framework.py 202.9 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
            "'%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 "
251 252 253
            "ProgramTranslator if you are using @paddle.jit.to_static. If you have to run ProgramTranslator, "
            "please use other API to replace '%s'" % (func.__name__,
                                                      func.__name__))
254 255 256 257

    return __impl__


T
tangwei12 已提交
258 259
# 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
260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276
# 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


277 278
dygraph_not_support = wrap_decorator(_dygraph_not_support_)
dygraph_only = wrap_decorator(_dygraph_only_)
279
static_only = wrap_decorator(_static_only_)
280
fake_interface_only = wrap_decorator(_fake_interface_only_)
281 282


L
lujun 已提交
283 284
def _dygraph_tracer():
    return _dygraph_tracer_
285

W
Wu Yi 已提交
286

M
minqiyang 已提交
287
def _current_expected_place():
288 289 290
    global _global_expected_place_
    if _global_expected_place_ is None:
        if core.is_compiled_with_cuda():
291 292 293 294 295 296 297 298 299 300 301
            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()
302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317
        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
318 319


L
Leo Chen 已提交
320 321 322 323
# TODO(zhiqiu): remove this function.
def _var_base_to_np(var_base):
    """	
    convert VarBase tp numpy	
T
tangwei12 已提交
324

L
Leo Chen 已提交
325 326 327 328 329 330 331 332 333 334 335 336
    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
337
def _cpu_num():
338
    if "CPU_NUM" not in os.environ.keys():
C
chengduo 已提交
339 340 341 342 343 344 345 346
        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
347
        os.environ['CPU_NUM'] = str(1)
348
    cpu_num = os.environ.get('CPU_NUM')
C
chengduo 已提交
349 350 351 352 353 354 355 356 357 358
    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
359 360


361 362 363 364 365 366 367 368 369
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


370 371 372 373 374 375 376 377 378 379 380 381 382 383 384
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
385 386 387 388
def is_compiled_with_cuda():
    """
    Whether this whl package can be used to run the model on GPU.

389
    Returns (bool): `True` if CUDA is currently available, otherwise `False`.
C
chengduo 已提交
390 391 392 393

    Examples:
        .. code-block:: python

394 395
            import paddle
            support_gpu = paddle.is_compiled_with_cuda()
C
chengduo 已提交
396 397 398 399
    """
    return core.is_compiled_with_cuda()


S
sneaxiy 已提交
400
def cuda_places(device_ids=None):
L
lujun 已提交
401
    """
402 403 404 405
    **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
406
    This function creates a list of :code:`paddle.CUDAPlace` objects.
S
add doc  
sneaxiy 已提交
407 408

    If :code:`device_ids` is None, environment variable of
409
    :code:`FLAGS_selected_gpus` would be checked first. For example, if
S
add doc  
sneaxiy 已提交
410
    :code:`FLAGS_selected_gpus=0,1,2`, the returned list would
C
Chen Weihang 已提交
411
    be [paddle.CUDAPlace(0), paddle.CUDAPlace(1), paddle.CUDAPlace(2)].
S
add doc  
sneaxiy 已提交
412
    If :code:`FLAGS_selected_gpus` is not set, all visible
413
    gpu places would be returned according to the :code:`CUDA_VISIBLE_DEVICES` environment variable.
S
add doc  
sneaxiy 已提交
414 415

    If :code:`device_ids` is not None, it should be the device
416
    ids of GPUs. For example, if :code:`device_ids=[0,1,2]`,
S
add doc  
sneaxiy 已提交
417
    the returned list would be 
C
Chen Weihang 已提交
418
    [paddle.CUDAPlace(0), paddle.CUDAPlace(1), paddle.CUDAPlace(2)].
T
tangwei12 已提交
419

420 421
    Parameters:
        device_ids (list or tuple of int, optional): list of GPU device ids.
S
add doc  
sneaxiy 已提交
422 423

    Returns:
C
Chen Weihang 已提交
424
        list of paddle.CUDAPlace: Created GPU place list.
L
lujun 已提交
425 426 427 428

    Examples:
        .. code-block:: python

C
Chen Weihang 已提交
429 430
            import paddle
            import paddle.static as static
T
tangwei12 已提交
431

C
Chen Weihang 已提交
432 433 434
            paddle.enable_static()

            cuda_places = static.cuda_places()
L
lujun 已提交
435 436

    """
S
sneaxiy 已提交
437 438 439
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_ids is None:
C
chengduo 已提交
440
        device_ids = _cuda_ids()
S
sneaxiy 已提交
441 442 443 444 445
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.CUDAPlace(dev_id) for dev_id in device_ids]


446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467
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
468
        
469 470 471 472 473 474 475 476 477 478 479 480 481 482 483
            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
484
def cpu_places(device_count=None):
L
lujun 已提交
485
    """
C
Chen Weihang 已提交
486
    This function creates a list of :code:`paddle.CPUPlace` objects, and returns the created list.
T
tangwei12 已提交
487

S
add doc  
sneaxiy 已提交
488 489
    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
C
chengduo 已提交
490 491
    If :code:`CPU_NUM` is not set, the default value is 1,
    i.e. CPU_NUM=1.
492 493
    :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
494

495 496
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
497 498

    Returns:
C
Chen Weihang 已提交
499
        list of paddle.CPUPlace: Created list of CPU places.
L
lujun 已提交
500 501 502 503

    Examples:
        .. code-block:: python

C
Chen Weihang 已提交
504 505
            import paddle
            import paddle.static as static
T
tangwei12 已提交
506

C
Chen Weihang 已提交
507 508 509
            paddle.enable_static()

            cpu_places = static.cpu_places()
L
lujun 已提交
510 511
    """

S
sneaxiy 已提交
512 513 514 515 516 517
    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
518
    """
519
    This function creates a list of :code:`fluid.CUDAPinnedPlace` objects.
S
add doc  
sneaxiy 已提交
520 521 522

    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
523 524 525 526
    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
527

528 529
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
530 531

    Returns:
532
        list of fluid.CUDAPinnedPlace: Created list of CUDA pinned places.
L
lujun 已提交
533 534 535 536

    Examples:
        .. code-block:: python

537
            import paddle.fluid as fluid
L
lujun 已提交
538 539 540 541 542
            cuda_pinned_places_cpu_num = fluid.cuda_pinned_places()
            # or
            cuda_pinned_places = fluid.cuda_pinned_places(1)

    """
S
sneaxiy 已提交
543 544 545
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_count is None:
546 547
        device_count = len(_cuda_ids())
    return [core.CUDAPinnedPlace()] * device_count
S
sneaxiy 已提交
548 549


550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575
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
576
@signature_safe_contextmanager
577 578
def name_scope(prefix=None):
    """
579 580
    :api_attr: Static Graph

581
    Generate hierarchical name prefix for the operators in Static Graph.
582

T
Tao Luo 已提交
583 584 585
    Note: 
        This should only used for debugging and visualization purpose.
        Don't use it for serious analysis such as graph/program transformations.
586
        Don't use it in dygraph, since it will cause memory leak.
587 588

    Args:
T
Tao Luo 已提交
589
        prefix(str, optional): prefix. Default is none.
590 591 592

    Examples:
        .. code-block:: python
T
Tink_Y 已提交
593

594 595 596
          import paddle
          paddle.enable_static()
          with paddle.static.name_scope("s1"):
597
             a = paddle.static.data(name='data', shape=[None, 1], dtype='int32')
T
Tao Luo 已提交
598
             b = a + 1
599
             with paddle.static.name_scope("s2"):
T
Tao Luo 已提交
600
                c = b * 1
601
             with paddle.static.name_scope("s3"):
T
Tao Luo 已提交
602
                d = c / 1
603 604 605
          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
606 607 608
                g = f - 1

          # Op are created in the default main program.  
609
          for op in paddle.static.default_main_program().block(0).ops:
T
Tao Luo 已提交
610 611 612 613 614 615 616 617 618 619 620 621 622 623 624
              # 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/'
625 626
    """
    # TODO(panyx0718): Only [0-9a-z].
627
    # in dygraph we don't need namescope since it will cause mem leak
L
Leo Chen 已提交
628 629 630
    if in_dygraph_mode():
        yield
    else:
T
tianshuo78520a 已提交
631
        assert prefix, "namescope prefix can not be empty."
632 633
        global _name_scope
        _name_scope = _name_scope.child(prefix)
634 635 636 637
        try:
            yield
        finally:
            _name_scope = _name_scope.parent()
638 639 640 641 642 643 644 645 646 647 648 649


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
650 651 652
def generate_control_dev_var_name():
    import random
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
653 654 655 656


def grad_var_name(var_name):
    """
657 658
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
659 660 661
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
662

663
def convert_np_dtype_to_dtype_(np_dtype):
664 665
    """
    Convert the data type in numpy to the data type in Paddle
666

667
    Args:
668
        np_dtype(np.dtype): the data type in numpy.
669

670 671
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
672 673

    """
674 675
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
676
        return core.VarDesc.VarType.FP32
677
    elif dtype == np.float64:
678
        return core.VarDesc.VarType.FP64
679
    elif dtype == np.float16:
680
        return core.VarDesc.VarType.FP16
681
    elif dtype == np.int32:
682
        return core.VarDesc.VarType.INT32
683
    elif dtype == np.int16:
684
        return core.VarDesc.VarType.INT16
685
    elif dtype == np.int64:
686
        return core.VarDesc.VarType.INT64
687
    elif dtype == np.bool:
688
        return core.VarDesc.VarType.BOOL
689
    elif dtype == np.uint16:
690 691 692
        # since there is still no support for bfloat16 in NumPy,
        # uint16 is used for casting bfloat16
        return core.VarDesc.VarType.BF16
693 694
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
695 696
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
697 698 699 700
    elif dtype == np.complex64:
        return core.VarDesc.VarType.COMPLEX64
    elif dtype == np.complex128:
        return core.VarDesc.VarType.COMPLEX128
701
    else:
M
minqiyang 已提交
702
        raise ValueError("Not supported numpy dtype %s" % dtype)
703 704 705


def dtype_is_floating(dtype):
706 707 708
    """
    Check the data type is floating or not.
    Args:
709
        dtype(np.dtype|core.VarDesc.VarType): data type.
710 711 712 713 714
            Could be numpy format or Paddle format

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

    """
715
    if not isinstance(dtype, core.VarDesc.VarType):
716 717
        dtype = convert_np_dtype_to_dtype_(dtype)

718 719 720 721
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
722 723


Y
Yang Yang(Tony) 已提交
724
def _debug_string_(proto, throw_on_error=True):
725 726 727 728 729 730 731 732 733 734 735
    """
    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
736
    error_fields = list()
Y
Yang Yang(Tony) 已提交
737
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
738 739
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
740 741 742
    return proto.__str__()


743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799
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 = []
800
    target_block = default_main_program().current_block()
801 802 803 804 805 806 807 808 809 810 811

    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
812
            })
813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851
        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):
852
                temp_1 = var.block.create_var(dtype=slice_item.dtype)
853
                fill_constant([1], 1, force_cpu=True, out=temp_1)
854
                temp_end = target_block.create_var(dtype=slice_item.dtype)
855
                target_block.append_op(
856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894
                    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
895

896
    # starts
L
Leo Chen 已提交
897
    if contain_var(slice_start):
898 899 900 901 902 903 904 905
        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
906 907 908 909
        attrs['starts'] = slice_start

    # ends
    if contain_var(slice_end):
910 911 912 913 914 915 916
        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
917 918 919
    else:
        attrs['ends'] = slice_end

920 921
    # strides
    if use_strided_slice == True:
L
Leo Chen 已提交
922
        if contain_var(slice_step):
923 924 925 926 927 928 929
            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
930 931
        else:
            attrs['strides'] = slice_step
932 933 934 935 936 937
    # infer_flags
    attrs['infer_flags'] = infer_flags

    out = var
    if use_strided_slice == False and len(slice_axis) > 0:
        # append slice_op here
938
        slice_out_var = target_block.create_var(
939 940 941
            name=unique_name.generate_with_ignorable_key(var.name + "_slice"),
            dtype=var.dtype)

942
        target_block.append_op(
943 944 945 946 947 948 949
            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:
950
        strided_slice_out_var = target_block.create_var(
951 952 953
            name=unique_name.generate_with_ignorable_key(var.name +
                                                         "_strided_slice"),
            dtype=var.dtype)
954
        target_block.append_op(
955 956 957 958 959 960 961 962
            type="strided_slice",
            inputs=inputs,
            outputs={'Out': [strided_slice_out_var]},
            attrs=attrs)

        out = strided_slice_out_var

    if len(reverse_axis) > 0:
963
        reverse_out_var = target_block.create_var(
964 965 966
            name=unique_name.generate_with_ignorable_key(var.name +
                                                         "_slice_reverse"),
            dtype=var.dtype)
967
        target_block.append_op(
968 969 970 971 972 973 974 975 976 977 978
            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
979
class Variable(object):
980
    """
J
Jiabin Yang 已提交
981
    **Notes**:
982
        **The constructor of Variable should not be invoked directly.**
J
Jiabin Yang 已提交
983

984 985
        **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
986 987 988
        **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
989
    cases, variables are used for holding different kinds of data or training
J
Jiabin Yang 已提交
990 991
    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.
992

993
    There are many kinds of variables. Each kind of them has its own attributes
J
Jiabin Yang 已提交
994
    and usages. Please refer to the `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_ for details.
995

T
tianshuo78520a 已提交
996
    Most of a Variable's member variables can be set to be None. It mean
997
    it is not available or will be specified later.
998

