framework.py 260.5 KB
Newer Older
G
gouzil 已提交
1
#   Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

Y
Yu Yang 已提交
15
import collections
X
Xin Pan 已提交
16
from collections import defaultdict
H
huzhiqiang 已提交
17
from collections.abc import Iterable
Q
qiaolongfei 已提交
18
import contextlib
19
from .wrapped_decorator import signature_safe_contextmanager, wrap_decorator
P
peizhilin 已提交
20
import os
F
fengjiayi 已提交
21
import re
22
import traceback
23
import copy
24
from types import MethodType, FunctionType
25

Y
Yu Yang 已提交
26
import numpy as np
27
import subprocess
S
sneaxiy 已提交
28
import multiprocessing
29
import sys
30
import logging
M
minqiyang 已提交
31
from .. import compat as cpt
32
from .proto import framework_pb2
33 34

from . import core
35
from . import unique_name
36 37
import paddle.version as fluid_version
import warnings
38
import functools
39
from .variable_index import _getitem_impl_, _setitem_impl_
Y
Yu Yang 已提交
40

41
__all__ = [
42 43 44 45
    'Program',
    'default_startup_program',
    'default_main_program',
    'program_guard',
46
    'name_scope',
J
jianghaicheng 已提交
47
    'ipu_shard_guard',
48
    'set_ipu_shard',
S
sneaxiy 已提交
49 50
    'cuda_places',
    'cpu_places',
51
    'xpu_places',
52
    'mlu_places',
S
sneaxiy 已提交
53
    'cuda_pinned_places',
J
Jiabin Yang 已提交
54
    '_non_static_mode',
L
lujun 已提交
55
    'in_dygraph_mode',
56
    'is_compiled_with_cinn',
C
chengduo 已提交
57
    'is_compiled_with_cuda',
58
    'is_compiled_with_rocm',
59
    'is_compiled_with_xpu',
60
    'is_compiled_with_npu',
61
    'Variable',
62
    'require_version',
63
    'device_guard',
G
guofei 已提交
64 65
    'set_flags',
    'get_flags',
66
]
Y
Yu Yang 已提交
67

Q
qiaolongfei 已提交
68 69 70 71
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
72 73
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()

L
lujun 已提交
74
_dygraph_tracer_ = None
75
_in_eager_mode_ = os.environ.get('FLAGS_enable_eager_mode', '1') == '1'
76
_global_expected_place_ = None
77
_current_device = None
78
global_prog_seed = 0
79
_current_pipeline_stage = None
80
_already_patch_eager_tensor = False
J
Jiabin Yang 已提交
81
_already_patch_varbase = False
82
_current_cuda_graph_mode = None
83
_global_flags_ = core.globals()
84 85 86 87 88 89
_enable_standalone_executor_ = os.environ.get(
    'FLAGS_USE_STANDALONE_EXECUTOR', None
)
_dy2st_enable_standalone_executor_ = os.environ.get(
    'FLAGS_DY2ST_USE_STANDALONE_EXECUTOR', 1
)
J
Jiabin Yang 已提交
90 91

# Some explanation of our execution system 2022.03
92
# For now we have 3 kinds of execution system, since we refactored dygraph mode to
J
Jiabin Yang 已提交
93
# build a fast execution system for dynamic mode. But we can't just remove all legacy
94
# code once we present the new system for some historical reason. That's why we have
J
Jiabin Yang 已提交
95
# these flags.
96
#
J
Jiabin Yang 已提交
97
# 1. _non_static_mode():
98
# _non_static_mode means  we are now running in legacy dygraph mode or dygraph mode.
J
Jiabin Yang 已提交
99 100 101 102
# 2. dygraph_mode():
# This flags inidicates we are now running in dygraph mode which called eager mode before.
# 3. _in_legacy_dygraph():
# This flags inidicates we are now running in legacy dygraph mode
103
#
J
Jiabin Yang 已提交
104
# They have a relation ship as below:
105
# Both dygraph_mode and _in_legacy_dygraph are _non_static_mode, but if you are running in
J
Jiabin Yang 已提交
106
# dygraph mode means you are not in _in_legacy_dygraph.
107
#
J
Jiabin Yang 已提交
108 109 110 111 112 113
# Why we have to make different of _in_legacy_dygraph and dygraph_mode?
# In some performance issue, we find that python if statement cause server performance problem
# and we need our new dygraph mode becomes as fast as it could be. That's why we make these flags
# to make sure in most case, we find new dygraph mode first with only one if statement.


114 115 116 117 118
def _update_monkey_methods(is_eager):
    """
    Update monkey methods of VarBase or eager.Tensor while
    switching eager mode and legacy mode.
    """
119
    from paddle import _C_ops, _legacy_C_ops
120 121 122
    from .dygraph.varbase_patch_methods import monkey_patch_varbase
    from .dygraph import monkey_patch_math_varbase

123 124 125
    global _already_patch_eager_tensor
    global _already_patch_varbase

126
    assert isinstance(is_eager, bool)
127
    # switch into eager mode
128
    if is_eager:
129
        _legacy_C_ops.switch_to_eager_ops()
130 131 132 133 134 135
        if not _already_patch_eager_tensor:
            monkey_patch_varbase()
            monkey_patch_math_varbase()

            _already_patch_eager_tensor = True
    # switch back into legacy mode
136
    else:
137
        _legacy_C_ops.switch_to_core_ops()
138 139 140 141 142
        if not _already_patch_varbase:
            monkey_patch_varbase()
            monkey_patch_math_varbase()

            _already_patch_varbase = True
143

144 145 146 147 148 149
    # switch Paddle.Tensor bind type
    _switch_tensor_bind_type(is_eager)


def _switch_tensor_bind_type(is_eager):
    import paddle
150

151 152 153 154 155
    if is_eager:
        paddle.Tensor = core.eager.Tensor
    else:
        paddle.Tensor = core.VarBase
    paddle.Tensor.__qualname__ = 'Tensor'
156 157


J
Jiabin Yang 已提交
158 159 160
def _enable_legacy_dygraph():
    global _in_eager_mode_
    _in_eager_mode_ = False
161
    _update_monkey_methods(is_eager=False)
J
Jiabin Yang 已提交
162 163 164 165 166


def _disable_legacy_dygraph():
    global _in_eager_mode_
    _in_eager_mode_ = True
167
    _update_monkey_methods(is_eager=True)
J
Jiabin Yang 已提交
168 169 170 171 172 173 174


def _in_eager_without_dygraph_check():
    global _in_eager_mode_
    return _in_eager_mode_


175 176 177 178 179 180 181 182 183
# FIXME(dev): We haven't fully verified eager mode on XPU/NPU et.al but
# only GPU/CPU. Remove this after we improve this feature.
_is_first_import_ = True


def _fallback_legacy_dygraph():
    global _in_eager_mode_
    global _is_first_import_
    need_fallback = False
C
Chen Weihang 已提交
184
    # Only enable eager on CPU/GPU/XPU
185 186 187 188 189
    is_not_support = (
        core.is_compiled_with_npu()
        or core.is_compiled_with_ipu()
        or core.is_compiled_with_mlu()
    )
190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210

    if _in_eager_mode_ and is_not_support:
        # switch into legacy dygraph mode
        warnings.warn(
            "We will fallback into legacy dygraph on NPU/XPU/MLU/IPU/ROCM devices. Because we only support new eager dygraph mode on CPU/GPU currently. "
        )
        _in_eager_mode_ = False
        if not _is_first_import_:
            _enable_legacy_dygraph()
        need_fallback = True

    need_fallback = False
    _is_first_import_ = False

    return need_fallback


# switch into legacy mode if need while import paddle
_fallback_legacy_dygraph()


J
Jiabin Yang 已提交
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246
def in_dygraph_mode():
    """

    .. 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>`_  .

    Returns:
        bool: Whether paddle runs in dynamic graph mode.

    Examples:
        .. code-block:: python

            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

    """
    return (_dygraph_tracer_ is not None) and _in_eager_mode_


def _in_legacy_dygraph():
    return (not _in_eager_mode_) and (_dygraph_tracer_ is not None)


def _non_static_mode():
    return _dygraph_tracer_ is not None
247 248 249


@signature_safe_contextmanager
J
Jiabin Yang 已提交
250
def _test_eager_guard(place=None):
C
Chen Weihang 已提交
251 252
    # FIXME(dev): We haven't fully verified eager mode on NPU et.al but
    # only GPU/CPU/XPU. Remove this after we improve this feature.
253 254 255
    already_fallback = _fallback_legacy_dygraph()
    if not already_fallback:
        _disable_legacy_dygraph()
256
    try:
J
Jiabin Yang 已提交
257
        yield
258
    finally:
259 260
        if not already_fallback:
            _enable_legacy_dygraph()
261 262


263 264
global_ipu_index = -1
global_ipu_stage = -1
J
jianghaicheng 已提交
265 266 267 268
ipu_index_attr_name = 'ipu_index'
ipu_stage_attr_name = 'ipu_stage'


L
Leo Chen 已提交
269 270 271 272 273 274 275 276 277 278 279
@signature_safe_contextmanager
def _enable_standalone_executor(enable=True):
    global _enable_standalone_executor_
    original_ = _enable_standalone_executor_
    _enable_standalone_executor_ = enable
    try:
        yield
    finally:
        _enable_standalone_executor_ = original_


J
jianghaicheng 已提交
280
@signature_safe_contextmanager
281
def ipu_shard_guard(index=-1, stage=-1):
J
jianghaicheng 已提交
282 283 284 285
    """
    Used to shard the graph on IPUs. Set each Op run on which IPU in the sharding and which stage in the pipelining.

