return new_param

5481 5482 5483
    __repr__ = __str__


Y
Yu Yang 已提交
5484
# program is a global instance.
Y
Yu Yang 已提交
5485 5486
_main_program_ = Program()
_startup_program_ = Program()
5487

5488

5489
def default_startup_program():
Y
Yu Yang 已提交
5490
    """
Y
yuyang18 已提交
5491 5492
    Get default/global startup program.

5493 5494
    The :code:`paddle.nn` function will append the initialization operators into startup program.
    The :code:`startup_program` will initialize the parameters by the OPs. 
T
tangwei12 已提交
5495

5496 5497
    This method will return the default or the current startup program. Users can use
    :ref:`api_paddle_fluid_framework_program_guard`  to switch :ref:`api_paddle_fluid_framework_Program` .
Y
yuyang18 已提交
5498

5499 5500
    Returns:
        Program: current default startup program.
5501

5502
    Returns type: 
5503 5504 5505 5506

    Examples:
        .. code-block:: python

5507
            import paddle
5508

5509
            paddle.enable_static()
5510 5511 5512 5513
            x = paddle.static.data(name="x", shape=[-1, 784], dtype='float32')
            out = paddle.static.nn.fc(name="fc", x=x, size=10, activation="relu")
            print("main program is: {}".format(paddle.static.default_main_program()))
            print("start up program is: {}".format(paddle.static.default_startup_program()))
Y
Yu Yang 已提交
5514
    """
Y
Yu Yang 已提交
5515
    return _startup_program_
5516

5517

5518
def default_main_program():
Y
Yu Yang 已提交
5519
    """
5520
    This API can be used to get ``default main program`` which store the 
5521
    descriptions of Ops and tensors.
T
tangwei12 已提交
5522

5523
    For example ``z = paddle.add(x, y)`` will create a new ``add`` 
5524
    Op and a new ``z`` tensor, and they will be recorded in ``default main program`` . 
Y
yuyang18 已提交
5525

5526 5527
    The ``default main program`` is the default value for ``Program`` parameter in 
    a lot of APIs. For example, the :code:`Executor.run()` will execute the
Y
yuyang18 已提交
5528
    :code:`default_main_program` when the program is not specified.
5529

5530
    If you want to switch the ``default main program``, you can use :ref:`api_paddle_fluid_framework_program_guard` .
T
tangwei12 已提交
5531

Y
Yu Yang 已提交
5532
    Returns:
5533
        Program: A ``Program`` which holding the descriptions of OPs and tensors in the network.
5534 5535 5536 5537

    Examples:
        ..  code-block:: python

5538
            import paddle
5539

5540
            paddle.enable_static()
5541
            # Sample Network:
5542 5543 5544
            x = paddle.static.data(name='x', shape=[100, 100], dtype='float32')
            y = paddle.static.data(name='x', shape=[100, 100], dtype='float32')
            out = paddle.add(x, y)
5545

5546 5547 5548
            #print the number of blocks in the program, 1 in this case
            print(paddle.static.default_main_program().num_blocks) # 1
            #print the default_main_program
5549
            print(paddle.static.default_main_program())
Y
Yu Yang 已提交
5550
    """
Y
Yu Yang 已提交
5551
    return _main_program_
Y
Yu Yang 已提交
5552 5553 5554 5555 5556


def switch_main_program(program):
    """
    Switch the main program to a new program.
5557

Y
Yu Yang 已提交
5558 5559 5560 5561 5562 5563 5564 5565 5566 5567 5568 5569 5570 5571
    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):
    """
5572
    Switch the startup program to a new program
Y
Yu Yang 已提交
5573 5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584
    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
5585
@signature_safe_contextmanager
Y
Yu Yang 已提交
5586 5587
def program_guard(main_program, startup_program=None):
    """
5588 5589
    :api_attr: Static Graph

5590 5591 5592
    Change the global main program and startup program with ``with`` statement.
    Layer functions in the Python ``with`` block will append operators and
    Tensors to the new main programs.
5593

G
guofei 已提交
5594
    Args:
5595 5596
        main_program(Program): New main program inside ``with`` statement.
        startup_program(Program, optional): New startup program inside ``with`` 
G
guofei 已提交
5597 5598 5599 5600
            statement. :code:`None` means not changing startup program, 
            default_startup_program is still used.
            Default: None.