999
    Examples:
1000 1001
        In Static Graph Mode:

1002 1003
        .. code-block:: python

1004
            import paddle.fluid as fluid
1005
            cur_program = fluid.Program()
1006 1007 1008 1009
            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
1010
        In `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_  Mode:
1011 1012 1013 1014 1015 1016 1017 1018 1019

        .. 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))

1020 1021
    """

Y
Yu Yang 已提交
1022 1023
    def __init__(self,
                 block,
Y
Yu Yang 已提交
1024
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
1025 1026 1027 1028
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
1029
                 capacity=None,
Q
QI JUN 已提交
1030
                 persistable=None,
F
fengjiayi 已提交
1031
                 error_clip=None,
Y
Yu Yang 已提交
1032
                 stop_gradient=False,
F
fengjiayi 已提交
1033
                 is_data=False,
H
Huihuang Zheng 已提交
1034
                 need_check_feed=False,
H
hong 已提交
1035
                 belong_to_optimizer=False,
Y
Yu Yang 已提交
1036
                 **kwargs):
Y
Yu Yang 已提交
1037 1038
        self.block = block
        if name is None:
Y
Yu Yang 已提交
1039
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
1040

Y
Yu Yang 已提交
1041
        if dtype is not None:
1042
            if not isinstance(dtype, core.VarDesc.VarType):
1043
                dtype = convert_np_dtype_to_dtype_(dtype)
1044

H
hong 已提交
1045 1046
        self.belong_to_optimizer = belong_to_optimizer

1047 1048 1049 1050 1051
        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))
1052

1053 1054 1055
        if self.desc is None:
            self.desc = self.block.desc.var(cpt.to_bytes(name))
            is_new_var = True
1056

1057 1058 1059
        if is_new_var:
            self.desc.set_type(type)
        elif self.desc.type() != type:
L
Leo Chen 已提交
1060 1061
            raise ValueError("Variable '{0}' has been created before. The "
                             "previous type is {1}, the new type is {2}. They"
1062 1063
                             " are not matched".format(self.name,
                                                       self.desc.type(), type))
1064

1065
        if shape is not None:
1066
            if is_new_var:
1067 1068 1069 1070 1071 1072
                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
1073 1074
                        "Variable '{0}' has been created before. The previous "
                        "shape is {1}, the new shape is {2}. They are not "
1075 1076 1077 1078 1079 1080 1081
                        "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
1082 1083
                    raise ValueError("Variable '{0}' has been created before. "
                                     "The previous data type is {1}, the new "
1084 1085 1086 1087 1088 1089 1090 1091 1092
                                     "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
1093 1094
                    raise ValueError("Variable '{0}' has been created before. "
                                     "The previous lod_level is {1}, the new "
1095 1096 1097 1098 1099 1100 1101 1102 1103
                                     "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
1104 1105
                        "Variable '{0}' has been created before."
                        "The previous persistable is {1}, the new "
1106 1107
                        "persistable is {2}. They are not matched".format(
                            self.name, self.persistable, persistable))
1108

1109 1110
        if need_check_feed and is_new_var:
            self.desc.set_need_check_feed(need_check_feed)
H
Huihuang Zheng 已提交
1111

1112 1113 1114 1115 1116 1117 1118
        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
1119

1120 1121 1122 1123
        self.block.vars[name] = self
        self.op = None
        self._stop_gradient = stop_gradient
        self.is_data = is_data
Y
Yu Yang 已提交
1124

1125
    @fake_interface_only
1126 1127
    def detach(self):
        """
J
Jiabin Yang 已提交
1128
        **Notes**:
T
tianshuo78520a 已提交
1129
            **This API is ONLY available in Dygraph mode**
1130

1131
        Returns a new Variable, detached from the current graph.
1132

1133
        Returns:
J
Jiabin Yang 已提交
1134
             ( :ref:`api_guide_Variable_en` | dtype is same as current Variable): The detached Variable.
1135

1136

1137 1138 1139 1140 1141
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1142
                from paddle.fluid.dygraph import Linear
1143 1144 1145 1146
                import numpy as np

                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
                with fluid.dygraph.guard():
1147
                    linear = Linear(32, 64)
1148
                    data = to_variable(data)
1149
                    x = linear(data)
1150 1151 1152
                    y = x.detach()

        """
1153
        pass
1154

1155
    @fake_interface_only
1156
    def numpy(self):
1157
        """
J
Jiabin Yang 已提交
1158
        **Notes**:
T
tianshuo78520a 已提交
1159
            **This API is ONLY available in Dygraph mode**
1160

J
Jiabin Yang 已提交
1161
        Returns a numpy array shows the value of current :ref:`api_guide_Variable_en`
1162 1163 1164 1165 1166

        Returns:
            ndarray: The numpy value of current Variable.

        Returns type:
J
Jiabin Yang 已提交
1167
            ndarray: dtype is same as current Variable
1168 1169 1170 1171 1172 1173

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1174
                from paddle.fluid.dygraph import Linear
1175 1176 1177 1178
                import numpy as np

                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
                with fluid.dygraph.guard():
1179
                    linear = Linear(32, 64)
1180
                    data = to_variable(data)
1181
                    x = linear(data)
1182 1183 1184
                    print(x.numpy())

        """
1185
        pass
1186

1187
    @fake_interface_only
1188 1189
    def set_value(self, value):
        """
J
Jiabin Yang 已提交
1190
        **Notes**:
T
tianshuo78520a 已提交
1191
            **This API is ONLY available in Dygraph mode**
J
Jiabin Yang 已提交
1192

1193 1194 1195 1196 1197 1198 1199 1200 1201 1202
        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
1203
                from paddle.fluid.dygraph import Linear
1204 1205
                import numpy as np

1206
                data = np.ones([3, 1024], dtype='float32')
1207
                with fluid.dygraph.guard():
1208
                    linear = fluid.dygraph.Linear(1024, 4)
1209
                    t = to_variable(data)
1210
                    linear(t)  # call with default weight
1211
                    custom_weight = np.random.randn(1024, 4).astype("float32")
1212 1213
                    linear.weight.set_value(custom_weight)  # change existing weight
                    out = linear(t)  # call with different weight
1214 1215

        """
1216
        pass
1217

1218
    @fake_interface_only
1219
    def backward(self, retain_graph=False):
1220
        """
J
Jiabin Yang 已提交
1221
        **Notes**:
T
tianshuo78520a 已提交
1222
            **This API is ONLY available in Dygraph mode**
1223

1224
        Run backward of current Graph which starts from current Tensor.
1225

J
Jiabin Yang 已提交
1226
        Args:
1227 1228 1229 1230
            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.
1231

J
Jiabin Yang 已提交
1232 1233
        Returns:
            NoneType: None
1234 1235 1236 1237 1238

        Examples:
            .. code-block:: python

                import numpy as np
1239 1240
                import paddle
                paddle.disable_static()
1241 1242

                x = np.ones([2, 2], np.float32)
1243 1244 1245 1246 1247 1248 1249
                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)
1250 1251
                ret = paddle.add_n(inputs)
                loss = paddle.sum(ret)
1252
                loss.backward()
1253 1254

        """
1255
        pass
1256

1257
    @fake_interface_only
1258
    def gradient(self):
1259
        """
J
Jiabin Yang 已提交
1260
        **Notes**:
T
tianshuo78520a 已提交
1261
            **This API is ONLY available in Dygraph mode**
1262 1263 1264

        Get the Gradient of Current Variable

J
Jiabin Yang 已提交
1265
        Returns:
1266
            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.
1267 1268 1269 1270 1271 1272 1273

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

1274
                # example1: return ndarray
1275 1276 1277 1278 1279 1280 1281 1282 1283
                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)
1284
                    loss2.backward()
1285 1286
                    print(loss2.gradient())

1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299
                # 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())

1300
        """
1301
        pass
1302

1303
    @fake_interface_only
1304
    def clear_gradient(self):
1305
        """
J
Jiabin Yang 已提交
1306
        **Notes**:
T
tianshuo78520a 已提交
1307
            **1. This API is ONLY available in Dygraph mode**
J
Jiabin Yang 已提交
1308 1309

            **2. Use it only Variable has gradient, normally we use this for Parameters since other temporal Variable will be deleted by Python's GC**
1310

J
Jiabin Yang 已提交
1311
        Clear  (set to ``0`` ) the Gradient of Current Variable
1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329

        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)
1330
                    loss2.backward()
1331 1332 1333 1334 1335
                    print(loss2.gradient())
                    loss2.clear_gradient()
                    print("After clear {}".format(loss2.gradient()))

        """
1336
        pass
X
Xin Pan 已提交
1337

1338
    def __str__(self):
1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354
        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

1355 1356
                import paddle
                import paddle.static as static
1357

1358 1359 1360
                paddle.enable_static()

                cur_program = static.Program()
1361 1362 1363 1364 1365 1366
                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())
        """
1367 1368
        # VarType.LOD_TENSOR -> LOD_TENSOR
        type_str = str(self.type).split('.')[1]
1369
        if self.type == core.VarDesc.VarType.SELECTED_ROWS or self.type == core.VarDesc.VarType.LOD_TENSOR:
1370 1371
            dtype_str = str(self.dtype).split('.')[1]
            var_str = "{name} : {type}.shape{shape}.dtype({dtype}).stop_gradient({stop_gradient})".\
T
tangwei12 已提交
1372 1373
                format(name=self.name, type=type_str, shape=self.shape,
                       dtype=dtype_str, stop_gradient=self.stop_gradient)
1374
        else:
1375 1376
            var_str = "{name} : {type})".\
                format(name=self.name, type=type_str)
1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389

        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
1390

F
update  
fengjiayi 已提交
1391
    def to_string(self, throw_on_error, with_details=False):
1392 1393 1394
        """
        Get debug string.

J
Jiabin Yang 已提交
1395 1396 1397 1398 1399
        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;
1400

1401 1402
        Returns:
            str: The debug string.
1403 1404 1405 1406 1407

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
1408
                import paddle
1409

1410
                paddle.enable_static()
1411 1412 1413 1414 1415
                cur_program = fluid.Program()
                cur_block = cur_program.current_block()
                new_variable = cur_block.create_var(name="X",
                                                    shape=[-1, 23, 48],
                                                    dtype='float32')
1416
                print(new_variable.to_string(True))
J
Jiabin Yang 已提交
1417
                print("=============with detail===============")
1418
                print(new_variable.to_string(True, True))
1419
        """
F
update  
fengjiayi 已提交
1420 1421
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
1422
        protostr = self.desc.serialize_to_string()
1423
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
1424 1425 1426 1427
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
1428 1429 1430
                res_str += "%s: %s\n" % (attr_name,
                                         cpt.to_text(getattr(self, attr_name)))

F
update  
fengjiayi 已提交
1431
        return res_str
1432 1433 1434

    __repr__ = __str__

1435
    @property
1436
    def stop_gradient(self):
J
Jiabin Yang 已提交
1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451
        """
        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")
1452 1453
                linear = fluid.Linear(13, 5, dtype="float32")
                linear2 = fluid.Linear(3, 3, dtype="float32")
J
Jiabin Yang 已提交
1454 1455 1456
                a = fluid.dygraph.to_variable(value0)
                b = fluid.dygraph.to_variable(value1)
                c = fluid.dygraph.to_variable(value2)
1457 1458
                out1 = linear(a)
                out2 = linear2(b)
J
Jiabin Yang 已提交
1459 1460 1461 1462
                out1.stop_gradient = True
                out = fluid.layers.concat(input=[out1, out2, c], axis=1)
                out.backward()

1463
                assert linear.weight.gradient() is None
J
Jiabin Yang 已提交
1464 1465
                assert (out1.gradient() == 0).all()
        """
1466
        return self._stop_gradient
1467

1468 1469
    @stop_gradient.setter
    def stop_gradient(self, s):
1470
        self._stop_gradient = s
1471

1472 1473
    @property
    def persistable(self):
J
Jiabin Yang 已提交
1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494
        """
        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))
        """
1495
        return self.desc.persistable()
1496

Y
Yu Yang 已提交
1497 1498
    @persistable.setter
    def persistable(self, p):
1499
        self.desc.set_persistable(p)
Y
Yu Yang 已提交
1500

Y
Yu Yang 已提交
1501 1502
    @property
    def name(self):
J
Jiabin Yang 已提交
1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518
        """
        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))
        """
1519
        return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
1520

1521 1522 1523 1524 1525 1526 1527 1528
    @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
1529

1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540
        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
1541 1542
    @name.setter
    def name(self, new_name):
1543
        self.desc.set_name(new_name)
T
typhoonzero 已提交
1544

Y
Yu Yang 已提交
1545 1546
    @property
    def shape(self):
J
Jiabin Yang 已提交
1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563
        """
        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
1564
        # convert to tuple, make it as same as numpy API.
1565
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
1566 1567

    @property
F
fengjiayi 已提交
1568
    def dtype(self):
J
Jiabin Yang 已提交
1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584
        """
        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))
        """
1585
        return self.desc.dtype()
Y
Yu Yang 已提交
1586 1587 1588

    @property
    def lod_level(self):
J
Jiabin Yang 已提交
1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609
        """
        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))
        """
1610 1611 1612
        if self.type == core.VarDesc.VarType.SELECTED_ROWS:
            raise Exception("SelectedRows DO NOT supprt lod")