    Args:
W
Weilong Wu 已提交
286
        index(int, optional): Specify which ipu the Tensor is computed on, (such as '0, 1, 2, 3').
287
            The default value is -1, which means the Op only run on IPU 0.
W
Weilong Wu 已提交
288
        stage(int, optional): Specify the computation order of the sharded model(such as '0, 1, 2, 3').
289
            The sharded model will be computed from small to large. The default value is -1,
J
jianghaicheng 已提交
290
            which means no pipelining computation order and run Ops in terms of graph.
291

G
gouzil 已提交
292 293 294 295 296 297 298
    Note:
        Only if the enable_manual_shard=True, the 'index' is able to be set not -1. Please refer
        to :ref:`api_paddle_static_IpuStrategy`.
        Only if the enable_pipelining=True, the 'stage' is able to be set not -1. Please refer
        to :ref:`api_paddle_static_IpuStrategy`.
        A index is allowed to match none stage or a stage. A stage is only allowed to match a new or
        duplicated index.
J
jianghaicheng 已提交
299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332

    Examples:
        .. code-block:: python

            # required: ipu

            import paddle
            paddle.enable_static()
            a = paddle.static.data(name='data', shape=[None, 1], dtype='int32')
            with paddle.static.ipu_shard_guard(index=0, stage=0):
                b = a + 1
            with paddle.static.ipu_shard_guard(index=1, stage=1):
                c = b + 1
            with paddle.static.ipu_shard_guard(index=0, stage=2):
                d = c + 1
    """
    if not core.is_compiled_with_ipu():
        raise ValueError(
            "Can not use this function since PaddlePaddle is not compiled with IPU"
        )

    global global_ipu_index
    global global_ipu_stage
    prev_ipu_index = global_ipu_index
    prev_ipu_stage = global_ipu_stage
    global_ipu_index = index
    global_ipu_stage = stage
    try:
        yield
    finally:
        global_ipu_index = prev_ipu_index
        global_ipu_stage = prev_ipu_stage


333 334 335 336
def set_ipu_shard(call_func, index=-1, stage=-1):
    """
    Shard the ipu with the given call function. Set every ops in call function to the given ipu sharding.

G
gouzil 已提交
337 338 339 340 341
    Note:
        Only when enable_manual_shard=True to set the index to a value other than -1. please refer to :ref:`api_paddle_static_IpuStrategy` .
        Only when enable_pipelining=True to set stage to a value other than -1. please refer to :ref:`api_paddle_static_IpuStrategy` .
        An index supports a corresponding None stage or a stage, and a stage only supports a new index or a duplicate index.

342 343 344 345 346
    Args:
        call_func(Layer|function): Specify the call function to be wrapped.
        index(int, optional): Specify which ipu the Tensor is computed on, (such as ‘0, 1, 2, 3’).
            The default value is -1, which means the Op only run on IPU 0.
        stage(int, optional): Specify the computation order of the sharded model(such as ‘0, 1, 2, 3’).
347
            The sharded model will be computed from small to large. The default value is -1,
348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373
            which means no pipelining computation order and run Ops in terms of graph.

    Returns:
        The wrapped call function.

    Examples:
        .. code-block:: python

            # required: ipu

            import paddle
            paddle.enable_static()
            a = paddle.static.data(name='data', shape=[None, 1], dtype='float32')
            relu = paddle.nn.ReLU()
            relu = paddle.static.set_ipu_shard(relu, index=1, stage=1)
            relu(a)
    """

    def decorate(func):
        def wrapper(*args, **kwargs):
            with ipu_shard_guard(index=index, stage=stage):
                return func(*args, **kwargs)

        return wrapper

    from .dygraph.layers import Layer
374

375 376 377 378 379
    if not isinstance(call_func, Layer):
        if callable(call_func):
            return decorate(call_func)
        else:
            raise TypeError(
380 381
                "Unsupported type. Only accept paddle.nn.Layer or function."
            )
382 383 384 385 386 387 388 389 390 391 392 393

    # patch paddle.nn.Layer
    class BlockFn(type(call_func)):
        def __call__(self, *args, **kwargs):
            with ipu_shard_guard(index=index, stage=stage):
                return super().__call__(*args, **kwargs)

    BlockFn.__name__ = type(call_func).__name__
    call_func.__class__ = BlockFn
    return call_func


394 395
def require_version(min_version, max_version=None):
    """
396 397 398
    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.
399

400 401 402 403
    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.
404

405 406
    Returns:
        None.
407

408 409 410 411 412 413
    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``.
414

415 416
    Examples:
        .. code-block:: python
417

418
            import paddle.fluid as fluid
419

420 421
            # any version >= 0.1.0 is acceptable.
            fluid.require_version('0.1.0')
422

423 424 425
            # if 0.1.0 <= version <= 10.0.0, it is acceptable.
            fluid.require_version(min_version='0.1.0', max_version='10.0.0')
    """
426 427 428
    if not isinstance(min_version, str):
        raise TypeError(
            "The type of 'min_version' in require_version must be str, but received %s."
429 430
            % (type(min_version))
        )
431 432 433 434

    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."
435 436
            % (type(max_version))
        )
437 438 439 440 441

    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}', "
442 443
            "like '1.5.2.0', but received %s" % min_version
        )
444 445 446 447 448 449

    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}', "
450 451
                "like '1.5.2.0', but received %s" % max_version
            )
452 453

    version_installed = [
454 455 456 457
        fluid_version.major,
        fluid_version.minor,
        fluid_version.patch,
        fluid_version.rc,
458 459 460 461
    ]
    zero_version = ['0', '0', '0', '0']

    def version_cmp(ver_a, ver_b):
462
        for i in range(len(ver_a)):
463 464 465 466 467 468 469 470 471 472 473
            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, "
474 475 476
                "please make sure the version is good with your code."
                % (min_version, max_version, fluid_version.full_version)
            )
477 478 479 480
        else:
            warnings.warn(
                "PaddlePaddle version %s or higher is required, but %s installed, "
                "Maybe you are using a develop version, "
481 482 483
                "please make sure the version is good with your code."
                % (min_version, fluid_version.full_version)
            )
484 485 486
        return

    min_version_split = min_version.split('.')
487 488 489
    min_version_to_check = (
        min_version_split + zero_version[len(min_version_split) :]
    )
490 491 492

    if max_version is not None:
        max_version_split = max_version.split('.')
493 494 495
        max_version_to_check = (
            max_version_split + zero_version[len(max_version_split) :]
        )
496

497 498 499 500
        if (
            version_cmp(version_installed, max_version_to_check) > 0
            or version_cmp(version_installed, min_version_to_check) < 0
        ):
501 502
            raise Exception(
                "VersionError: PaddlePaddle version in [%s, %s] required, but %s installed."
503 504
                % (min_version, max_version, fluid_version.full_version)
            )
505 506 507 508 509
    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."
510 511
                % (min_version, fluid_version.full_version, min_version)
            )
512 513


514 515
def _dygraph_not_support_(func):
    def __impl__(*args, **kwargs):
516 517 518
        assert not _non_static_mode(), (
            "We don't support %s in dynamic graph mode" % func.__name__
        )
519 520 521 522 523 524 525
        return func(*args, **kwargs)

    return __impl__


def _dygraph_only_(func):
    def __impl__(*args, **kwargs):
526 527 528 529
        assert _non_static_mode(), (
            "We only support '%s()' in dynamic graph mode, please call 'paddle.disable_static()' to enter dynamic graph mode."
            % func.__name__
        )
530 531 532 533 534
        return func(*args, **kwargs)

    return __impl__


535 536 537
def _non_static_only_(func):
    def __impl__(*args, **kwargs):
        from .dygraph.base import in_declarative_mode
538 539 540 541 542

        assert _non_static_mode() or in_declarative_mode(), (
            "We only support '%s()' in dynamic graph mode, please call 'paddle.disable_static()' to enter dynamic graph mode."
            % func.__name__
        )
543 544 545 546 547
        return func(*args, **kwargs)

    return __impl__


548 549
def _static_only_(func):
    def __impl__(*args, **kwargs):
550 551 552 553
        assert not _non_static_mode(), (
            "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__
        )
554 555 556 557 558
        return func(*args, **kwargs)

    return __impl__


559 560 561 562 563
def _set_pipeline_stage(stage):
    global _current_pipeline_stage
    _current_pipeline_stage = stage