Y
Yu Yang 已提交
5601
    Examples:
5602
       .. code-block:: python
T
tangwei12 已提交
5603

5604
          import paddle
Y
yuyang18 已提交
5605

5606 5607 5608 5609 5610
          paddle.enable_static()
          main_program = paddle.static.Program()
          startup_program = paddle.static.Program()
          with paddle.static.program_guard(main_program, startup_program):
              data = paddle.static.data(name='image', shape=[None, 784, 784], dtype='float32')
5611
              hidden = paddle.static.nn.fc(x=data, size=10, activation='relu')
Y
yuyang18 已提交
5612 5613 5614

    Notes: The temporary :code:`Program` can be used if the user does not need
    to construct either of startup program or main program.
5615

Y
Yu Yang 已提交
5616
    Examples:
5617
       .. code-block:: python
Y
yuyang18 已提交
5618

5619
          import paddle
5620

5621 5622 5623 5624 5625
          paddle.enable_static()
          main_program = paddle.static.Program()
          # does not care about startup program. Just pass a temporary value.
          with paddle.static.program_guard(main_program, paddle.static.Program()):
              data = paddle.static.data(name='image', shape=[None, 784, 784], dtype='float32')
T
tangwei12 已提交
5626

Y
Yu Yang 已提交
5627
    """
5628
    from .data_feeder import check_type
5629 5630
    check_type(main_program, 'main_program', Program,
               'paddle.static.program_guard')
Y
Yu Yang 已提交
5631 5632
    main_program = switch_main_program(main_program)
    if startup_program is not None:
5633
        check_type(startup_program, 'startup_program', Program,
5634
                   'paddle.static.program_guard')
Y
Yu Yang 已提交
5635
        startup_program = switch_startup_program(startup_program)
5636 5637 5638 5639 5640 5641
    try:
        yield
    finally:
        switch_main_program(main_program)
        if startup_program is not None:
            switch_startup_program(startup_program)
X
xuwei06 已提交
5642 5643


W
Wu Yi 已提交
5644
def _get_var(name, program=None):
X
xuwei06 已提交
5645
    """
Y
yuyang18 已提交
5646
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
5647

X
xuwei06 已提交
5648 5649 5650
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
5651
        If None, default_global_program() will be used.
X
xuwei06 已提交
5652 5653 5654 5655 5656 5657 5658

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
5659
    assert isinstance(program, Program)
X
xuwei06 已提交
5660 5661

    return program.global_block().var(name)
5662 5663


S
rename  
sneaxiy 已提交
5664
@signature_safe_contextmanager
L
lujun 已提交
5665 5666
def _dygraph_guard(tracer):
    global _dygraph_tracer_
5667
    tmp_tracer = _dygraph_tracer_
L
lujun 已提交
5668
    _dygraph_tracer_ = tracer
5669
    core._switch_tracer(tracer)
M
minqiyang 已提交
5670

5671 5672 5673
    try:
        yield
    finally:
5674 5675
        core._switch_tracer(tmp_tracer)
        _dygraph_tracer_ = tmp_tracer
P
Paddle CI 已提交
5676 5677


S
rename  
sneaxiy 已提交
5678
@signature_safe_contextmanager
L
lujun 已提交
5679
def _dygraph_place_guard(place):
5680 5681 5682
    global _global_expected_place_
    tmp_place = _global_expected_place_
    _global_expected_place_ = place
M
minqiyang 已提交
5683

5684 5685
    _set_dygraph_tracer_expected_place(place)