1613
        return self.desc.lod_level()
Y
Yu Yang 已提交
1614

Y
Yu Yang 已提交
1615 1616
    @property
    def type(self):
J
Jiabin Yang 已提交
1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632
        """
        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))
        """
1633
        return self.desc.type()
Y
Yu Yang 已提交
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 1668
    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
1669
    def _set_error_clip(self, error_clip):
1670 1671 1672 1673 1674 1675 1676 1677 1678
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
1679 1680
        self.error_clip = error_clip

1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709
    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

1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720
    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
1721
            raise ValueError("slice step can not be zero")
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 1796

        # 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
1797
    def _cloneVar(self, copy=False):
1798 1799
        if not copy:
            return self.block.create_var(
H
Hongyu Liu 已提交
1800 1801
                name=unique_name.generate_with_ignorable_key(self.name),
                dtype=self.dtype)
1802 1803 1804 1805
        else:
            return self

    def _sliceVar(self, axes, starts, ends):
L
lujun 已提交
1806
        new_var = self._cloneVar()
1807 1808 1809 1810 1811 1812 1813 1814 1815 1816
        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
1817
        new_var = self._cloneVar()
1818 1819 1820 1821 1822 1823 1824 1825 1826 1827
        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
1828
                return self._cloneVar(True)
1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846
            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
1847
                return self._cloneVar(True)
1848
            index = int(item)
1849
            if (index > 0 and index >= self.shape[axis]) \
1850 1851 1852 1853 1854 1855 1856
                    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):
1857
        return _getitem_impl_(self, item)
1858

1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869
    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
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

        def replace_ellipsis(item):
            # Use slice(None) to replace Ellipsis.
            # For var, var.shape = [3,4,5,6]
            #
            #   var[..., 1:2] -> var[:, :, :, 1:2]
            #   var[0, ...] -> var[0]
            #   var[0, ..., 1:2] -> var[0, :, :, 1:2]

            item = list(item)
            ell_count = item.count(Ellipsis)
            if ell_count == 0:
                return item
            elif ell_count > 1:
                raise IndexError(
                    "An index can only have a single ellipsis ('...')")

            ell_idx = item.index(Ellipsis)

            if ell_idx == len(item) - 1:
                return item[:-1]
            else:
                item[ell_idx:ell_idx + 1] = [slice(None)] * (
                    len(self.shape) - len(item) + 1)

            return item

        item = replace_ellipsis(item)

1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929
        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

1930
        from .data_feeder import convert_dtype
1931 1932
        #  2.1 value is an integer of float
        if isinstance(value, (int, float)):
1933
            value = np.array([value]).astype(convert_dtype(dtype))
1934 1935 1936 1937 1938 1939 1940 1941 1942 1943

        #  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]
1944 1945 1946
            elif dtype == core.VarDesc.VarType.FP64:
                value_name = "fp64_values"
                values = [float(v) for v in value.flat]
1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972
            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
1973

F
fengjiayi 已提交
1974 1975 1976
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
1977

1978 1979
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
1980 1981 1982 1983
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
1984
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
1985 1986 1987 1988 1989
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
1990 1991 1992 1993
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
1994 1995 1996 1997 1998 1999 2000 2001 2002
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
2003
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
2004 2005 2006 2007 2008 2009
        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):
2010 2011 2012 2013 2014 2015 2016 2017
        """
        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
2018 2019
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
2020 2021
        return self.op_proto_map[type]

2022 2023
    def update_op_proto(self):
        op_protos = get_all_op_protos()
2024
        custom_op_names = []
2025 2026 2027
        for proto in op_protos:
            if proto.type not in self.op_proto_map:
                self.op_proto_map[proto.type] = proto
2028 2029 2030
                custom_op_names.append(proto.type)

        return custom_op_names
2031

2032 2033 2034 2035
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
2036
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
2037
            core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
2038 2039
            core.op_proto_and_checker_maker.kOpCreationCallstackAttrName(),
            core.op_proto_and_checker_maker.kOpDeviceAttrName()
2040 2041
        }

F
fengjiayi 已提交
2042

X
Xin Pan 已提交
2043
class Operator(object):
2044
    """
2045 2046 2047 2048 2049 2050 2051
    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
2052
        type(str): The type of operator. Default None.
2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072
        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
2073
        Block.append_op or Block._prepend_op instead.
2074 2075 2076 2077

    Examples:
        .. code-block:: python

2078
            import paddle.fluid as fluid
2079
            cur_program = fluid.Program()
2080 2081 2082 2083 2084
            cur_block = cur_program.current_block()
            # var1 += var2 + var3
            cur_block.append_op(type="sum",
                                inputs={"X": [var1, var2, var3]},
                                outputs={"Out": [var1]})
2085
    """
2086
    OP_WITHOUT_KERNEL_SET = {
2087 2088
        'feed', 'fetch', 'recurrent', 'go', 'rnn_memory_helper_grad',
        'conditional_block', 'while', 'send', 'recv', 'listen_and_serv',
2089
        'fl_listen_and_serv', 'ncclInit', 'select', 'checkpoint_notify',
2090 2091 2092
        'gen_bkcl_id', 'c_gen_bkcl_id', 'gen_nccl_id', 'c_gen_nccl_id',
        'c_comm_init', 'c_sync_calc_stream', 'c_sync_comm_stream',
        'queue_generator', 'dequeue', 'enqueue', 'heter_listen_and_serv'
2093
    }
2094

Y
Yu Yang 已提交
2095 2096
    def __init__(self,
                 block,
Y
Yu Yang 已提交
2097
                 desc,
Y
Yu Yang 已提交
2098 2099 2100
                 type=None,
                 inputs=None,
                 outputs=None,
M
minqiyang 已提交
2101
                 attrs=None):
L
lujun 已提交
2102
        if in_dygraph_mode():
2103 2104
            if type is None:
                raise ValueError(
2105
                    "`type` to initialized an Operator can not be None.")
J
Jiabin Yang 已提交
2106
            self._type = type
M
minqiyang 已提交
2107
            self.attrs = attrs if attrs else {}
2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121
        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(
2122
                )] = self.block.program._op_role
2123 2124 2125

            role_var_name = op_maker.kOpRoleVarAttrName()
            if len(self.block.program.
2126 2127
                   _op_role_var) != 0 and role_var_name not in op_attrs:
                op_attrs[role_var_name] = self.block.program._op_role_var
2128 2129 2130 2131 2132 2133 2134 2135

            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(
2136
                    "`type` to initialized an Operator can not be None.")
2137 2138
            else:
                callstack_var_name = op_maker.kOpCreationCallstackAttrName()
2139 2140 2141 2142 2143 2144 2145
                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]))
2146 2147 2148 2149 2150 2151 2152

            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()

2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170
            # 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)

2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183
            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]
2184
                        if not isinstance(in_args, (list, tuple)):
2185 2186 2187 2188 2189 2190
                            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 = []
2191
                        for index, arg in enumerate(in_args):
2192 2193 2194 2195
                            if isinstance(arg, six.string_types):
                                in_arg_names.append(arg)
                            elif isinstance(arg, six.binary_type):
                                in_arg_names.append(arg.decode())
2196
                            elif isinstance(arg, (Variable, core.VarBase)):
2197
                                in_arg_names.append(cpt.to_text(arg.name))
2198
                            else:
2199 2200 2201 2202
                                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."
2203 2204
                                    "but received : %s" %
                                    (in_proto.name, type, arg))
2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228
                        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:
2229 2230 2231 2232
                        if isinstance(arg, six.string_types):
                            out_arg_names.append(arg)
                        else:
                            out_arg_names.append(cpt.to_text(arg.name))
2233
                        # TODO(minqiyang): could we remove variable's op in static mode?
L
lujun 已提交
2234
                        if not in_dygraph_mode():
2235 2236 2237 2238
                            if isinstance(arg, six.string_types):
                                block.var(arg).op = self
                            else:
                                arg.op = self
2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256
                    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
2257
    def _has_kernel(self, op_type):
2258 2259
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
2260
    def to_string(self, throw_on_error):
2261
        """
2262 2263
        Get debug string.

2264
        Args:
2265 2266
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
2267

2268 2269
        Returns:
            str: The debug string.
2270 2271

        """
2272
        protostr = self.desc.serialize_to_string()
2273
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
2274 2275
        return _debug_string_(proto, throw_on_error)

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 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362
    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
2363 2364
                format(outputs=outputs_str, op_type=self.type,
                       inputs=inputs_str, attrs=attrs_str)
2365 2366 2367 2368 2369
        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
2370
    def __str__(self):
2371
        return self._to_readable_code()
2372 2373 2374

    __repr__ = __str__

F
fengjiayi 已提交
2375 2376
    @property
    def type(self):
2377
        return self.desc.type()
F
fengjiayi 已提交
2378 2379

    def input(self, name):
2380
        r"""
2381
        Get the input arguments according to the input parameter name.
2382

2383 2384
        Args:
            name(str): The input parameter name.
2385

2386 2387 2388
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
2389
        """
F
fengjiayi 已提交
2390 2391
        return self.desc.input(name)

W
Wu Yi 已提交
2392
    def _rename_input(self, old_name, new_name):
2393 2394 2395 2396 2397 2398 2399 2400 2401 2402
        """
        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
2403
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
2404

W
Wu Yi 已提交
2405
    def _rename_output(self, old_name, new_name):
2406 2407 2408 2409 2410 2411 2412 2413 2414 2415
        """
        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
2416
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
2417

F
fengjiayi 已提交
2418 2419 2420 2421
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
2422 2423 2424 2425 2426 2427 2428 2429
    @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
2430
    def output(self, name):
2431
        r"""
2432
        Get output arguments by the output parameter name.
2433

2434 2435
        Args:
            name(str): The output parameter name.
2436

2437 2438 2439
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
2440
        """
F
fengjiayi 已提交
2441 2442 2443 2444 2445 2446
        return self.desc.output(name)

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

2447 2448 2449 2450 2451 2452 2453 2454
    @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
2455
    def has_attr(self, name):
2456
        """
2457 2458
        Whether this Operator has the attribute with name or not.

2459
        Args:
2460
            name(str): the attribute name.
2461

2462 2463
        Returns:
            bool: True if has this attribute.
2464 2465

        """
F
fengjiayi 已提交
2466 2467 2468
        return self.desc.has_attr(name)

    def attr_type(self, name):
2469
        """
2470
        Get the type of attribute by attribute's name.
2471

2472 2473
        Args:
            name(str): the attribute name.
2474

2475 2476
        Returns:
            core.AttrType: the attribute type.
2477
        """
F
fengjiayi 已提交
2478 2479
        return self.desc.attr_type(name)

W
Wu Yi 已提交
2480
    def _set_attr(self, name, val):
2481 2482 2483 2484 2485 2486 2487 2488 2489 2490
        """
        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
2491 2492
        self._update_desc_attr(name, val)

2493 2494 2495
    def _remove_attr(self, name):
        self.desc.remove_attr(name)

G
gongweibao 已提交
2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506
    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
2507 2508
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
2509 2510
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
2511
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
2512 2513 2514 2515
        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
2516
            self.desc._set_attr(name, val)
Y
yuyang18 已提交
2517

F
fengjiayi 已提交
2518 2519 2520 2521 2522
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
2523
        """
2524 2525
        Get the attribute by name.

2526
        Args:
2527
            name(str): the attribute name.
2528

2529 2530
        Returns:
            bool|int|str|float|list: The attribute value. The return value
2531 2532
            can be any valid attribute type.
        """
F
fengjiayi 已提交
2533
        return self.desc.attr(name)
Y
Yu Yang 已提交
2534

W
Wu Yi 已提交
2535
    def _block_attr_id(self, name):
2536
        """
G
gongweibao 已提交
2537
        Get the block attribute's id by name.
2538

2539 2540
        Args:
            name(str): the attribute name.
2541

2542 2543
        Returns:
            int: the block index.
2544
        """
W
Wu Yi 已提交
2545
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
2546

W
Wu Yi 已提交
2547
    def _block_attr(self, name):
G
gongweibao 已提交
2548 2549 2550 2551 2552 2553 2554 2555 2556 2557
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
2558
        id = self._block_attr_id(name)
G
gongweibao 已提交
2559 2560 2561
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
2562
    def _blocks_attr(self, name):
G
gongweibao 已提交
2563 2564 2565 2566 2567 2568 2569 2570 2571 2572
        """
        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
2573
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
2574 2575 2576 2577 2578
            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
2579
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
2580 2581 2582 2583 2584 2585 2586 2587 2588 2589
        """
        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
2590
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
2591

J
JiayiFeng 已提交
2592
    def all_attrs(self):
F
fengjiayi 已提交
2593
        """
2594 2595 2596
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
2597
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
2598 2599 2600 2601
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
G
gongweibao 已提交
2602 2603
            attr_type = self.desc.attr_type(n)
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
2604
                attr_map[n] = self._block_attr(n)
G
gongweibao 已提交
2605 2606 2607
                continue

            if attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
2608
                attr_map[n] = self._blocks_attr(n)
G
gongweibao 已提交
2609 2610 2611 2612
                continue

            attr_map[n] = self.attr(n)

F
fengjiayi 已提交
2613 2614
        return attr_map

2615 2616 2617
    def _is_optimize_op(self):
        op_maker = core.op_proto_and_checker_maker
        OPTIMIZE = core.op_proto_and_checker_maker.OpRole.Optimize
2618 2619 2620 2621

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

2622 2623 2624
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(OPTIMIZE):
            return True
2625 2626 2627 2628 2629 2630 2631 2632

        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()):
2633 2634
            return False

2635 2636 2637 2638 2639 2640
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(BACKWARD):
            return True

        return False

Y
Yu Yang 已提交
2641

Y
Yu Yang 已提交
2642
class Block(object):
2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656
    """
    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
2657
        use `Program._create_block()` to create a block.
2658 2659 2660 2661

    Examples:
        .. code-block:: python

2662 2663 2664
            import paddle.fluid as fluid

            cur_program = fluid.Program()
2665 2666 2667 2668 2669 2670 2671 2672 2673
            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
2674
    def __init__(self, program, idx):
Y
Yu Yang 已提交
2675
        self.desc = program.desc.block(idx)
2676
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
2677
        self.ops = list()  # operator list
Y
Yu Yang 已提交
2678
        self.program = program
2679
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
2680

2681
    def __str__(self):
2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727
        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
2728

F
fengjiayi 已提交
2729 2730
    def to_string(self, throw_on_error, with_details=False):
        """
2731 2732
        Get debug string.