564 565 566 567 568 569
# 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
570
# same base class.
571 572 573
def _fake_interface_only_(func):
    def __impl__(*args, **kwargs):
        raise AssertionError(
574 575 576 577
            "'%s' only can be called by `paddle.Tensor` in dynamic graph mode. Suggestions:\n"
            "  1. If you are in static graph mode, you can switch to dynamic graph mode by turning off `paddle.enable_static()` or calling `paddle.disable_static()`.\n"
            "  2. If you are using `@paddle.jit.to_static`, you can turn off ProgramTranslator by calling `paddle.jit.ProgramTranslator().enable(False)`. "
            "If you have to translate dynamic graph to static graph, please use other API to replace '%s'."
578 579
            % (func.__name__, func.__name__)
        )
580 581 582 583

    return __impl__


T
tangwei12 已提交
584 585
# 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
586 587 588 589 590 591 592 593 594
# 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`.",
595 596
                DeprecationWarning,
            )
597 598 599 600 601 602 603
            kwargs['state_dict'] = kwargs['stat_dict']
            kwargs.pop('stat_dict')
        return func(*args, **kwargs)

    return wrapper


604 605
dygraph_not_support = wrap_decorator(_dygraph_not_support_)
dygraph_only = wrap_decorator(_dygraph_only_)
606
static_only = wrap_decorator(_static_only_)
607
fake_interface_only = wrap_decorator(_fake_interface_only_)
608
non_static_only = wrap_decorator(_non_static_only_)
609 610


L
lujun 已提交
611 612
def _dygraph_tracer():
    return _dygraph_tracer_
613

W
Wu Yi 已提交
614

615 616 617 618
def _global_flags():
    return _global_flags_


M
minqiyang 已提交
619
def _current_expected_place():
620 621 622
    global _global_expected_place_
    if _global_expected_place_ is None:
        if core.is_compiled_with_cuda():
623 624 625 626 627
            try:
                device_count = core.get_cuda_device_count()
            except Exception as e:
                device_count = 0
            if device_count > 0:
628
                _global_expected_place_ = core.CUDAPlace(_cuda_ids()[0])
629 630 631 632 633
            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()
634 635 636 637 638 639
        elif core.is_compiled_with_xpu():
            try:
                device_count = core.get_xpu_device_count()
            except Exception as e:
                device_count = 0
            if device_count > 0:
640
                _global_expected_place_ = core.XPUPlace(_xpu_ids()[0])
641 642 643 644 645
            else:
                warnings.warn(
                    "You are using XPU version Paddle, but your XPU device is not set properly. CPU device will be used by default."
                )
                _global_expected_place_ = core.CPUPlace()
646 647 648 649 650 651
        elif core.is_compiled_with_mlu():
            try:
                device_count = core.get_mlu_device_count()
            except Exception as e:
                device_count = 0
            if device_count > 0:
652
                _global_expected_place_ = core.MLUPlace(_mlu_ids()[0])
653 654 655 656 657
            else:
                warnings.warn(
                    "You are using MLU version Paddle, but your MLU device is not set properly. CPU device will be used by default."
                )
                _global_expected_place_ = core.CPUPlace()
658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
        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
J
Jiabin Yang 已提交
673
    _set_dygraph_tracer_expected_place(place)
M
minqiyang 已提交
674 675


L
Leo Chen 已提交
676 677
# TODO(zhiqiu): remove this function.
def _var_base_to_np(var_base):
678 679
    """
    convert VarBase tp numpy
T
tangwei12 已提交
680

681 682 683
    Args:
        var_base(VarBase) : the VarBase to convert
    Returns (np.ndarray): the np.ndarray contain the value of VarBase
L
Leo Chen 已提交
684 685 686 687 688 689 690 691 692
    """

    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
693
def _cpu_num():
694
    if "CPU_NUM" not in os.environ.keys():
C
chengduo 已提交
695 696 697 698 699 700 701
        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(
702 703 704
                    multiprocessing.cpu_count(), multiprocessing.cpu_count()
                )
            )
C
chengduo 已提交
705
        os.environ['CPU_NUM'] = str(1)
706
    cpu_num = os.environ.get('CPU_NUM')
C
chengduo 已提交
707 708 709 710 711 712 713 714
    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:
715
        device_ids = range(core.get_cuda_device_count())
C
chengduo 已提交
716
    return device_ids
S
sneaxiy 已提交
717 718


719 720 721 722 723
def _xpu_ids():
    xpus_env = os.getenv("FLAGS_selected_xpus")
    if xpus_env:
        device_ids = [int(s) for s in xpus_env.split(",")]
    else:
724
        device_ids = range(core.get_xpu_device_count())
725 726 727
    return device_ids


728 729 730 731 732
def _npu_ids():
    npus_env = os.getenv("FLAGS_selected_npus")
    if npus_env:
        device_ids = [int(s) for s in npus_env.split(",")]
    else:
733
        device_ids = range(core.get_npu_device_count())
734 735 736
    return device_ids


737 738 739 740 741
def _mlu_ids():
    mlus_env = os.getenv("FLAGS_selected_mlus")
    if mlus_env:
        device_ids = [int(s) for s in mlus_env.split(",")]
    else:
742
        device_ids = range(core.get_mlu_device_count())
743 744 745
    return device_ids


746 747 748 749 750 751 752 753 754 755 756 757 758 759 760
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()


761 762 763 764 765 766 767 768 769 770 771 772 773 774 775
def is_compiled_with_npu():
    """
    Whether this whl package can be used to run the model on NPU.

    Returns (bool): support npu or not.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            support_npu = fluid.is_compiled_with_npu()
    """
    return core.is_compiled_with_npu()


776 777 778 779 780 781 782
def disable_signal_handler():
    """
    Reset signal handler registered by Paddle.

    Paddle installs signal handlers at C++ level to log debug information upon failing.
    However, conflicts can happen if another python module is making use of such signal.
    Such being the case, one may disblae paddle signal handler via this interface.
783

784 785 786 787 788 789
    Known frameworks that require disabling signal handler includes:
    1. TVM
    2. ADLIK

    Make sure you called paddle.disable_signal_handler() before using above mentioned frameworks.

790
    Returns: None
791 792 793 794 795 796 797 798 799 800

    Examples:
        .. code-block:: python

            import paddle
            paddle.disable_signal_handler()
    """
    core.disable_signal_handler()


801 802 803 804 805 806 807 808 809 810 811 812 813 814 815
def is_compiled_with_cinn():
    """
    Whether this whl package can be used to run the model on CINN.

    Returns (bool): `True` if CINN is currently available, otherwise `False`.

    Examples:
        .. code-block:: python

            import paddle
            support_cinn = paddle.device.is_compiled_with_cinn()
    """
    return core.is_compiled_with_cinn()


C
chengduo 已提交
816 817 818 819
def is_compiled_with_cuda():
    """
    Whether this whl package can be used to run the model on GPU.

820
    Returns (bool): `True` if CUDA is currently available, otherwise `False`.
C
chengduo 已提交
821 822 823 824

    Examples:
        .. code-block:: python

825
            import paddle
826
            support_gpu = paddle.device.is_compiled_with_cuda()
C
chengduo 已提交
827 828 829 830
    """
    return core.is_compiled_with_cuda()


831 832 833 834 835 836 837 838 839 840
def is_compiled_with_rocm():
    """
    Whether this whl package can be used to run the model on AMD or Hygon GPU(ROCm).

    Returns (bool): `True` if ROCm is currently available, otherwise `False`.

    Examples:
        .. code-block:: python

            import paddle
841
            support_gpu = paddle.device.is_compiled_with_rocm()
842 843 844 845
    """
    return core.is_compiled_with_rocm()


S
sneaxiy 已提交
846
def cuda_places(device_ids=None):
L
lujun 已提交
847
    """
848
    Note:
849 850 851
        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
852
    This function creates a list of :code:`paddle.CUDAPlace` objects.
S
add doc  
sneaxiy 已提交
853 854

    If :code:`device_ids` is None, environment variable of
855
    :code:`FLAGS_selected_gpus` would be checked first. For example, if
S
add doc  
sneaxiy 已提交
856
    :code:`FLAGS_selected_gpus=0,1,2`, the returned list would
C
Chen Weihang 已提交
857
    be [paddle.CUDAPlace(0), paddle.CUDAPlace(1), paddle.CUDAPlace(2)].
S
add doc  
sneaxiy 已提交
858
    If :code:`FLAGS_selected_gpus` is not set, all visible
859
    gpu places would be returned according to the :code:`CUDA_VISIBLE_DEVICES` environment variable.
S
add doc  
sneaxiy 已提交
860 861

    If :code:`device_ids` is not None, it should be the device
862
    ids of GPUs. For example, if :code:`device_ids=[0,1,2]`,
863
    the returned list would be
C
Chen Weihang 已提交
864
    [paddle.CUDAPlace(0), paddle.CUDAPlace(1), paddle.CUDAPlace(2)].
T
tangwei12 已提交
865

866
    Parameters:
867
        device_ids (list|tuple, optional): A list/tuple of int of GPU device ids.
S
add doc  
sneaxiy 已提交
868 869

    Returns:
C
Chen Weihang 已提交
870
        list of paddle.CUDAPlace: Created GPU place list.
L
lujun 已提交
871 872

    Examples:
873

L
lujun 已提交
874 875
        .. code-block:: python

C
Chen Weihang 已提交
876 877
            import paddle
            import paddle.static as static
T
tangwei12 已提交
878