5686 5687 5688
    try:
        yield
    finally:
5689
        _global_expected_place_ = tmp_place
5690
        _set_dygraph_tracer_expected_place(tmp_place)
5691 5692 5693 5694


def load_op_library(lib_filename):
    """
5695
    :api_attr: Static Graph
T
tangwei12 已提交
5696

5697 5698 5699
    Load a dynamic library, including custom operators and kernels.
    When library is loaded, ops and kernels registered in the library
    will be available in PaddlePaddle main process.
T
tianshuo78520a 已提交
5700
    Please note, the type of custom operators can't have the same type
5701 5702 5703 5704 5705 5706 5707 5708 5709 5710 5711 5712 5713 5714
    with the existing operators in the framework.

    Args:
        lib_filename (str): name of dynamic library.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            #fluid.load_op_library('custom_op.so')

    """
    core.load_op_library(lib_filename)
    OpProtoHolder.instance().update_op_proto()
5715 5716 5717 5718 5719 5720 5721 5722 5723 5724 5725 5726 5727 5728 5729 5730 5731 5732 5733 5734 5735 5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751 5752 5753 5754 5755 5756 5757 5758 5759 5760 5761 5762 5763 5764 5765 5766


def switch_device(device):
    global _current_device
    pre_device = _current_device
    _current_device = device
    return pre_device


@signature_safe_contextmanager
def device_guard(device=None):
    """
    **Notes**:
        **The API only supports static mode.**

    A context manager that specifies the device on which the OP will be placed.

    Args:
        device(str|None): Specify the device to use in the context. It should be 'cpu' or 'gpu',
            When it is set to 'cpu' or 'gpu', all OPs created in the context will be
            placed on CPUPlace or CUDAPlace. When 'gpu' is set and the program runs on
            single-card, the device index will be the same as the device on which the
            executor runs. Default: None, OPs in this context will be automatically
            assigned devices.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            support_gpu = fluid.is_compiled_with_cuda()
            place = fluid.CPUPlace()
            if support_gpu:
                place = fluid.CUDAPlace(0)

            # if GPU is supported, the three OPs below will be automatically assigned to CUDAPlace(0)
            data1 = fluid.layers.fill_constant(shape=[1, 3, 8, 8], value=0.5, dtype='float32')
            data2 = fluid.layers.fill_constant(shape=[1, 3, 5, 5], value=0.5, dtype='float32')
            shape = fluid.layers.shape(data2)

            with fluid.device_guard("cpu"):
                # Ops created here will be placed on CPUPlace
                shape = fluid.layers.slice(shape, axes=[0], starts=[0], ends=[4])
            with fluid.device_guard('gpu'):
                # if GPU is supported, OPs created here will be placed on CUDAPlace(0), otherwise on CPUPlace
                out = fluid.layers.crop_tensor(data1, shape=shape)

            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())
            result = exe.run(fetch_list=[out])
    """

5767 5768 5769 5770 5771
    index = None
    if device and ':' in device:
        device, index = device.split(':')
        if device == 'cpu':
            raise ValueError("Should not set device id for cpu.")
5772 5773 5774 5775
    if device not in ['cpu', 'gpu', '', None]:
        raise ValueError(
            "The Attr(device) should be 'cpu' or 'gpu', and it can also be empty string or None "
            "when there is no need to specify device. But received %s" % device)
5776 5777
    if index:
        device = ":".join([device, index])
5778
    pre_device = switch_device(device)
5779 5780 5781 5782
    try:
        yield
    finally:
        switch_device(pre_device)
G
guofei 已提交
5783 5784 5785 5786 5787 5788 5789 5790 5791 5792 5793 5794 5795 5796 5797 5798 5799 5800 5801 5802 5803 5804 5805 5806 5807 5808 5809 5810 5811 5812 5813 5814 5815 5816 5817 5818 5819 5820 5821 5822 5823 5824 5825 5826 5827 5828 5829 5830 5831 5832 5833 5834 5835 5836 5837 5838 5839 5840 5841 5842 5843 5844 5845 5846 5847 5848 5849


def set_flags(flags):
    """
    This function sets the GFlags value in Paddle.