F
fengjiayi 已提交
2733 2734
        Args:
            throw_on_error(bool): raise exception when self is not initialized
2735
                when throw_on_error is True.
F
update  
fengjiayi 已提交
2736
            with_details(bool): more details about variables and parameters
2737 2738
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
2739

2740 2741
        Returns:
            str: The debug string.
F
fengjiayi 已提交
2742 2743 2744 2745
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
2746
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
2747 2748
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
2749
            for var in list(self.vars.values()):
F
fengjiayi 已提交
2750
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
2751
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
2752
            for op in self.ops:
F
fengjiayi 已提交
2753 2754
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
2755 2756 2757
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
2758 2759
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
2760 2761
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
2762 2763 2764

    __repr__ = __str__

Y
Yu Yang 已提交
2765 2766
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
2767
        return self.desc.parent
Y
Yu Yang 已提交
2768

Y
Yu Yang 已提交
2769 2770 2771 2772
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
2773
    def _set_forward_block_idx(self, idx):
2774 2775 2776 2777 2778 2779 2780 2781 2782
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
2783
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
2784

2785 2786 2787 2788 2789 2790 2791 2792
    @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
2793 2794
    @property
    def idx(self):
Y
Yu Yang 已提交
2795
        return self.desc.id
Y
Yu Yang 已提交
2796

Q
Qiao Longfei 已提交
2797
    def var(self, name):
2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810
        """
        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.
        """
2811
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
2812 2813 2814
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
2815 2816
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
2817
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
2818
        return v
Q
Qiao Longfei 已提交
2819

X
Xin Pan 已提交
2820
    def _find_var_recursive(self, name):
2821 2822 2823 2824 2825 2826 2827
        """
        Get a Variable by name from this block recursively.

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

        Returns:
X
Xin Pan 已提交
2828
            Variable: the Variable with the giving name. Or None if not found.
2829
        """
Y
Yu Yang 已提交
2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853
        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
2854
        return None
Y
Yu Yang 已提交
2855

X
Xin Pan 已提交
2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874
    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
2875

Q
Qiao Longfei 已提交
2876
    def all_parameters(self):
2877
        return list(self.iter_parameters())
2878

2879
    def iter_parameters(self):
M
minqiyang 已提交
2880
        return (item[1] for item in six.iteritems(self.vars)
2881
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
2882

Y
Yu Yang 已提交
2883
    def create_var(self, *args, **kwargs):
L
Leo Chen 已提交
2884 2885 2886
        if in_dygraph_mode():
            var = _varbase_creator(*args, **kwargs)
        else:
2887 2888 2889
            var = Variable(block=self, *args, **kwargs)
            if 'initializer' in kwargs:
                kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
2890
        return var
Y
Yu Yang 已提交
2891

Q
Qiao Longfei 已提交
2892 2893 2894
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
2895
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
2896 2897
        """
        Rename variable in vars and ops' inputs and outputs
2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909

        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
2910
        """
M
minqiyang 已提交
2911 2912
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
2913

T
typhoonzero 已提交
2914
        if not self.has_var(name):
2915
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
2916 2917
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
2918
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
2919 2920 2921 2922 2923 2924
            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
2925
            var_type = "Variable"
T
wip  
typhoonzero 已提交
2926 2927 2928 2929
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
2930
        orig_var_type = v.type
M
minqiyang 已提交
2931
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
2932
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
2933
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
2934
        if var_type == "Parameter":
L
Leo Chen 已提交
2935 2936
            if in_dygraph_mode():
                var = ParamBase(
2937 2938 2939 2940 2941 2942 2943 2944 2945 2946
                    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
2947 2948
                var = Parameter(
                    self,
2949 2950 2951 2952 2953 2954 2955 2956 2957
                    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
2958
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
2959 2960
            var = Variable(
                self,
T
typhoonzero 已提交
2961
                type=orig_var_type,
T
wip  
typhoonzero 已提交
2962 2963 2964 2965
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
2966
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
2967 2968 2969
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
2970
        self._sync_with_cpp()
2971
        return var
T
typhoonzero 已提交
2972

2973 2974 2975
    def _remove_var(self, name, sync=True):
        if sync == True:
            self._sync_with_cpp()
M
minqiyang 已提交
2976
        self.desc._remove_var(cpt.to_bytes(name))
2977 2978
        del self.vars[name]

Y
Yu Yang 已提交
2979 2980
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
2981
        param = None
L
Leo Chen 已提交
2982
        if in_dygraph_mode():
2983
            param = ParamBase(*args, **kwargs)
L
Leo Chen 已提交
2984 2985
        else:
            param = Parameter(global_block, *args, **kwargs)
2986 2987 2988 2989 2990 2991
            # 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
2992
        if 'initializer' in kwargs:
2993 2994 2995 2996 2997

            def _is_inited_by(block, var):
                init_ops = []
                for op in block.ops:
                    if var.name in op.output_arg_names:
2998
                        # In startup_program, "c_broadcast" and "c_sync_comm_stream"
T
tangwei12 已提交
2999
                        # are treated as initialization ops that cause error.
3000 3001 3002
                        # 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
3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013
                        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:
3014
                # TODO already inited, do nothing, should log a warning
3015 3016 3017
                pass
            else:
                initializer(param, self)
Q
Qiao Longfei 已提交
3018
        return param
Y
Yu Yang 已提交
3019

Y
Yu Yang 已提交
3020
    def append_op(self, *args, **kwargs):
3021 3022 3023 3024 3025 3026
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
L
lujun 已提交
3027
        if in_dygraph_mode():
3028
            attrs = kwargs.get("attrs", {})
J
Jiabin Yang 已提交
3029
            type = kwargs.get("type", None)
3030 3031 3032
            op = Operator(
                block=self,
                desc=None,
J
Jiabin Yang 已提交
3033
                type=type,
M
minqiyang 已提交
3034 3035
                inputs=None,
                outputs=None,
3036
                attrs=attrs)
3037

M
minqiyang 已提交
3038 3039 3040
            # record ops in tracer rather than blocks
            #
            # TODO(minqiyang): add op stop_gradient support in static mode too.
L
lujun 已提交
3041
            # currently, we only support stop_gradient in dygraph mode.
J
Jiabin Yang 已提交
3042 3043

            _dygraph_tracer().trace_op(type,
M
minqiyang 已提交
3044
                                       kwargs.get("inputs", {}),
J
Jiabin Yang 已提交
3045 3046
                                       kwargs.get("outputs", {}), attrs
                                       if attrs else {},
M
minqiyang 已提交
3047
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
3048
        else:
3049 3050 3051 3052 3053 3054 3055 3056 3057
            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
3058
            self.ops.append(op)
M
minqiyang 已提交
3059

3060 3061
        return op

W
Wu Yi 已提交
3062
    def _insert_op(self, index, *args, **kwargs):
3063 3064 3065 3066 3067 3068 3069 3070 3071
        """
        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
3072 3073
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
3074 3075 3076 3077
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094
    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):
3095 3096 3097 3098 3099 3100 3101 3102 3103
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
3104 3105
        if sync == True:
            self._sync_with_cpp()
W
Wu Yi 已提交
3106
        self.desc._remove_op(index, index + 1)
3107 3108
        del self.ops[index]

W
Wu Yi 已提交
3109
    def _slice_ops(self, start, end):
3110 3111 3112 3113 3114 3115 3116 3117 3118 3119
        """
        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
3120
        return self.ops[start:end]
Y
Yancey1989 已提交
3121

W
Wu Yi 已提交
3122
    def _prepend_op(self, *args, **kwargs):
L
lujun 已提交
3123
        if in_dygraph_mode():
J
Jiabin Yang 已提交
3124 3125
            type = kwargs.get("type", None)
            attrs = kwargs.get("attrs", {})
3126
            op = Operator(
J
Jiabin Yang 已提交
3127
                self, None, type=type, inputs=None, outputs=None, attrs=attrs)
M
minqiyang 已提交
3128

J
Jiabin Yang 已提交
3129
            _dygraph_tracer().trace_op(type,
M
minqiyang 已提交
3130
                                       kwargs.get("inputs", {}),
J
Jiabin Yang 已提交
3131 3132
                                       kwargs.get("outputs", {}), attrs
                                       if attrs else {},
M
minqiyang 已提交
3133
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
3134
        else:
3135 3136 3137 3138 3139 3140 3141 3142
            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
3143
            self.ops.insert(0, op)
3144

Y
Yu Yang 已提交
3145 3146
        return op

W
Wu Yi 已提交
3147
    def _sync_with_cpp(self):
3148
        """
3149 3150
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
3151
        """
Q
Qiao Longfei 已提交
3152 3153 3154 3155 3156
        # 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())

3157
        # sync variables removed from c++ end
3158
        for var in list(self.vars.keys()):
M
minqiyang 已提交
3159
            if not self.desc.find_var(cpt.to_bytes(var)):
3160 3161
                self.vars.pop(var)

Q
Qiao Longfei 已提交
3162
        # sync operators from cpp
3163 3164 3165 3166
        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
3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182
        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
3183 3184 3185 3186 3187

        # 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
3188
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
3189 3190 3191 3192 3193 3194 3195

        # 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)

3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208
        # 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
3209 3210 3211 3212
        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
3213
    def _copy_param_info_from(self, other):
3214
        """
3215 3216
        Copy the information of parameters from the other block.

3217
        Args:
3218 3219 3220 3221 3222
            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.
3223 3224 3225 3226 3227

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
3228 3229
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
3230
        for p in other.iter_parameters():
3231 3232 3233
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
3234 3235
                # if the Parameter is pruned, v may be None
                continue
3236
            assert isinstance(v, Variable)
3237
            new_p = None
L
Leo Chen 已提交
3238 3239
            if in_dygraph_mode():
                new_p = ParamBase(
3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250
                    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
3251 3252
                new_p = Parameter(
                    block=self,
3253 3254 3255
                    shape=v.shape,
                    dtype=v.dtype,
                    type=v.type,
3256 3257
                    lod_level=v.lod_level
                    if v.type == core.VarDesc.VarType.LOD_TENSOR else None,
3258 3259 3260 3261 3262 3263
                    stop_gradient=p.stop_gradient,
                    trainable=p.trainable,
                    optimize_attr=p.optimize_attr,
                    regularizer=p.regularizer,
                    error_clip=p.error_clip,
                    name=v.name)
3264 3265
            self.vars[new_p.name] = new_p

3266
    def _clone_variable(self, var, force_persistable=True):
3267 3268
        """
        Clone a variable into current block.
3269

3270 3271
        Args:
            var: the variable to be cloned.
3272 3273 3274
            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.
3275 3276

        Returns:
3277
            Variable: the new  variable cloned from 'var' in current block.
3278 3279
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
3280 3281 3282 3283 3284
        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
3285 3286
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
3287
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
3288 3289 3290 3291 3292 3293
        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,
3294
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
3295 3296
                is_data=var.is_data,
                need_check_feed=var.desc.need_check_feed())
T
update  
typhoonzero 已提交
3297 3298 3299 3300 3301 3302 3303
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
3304
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
3305 3306
                is_data=var.is_data,
                need_check_feed=var.desc.need_check_feed())
T
update  
typhoonzero 已提交
3307
        return ret_var
3308

Y
Yu Yang 已提交
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 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404
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()

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

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

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

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

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

        Args:
3428
            node(IrNode): the node being appended.
3429
        """
3430
        self.node.append_input(node.node)
3431 3432 3433 3434 3435 3436 3437 3438

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

3439
    def remove_output_by_id(self, node_id):
3440 3441 3442 3443 3444 3445
        """
        Remove a node from outputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
3446
        self.node.remove_output(node_id)
3447

3448
    def remove_output(self, node):
3449 3450 3451 3452
        """
        Remove a node from outputs.

        Args:
3453
            node(IrNode): the node being removed.
3454
        """
3455
        self.node.remove_output(node.node)
3456

3457
    def append_output(self, node):
3458 3459 3460 3461
        """
        Append a node in outputs.