879
            # required: gpu
880

C
Chen Weihang 已提交
881 882 883
            paddle.enable_static()

            cuda_places = static.cuda_places()
L
lujun 已提交
884 885

    """
886
    assert core.is_compiled_with_cuda(), "Not compiled with CUDA"
S
sneaxiy 已提交
887
    if device_ids is None:
C
chengduo 已提交
888
        device_ids = _cuda_ids()
S
sneaxiy 已提交
889 890 891 892 893
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.CUDAPlace(dev_id) for dev_id in device_ids]


894 895 896 897
def xpu_places(device_ids=None):
    """
    **Note**:
        For multi-card tasks, please use `FLAGS_selected_xpus` environment variable to set the visible XPU device.
S
sunzhongkai588 已提交
898 899 900 901 902 903 904 905 906
        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]`,
907
        the returned list would be
S
sunzhongkai588 已提交
908
        [paddle.XPUPlace(0), paddle.XPUPlace(1), paddle.XPUPlace(2)].
909

910 911 912 913 914 915
    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
S
sunzhongkai588 已提交
916

917 918
            # required: xpu

919 920
            import paddle
            import paddle.static as static
921

922 923 924
            paddle.enable_static()
            xpu_places = static.xpu_places()
    """
925
    assert core.is_compiled_with_xpu(), "Not compiled with XPU"
926 927 928 929 930 931 932
    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]


933 934 935 936
def npu_places(device_ids=None):
    """
    **Note**:
        For multi-card tasks, please use `FLAGS_selected_npus` environment variable to set the visible NPU device.
937

938 939 940 941 942 943 944 945 946
    This function creates a list of :code:`paddle.NPUPlace` objects.
    If :code:`device_ids` is None, environment variable of
    :code:`FLAGS_selected_npus` would be checked first. For example, if
    :code:`FLAGS_selected_npus=0,1,2`, the returned list would
    be [paddle.NPUPlace(0), paddle.NPUPlace(1), paddle.NPUPlace(2)].
    If :code:`FLAGS_selected_npus` is not set, all visible
    npu places would be returned.
    If :code:`device_ids` is not None, it should be the device
    ids of NPUs. For example, if :code:`device_ids=[0,1,2]`,
947
    the returned list would be
948
    [paddle.NPUPlace(0), paddle.NPUPlace(1), paddle.NPUPlace(2)].
949

950 951 952 953 954 955 956 957 958 959 960
    Parameters:
        device_ids (list or tuple of int, optional): list of NPU device ids.
    Returns:
        list of paddle.NPUPlace: Created NPU place list.
    Examples:
        .. code-block:: python

            # required: npu

            import paddle
            import paddle.static as static
961

962 963 964
            paddle.enable_static()
            npu_places = static.npu_places()
    """
965
    assert core.is_compiled_with_npu(), "Not compiled with NPU"
966 967 968 969 970 971 972
    if device_ids is None:
        device_ids = _npu_ids()
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.NPUPlace(dev_id) for dev_id in device_ids]


S
sneaxiy 已提交
973
def cpu_places(device_count=None):
L
lujun 已提交
974
    """
C
Chen Weihang 已提交
975
    This function creates a list of :code:`paddle.CPUPlace` objects, and returns the created list.
T
tangwei12 已提交
976

S
add doc  
sneaxiy 已提交
977
    If :code:`device_count` is None, the device count would
978
    be determined by environment variable :code:`CPU_NUM`.
C
chengduo 已提交
979 980
    If :code:`CPU_NUM` is not set, the default value is 1,
    i.e. CPU_NUM=1.
981 982
    :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
983

984 985
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
986 987

    Returns:
C
Chen Weihang 已提交
988
        list of paddle.CPUPlace: Created list of CPU places.
L
lujun 已提交
989 990

    Examples:
991

L
lujun 已提交
992 993
        .. code-block:: python

C
Chen Weihang 已提交
994 995
            import paddle
            import paddle.static as static
T
tangwei12 已提交
996

C
Chen Weihang 已提交
997 998 999
            paddle.enable_static()

            cpu_places = static.cpu_places()
L
lujun 已提交
1000 1001
    """

S
sneaxiy 已提交
1002 1003 1004 1005 1006 1007
    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
1008
    """
1009
    This function creates a list of :code:`fluid.CUDAPinnedPlace` objects.
S
add doc  
sneaxiy 已提交
1010 1011

    If :code:`device_count` is None, the device count would
1012
    be determined by environment variable :code:`CPU_NUM`.
1013 1014 1015 1016
    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
1017

1018 1019
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
1020 1021

    Returns:
1022
        list of fluid.CUDAPinnedPlace: Created list of CUDA pinned places.
L
lujun 已提交
1023 1024 1025 1026

    Examples:
        .. code-block:: python

1027
            import paddle.fluid as fluid
L
lujun 已提交
1028 1029 1030 1031 1032
            cuda_pinned_places_cpu_num = fluid.cuda_pinned_places()
            # or
            cuda_pinned_places = fluid.cuda_pinned_places(1)

    """
1033
    assert core.is_compiled_with_cuda(), "Not compiled with CUDA"
S
sneaxiy 已提交
1034
    if device_count is None:
1035 1036
        device_count = len(_cuda_ids())
    return [core.CUDAPinnedPlace()] * device_count
S
sneaxiy 已提交
1037 1038


1039 1040
def mlu_places(device_ids=None):
    """
G
gouzil 已提交
1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053
    This function creates a list of :code:`paddle.device.MLUPlace` objects.
    If :code:`device_ids` is None, environment variable of
    :code:`FLAGS_selected_mlus` would be checked first. For example, if
    :code:`FLAGS_selected_mlus=0,1,2`, the returned list would
    be [paddle.device.MLUPlace(0), paddle.device.MLUPlace(1), paddle.device.MLUPlace(2)].
    If :code:`FLAGS_selected_mlus` is not set, all visible
    mlu places would be returned.
    If :code:`device_ids` is not None, it should be the device
    ids of MLUs. For example, if :code:`device_ids=[0,1,2]`,
    the returned list would be
    [paddle.device.MLUPlace(0), paddle.device.MLUPlace(1), paddle.device.MLUPlace(2)].

    Note:
1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072
        For multi-card tasks, please use `FLAGS_selected_mlus` environment variable to set the visible MLU device.

    Parameters:
        device_ids (list or tuple of int, optional): list of MLU device ids.

    Returns:
        list of paddle.device.MLUPlace: Created MLU place list.

    Examples:
        .. code-block:: python

            # required: mlu

            import paddle
            import paddle.static as static

            paddle.enable_static()
            mlu_places = static.mlu_places()
    """
1073
    assert core.is_compiled_with_mlu(), "Not compiled with MLU"
1074 1075 1076 1077 1078 1079 1080
    if device_ids is None:
        device_ids = _mlu_ids()
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.MLUPlace(dev_id) for dev_id in device_ids]


1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091
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:
1092 1093 1094
            new_child = NameScope(
                prefix + "_%d" % len(self._children[prefix]), self
            )
1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107
            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
1108
@signature_safe_contextmanager
1109 1110
def name_scope(prefix=None):
    """
1111

1112
    Generate hierarchical name prefix for the operators in Static Graph.
1113

1114
    Note:
T
Tao Luo 已提交
1115 1116
        This should only used for debugging and visualization purpose.
        Don't use it for serious analysis such as graph/program transformations.
1117
        Don't use it in dygraph, since it will cause memory leak.
1118 1119

    Args:
T
Tao Luo 已提交
1120
        prefix(str, optional): prefix. Default is none.
1121 1122

    Examples:
1123

1124
        .. code-block:: python
T
Tink_Y 已提交
1125

1126 1127 1128
          import paddle
          paddle.enable_static()
          with paddle.static.name_scope("s1"):
1129
             a = paddle.static.data(name='data', shape=[None, 1], dtype='int32')
T
Tao Luo 已提交
1130
             b = a + 1
1131
             with paddle.static.name_scope("s2"):
T
Tao Luo 已提交
1132
                c = b * 1
1133
             with paddle.static.name_scope("s3"):
T
Tao Luo 已提交
1134
                d = c / 1
1135 1136 1137
          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
1138 1139
                g = f - 1

1140
          # Op are created in the default main program.
1141
          for op in paddle.static.default_main_program().block(0).ops:
T
Tao Luo 已提交
1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156
              # 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/'
1157 1158
    """
    # TODO(panyx0718): Only [0-9a-z].
1159
    # in dygraph we don't need namescope since it will cause mem leak
J
Jiabin Yang 已提交
1160
    if _non_static_mode():
L
Leo Chen 已提交
1161 1162
        yield
    else:
T
tianshuo78520a 已提交
1163
        assert prefix, "namescope prefix can not be empty."
1164 1165
        global _name_scope
        _name_scope = _name_scope.child(prefix)
1166 1167 1168 1169
        try:
            yield
        finally:
            _name_scope = _name_scope.parent()
1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181


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
1182 1183
def generate_control_dev_var_name():
    import random
1184

W
Wu Yi 已提交
1185
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
1186 1187 1188 1189


def grad_var_name(var_name):
    """
1190 1191
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
1192 1193 1194
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
1195