    Args:
        flags (dict): A dict contains flags and its value.

    Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                fluid.set_flags({'FLAGS_eager_delete_tensor_gb': 1.0})
    """
    if not isinstance(flags, dict):
        raise TypeError('flags in set_flags should be a dict')
    for key, value in flags.items():
        if core.globals().is_public(key):
            core.globals()[key] = value
        else:
            raise ValueError(
                "Flag %s cannot set its value through this function." % (key))


def get_flags(flags):
    """
    This function gets the GFlags value in Paddle.

    Args:
        flags(list|tuple|str): A list/tuple of string or a string which is the flag's name.

    Returns:
        flag's value in Paddle.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            flags = ['FLAGS_eager_delete_tensor_gb', 'FLAGS_check_nan_inf']
            res = fluid.get_flags(flags)
            print(res)
            # {'FLAGS_eager_delete_tensor_gb': 0.0, 'FLAGS_check_nan_inf': False}
    """
    flags_value = {}
    if isinstance(flags, (list, tuple)):
        for key in flags:
            if (core.globals().is_public(key)):
                value = core.globals()[key]
                temp = {key: value}
                flags_value.update(temp)
            else:
                raise ValueError(
                    'Flag %s cannot get its value through this function.' %
                    (key))
    elif isinstance(flags, str):
        if (core.globals().is_public(flags)):
            value = core.globals()[flags]
            temp = {flags: value}
            flags_value.update(temp)
        else:
            raise ValueError(
                'Flag %s cannot get its value through this function.' % (flags))
    else:
        raise TypeError('Flags in get_flags should be a list, tuple or string.')
    return flags_value
5850 5851 5852 5853 5854 5855 5856 5857 5858 5859 5860 5861 5862 5863 5864 5865 5866 5867 5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880 5881 5882 5883 5884 5885 5886 5887 5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905 5906 5907 5908 5909 5910


def _get_paddle_place(place):
    "convert the string to paddle Place"
    if place is None:
        return place
    if isinstance(place, (core.Place, core.XPUPlace, core.CPUPlace,
                          core.CUDAPinnedPlace, core.CUDAPlace)):
        return place

    if not isinstance(place, str):
        raise ValueError(
            "place only support string which is 'Place' and so on.")

    place = place.lower()
    if (place == "cpu"):
        return core.CPUPlace()
    if (place == "device"):
        return core.Place()

    avaliable_gpu_place = re.match(r'gpu:\d+', place)
    if place == "gpu_pinned" or place == "gpu" or avaliable_gpu_place:
        if not core.is_compiled_with_cuda():
            raise ValueError(
                "The device should not be {}, since PaddlePaddle is " \
                "not compiled with CUDA".format(avaliable_gpu_place))
        if place == "gpu_pinned":
            return core.CUDAPinnedPlace()
        elif place == "gpu":
            return core.CUDAPlace(0)
        else:
            place_info_list = place.split(':', 1)
            device_id = place_info_list[1]
            device_id = int(device_id)
            return core.CUDAPlace(device_id)
    avaliable_xpu_place = re.match(r'xpu:\d+', place)
    if avaliable_xpu_place:
        if not core.is_compiled_with_xpu():
            raise ValueError(
                "The device should not be {}, since PaddlePaddle is " \
                "not compiled with XPU".format(avaliable_xpu_place))
        place_info_list = place.split(':', 1)
        device_id = place_info_list[1]
        device_id = int(device_id)
        return core.XPUPlace(device_id)
    raise ValueError(
        "paddle support CPUPlace, CUDAPlace,CUDAPinnedPlace and XPUPlace, Please check your Place Input"
    )


def _get_paddle_place_list(places):

    if not isinstance(places, (list, tuple)):
        raise TypeError("places must to be List or Tuple")

    ret = []
    for p in places:
        p = _get_paddle_place(p)
        ret.append(p)

    return ret