        Args:
3462
            node(IrNode): the node being appended.
3463
        """
3464
        self.node.append_output(node.node)
3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511

    @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
3512
            "The node variable description can not be None."
3513 3514 3515 3516 3517 3518 3519 3520 3521 3522
        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
3523
            "The node variable description can not be None."
3524 3525
        return self.node.var().persistable()

3526 3527 3528 3529 3530 3531 3532 3533
    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
3534
            "The node variable description can not be None."
3535 3536 3537 3538 3539 3540 3541 3542 3543 3544
        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
3545
            "The node variable description can not be None."
3546 3547 3548 3549 3550 3551 3552 3553 3554 3555
        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
3556
            "The node variable description can not be None."
3557 3558
        return self.node.var().shape()

3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605
    @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
3606
            "The node operator description can not be None."
3607 3608
        self.node.op()._rename_input(old_input_name, new_input_name)

3609 3610 3611 3612 3613 3614 3615 3616 3617
    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
3618
            "The node operator description can not be None."
3619 3620
        self.node.op()._rename_output(old_output_name, new_output_name)

3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631
    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
3632
            "The node operator description can not be None."
3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645
        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
3646
            "The node operator description can not be None."
3647 3648 3649 3650 3651 3652 3653 3654 3655 3656
        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
3657
            "The node operator description can not be None."
3658 3659
        return self.node.op().set_type(new_type)

3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674
    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
3675
            "The node operator description can not be None."
3676 3677 3678 3679
        desc = self.node.op()
        if isinstance(val, Block):
            desc.set_block_attr(name, val.desc)
        elif isinstance(val, list) and val and \
3680
                all(isinstance(v, Block) for v in val):
3681 3682
            desc.set_blocks_attr(name, [v.desc for v in val])
        elif isinstance(val, core.BlockDesc) or \
3683
                isinstance(val, core.ProgramDesc):
3684 3685 3686 3687
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)

3688 3689 3690 3691 3692 3693 3694 3695
    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
3696
            "The node operator description can not be None."
3697 3698 3699 3700 3701 3702 3703 3704 3705 3706
        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
3707
            "The node operator description can not be None."
3708 3709
        return self.node.op().output_arg_names()

3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730
    @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]


3731 3732
class IrGraph(object):
    """
3733
    Python IrGraph. Beneath it is a core.Graph, which is used for
3734
    creating a c++ Ir Pass Graph. An IrGraph is just a graph view of
3735 3736
    a Program. In an IrGraph, both Variables and Operators are graph
    nodes.
3737 3738 3739 3740
    """

    def __init__(self, graph, for_test=False):
        """
3741 3742
        Construct an IrGraph using core.Graph.

3743 3744 3745 3746 3747 3748 3749 3750 3751
        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

3752 3753 3754 3755
    def clone(self):
        """
        Create a new and duplicated IrGraph.

3756 3757 3758
        Warns:
            The method only clones the graph structure, not its attributes.

3759 3760 3761
        Returns:
            IrGraph: A new and duplicated graph.
        """
3762
        g = self.graph.clone()
3763 3764
        return IrGraph(g, self._for_test)

3765
    def is_test(self):
3766 3767 3768
        """
        If the graph is used for testing, the function returns true. Otherwise, returns false.
        """
3769 3770
        return self._for_test

W
WangZhen 已提交
3771
    def all_nodes(self):
3772 3773 3774
        """
        Return all nodes included in the graph as a set.
        """
3775
        return {IrNode(node) for node in self.graph.nodes()}
3776

3777
    def all_var_nodes(self):
3778 3779 3780
        """
        Return all variable nodes included in the graph as a set.
        """
3781
        return {IrVarNode(node) for node in self.graph.nodes() if node.is_var()}
3782

3783
    def all_persistable_nodes(self):
3784 3785 3786
        """
        Return all persistable variable nodes included in the graph as a set.
        """
W
WangZhen 已提交
3787 3788 3789 3790 3791
        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)
3792
        return {IrVarNode(p) for p in persistable_nodes}
W
WangZhen 已提交
3793

3794
    def all_op_nodes(self):
3795 3796 3797
        """
        Return all operator nodes included in the graph as a set.
        """
3798
        return {IrOpNode(node) for node in self.graph.nodes() if node.is_op()}
3799

3800
    def create_persistable_node(self, name, var_type, shape, var_dtype):
3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811
        """
        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:
3812
            IrVarNode: the created persistable variable node.
3813
        """
3814 3815 3816 3817 3818
        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)
3819
        return IrVarNode(self.graph.create_var_node(var_desc))
3820 3821

    def create_var_node(self, name, var_type, shape, var_dtype):
3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832
        """
        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:
3833
            IrVarNode: the created variable node.
3834 3835
        """

3836 3837 3838 3839
        var_desc = core.VarDesc(name)
        var_desc.set_type(var_type)
        var_desc.set_shape(shape)
        var_desc.set_dtype(var_dtype)
3840
        return IrVarNode(self.graph.create_var_node(var_desc))
3841

3842 3843 3844 3845 3846 3847
    def create_control_dep_var(self):
        """
        create a control var
        """
        return IrVarNode(self.graph.create_control_dep_var())

3848
    def create_var_node_from_desc(self, var_desc):
3849 3850 3851 3852 3853 3854 3855 3856
        """
        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:
3857
            IrVarNode: the created variable node.
3858
        """
3859
        return IrVarNode(self.graph.create_var_node(var_desc))
3860 3861

    def create_op_node(self, op_type, attrs, inputs, outputs):
3862 3863 3864 3865 3866 3867 3868
        """
        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
3869
            outputs(dict): the outputs of the operator node.
3870 3871

        Returns:
3872
            IrOpNode: the created operator node.
3873
        """
3874 3875
        op_desc = core.OpDesc()
        op_desc.set_type(op_type)
3876
        for attr, value in six.iteritems(attrs):
3877
            self._update_desc_attr(op_desc, attr, value)
3878
        for input_name, var_nodes in six.iteritems(inputs):
3879 3880 3881 3882
            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])
3883
        for output_name, var_nodes in six.iteritems(outputs):
3884 3885 3886 3887
            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])
3888
        return IrOpNode(self.graph.create_op_node(op_desc))
3889 3890

    def create_op_node_from_desc(self, op_desc):
3891 3892 3893 3894 3895 3896 3897
        """
        Create a operator node by using an existing OpDesc in the graph.

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

        Returns:
3898
            IrOpNode: the created operator node.
3899
        """
3900
        return IrOpNode(self.graph.create_op_node(op_desc))
3901 3902

    def update_input_link(self, old_input_node, new_input_node, op_node):
3903 3904 3905 3906
        """
        Update the input's link of a operator node.

        Args:
3907 3908 3909
            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.
3910
        """
3911
        assert old_input_node.node in self.graph.nodes() and new_input_node.node in \
T
tangwei12 已提交
3912
            self.graph.nodes() and op_node.node in self.graph.nodes(), \
3913
            'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.'
3914 3915 3916 3917
        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)
3918
        op_node.rename_input(old_input_node.name(), new_input_node.name())
3919

3920 3921 3922 3923 3924 3925 3926 3927 3928 3929
    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
3930
            self.graph.nodes() and op_node.node in self.graph.nodes(), \
3931
            'The three arguments(old_output_node &new_output_node &op_node) must be in the graph nodes.'
3932 3933 3934 3935 3936 3937
        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())

3938
    def link_to(self, node_in, node_out):
3939 3940 3941 3942
        """
        Connect two nodes.

        Args:
3943 3944
            node_in(IrNode): the input node.
            node_out(IrNode): the output node.
3945
        """
3946
        assert node_in.node in self.graph.nodes() and node_out.node in self.graph.nodes(), \
W
WangZhen 已提交
3947
            'The two arguments(node_in&node_out) must be in the graph nodes.'
3948 3949
        node_in.append_output(node_out)
        node_out.append_input(node_in)
3950 3951

    def safe_remove_nodes(self, remove_nodes):
3952 3953 3954 3955 3956 3957 3958
        """
        Remove nodes safely since links connected to these removed nodes are
        also removed.

        Args:
            remove_nodes(set): the nodes prepared to be removed.
        """
3959
        if not isinstance(remove_nodes, set):
W
WangZhen 已提交
3960 3961 3962 3963
            if isinstance(remove_nodes, Iterable):
                remove_nodes = set(remove_nodes)
            else:
                remove_nodes = {remove_nodes}
3964 3965
        original_nodes = {n.node for n in remove_nodes}
        core.graph_safe_remove_nodes(self.graph, original_nodes)
3966

Z
Zhen Wang 已提交
3967 3968 3969 3970 3971 3972 3973 3974
    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] = [
3975
                            self._find_node_by_name(node.inputs, each_var_name)
Z
Zhen Wang 已提交
3976 3977 3978 3979
                        ]
                for each_var_name in node.op().output_arg_names():
                    if each_var_name not in var_nodes:
                        var_nodes[each_var_name] = [
3980
                            self._find_node_by_name(node.outputs, each_var_name)
Z
Zhen Wang 已提交
3981 3982 3983
                        ]
                    else:
                        var_nodes[each_var_name].append(
3984 3985
                            self._find_node_by_name(node.outputs,
                                                    each_var_name))
Z
Zhen Wang 已提交
3986 3987
        self.graph.resolve_hazard(var_nodes)

W
WangZhen 已提交
3988
    def has_circle(self):
3989 3990 3991 3992 3993 3994
        """
        Check if the graph has a circle.

        Returns:
            bool: True if the graph has a circle else False.
        """
W
WangZhen 已提交
3995 3996 3997
        return core.has_circle(self.graph)

    def graph_num(self):
3998 3999 4000 4001 4002 4003
        """
        Count the number of unconnected graphs in this graph.

        Returns:
            int: the number of unconnected graphs.
        """
W
WangZhen 已提交
4004 4005 4006
        return core.graph_num(self.graph)

    def topology_sort(self):
4007 4008 4009
        """
        Perform the topology sort operation on the graph.

T
tianshuo78520a 已提交
4010
        Notes: the `graph` can not contain a circle.
4011 4012

        Returns:
Z
Zhen Wang 已提交
4013
            list(IrNode): nodes in topology order.
4014
        """
4015
        ordered_nodes = core.topology_sort(self.graph)
Z
Zhen Wang 已提交
4016
        return [IrNode(n) for n in ordered_nodes]
W
WangZhen 已提交
4017 4018

    def build_adjacency_list(self):
4019 4020 4021 4022
        """
        Build an adjacency list of operations for the `graph`.

        Returns:
4023
            dict{IrNode: set(IrNode)}: the adjacency list.
4024
        """
4025 4026 4027 4028 4029
        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
4030

4031 4032 4033 4034 4035 4036 4037 4038
    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.
4039
            marked_nodes(set(IrNode)): nodes that are needed to be marked.
4040 4041 4042 4043 4044
            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.
        """

4045 4046
        def _convert_to_pdf(dot_file_path):
            pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf'
T
tangwei12 已提交
4047 4048 4049
            exited_code = subprocess.call(
                'dot -Tpdf ' + dot_file_path + ' -o ' + pdf_save_path,
                shell=True)
4050 4051 4052 4053 4054
            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))

4055
        remove_ctr_vars = set()
4056
        if remove_ctr_var:
4057
            for node in self.all_var_nodes():
4058 4059 4060
                if node.is_ctrl_var():
                    remove_ctr_vars.add(node)
            self.safe_remove_nodes(remove_ctr_vars)
4061 4062
        print('Total ops num = {}.'.format(len(self.all_op_nodes())))

4063 4064
        if marked_nodes is not None:
            if not isinstance(marked_nodes, set):
4065 4066 4067 4068 4069 4070
                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}
4071 4072 4073 4074
            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)
4075 4076
        if not os.path.exists(save_path):
            os.makedirs(save_path)
4077 4078 4079 4080 4081 4082 4083
        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):
4084 4085 4086
        """
        Convert the graph into a Program.

Z
Zhen Wang 已提交
4087
        WARN: When the graph includes backward operator nodes, the
4088 4089 4090 4091 4092 4093
        conversion process may be failed. Usually, this function is
        only used to convert a test graph.

        Returns:
            Program: a program converted from the graph.
        """
4094
        convert_pass = core.get_pass('graph_to_program_pass')
4095 4096
        desc = core.ProgramDesc()
        convert_pass.set_not_owned('program', desc)
4097 4098 4099 4100
        convert_pass.apply(self.graph)
        program = Program._construct_from_desc(desc)
        return program

4101 4102 4103 4104 4105 4106 4107 4108 4109 4110 4111
    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

4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127
    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
4128
class Program(object):
D
dzhwinter 已提交
4129
    """
4130
    Create Python Program.  It has at least one :ref:`api_guide_Block_en`, when the
4131
    control flow op like conditional_block, while :ref:`api_paddle_fluid_layers_While` is included,
J
Jiabin Yang 已提交
4132
    it will contain nested block.
4133

J
Jiabin Yang 已提交
4134 4135 4136
    Please reference the
    `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_
    for details.
D
dzhwinter 已提交
4137

J
Jiabin Yang 已提交
4138
    A set of Program usually contains startup program and main program.
J
Jiabin Yang 已提交
4139
    A startup program is set to contain some initial work, eg. initialize the ``Parameter``, and the main
J
Jiabin Yang 已提交
4140 4141 4142 4143 4144 4145 4146
    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
4147
    **Notes**:
4148 4149 4150
        **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
4151 4152

    Returns:
J
Jiabin Yang 已提交
4153
        Program: An empty Program.
D
dzhwinter 已提交
4154 4155

    Examples:
4156 4157
        .. code-block:: python

4158 4159 4160 4161
            import paddle
            import paddle.static as static

            paddle.enable_static()
4162

4163 4164 4165 4166 4167
            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')
4168
                z = static.nn.fc(name="fc", x=x, size=10, activation="relu")
4169 4170 4171

            print("main program is: {}".format(main_program))
            print("start up program is: {}".format(startup_program))
D
dzhwinter 已提交
4172 4173 4174

    """

4175 4176
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
4177 4178
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
4179 4180
        global global_prog_seed
        self._seed = global_prog_seed
Y
yuyang18 已提交
4181
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
4182
        self.__op_role_var = []
T
tangwei12 已提交
4183

4184 4185
        # for distribute training
        # _is_distributed = True if under distributed training
T
tangwei12 已提交
4186
        self._is_distributed = False
4187
        # _is_chief = True if the trainer is the first one, usually No.0
T
tangwei12 已提交
4188
        self._is_chief = False
4189 4190 4191
        # _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
4192
        self._endpoints = []
4193 4194 4195
        # if current role is parameter server, the _ps_endpoint is its "ip:port"
        self._ps_endpoint = None
        # trainers_endpoints, it is used for distribution.
4196
        self._trainers_endpoints = []
4197
        # the distributed lookup table names
T
tangwei12 已提交
4198
        self._distributed_lookup_table = None
4199 4200 4201

        # use Deep gradient comrepssion or not
        self._enable_dgc = False
4202 4203
        self._use_lamb = False

4204 4205 4206
        self._nccl_comm_num = 1
        self._use_hierarchical_allreduce = False
        self._hierarchical_allreduce_inter_nranks = 0
4207

4208 4209 4210
        # 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
4211
        self._program_config = None
4212

H
hutuxian 已提交
4213 4214 4215
        # assigned if this program has been parsed by a pipeline optimizer
        self._pipeline_opt = None

4216 4217 4218
        # appending gradients times
        self._appending_grad_times = 0

4219 4220 4221 4222
        # identifier for auto checkpoint
        self._auto_checkpoint_name = unique_name.generate(
            "__auto_checkpoint_program__")

4223 4224 4225
        # compiled program, i.e. Graph
        self._graph = None

4226 4227 4228 4229 4230 4231 4232 4233 4234 4235
    def global_seed(self, seed=0):
        """
        Set global seed for Program

        Returns:
            None.