1196
def convert_np_dtype_to_dtype_(np_dtype):
1197
    """
1198
    Convert the data type in numpy to the data type in Paddle.
1199

1200
    Args:
1201 1202
        np_dtype (np.dtype|str): The data type in numpy or valid data type
            string.
1203

1204
    Returns:
1205
        core.VarDesc.VarType: The data type in Paddle.
1206 1207

    """
1208 1209
    # Convert the data type string to numpy data type.
    if isinstance(np_dtype, str) and np_dtype == "bfloat16":
1210 1211 1212
        dtype = np.uint16
    else:
        dtype = np.dtype(np_dtype)
1213

1214
    if dtype == np.float32:
1215
        return core.VarDesc.VarType.FP32
1216
    elif dtype == np.float64:
1217
        return core.VarDesc.VarType.FP64
1218
    elif dtype == np.float16:
1219
        return core.VarDesc.VarType.FP16
1220
    elif dtype == np.int32:
1221
        return core.VarDesc.VarType.INT32
1222
    elif dtype == np.int16:
1223
        return core.VarDesc.VarType.INT16
1224
    elif dtype == np.int64:
1225
        return core.VarDesc.VarType.INT64
1226
    elif dtype == np.bool_:
1227
        return core.VarDesc.VarType.BOOL
1228
    elif dtype == np.uint16:
1229 1230 1231
        # since there is still no support for bfloat16 in NumPy,
        # uint16 is used for casting bfloat16
        return core.VarDesc.VarType.BF16
1232 1233
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
1234 1235
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
1236 1237 1238 1239
    elif dtype == np.complex64:
        return core.VarDesc.VarType.COMPLEX64
    elif dtype == np.complex128:
        return core.VarDesc.VarType.COMPLEX128
1240
    else:
M
minqiyang 已提交
1241
        raise ValueError("Not supported numpy dtype %s" % dtype)
1242 1243 1244


def dtype_is_floating(dtype):
1245 1246 1247
    """
    Check the data type is floating or not.
    Args:
1248
        dtype(np.dtype|core.VarDesc.VarType): data type.
1249 1250 1251 1252 1253
            Could be numpy format or Paddle format

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

    """
1254
    if not isinstance(dtype, core.VarDesc.VarType):
1255 1256
        dtype = convert_np_dtype_to_dtype_(dtype)

1257
    return dtype in [
1258 1259 1260
        core.VarDesc.VarType.FP16,
        core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64,
1261
    ]
1262 1263


Y
Yang Yang(Tony) 已提交
1264
def _debug_string_(proto, throw_on_error=True):
1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275
    """
    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
1276
    error_fields = list()
Y
Yang Yang(Tony) 已提交
1277
    if not proto.IsInitialized(error_fields) and throw_on_error:
1278 1279
        raise ValueError(
            "{0} are not initialized.\nThe message is {1}:\n".format(
1280 1281 1282
                error_fields, proto
            )
        )
Y
Yu Yang 已提交
1283 1284 1285
    return proto.__str__()


1286 1287 1288 1289 1290 1291 1292 1293
def _varbase_creator(
    type=core.VarDesc.VarType.LOD_TENSOR,
    name=None,
    shape=None,
    dtype=None,
    persistable=None,
    **kwargs
):
1294 1295 1296 1297
    if dtype is not None:
        if not isinstance(dtype, core.VarDesc.VarType):
            dtype = convert_np_dtype_to_dtype_(dtype)

J
Jiabin Yang 已提交
1298
    if _in_eager_mode_:
1299
        eager_tensor = core.eager.Tensor(
1300
            dtype if dtype else core.VarDesc.VarType.FP32,
1301 1302
            list(shape) if shape else [],
            name,
1303
            type if type else core.VarDesc.VarType.LOD_TENSOR,
1304 1305
            True if persistable else False,
        )
1306 1307
        eager_tensor.retain_grads()
        return eager_tensor
J
Jiabin Yang 已提交
1308
    else:
1309 1310 1311 1312 1313 1314 1315
        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,
        )
1316 1317


1318 1319 1320 1321 1322 1323 1324
def _all_is_type(vals, expected_type):
    """
    Return True if type of each element is expected_type.

    NOTE: BuiltIn all() will always return True if vals is empty.
    """
    assert isinstance(vals, (list, tuple))
1325 1326
    if not vals:
        return False
1327 1328 1329
    return all(isinstance(v, expected_type) for v in vals)


1330 1331 1332 1333 1334
class VariableMetaClass(type):
    @classmethod
    def __instancecheck__(cls, instance):
        t = type(instance)
        if in_dygraph_mode():
J
Jiabin Yang 已提交
1335
            return issubclass(t, core.eager.Tensor)
1336
        else:
J
Jiabin Yang 已提交
1337 1338
            if _in_legacy_dygraph():
                return issubclass(t, core.VarBase)
1339 1340 1341 1342 1343 1344 1345 1346
            return issubclass(t, Variable)


class ParameterMetaClass(VariableMetaClass):
    @classmethod
    def __instancecheck__(cls, instance):
        t = type(instance)
        if in_dygraph_mode():
J
Jiabin Yang 已提交
1347
            return issubclass(t, EagerParamBase)
1348
        else:
J
Jiabin Yang 已提交
1349 1350
            if _in_legacy_dygraph():
                return issubclass(t, ParamBase)
1351 1352 1353
            return issubclass(t, Parameter)


1354
class Variable(metaclass=VariableMetaClass):
1355
    """
J
Jiabin Yang 已提交
1356
    **Notes**:
1357
        **The constructor of Variable should not be invoked directly.**
J
Jiabin Yang 已提交
1358

1359 1360
        **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
1361 1362 1363
        **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
1364
    cases, variables are used for holding different kinds of data or training
J
Jiabin Yang 已提交
1365 1366
    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.
1367

1368
    There are many kinds of variables. Each kind of them has its own attributes
J
Jiabin Yang 已提交
1369
    and usages. Please refer to the `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_ for details.
1370

T
tianshuo78520a 已提交
1371
    Most of a Variable's member variables can be set to be None. It mean
1372
    it is not available or will be specified later.
1373

1374
    Examples:
1375 1376
        In Static Graph Mode:

1377 1378
        .. code-block:: python

1379
            import paddle.fluid as fluid
1380
            cur_program = fluid.Program()
1381 1382 1383 1384
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
S
sunzhongkai588 已提交
1385

J
Jiabin Yang 已提交
1386
        In `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_  Mode:
1387 1388 1389 1390 1391 1392 1393 1394 1395

        .. 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))

1396 1397
    """

1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414
    def __init__(
        self,
        block,
        type=core.VarDesc.VarType.LOD_TENSOR,
        name=None,
        shape=None,
        dtype=None,
        lod_level=None,
        capacity=None,
        persistable=None,
        error_clip=None,
        stop_gradient=False,
        is_data=False,
        need_check_feed=False,
        belong_to_optimizer=False,
        **kwargs
    ):
Y
Yu Yang 已提交
1415 1416
        self.block = block
        if name is None:
Y
Yu Yang 已提交
1417
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
1418

Y
Yu Yang 已提交
1419
        if dtype is not None:
1420
            if not isinstance(dtype, core.VarDesc.VarType):
1421
                dtype = convert_np_dtype_to_dtype_(dtype)
1422

S
Steffy-zxf 已提交
1423 1424 1425 1426
        if dtype == core.VarDesc.VarType.STRINGS:
            type = core.VarDesc.VarType.STRINGS
            lod_level = None

1427 1428 1429
        if type == core.VarDesc.VarType.SPARSE_COO:
            lod_level = None

H
hong 已提交
1430 1431
        self.belong_to_optimizer = belong_to_optimizer

1432 1433 1434
        self.error_clip = error_clip

        is_new_var = False
1435
        self.desc = self.block.desc.find_var(name.encode())
1436

1437
        if self.desc is None:
1438
            self.desc = self.block.desc.var(name.encode())
1439
            is_new_var = True
1440

1441 1442 1443
        if is_new_var:
            self.desc.set_type(type)
        elif self.desc.type() != type:
1444 1445 1446 1447 1448
            raise ValueError(
                "Variable '{0}' has been created before. The "
                "previous type is {1}, the new type is {2}. They"
                " are not matched".format(self.name, self.desc.type(), type)
            )
1449

1450
        if shape is not None:
1451
            if is_new_var:
1452 1453 1454 1455 1456 1457
                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
1458 1459
                        "Variable '{0}' has been created before. The previous "
                        "shape is {1}, the new shape is {2}. They are not "
1460 1461
                        "matched.".format(self.name, old_shape, shape)
                    )
1462 1463 1464 1465 1466 1467
        if dtype is not None:
            if is_new_var:
                self.desc.set_dtype(dtype)
            else:
                old_dtype = self.dtype
                if dtype != old_dtype:
1468 1469 1470 1471 1472 1473
                    raise ValueError(
                        "Variable '{0}' has been created before. "
                        "The previous data type is {1}, the new "
                        "data type is {2}. They are not "
                        "matched.".format(self.name, old_dtype, dtype)
                    )
1474 1475 1476 1477 1478 1479

        if lod_level is not None:
            if is_new_var:
                self.desc.set_lod_level(lod_level)
            else:
                if lod_level != self.lod_level:
1480 1481 1482 1483 1484 1485
                    raise ValueError(
                        "Variable '{0}' has been created before. "
                        "The previous lod_level is {1}, the new "
                        "lod_level is {2}. They are not "
                        "matched".format(self.name, self.lod_level, lod_level)
                    )
1486 1487 1488 1489 1490 1491
        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
1492 1493
                        "Variable '{0}' has been created before."
                        "The previous persistable is {1}, the new "
1494
                        "persistable is {2}. They are not matched".format(
1495 1496 1497
                            self.name, self.persistable, persistable
                        )
                    )
1498

1499 1500
        if need_check_feed and is_new_var:
            self.desc.set_need_check_feed(need_check_feed)
H
Huihuang Zheng 已提交
1501

1502 1503 1504 1505 1506 1507 1508
        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
1509

1510 1511
        self.block.vars[name] = self
        self.op = None
1512
        self.stop_gradient = stop_gradient
1513
        self.is_data = is_data
Y
Yu Yang 已提交
1514

1515 1516 1517
    def detach(self):
        """
        Returns a new Variable, detached from the current graph.
1518 1519
        It will share data with origin Variable and without tensor copy.
        In addition, the detached Variable doesn't provide gradient propagation.
1520

1521
        Returns:
J
Jiabin Yang 已提交
1522
             ( :ref:`api_guide_Variable_en` | dtype is same as current Variable): The detached Variable.
1523 1524 1525 1526

        Examples:
            .. code-block:: python

1527
                import paddle
1528

1529 1530 1531 1532
                paddle.enable_static()

                # create a static Variable
                x = paddle.static.data(name='x', shape=[3, 2, 1])
1533

1534 1535
                # create a detached Variable
                y = x.detach()
1536
        """
1537

1538 1539 1540 1541
        assert (
            self.type == core.VarDesc.VarType.SELECTED_ROWS
            or self.type == core.VarDesc.VarType.LOD_TENSOR
        ), "only support a variable with SELECTED_ROWS or LOD_TENSOR to be detached"
1542 1543 1544 1545 1546 1547

        output = self.block.create_var(
            name=unique_name.generate_with_ignorable_key("detach_" + self.name),
            dtype=self.dtype,
            type=self.type,
            persistable=self.persistable,
1548 1549
            stop_gradient=True,
        )
1550

1551 1552 1553
        self.block.append_op(
            type='share_data', inputs={'X': [self]}, outputs={'Out': [output]}
        )
1554
        return output
1555

1556
    @fake_interface_only
1557
    def numpy(self):
1558
        """
J
Jiabin Yang 已提交
1559
        **Notes**:
T
tianshuo78520a 已提交
1560
            **This API is ONLY available in Dygraph mode**
1561

J
Jiabin Yang 已提交
1562
        Returns a numpy array shows the value of current :ref:`api_guide_Variable_en`
1563 1564 1565 1566 1567

        Returns:
            ndarray: The numpy value of current Variable.

        Returns type:
J
Jiabin Yang 已提交
1568
            ndarray: dtype is same as current Variable
1569 1570 1571 1572 1573 1574

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1575
                from paddle.fluid.dygraph import Linear
1576 1577 1578 1579
                import numpy as np

                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
                with fluid.dygraph.guard():
1580
                    linear = Linear(32, 64)
1581
                    data = to_variable(data)
1582
                    x = linear(data)
1583 1584 1585
                    print(x.numpy())

        """
1586
        pass
1587

1588
    @fake_interface_only
1589
    def backward(self, retain_graph=False):
1590
        """
J
Jiabin Yang 已提交
1591
        **Notes**:
T
tianshuo78520a 已提交
1592
            **This API is ONLY available in Dygraph mode**
1593

1594
        Run backward of current Graph which starts from current Tensor.
1595

J
Jiabin Yang 已提交
1596
        Args:
1597 1598 1599 1600
            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.
1601

J
Jiabin Yang 已提交
1602 1603
        Returns:
            NoneType: None
1604 1605 1606 1607 1608

        Examples:
            .. code-block:: python

                import numpy as np
1609 1610
                import paddle
                paddle.disable_static()
1611 1612

                x = np.ones([2, 2], np.float32)
1613 1614 1615 1616 1617 1618 1619
                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)
1620 1621
                ret = paddle.add_n(inputs)
                loss = paddle.sum(ret)
1622
                loss.backward()
1623 1624

        """
1625
        pass
1626

1627
    @fake_interface_only
1628
    def gradient(self):
1629
        """
J
Jiabin Yang 已提交
1630
        **Notes**:
T
tianshuo78520a 已提交
1631
            **This API is ONLY available in Dygraph mode**
1632 1633 1634

        Get the Gradient of Current Variable

J
Jiabin Yang 已提交
1635
        Returns:
1636
            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.
1637 1638 1639 1640 1641 1642 1643

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

1644
                # example1: return ndarray
1645 1646 1647 1648 1649 1650 1651 1652 1653
                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)
1654
                    loss2.backward()
1655 1656
                    print(loss2.gradient())

1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669
                # 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())

1670
        """
1671
        pass
1672

1673
    @fake_interface_only
1674
    def clear_gradient(self):
1675
        """
J
Jiabin Yang 已提交
1676
        **Notes**:
T
tianshuo78520a 已提交
1677
            **1. This API is ONLY available in Dygraph mode**
J
Jiabin Yang 已提交
1678 1679

            **2. Use it only Variable has gradient, normally we use this for Parameters since other temporal Variable will be deleted by Python's GC**
1680

J
Jiabin Yang 已提交
1681
        Clear  (set to ``0`` ) the Gradient of Current Variable
1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699

        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)
1700
                    loss2.backward()
1701 1702 1703 1704 1705
                    print(loss2.gradient())
                    loss2.clear_gradient()
                    print("After clear {}".format(loss2.gradient()))

        """
1706
        pass
X
Xin Pan 已提交
1707

1708 1709 1710 1711
    @fake_interface_only
    def register_hook(self, hook):
        pass

1712
    def __str__(self):
1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728
        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

1729 1730
                import paddle
                import paddle.static as static
1731

1732 1733 1734
                paddle.enable_static()

                cur_program = static.Program()
1735 1736 1737 1738 1739 1740
                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())
        """
1741 1742
        # VarType.LOD_TENSOR -> LOD_TENSOR
        type_str = str(self.type).split('.')[1]
1743 1744 1745 1746
        if (
            self.type == core.VarDesc.VarType.SELECTED_ROWS
            or self.type == core.VarDesc.VarType.LOD_TENSOR
        ):
1747
            dtype_str = str(self.dtype).split('.')[1]
1748 1749 1750 1751 1752 1753 1754
            var_str = "{name} : {type}.shape{shape}.dtype({dtype}).stop_gradient({stop_gradient})".format(
                name=self.name,
                type=type_str,
                shape=self.shape,
                dtype=dtype_str,
                stop_gradient=self.stop_gradient,
            )
1755
        else:
1756
            var_str = "{name} : {type})".format(name=self.name, type=type_str)
1757

1758
        if self.is_parameter:
1759 1760 1761 1762 1763 1764 1765 1766 1767 1768
            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

1769 1770 1771 1772
        from paddle.distributed.auto_parallel.dist_context import (
            get_default_distributed_context,
        )

1773
        dist_context = get_default_distributed_context()
1774 1775
        dist_tensor = dist_context.get_dist_tensor_for_program(self)
        if dist_tensor is not None:
1776 1777 1778
            var_str += ", {name} = {value}".format(
                name="dist_attr", value=dist_tensor
            )
1779

1780
        return var_str
Y
Yang Yang(Tony) 已提交
1781

F
update  
fengjiayi 已提交
1782
    def to_string(self, throw_on_error, with_details=False):
1783 1784 1785
        """
        Get debug string.