        Examples:
            .. code-block:: python

4236 4237
                import paddle
                import paddle.static as static
4238

4239 4240 4241
                paddle.enable_static()

                prog = static.default_main_program()
4242 4243 4244 4245 4246
                print(prog.random_seed)
                ## 0
                ## the default random seed is 0

                prog.global_seed(102)
4247
                prog1 = static.default_main_program()
4248 4249 4250 4251 4252 4253 4254 4255
                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
4256
    @property
4257
    def _op_role(self):
Y
yuyang18 已提交
4258 4259 4260 4261 4262 4263 4264 4265
        """
        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
4266
        parameter gradient of backward (use :code:`_op_role_var` to get this
Y
yuyang18 已提交
4267 4268 4269 4270
        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
4271 4272
        return self._current_role

4273 4274
    @_op_role.setter
    def _op_role(self, role):
Y
yuyang18 已提交
4275 4276 4277
        self._current_role = role

    @property
4278
    def _op_role_var(self):
Y
yuyang18 已提交
4279
        """
4280
        The auxiliary variables for :code:`_op_role` property.
Y
yuyang18 已提交
4281

4282
        See Also: :code:`Program._op_role`'s documentation for details.
Y
yuyang18 已提交
4283 4284 4285

        Notes: This is a very low-level API. Users should not use it directly.
        """
4286
        return self.__op_role_var
Y
yuyang18 已提交
4287

4288
    @signature_safe_contextmanager
4289 4290 4291 4292 4293
    def _backward_role_guard(self):
        tmp_role = self._current_role

        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Backward
4294 4295 4296 4297
        try:
            yield
        finally:
            self._current_role = tmp_role
4298

S
rename  
sneaxiy 已提交
4299
    @signature_safe_contextmanager
W
Wu Yi 已提交
4300
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
4301 4302 4303 4304 4305 4306 4307
        """
        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:
4308
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
4309 4310 4311

        Examples:

4312
            >>> import paddle.fluid as fluid
Y
yuyang18 已提交
4313
            >>> p, g = backward(...)
W
Wu Yi 已提交
4314
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
4315 4316
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
4317
        tmp_role = self._current_role
4318
        tmp_var = self.__op_role_var
X
Xin Pan 已提交
4319

Y
yuyang18 已提交
4320 4321
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
4322
        self.__op_role_var = [
4323 4324 4325
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
4326 4327 4328 4329 4330
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
Y
Yu Yang 已提交
4331

S
rename  
sneaxiy 已提交
4332
    @signature_safe_contextmanager
X
Xin Pan 已提交
4333
    def _lr_schedule_guard(self, is_with_opt=False):
4334 4335 4336 4337 4338 4339 4340
        """
        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
4341 4342 4343 4344
        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.
4345 4346 4347

        Examples:

4348
            >>> import paddle.fluid as fluid
4349 4350 4351 4352
            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
4353 4354

        tmp_role = self._current_role
4355
        tmp_var = self.__op_role_var
4356

4357 4358
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
4359 4360
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
4361
        # TODO(typhoonzero): how to set target learning rate var
4362
        self.__op_role_var = []
4363 4364 4365 4366 4367
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
4368

4369
    def __str__(self):
Y
yuyang18 已提交
4370 4371 4372 4373 4374 4375 4376 4377 4378
        """
        Get the protobuf debug string of this Program.

        Returns:
            (str): The protobuf debug string.

        Raises:
            ValueError: If any of required fields is not set.
        """
4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398
        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

4399 4400
            import paddle
            import paddle.static as static
4401

4402 4403 4404
            paddle.enable_static()

            cur_program = static.Program()
4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420
            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)
4421
            program_str += '\n'
4422
        return program_str
Y
Yang Yang(Tony) 已提交
4423

F
fengjiayi 已提交
4424 4425 4426
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
4427

J
Jiabin Yang 已提交
4428 4429 4430
        Args:

            throw_on_error (bool): raise Value error when any of required fields is not set.
F
fengjiayi 已提交
4431

J
Jiabin Yang 已提交
4432
            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
4433

H
haowang101779990 已提交
4434
        Returns:
J
Jiabin Yang 已提交
4435
            str: The debug string describe current Program.
Y
yuyang18 已提交
4436 4437

        Raises:
J
Jiabin Yang 已提交
4438
            ValueError: If any of required fields is not set and throw_on_error is True.
F
fengjiayi 已提交
4439

4440 4441 4442
        Examples:
            .. code-block:: python

4443 4444 4445 4446
                import paddle
                import paddle.static as static

                paddle.enable_static()
4447

4448 4449 4450
                prog = static.default_main_program()
                x = static.data(name="X", shape=[2,3], dtype="float32")
                pred = static.nn.fc(x, size=3)
4451
                prog_string = prog.to_string(throw_on_error=True, with_details=False)
4452
                prog_string_with_details = prog.to_string(throw_on_error=False, with_details=True)
T
tianshuo78520a 已提交
4453
                print("program string without detail: {}".format(prog_string))
4454
                print("program string with detail: {}".format(prog_string_with_details))
F
fengjiayi 已提交
4455
        """
4456 4457 4458 4459 4460 4461 4462 4463 4464
        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
4465 4466 4467 4468 4469 4470
        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()
4471 4472
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
4473 4474
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
4475

W
Wu Yi 已提交
4476
    def _get_desc(self):
Y
yuyang18 已提交
4477 4478 4479 4480 4481 4482 4483
        """
        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.
        """
4484 4485
        return self.desc

X
version  
Xin Pan 已提交
4486 4487 4488
    def _version(self):
        return self.desc._version()

4489
    def clone(self, for_test=False):
Y
yuyang18 已提交
4490
        """
4491 4492 4493 4494
        .. 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
4495

4496
        Create a new Program with forward content of original one when ``for_test=True``.
4497
        Create a new Program as same as the original one when ``for_test=False``.
4498

4499
        Some operators, e.g., :ref:`api_paddle_fluid_layers_batch_norm` , behave differently between
Y
yuyang18 已提交
4500 4501 4502
        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`.
4503

4504 4505
        * 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.
4506 4507
          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
4508
          recommend you to use :code:`clone` before using :code:`Opimizer.minimize`.
Y
yuyang18 已提交
4509

J
Jiabin Yang 已提交
4510
        For Example:
4511
          ::
L
Luo Tao 已提交
4512

4513 4514 4515 4516 4517 4518
            import paddle
            import paddle.static as static

            paddle.enable_static()

            img = static.data(name='image', shape=[None, 784])
4519
            pred = static.nn.fc(x=img, size=10, actvation='relu')
4520
            loss = paddle.mean(pred)
4521
            # Here we use clone before Momentum
4522 4523
            test_program = static.default_main_program().clone(for_test=True)
            optimizer = paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
4524
            optimizer.minimize(loss)
4525

J
Jiabin Yang 已提交
4526
        Args:
4527

4528 4529
            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` .
4530

J
Jiabin Yang 已提交
4531
        Returns:
4532
            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``
4533

Y
yuyang18 已提交
4534 4535 4536

        Examples:

4537 4538 4539 4540 4541 4542 4543
            .. 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`:

4544 4545 4546 4547 4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559
            .. 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))


4560
            1. To clone a test program, the sample code is:
4561 4562 4563
                .. code-block:: python

                    import six
4564 4565 4566 4567 4568 4569
                    import paddle
                    import paddle.static as static
                    import paddle.utils as utils
                    import paddle.nn.functional as F

                    paddle.enable_static()
4570 4571 4572 4573 4574 4575 4576 4577 4578 4579 4580 4581

                    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))

4582 4583
                    train_program = static.Program()
                    startup_program = static.Program()
J
Jiabin Yang 已提交
4584 4585 4586

                    # startup_program is used to do some parameter init work,
                    # and main program is used to hold the network
4587 4588 4589
                    with static.program_guard(train_program, startup_program):
                        with utils.unique_name.guard():
                            img = static.data(name='image', shape=[None, 784])
4590
                            hidden = static.nn.fc(x=img, size=200, activation='relu')
4591 4592
                            hidden = F.dropout(hidden, p=0.5)
                            loss = F.cross_entropy(
4593
                                input=static.nn.fc(x=hidden, size=10, activation='softmax'),
4594 4595
                                label=static.data(name='label', shape=[1], dtype='int64'))
                            avg_loss = paddle.mean(loss)
4596
                            test_program = train_program.clone(for_test=True)
4597
                    print_prog(test_program)
J
Jiabin Yang 已提交
4598 4599 4600 4601

                    # 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

4602
                    # In Paddle we will share weights by using the same Tensor name. In train and test program
J
Jiabin Yang 已提交
4603 4604 4605 4606
                    # 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.

4607 4608 4609
                    with static.program_guard(train_program, startup_program):
                        with utils.unique_name.guard():
                            sgd = paddle.optimizer.SGD(learning_rate=1e-3)
4610 4611 4612
                            sgd.minimize(avg_loss)


4613
            2. The clone method can be avoid if you create program for training and program for testing individually.
4614 4615 4616
                .. code-block:: python

                    import six
4617 4618 4619 4620 4621 4622
                    import paddle
                    import paddle.static as static
                    import paddle.utils as utils
                    import paddle.nn.functional as F

                    paddle.enable_static()
4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633

                    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))
4634

4635
                    def network():
4636
                        img = static.data(name='image', shape=[None, 784])
4637
                        hidden = static.nn.fc(x=img, size=200, activation='relu')
4638 4639
                        hidden = F.dropout(hidden, p=0.5)
                        loss = F.cross_entropy(
4640
                            input=static.nn.fc(x=hidden, size=10, activation='softmax'),
4641 4642
                            label=static.data(name='label', shape=[1], dtype='int64'))
                        avg_loss = paddle.mean(loss)
4643 4644
                        return avg_loss

4645 4646 4647 4648 4649
                    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():
4650
                            avg_loss = network()
4651
                            sgd = paddle.optimizer.SGD(learning_rate=1e-3)
4652
                            sgd.minimize(avg_loss)
4653
                    # the test startup program is not used.
4654 4655
                    with static.program_guard(test_program_2, startup_program_2):
                        with utils.unique_name.guard():
4656 4657
                            avg_loss = network()
                    print_prog(test_program_2)
4658

4659
            The two code snippets above will generate and print same programs.
4660
        """
4661

T
tangwei12 已提交
4662
        # NOTE(zhiqiu): we sync the original program first, since its program may diff with
4663 4664 4665
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

4666
        pruned_origin_block_id_map = None
4667
        if for_test:
4668 4669 4670 4671 4672 4673 4674 4675 4676
            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)
4677
        else:
4678
            p = Program()
G
gongweibao 已提交
4679 4680
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
4681
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
4682 4683 4684
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
4685 4686

            p._current_role = self._current_role
4687
            p.__op_role_var = self.__op_role_var
4688
            p._appending_grad_times = self._appending_grad_times
4689 4690
            if hasattr(self, 'lr_sheduler'):
                p.lr_sheduler = self.lr_sheduler
G
gongweibao 已提交
4691

T
tangwei12 已提交
4692
            # NOTE(zhiqiu): we sync the cloned program, to update its program by
4693
            # its desc.
W
Wu Yi 已提交
4694
            p._sync_with_cpp()
4695

W
Wu Yi 已提交
4696
        p._copy_param_info_from(self)
4697
        p._copy_data_info_from(self, pruned_origin_block_id_map)
4698
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
4699
        return p
4700

4701
    def _prune(self, targets):
Y
yuyang18 已提交
4702 4703 4704 4705 4706 4707 4708 4709
        """
        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:
4710
            targets(list|Variable|Operator): A list of variables, operators, or variable names
Y
yuyang18 已提交
4711 4712 4713 4714
                need to be pruned

        Returns:
            Program:  A new, pruned program.
4715
        """
4716
        return self._prune_with_input([], targets)
4717 4718

    def _prune_with_input(self, feeded_var_names, targets):
Y
yuyang18 已提交
4719
        """
4720 4721 4722 4723 4724 4725 4726 4727 4728 4729
        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()
4730
            targets(list|Variable|Operator): A list of variables, operators, or variable names
4731 4732 4733 4734 4735 4736
                need to be pruned

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

T
tangwei12 已提交
4737
        # NOTE(zhiqiu): we sync the original program first, since its program may diff with
4738 4739 4740
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

4741 4742
        if not isinstance(feeded_var_names, list):
            feeded_var_names = [feeded_var_names]
4743 4744
        if not isinstance(targets, list):
            targets = [targets]
4745 4746 4747

        for var in feeded_var_names:
            if not isinstance(var, six.string_types):
4748 4749 4750
                raise ValueError(
                    "All feeded_var_names of Program._prune_with_input() can only be "
                    "str, but received %s." % type(var))
4751

4752 4753 4754 4755
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
4756 4757 4758
                    name = t.name
                elif isinstance(t, six.string_types):
                    name = str(t)
4759
                else:
4760 4761 4762
                    raise ValueError(
                        "All targets of Program._prune_with_input() can only be "
                        "Variable or Operator, but received %s." % type(t))
4763 4764 4765 4766 4767 4768 4769 4770

                # 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

4771 4772 4773 4774 4775 4776 4777 4778 4779
                # 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
4780
                        # Skip optimize op except for optimize op in targets,
4781 4782 4783 4784 4785 4786
                        # since optimize op generates parameters.
                        if op._is_optimize_op() and op not in targets:
                            continue
                        else:
                            target_op = op
                            break
4787 4788 4789 4790 4791 4792 4793 4794
                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])
4795

4796
        res = Program()
4797 4798 4799
        res.desc, pruned_origin_block_id_map = core.prune(self.desc,
                                                          set(feeded_var_names),
                                                          targets_idx)
M
minqiyang 已提交
4800 4801 4802
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
4803
        res._sync_with_cpp()
4804 4805 4806 4807 4808

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

4809 4810
        return res

X
Xin Pan 已提交
4811
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
4812
        """
F
fengjiayi 已提交
4813 4814 4815 4816 4817
        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.