J
Jiabin Yang 已提交
1786 1787 1788 1789 1790
        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;
1791

1792 1793
        Returns:
            str: The debug string.
1794 1795 1796 1797 1798

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
1799
                import paddle
1800

1801
                paddle.enable_static()
1802 1803 1804 1805 1806
                cur_program = fluid.Program()
                cur_block = cur_program.current_block()
                new_variable = cur_block.create_var(name="X",
                                                    shape=[-1, 23, 48],
                                                    dtype='float32')
1807
                print(new_variable.to_string(True))
J
Jiabin Yang 已提交
1808
                print("=============with detail===============")
1809
                print(new_variable.to_string(True, True))
1810
        """
1811
        assert isinstance(throw_on_error, bool) and isinstance(
1812 1813
            with_details, bool
        )
1814
        protostr = self.desc.serialize_to_string()
1815
        proto = framework_pb2.VarDesc.FromString(bytes(protostr))
F
update  
fengjiayi 已提交
1816 1817
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
1818
            additional_attr = ("error_clip",)
F
update  
fengjiayi 已提交
1819
            for attr_name in additional_attr:
1820
                res_str += "%s: %s\n" % (attr_name, getattr(self, attr_name))
1821

F
update  
fengjiayi 已提交
1822
        return res_str
1823 1824 1825

    __repr__ = __str__

1826 1827 1828
    def element_size(self):
        """
        Returns the size in bytes of an element in the Tensor.
1829

1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852
        Examples:
          .. code-block:: python

            import paddle
            paddle.enable_static()

            x = paddle.static.data(name='x1', shape=[3, 2], dtype='bool')
            x.element_size() # 1

            x = paddle.static.data(name='x2', shape=[3, 2], dtype='int16')
            x.element_size() # 2

            x = paddle.static.data(name='x3', shape=[3, 2], dtype='float16')
            x.element_size() # 2

            x = paddle.static.data(name='x4', shape=[3, 2], dtype='float32')
            x.element_size() # 4

            x = paddle.static.data(name='x5', shape=[3, 2], dtype='float64')
            x.element_size() # 8
        """
        return self.desc.element_size()

1853
    @property
1854
    def stop_gradient(self):
J
Jiabin Yang 已提交
1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869
        """
        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")
1870 1871
                linear = fluid.Linear(13, 5, dtype="float32")
                linear2 = fluid.Linear(3, 3, dtype="float32")
J
Jiabin Yang 已提交
1872 1873 1874
                a = fluid.dygraph.to_variable(value0)
                b = fluid.dygraph.to_variable(value1)
                c = fluid.dygraph.to_variable(value2)
1875 1876
                out1 = linear(a)
                out2 = linear2(b)
J
Jiabin Yang 已提交
1877 1878 1879 1880
                out1.stop_gradient = True
                out = fluid.layers.concat(input=[out1, out2, c], axis=1)
                out.backward()

1881
                assert linear.weight.gradient() is None
J
Jiabin Yang 已提交
1882 1883
                assert (out1.gradient() == 0).all()
        """
1884
        return self.desc.stop_gradient()
1885

1886 1887
    @stop_gradient.setter
    def stop_gradient(self, s):
1888
        self.desc.set_stop_gradient(s)
1889

1890 1891
    @property
    def persistable(self):
J
Jiabin Yang 已提交
1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912
        """
        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))
        """
1913
        return self.desc.persistable()
1914

Y
Yu Yang 已提交
1915 1916
    @persistable.setter
    def persistable(self, p):
1917
        self.desc.set_persistable(p)
Y
Yu Yang 已提交
1918

1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943
    @property
    def is_parameter(self):
        """
        Indicating if current Variable is a Parameter

        Examples:
          .. code-block:: python

            import paddle
            new_parameter = paddle.static.create_parameter(name="X",
                                                shape=[10, 23, 48],
                                                dtype='float32')
            if new_parameter.is_parameter:
                print("Current var is a Parameter")
            else:
                print("Current var is not a Parameter")

            # Current var is a Parameter
        """
        return self.desc.is_parameter()

    @is_parameter.setter
    def is_parameter(self, p):
        self.desc.set_is_parameter(p)

Y
Yu Yang 已提交
1944 1945
    @property
    def name(self):
J
Jiabin Yang 已提交
1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961
        """
        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))
        """
1962
        return self.desc.name()
Y
Yu Yang 已提交
1963

1964 1965 1966 1967 1968 1969
    @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
S
sunzhongkai588 已提交
1970 1971
        gradient Variable from a naming convention but doesn't guarantee
        the gradient exists.**
T
tangwei12 已提交
1972

1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983
        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
1984 1985
    @name.setter
    def name(self, new_name):
1986
        self.desc.set_name(new_name)
T
typhoonzero 已提交
1987

Y
Yu Yang 已提交
1988 1989
    @property
    def shape(self):
J
Jiabin Yang 已提交
1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006
        """
        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
2007
        # convert to tuple, make it as same as numpy API.
2008
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
2009 2010

    @property
F
fengjiayi 已提交
2011
    def dtype(self):
J
Jiabin Yang 已提交
2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027
        """
        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))
        """
2028
        return self.desc.dtype()
Y
Yu Yang 已提交
2029 2030 2031

    @property
    def lod_level(self):
J
Jiabin Yang 已提交
2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044
        """
        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

2045
            import paddle
J
Jiabin Yang 已提交
2046
            import paddle.fluid as fluid
2047 2048

            paddle.enable_static()
J
Jiabin Yang 已提交
2049 2050 2051 2052 2053 2054 2055
            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))
        """
2056 2057
        if self.type == core.VarDesc.VarType.SELECTED_ROWS:
            raise Exception("SelectedRows DO NOT supprt lod")
2058 2059
        if self.type == core.VarDesc.VarType.STRINGS:
            return None
2060
        return self.desc.lod_level()
Y
Yu Yang 已提交
2061

Y
Yu Yang 已提交
2062 2063
    @property
    def type(self):
J
Jiabin Yang 已提交
2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079
        """
        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))
        """
2080
        return self.desc.type()
Y
Yu Yang 已提交
2081

2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114
    @property
    def T(self):
        """
        Permute current Variable with its dimensions reversed.

        If `n` is the dimensions of `x` , `x.T` is equivalent to `x.transpose([n-1, n-2, ..., 0])`.

        Examples:

            .. code-block:: python

                import paddle
                paddle.enable_static()

                x = paddle.ones(shape=[2, 3, 5])
                x_T = x.T

                exe = paddle.static.Executor()
                x_T_np = exe.run(paddle.static.default_main_program(), fetch_list=[x_T])[0]
                print(x_T_np.shape)
                # (5, 3, 2)
        """
        if len(self.shape) == 1:
            return self
        perm = []
        for i in range(len(self.shape)):
            perm.insert(0, i)

        out = self.block.create_var(
            name=unique_name.generate_with_ignorable_key(self.name + '.tmp'),
            dtype=self.dtype,
            type=self.type,
            persistable=False,
2115 2116
            stop_gradient=False,
        )
2117 2118 2119 2120 2121
        input_shape = self.block.create_var(
            name=unique_name.generate_with_ignorable_key(self.name + '.tmp'),
            dtype=self.dtype,
            type=core.VarDesc.VarType.LOD_TENSOR,
            persistable=False,
2122 2123 2124 2125 2126 2127 2128 2129 2130
            stop_gradient=False,
        )

        self.block.append_op(
            type='transpose2',
            inputs={'X': [self]},
            outputs={'Out': [out], 'XShape': [input_shape]},
            attrs={'axis': perm},
        )
2131 2132
        return out

2133 2134 2135
    def clone(self):
        """
        Returns a new static Variable, which is the clone of the original static
2136
        Variable. It remains in the current graph, that is, the cloned Variable
2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160
        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,
2161 2162
            stop_gradient=self.stop_gradient,
        )
2163

2164 2165 2166
        self.block.append_op(
            type='assign', inputs={'X': [self]}, outputs={'Out': [output]}
        )
2167 2168
        return output

W
Wu Yi 已提交
2169
    def _set_error_clip(self, error_clip):
2170 2171 2172 2173 2174 2175 2176 2177 2178
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
2179 2180
        self.error_clip = error_clip

2181 2182 2183 2184 2185 2186 2187 2188
    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.

2189
        Returns:
2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202
            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.

2203
        Returns:
2204 2205 2206 2207 2208 2209
            object
        """
        if hasattr(self, "_info") and key in self._info:
            return self._info[key]
        return None

2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220
    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
2221
            raise ValueError("slice step can not be zero")
2222 2223 2224 2225 2226 2227 2228 2229 2230 2231

        # 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
2232 2233 2234
            start = (
                max(start + length, lower) if start < 0 else min(start, upper)
            )
2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279

        # 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)
2280 2281 2282
                if (index > 0 and index >= self.shape[index]) or (
                    index < 0 and (index + self.shape[index]) < 0
                ):
2283
                    raise IndexError("invalid index")
2284 2285 2286 2287 2288
                start = (
                    max(start + self.shape[index], 0)
                    if start < 0
                    else min(start, self.shape[index])
                )
2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301
                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
2302
    def _cloneVar(self, copy=False):
2303 2304
        if not copy:
            return self.block.create_var(
H
Hongyu Liu 已提交
2305
                name=unique_name.generate_with_ignorable_key(self.name),
2306 2307
                dtype=self.dtype,
            )
2308 2309 2310 2311
        else:
            return self

    def _sliceVar(self, axes, starts, ends):
L
lujun 已提交
2312
        new_var = self._cloneVar()
2313 2314 2315 2316 2317 2318
        self.block.append_op(
            type="slice",
            inputs={'Input': [self]},
            outputs={'Out': [new_var]},
            attrs={'axes': axes, 'starts': starts, 'ends': ends},
        )
2319 2320 2321
        return new_var

    def _concatVar(self, inputs, axis):
L
lujun 已提交
2322
        new_var = self._cloneVar()
2323 2324 2325 2326 2327 2328 2329 2330
        self.block.append_op(
            type="concat",
            inputs={'X': inputs},
            outputs={'Out': [new_var]},
            attrs={
                'axis': axis,
            },
        )
2331 2332 2333 2334 2335
        return new_var

    def _sliceAndConcatVar(self, item, axis):
        if isinstance(item, slice):
            if self.shape[axis] < 0:
L
lujun 已提交
2336
                return self._cloneVar(True)
2337 2338 2339 2340 2341 2342 2343
            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:
2344 2345 2346
                        vars.append(
                            self._sliceVar([axis], [start], [start + 1])
                        )
2347 2348 2349
                        start += step
                else:
                    while start > stop:
2350 2351 2352
                        vars.append(
                            self._sliceVar([axis], [start], [start + 1])
                        )
2353 2354 2355 2356
                        start += step
                return self._concatVar(vars, axis)
        elif isinstance(item, int):
            if self.shape[axis] < 0:
L
lujun 已提交
2357
                return self._cloneVar(True)
2358
            index = int(item)
2359 2360 2361
            if (index > 0 and index >= self.shape[axis]) or (
                index < 0 and (index + self.shape[axis]) < 0
            ):
2362 2363 2364 2365 2366 2367
                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):
2368
        return _getitem_impl_(self, item)
2369