4818
        3. change the :code:`is_test`
Y
yuyang18 已提交
4819 4820 4821
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

4822
        Args:
X
Xin Pan 已提交
4823 4824
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
4825

Y
yuyang18 已提交
4826 4827 4828 4829 4830 4831
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
4832
        res = Program()
4833
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
4834 4835 4836 4837

        # 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
4838
        if prune_read_op:
4839 4840 4841 4842 4843 4844 4845 4846 4847
            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
4848
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
4849 4850

        # change all `is_test` attributes to True
M
minqiyang 已提交
4851
        for i in six.moves.range(res.desc.num_blocks()):
4852
            block = res.desc.block(i)
M
minqiyang 已提交
4853
            for j in six.moves.range(block.op_size()):
4854 4855
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
4856
                    op._set_attr('is_test', True)
M
minqiyang 已提交
4857 4858 4859
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
4860
        res._sync_with_cpp()
4861 4862
        return res

4863 4864
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
4865
        """
4866 4867 4868
        .. 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
4869

4870 4871
        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
4872

J
Jiabin Yang 已提交
4873
        Args:
Y
yuyang18 已提交
4874

J
Jiabin Yang 已提交
4875
            binary_str_type (str): the binary prootbuf string.
4876

J
Jiabin Yang 已提交
4877 4878
        Returns:
            Program: A deserialized Program.
4879 4880 4881 4882

        Examples:
            .. code-block:: python

4883 4884 4885 4886
                import paddle
                import paddle.static as static

                paddle.enable_static()
4887

4888 4889 4890 4891
                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')
4892

4893
                    y = static.data(name='Y', shape=[784, 100], dtype='float32')
4894

4895
                    z = paddle.matmul(x=x, y=y)
4896

4897 4898
                    binary_str = static.default_main_program().desc.serialize_to_string()
                    prog_restored = static.default_main_program().parse_from_string(binary_str)
4899

4900
                    print(static.default_main_program())
4901
                    print(prog_restored)
Y
yuyang18 已提交
4902
        """
4903 4904
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
4905
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
4906
        p._sync_with_cpp()
4907
        return p
Y
Yu Yang 已提交
4908

4909
    @staticmethod
4910
    def _construct_from_desc(desc):
4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925
        """
        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
4926 4927
    @property
    def random_seed(self):
Y
yuyang18 已提交
4928
        """
J
Jiabin Yang 已提交
4929
        The default random seed for random operators in Program. ``0`` means get
Y
yuyang18 已提交
4930 4931
        the random seed from random device.

4932 4933
        .. note:: 
            It must be set before the operators have been added.
J
Jiabin Yang 已提交
4934 4935 4936

        Returns:
            int64: Random seed in current Program
4937

4938 4939 4940 4941

        Examples:
            .. code-block:: python

4942 4943 4944
                import paddle
                import paddle.static as static
                import paddle.nn.functional as F
4945

4946 4947 4948
                paddle.enable_static()

                prog = static.default_main_program()
4949
                random_seed = prog.random_seed
4950
                x_var = static.data(name="X", shape=[3,3], dtype="float32")
4951 4952 4953
                print(random_seed)
                ## 0
                ## the default random seed is 0
4954

4955
                # Here we need to set random seed before we use paddle.nn.functional.dropout
4956
                prog.random_seed = 1
4957
                z_var = F.dropout(x_var, 0.7)
4958

4959
                print(prog.random_seed)
4960 4961
                ## 1
                ## the random seed is change to 1
Y
yuyang18 已提交
4962
        """
D
dzhwinter 已提交
4963 4964
        return self._seed

Q
qiaolongfei 已提交
4965 4966
    @property
    def num_blocks(self):
Y
yuyang18 已提交
4967
        """
4968 4969
        The number of :ref:`api_guide_Block_en`  in this Program.

4970 4971
        .. note:: 
            This API has no effect in Dygraph mode.
J
Jiabin Yang 已提交
4972 4973 4974

        Returns:
            int(Platform-dependent size): num of :ref:`api_guide_Block_en`  in current Program
4975

4976 4977 4978 4979

        Examples:
            .. code-block:: python

4980 4981 4982 4983
                import paddle
                import paddle.static as static

                paddle.enable_static()
4984

4985
                prog = static.default_main_program()
4986 4987
                num_blocks = prog.num_blocks
                print(num_blocks)
4988

4989 4990
                # print result:
                # 1
Y
yuyang18 已提交
4991
        """
Q
qiaolongfei 已提交
4992 4993
        return self.desc.num_blocks()

D
dzhwinter 已提交
4994 4995 4996
    @random_seed.setter
    def random_seed(self, seed):
        if not isinstance(seed, int):
4997 4998 4999
            raise ValueError(
                "Program.random_seed's input seed must be an integer, but received %s."
                % type(seed))
D
dzhwinter 已提交
5000 5001
        self._seed = seed

Y
Yu Yang 已提交
5002
    def __repr__(self):
5003
        return self.__str__()
5004

Y
Yu Yang 已提交
5005
    def global_block(self):
Y
yuyang18 已提交
5006
        """
5007 5008
        .. note::
            This API has no effect in Dygraph mode.
5009 5010 5011

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

J
Jiabin Yang 已提交
5012 5013
        Returns:
            :ref:`api_guide_Block_en`: The first :ref:`api_guide_Block_en`  of this Program.
5014

5015 5016 5017 5018

        Examples:
            .. code-block:: python

5019 5020 5021 5022
                import paddle
                import paddle.static as static

                paddle.enable_static()
5023

5024
                prog = static.default_main_program()
5025 5026
                gb_block = prog.global_block()
                print(gb_block)
5027

Y
yuyang18 已提交
5028
        """
Y
Yu Yang 已提交
5029 5030
        return self.blocks[0]

Q
Qiao Longfei 已提交
5031
    def block(self, index):
Y
yuyang18 已提交
5032
        """
5033 5034
        .. note::
            This API has no effect in Dygraph mode.
Y
yuyang18 已提交
5035

5036 5037
        Get the :code:`index`  :ref:`api_guide_Block_en`  of this Program

J
Jiabin Yang 已提交
5038 5039
        Args:
            index (int) - The index of  :ref:`api_guide_Block_en`  to get
5040

J
Jiabin Yang 已提交
5041 5042
        Returns:
            :ref:`api_guide_Block_en`: The :code:`index` block
5043 5044 5045 5046

        Examples:
            .. code-block:: python

5047 5048 5049 5050
                import paddle
                import paddle.static as static

                paddle.enable_static()
5051

5052
                prog = static.default_main_program()
5053 5054
                block_0 = prog.block(0)
                print(block_0)
Y
yuyang18 已提交
5055
        """
Q
Qiao Longfei 已提交
5056 5057
        return self.blocks[index]

Y
Yu Yang 已提交
5058
    def current_block(self):
Y
yuyang18 已提交
5059
        """
5060 5061
        .. note::
            This API has no effect in Dygraph mode.
5062

J
Jiabin Yang 已提交
5063 5064
        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.
5065

J
Jiabin Yang 已提交
5066 5067
        Returns:
             :ref:`api_guide_Block_en`: The :code:`index`  :ref:`api_guide_Block_en`
5068

5069 5070 5071
        Examples:
            .. code-block:: python

5072 5073 5074 5075
                import paddle
                import paddle.static as static

                paddle.enable_static()
5076

5077
                prog = static.default_main_program()
5078 5079
                current_blk = prog.current_block()
                print(current_blk)
Y
yuyang18 已提交
5080
        """
Y
Yu Yang 已提交
5081 5082
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
5083
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
5084 5085 5086 5087 5088
        """
        Create a new block with the :code:`parent_idx` and change the current block
        to new block.

        Args:
J
Jiabin Yang 已提交
5089

Y
yuyang18 已提交
5090 5091 5092 5093 5094
            parent_idx(int): The parent block index.

        Returns:
            Block: The new block.
        """
Y
Yu Yang 已提交
5095
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
5096 5097 5098
        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
5099 5100 5101 5102
        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
5103
    def _rollback(self):
Y
yuyang18 已提交
5104 5105 5106 5107 5108
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
5109 5110
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
5111
    def _sync_with_cpp(self):
Y
yuyang18 已提交
5112 5113 5114 5115 5116 5117 5118 5119 5120 5121
        """
        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
5122 5123 5124
        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
5125
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
5126

W
Wu Yi 已提交
5127
    def _copy_param_info_from(self, other):
5128
        """
5129
        Copy the information of parameters from other program.
D
dzhwinter 已提交
5130

Y
yuyang18 已提交
5131 5132 5133
        Notes: This is a very low level API. Users should not invoke it
        directly.

5134 5135 5136 5137 5138 5139 5140
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
5141 5142 5143
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
5144

W
Wu Yi 已提交
5145
        self.global_block()._copy_param_info_from(other.global_block())
5146

5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157
    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):
5158 5159 5160
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
5161 5162
        self._is_distributed = other._is_distributed
        self._is_chief = other._is_chief
5163
        self._parameters_on_pservers = other._parameters_on_pservers
5164
        self._endpoints = other._endpoints
5165
        self._ps_endpoint = other._ps_endpoint
5166 5167
        self._distributed_lookup_table = other._distributed_lookup_table

5168
    def _copy_data_info_from(self, other, pruned_origin_block_id_map=None):
F
fengjiayi 已提交
5169 5170
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
5171

Y
yuyang18 已提交
5172 5173 5174
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
5175 5176
        Args:
            other(Program): Other program
5177 5178 5179 5180
            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
5181 5182 5183 5184 5185

        Returns:
            None
        """
        if not isinstance(other, Program):
5186 5187 5188
            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
5189

5190 5191 5192 5193 5194
        if not pruned_origin_block_id_map:
            pruned_origin_block_id_map = {
                i: i
                for i in six.moves.range(self.desc.num_blocks())
            }
5195 5196 5197

        # NOTE(zhiqiu): All vars in cloned program exist in original program.
        # The reverse is not true, due to backward pruning.
5198 5199
        for i, block in enumerate(self.blocks):
            other_block = other.blocks[pruned_origin_block_id_map[i]]
5200
            for var in list(block.vars.values()):
5201 5202 5203 5204 5205 5206 5207
                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
5208

5209
    def list_vars(self):
Y
yuyang18 已提交
5210
        """
5211
        Get all Tensors from this Program. A iterable object is returned.
Y
yuyang18 已提交
5212

J
Jiabin Yang 已提交
5213
        Returns:
5214
            iterable Tensors: The Generator will yield every Tensor in this program.
5215 5216 5217 5218

        Examples:
            .. code-block:: python

5219 5220
                import paddle
                import paddle.static as static
5221

5222 5223 5224 5225 5226
                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')
5227 5228
                for var in prog.list_vars():
                    print(var)
T
tangwei12 已提交
5229

5230 5231
                # 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
5232
        """
5233
        for each_block in self.blocks:
5234
            for each_var in list(each_block.vars.values()):
5235 5236
                yield each_var

5237 5238 5239 5240 5241 5242 5243 5244 5245 5246
    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

5247 5248 5249 5250
                import paddle
                import paddle.static as static

                paddle.enable_static()
5251

5252 5253
                program = static.default_main_program()
                data = static.data(name='x', shape=[None, 13], dtype='float32')
5254
                hidden = static.nn.fc(x=data, size=10)
5255 5256
                loss = paddle.mean(hidden)
                paddle.optimizer.SGD(learning_rate=0.01).minimize(loss)
5257 5258 5259 5260 5261 5262 5263

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

                # Here will print all parameters in current program, in this example,
                # the result is like:
                #
5264 5265
                # 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)
5266 5267 5268 5269 5270 5271 5272 5273 5274 5275
                #
                # 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
5276

5277
@six.add_metaclass(ParameterMetaClass)
Y
Yu Yang 已提交
5278
class Parameter(Variable):
5279
    """
5280
    Parameter is derived from Variable. A parameter is a persistable
5281
    Variable, and will be updated by optimizers after each iteration.
5282
    The training of a neural network is essentially the updating of
5283 5284
    its parameters.

5285
    Relative to a general Variable, a Parameter has several its own
5286 5287
    member variables:

5288 5289 5290 5291 5292 5293 5294 5295 5296 5297
    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.
5298 5299
        need_clip (bool): Whether the parameter gradient need to be cliped 
            in optimizer. Default is True.
5300 5301
    """

5302 5303 5304 5305 5306 5307
    def __init__(self,
                 block,
                 shape,
                 dtype,
                 type=core.VarDesc.VarType.LOD_TENSOR,
                 **kwargs):
5308 5309 5310 5311 5312
        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
5313
        if len(shape) == 0:
5314 5315
            raise ValueError(
                "The dimensions of shape for Parameter must be greater than 0")
Y
Yu Yang 已提交
5316 5317 5318

        for each in shape:
            if each < 0:
5319 5320 5321
                raise ValueError(
                    "Each dimension of shape for Parameter must be greater than 0, but received %s"
                    % list(shape))
5322 5323

        Variable.__init__(
5324 5325 5326 5327 5328 5329 5330
            self,
            block,
            persistable=True,
            shape=shape,
            dtype=dtype,
            type=type,
            **kwargs)
Y
Yu Yang 已提交
5331 5332 5333 5334
        self.trainable = kwargs.get('trainable', True)

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

5335 5336
        self.regularizer = kwargs.get('regularizer', None)

W
wanghaoshuang 已提交
5337
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
5338

5339 5340
        self.need_clip = kwargs.get('need_clip', True)

5341 5342
        self.is_distributed = False

F
fengjiayi 已提交
5343
    def __str__(self):
5344
        return self._to_readable_code()
F
fengjiayi 已提交
5345

F
update  
fengjiayi 已提交
5346 5347 5348
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
5349

F
update  
fengjiayi 已提交
5350 5351 5352 5353 5354 5355 5356 5357
        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.

5358 5359 5360 5361 5362 5363 5364 5365 5366
        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
5367 5368 5369 5370 5371 5372
        """
        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",
5373
                               "do_model_average", "need_clip")
F
update  
fengjiayi 已提交
5374
            for attr_name in additional_attr:
5375 5376
                res_str += "%s: %s\n" % (attr_name,
                                         cpt.to_text(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
5377 5378
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
5379 5380 5381 5382
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
5383

5384 5385
class ParamBase(core.VarBase):
    """
5386 5387 5388
    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.
5389 5390 5391
    The training of a neural network is essentially the updating of
    its ParamBase.

5392
    Relative to a general Tensor, a ParamBase has several its own
5393 5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404
    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.
5405 5406
        need_clip (bool): Whether the parameter gradient need to be cliped 
            in optimizer. Default is True.
5407 5408 5409 5410 5411 5412 5413 5414 5415 5416 5417 5418 5419 5420 5421 5422 5423 5424 5425 5426 5427 5428 5429 5430 5431 5432 5433 5434 5435 5436
    """

    @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)

5437 5438
        trainable = kwargs.get('trainable', True)
        self.stop_gradient = not trainable
5439 5440 5441 5442 5443 5444 5445

        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)

5446 5447
        self.need_clip = kwargs.get('need_clip', True)

5448
        self.is_distributed = False
5449
        # self.block = default_main_program().global_block()
5450

5451 5452 5453 5454 5455 5456 5457 5458 5459 5460 5461 5462 5463
    @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))

5464
    def __str__(self):
5465
        """
5466
        Convert a ParamBase object to a readable string.
5467

5468
        Returns(str): A readable string.
5469 5470 5471 5472

        Examples:
            .. code-block:: python

5473
                import paddle
5474 5475 5476 5477 5478 5479 5480
                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]])
5481
        """
5482 5483
        return "Parameter containing:\n{tensor}".format(
            tensor=super(ParamBase, self).__str__())
5484

5485 5486 5487 5488 5489 5490 5491 5492 5493 5494 5495
    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
5496

5497 5498 5499 5500 5501 5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 5512 5513 5514
                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

5515 5516 5517
    __repr__ = __str__


Y
Yu Yang 已提交
5518
# program is a global instance.
Y
Yu Yang 已提交
5519 5520
_main_program_ = Program()
_startup_program_ = Program()
5521

5522

5523
def default_startup_program():
Y
Yu Yang 已提交
5524
    """
Y
yuyang18 已提交
5525 5526
    Get default/global startup program.

5527 5528
    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
5529

5530 5531
    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
5532

5533 5534
    Returns:
        Program: current default startup program.
5535

5536
    Returns type: 
5537 5538 5539 5540

    Examples:
        .. code-block:: python

5541
            import paddle
5542

5543
            paddle.enable_static()
5544 5545 5546 5547
            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
5548
    """
Y
Yu Yang 已提交
5549
    return _startup_program_
5550

5551

5552
def default_main_program():
Y
Yu Yang 已提交
5553
    """
5554
    This API can be used to get ``default main program`` which store the 
5555
    descriptions of Ops and tensors.
T
tangwei12 已提交
5556

5557
    For example ``z = paddle.add(x, y)`` will create a new ``add`` 
5558
    Op and a new ``z`` tensor, and they will be recorded in ``default main program`` . 
Y
yuyang18 已提交
5559

5560 5561
    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
5562
    :code:`default_main_program` when the program is not specified.
5563

5564
    If you want to switch the ``default main program``, you can use :ref:`api_paddle_fluid_framework_program_guard` .
T
tangwei12 已提交
5565

Y
Yu Yang 已提交
5566
    Returns:
5567
        Program: A ``Program`` which holding the descriptions of OPs and tensors in the network.
5568 5569 5570 5571

    Examples:
        ..  code-block:: python

5572
            import paddle
5573

5574
            paddle.enable_static()
5575
            # Sample Network:
5576 5577 5578
            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)
5579

5580 5581 5582
            #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
5583
            print(paddle.static.default_main_program())
Y
Yu Yang 已提交
5584
    """
Y
Yu Yang 已提交
5585
    return _main_program_
Y
Yu Yang 已提交
5586 5587 5588 5589 5590


def switch_main_program(program):
    """
    Switch the main program to a new program.
5591

Y
Yu Yang 已提交
5592 5593 5594 5595 5596 5597 5598 5599 5600 5601 5602 5603 5604 5605
    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):
    """
5606
    Switch the startup program to a new program
Y
Yu Yang 已提交
5607 5608 5609 5610 5611 5612 5613 5614 5615 5616 5617 5618
    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
5619
@signature_safe_contextmanager
Y
Yu Yang 已提交
5620 5621
def program_guard(main_program, startup_program=None):
    """
5622 5623
    :api_attr: Static Graph

5624 5625 5626
    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.
5627

G
guofei 已提交
5628
    Args:
5629 5630
        main_program(Program): New main program inside ``with`` statement.
        startup_program(Program, optional): New startup program inside ``with`` 
G
guofei 已提交
5631 5632 5633 5634
            statement. :code:`None` means not changing startup program, 
            default_startup_program is still used.
            Default: None.

Y
Yu Yang 已提交
5635
    Examples:
5636
       .. code-block:: python
T
tangwei12 已提交
5637

5638
          import paddle
Y
yuyang18 已提交
5639

5640 5641 5642 5643 5644
          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')
5645
              hidden = paddle.static.nn.fc(x=data, size=10, activation='relu')
Y
yuyang18 已提交
5646 5647 5648

    Notes: The temporary :code:`Program` can be used if the user does not need
    to construct either of startup program or main program.
5649

Y
Yu Yang 已提交
5650
    Examples:
5651
       .. code-block:: python
Y
yuyang18 已提交
5652

5653
          import paddle
5654

5655 5656 5657 5658 5659
          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
5660

Y
Yu Yang 已提交
5661
    """
5662
    from .data_feeder import check_type
5663 5664
    check_type(main_program, 'main_program', Program,
               'paddle.static.program_guard')
Y
Yu Yang 已提交
5665 5666
    main_program = switch_main_program(main_program)
    if startup_program is not None:
5667
        check_type(startup_program, 'startup_program', Program,
5668
                   'paddle.static.program_guard')
Y
Yu Yang 已提交
5669
        startup_program = switch_startup_program(startup_program)
5670 5671 5672 5673 5674 5675
    try:
        yield
    finally:
        switch_main_program(main_program)
        if startup_program is not None:
            switch_startup_program(startup_program)
X
xuwei06 已提交
5676 5677


W
Wu Yi 已提交
5678
def _get_var(name, program=None):
X
xuwei06 已提交
5679
    """
Y
yuyang18 已提交
5680
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
5681

X
xuwei06 已提交
5682 5683 5684
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
5685
        If None, default_global_program() will be used.
X
xuwei06 已提交
5686 5687 5688 5689 5690 5691 5692

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
5693
    assert isinstance(program, Program)
X
xuwei06 已提交
5694 5695

    return program.global_block().var(name)
5696 5697


S
rename  
sneaxiy 已提交
5698
@signature_safe_contextmanager
L
lujun 已提交
5699 5700
def _dygraph_guard(tracer):
    global _dygraph_tracer_
5701
    tmp_tracer = _dygraph_tracer_
L
lujun 已提交
5702
    _dygraph_tracer_ = tracer
5703
    core._switch_tracer(tracer)
M
minqiyang 已提交
5704

5705 5706 5707
    try:
        yield
    finally:
5708 5709
        core._switch_tracer(tmp_tracer)
        _dygraph_tracer_ = tmp_tracer
P
Paddle CI 已提交
5710 5711


S
rename  
sneaxiy 已提交
5712
@signature_safe_contextmanager
L
lujun 已提交
5713
def _dygraph_place_guard(place):
5714 5715 5716
    global _global_expected_place_
    tmp_place = _global_expected_place_
    _global_expected_place_ = place
M
minqiyang 已提交
5717

5718 5719
    _set_dygraph_tracer_expected_place(place)

5720 5721 5722
    try:
        yield
    finally:
5723
        _global_expected_place_ = tmp_place
5724
        _set_dygraph_tracer_expected_place(tmp_place)
5725 5726 5727 5728


def load_op_library(lib_filename):
    """
5729
    :api_attr: Static Graph
T
tangwei12 已提交
5730

5731 5732 5733
    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
5734
    Please note, the type of custom operators can't have the same type
5735 5736 5737 5738
    with the existing operators in the framework.

    Args:
        lib_filename (str): name of dynamic library.
5739 5740 5741
    
    Returns:
        list[str]: new registered custom op names.
5742 5743 5744 5745 5746 5747 5748 5749 5750

    Examples:
        .. code-block:: python

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

    """
    core.load_op_library(lib_filename)
5751
    return OpProtoHolder.instance().update_op_proto()
5752 5753 5754 5755 5756 5757 5758 5759 5760 5761 5762 5763 5764 5765 5766 5767 5768 5769 5770 5771 5772 5773 5774 5775 5776 5777 5778 5779


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

Z
Zhang Ting 已提交
5780
            import paddle
5781

Z
Zhang Ting 已提交
5782 5783 5784
            paddle.enable_static()
            support_gpu = paddle.is_compiled_with_cuda()
            place = paddle.CPUPlace()
5785
            if support_gpu:
Z
Zhang Ting 已提交
5786
                place = paddle.CUDAPlace(0)
5787 5788

            # if GPU is supported, the three OPs below will be automatically assigned to CUDAPlace(0)
Z
Zhang Ting 已提交
5789 5790 5791
            data1 = paddle.full(shape=[1, 3, 8, 8], fill_value=0.5, dtype='float32')
            data2 = paddle.full(shape=[1, 3, 64], fill_value=0.5, dtype='float32')
            shape = paddle.shape(data2)
5792

Z
Zhang Ting 已提交
5793
            with paddle.static.device_guard("cpu"):
5794
                # Ops created here will be placed on CPUPlace
Z
Zhang Ting 已提交
5795 5796
                shape = paddle.slice(shape, axes=[0], starts=[0], ends=[4])
            with paddle.static.device_guard('gpu'):
5797
                # if GPU is supported, OPs created here will be placed on CUDAPlace(0), otherwise on CPUPlace
Z
Zhang Ting 已提交
5798
                out = paddle.reshape(data1, shape=shape)
5799

Z
Zhang Ting 已提交
5800 5801
            exe = paddle.static.Executor(place)
            exe.run(paddle.static.default_startup_program())
5802 5803 5804
            result = exe.run(fetch_list=[out])
    """

5805 5806 5807 5808 5809
    index = None
    if device and ':' in device:
        device, index = device.split(':')
        if device == 'cpu':
            raise ValueError("Should not set device id for cpu.")
5810 5811 5812 5813
    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)
5814 5815
    if index:
        device = ":".join([device, index])
5816
    pre_device = switch_device(device)
5817 5818 5819 5820
    try:
        yield
    finally:
        switch_device(pre_device)
G
guofei 已提交
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 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


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
5888 5889 5890 5891 5892 5893 5894 5895 5896 5897 5898 5899 5900 5901 5902 5903 5904 5905 5906 5907 5908 5909 5910 5911 5912 5913 5914 5915 5916 5917 5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929 5930 5931 5932 5933 5934 5935 5936 5937 5938 5939 5940 5941 5942 5943 5944 5945 5946 5947 5948


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