2370
    def __setitem__(self, item, value):
2371
        return _setitem_impl_(self, item, value)
2372

2373 2374
    def get_value(self, scope=None):
        """
2375
        Get the value of variable in given scope.
2376 2377

        Args:
2378
            scope(Scope, optional) : If `scope` is None, it will be set to global scope
2379 2380 2381 2382 2383 2384 2385 2386 2387 2388
                obtained through 'paddle.static.global_scope()'. Otherwise, use `scope`.
                Default: None

        Returns:
            Tensor: the value in given scope.

        Examples:
            .. code-block:: python

                import paddle
2389
                import paddle.static as static
2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413
                import numpy as np

                paddle.enable_static()

                x = static.data(name="x", shape=[10, 10], dtype='float32')

                y = static.nn.fc(x, 10, name='fc')
                place = paddle.CPUPlace()
                exe = static.Executor(place)
                prog = paddle.static.default_main_program()
                exe.run(static.default_startup_program())
                inputs = np.ones((10, 10), dtype='float32')
                exe.run(prog, feed={'x': inputs}, fetch_list=[y, ])
                path = 'temp/tensor_'
                for var in prog.list_vars():
                    if var.persistable:
                        t = var.get_value()
                        paddle.save(t, path+var.name+'.pdtensor')

                for var in prog.list_vars():
                    if var.persistable:
                        t_load = paddle.load(path+var.name+'.pdtensor')
                        var.set_value(t_load)
        """
2414 2415
        # The 'framework' is a low-level module, and 'executor'
        # can not be imported at the begainning of this file.
2416 2417
        # Therefore, the above two modules are dynamically imported.
        from .executor import global_scope
2418

2419 2420
        if scope is not None and not isinstance(scope, core._Scope):
            raise TypeError(
2421 2422 2423 2424
                "`scope` should be None or `paddle.static.Scope` type, but received {}.".format(
                    type(scope)
                )
            )
2425 2426 2427 2428 2429

        if scope is None:
            scope = global_scope()
        var_temp = scope.find_var(self.name)
        if var_temp is None:
2430 2431 2432
            raise ValueError(
                "Can not find Variable '{}' in the Scope.".format(self.name)
            )
2433 2434 2435 2436 2437
        t = var_temp.get_tensor()
        return t

    def set_value(self, value, scope=None):
        '''
2438
        Set the value to the tensor in given scope.
2439 2440 2441

        Args:
            value(Tensor/ndarray) : The value to be set.
2442
            scope(Scope, optional) : If `scope` is None, it will be set to global scope
2443 2444 2445 2446 2447
                obtained through 'paddle.static.global_scope()'. Otherwise, use `scope`.
                Default: None

        Returns:
            None
2448

2449 2450 2451 2452
        Examples:
            .. code-block:: python

                import paddle
2453
                import paddle.static as static
2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479
                import numpy as np

                paddle.enable_static()

                x = static.data(name="x", shape=[10, 10], dtype='float32')

                y = static.nn.fc(x, 10, name='fc')
                place = paddle.CPUPlace()
                exe = static.Executor(place)
                prog = paddle.static.default_main_program()
                exe.run(static.default_startup_program())
                inputs = np.ones((10, 10), dtype='float32')
                exe.run(prog, feed={'x': inputs}, fetch_list=[y, ])
                path = 'temp/tensor_'
                for var in prog.list_vars():
                    if var.persistable:
                        t = var.get_value()
                        paddle.save(t, path+var.name+'.pdtensor')

                for var in prog.list_vars():
                    if var.persistable:
                        t_load = paddle.load(path+var.name+'.pdtensor')
                        var.set_value(t_load)
        '''

        # The 'framework' is a low-level module, and 'executor'
2480
        # can not be imported at the begainning of this file.
2481 2482 2483 2484 2485
        # Therefore, the above two modules are dynamically imported.
        from .executor import global_scope

        if not (isinstance(value, np.ndarray) or hasattr(value, '__array__')):
            raise TypeError(
2486 2487 2488 2489
                "`value` should be `numpy.ndarray` or `LoDTensor`, but received {}.".format(
                    type(value)
                )
            )
2490 2491 2492

        if scope is not None and not isinstance(scope, core._Scope):
            raise TypeError(
2493 2494 2495 2496
                "`scope` should be None or `paddle.static.Scope` type, but received {}.".format(
                    type(scope)
                )
            )
2497 2498 2499 2500 2501 2502

        if scope is None:
            scope = global_scope()

        var_temp = scope.find_var(self.name)
        if var_temp is None:
2503 2504 2505
            raise ValueError(
                "Can not find Variable '{}' in the Scope.".format(self.name)
            )
2506 2507 2508 2509 2510 2511 2512 2513 2514 2515

        t = var_temp.get_tensor()

        if hasattr(value, 'shape'):
            if isinstance(value.shape, (MethodType, FunctionType)):
                value_shape = value.shape()
            else:
                value_shape = value.shape
            if list(t.shape()) != list(value_shape):
                raise ValueError(
2516 2517 2518 2519
                    "{} expected a shape {}, but the received shape is {}.".format(
                        self.name, list(t.shape()), list(value_shape)
                    )
                )
2520 2521 2522 2523 2524 2525 2526 2527 2528 2529

        p = t._place()
        if p.is_cpu_place():
            place = core.CPUPlace()
        elif p.is_cuda_pinned_place():
            place = core.CUDAPinnedPlace()
        elif p.is_xpu_place():
            p = core.Place()
            p.set_place(t._place())
            place = core.XPUPlace(p.xpu_device_id())
2530 2531 2532 2533
        elif p.is_npu_place():
            p = core.Place()
            p.set_place(t._place())
            place = core.NPUPlace(p.npu_device_id())
2534 2535 2536 2537
        elif p.is_mlu_place():
            p = core.Place()
            p.set_place(t._place())
            place = core.MLUPlace(p.mlu_device_id())
2538 2539 2540 2541 2542 2543 2544
        else:
            p = core.Place()
            p.set_place(t._place())
            place = core.CUDAPlace(p.gpu_device_id())

        t.set(value, place)

2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567
    def size(self):
        """
        Returns the number of elements for current Variable, which is a int64 Variable with shape [1]

        Returns:
            Variable: the number of elements for current Variable

        Examples:
            .. code-block:: python

                import paddle

                paddle.enable_static()

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

                # get the number of elements of the Variable
                y = x.size()
        """

        output = self.block.create_var(
            name=unique_name.generate_with_ignorable_key(self.name + "_size"),
2568 2569
            dtype=core.VarDesc.VarType.INT64,
        )
2570

2571 2572 2573
        self.block.append_op(
            type='size', inputs={'Input': [self]}, outputs={'Out': [output]}
        )
2574 2575
        return output

2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629
    def _set_attr(self, name, val):
        """
        Set the value of attribute by attribute's name.

        Args:
            name(str): the attribute name.
            val(int|str|list): the value of the attribute.
        """
        self._update_desc_attr(name, val)

    def _has_attr(self, name):
        """
        Whether this Variable has the attribute with the name `name` or not.

        Args:
            name(str): the attribute name.

        Returns:
            bool: True if has this attribute.
        """
        return self.desc.has_attr(name)

    def _remove_attr(self, name):
        self.desc.remove_attr(name)

    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(int|str|list): the value of the attribute.
        """
        self.desc._set_attr(name, val)

    @property
    def attr_names(self):
        """Get the names of all attributes defined."""
        return self.desc.attr_names()

    def _get_attr(self, name):
        """
        Get the attribute by name.

        Args:
            name(str): the attribute name.

        Returns:
            int|str|list: The attribute value. The return value
            can be any valid attribute type.
        """
        return self.desc.attr(name)

    @property
2630
    def dist_attr(self):
2631
        """
2632
        Get distributed attribute of this Variable.
2633
        """
2634
        return self.desc.dist_attr
2635

2636 2637
    @dist_attr.setter
    def dist_attr(self, dist_attr):
2638
        """
2639
        Set distributed attribute of this Variable.
2640
        """
2641
        self.desc.dist_attr = dist_attr
2642

Y
Yu Yang 已提交
2643

F
fengjiayi 已提交
2644 2645 2646
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
2647

2648 2649
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
2650 2651 2652 2653
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
2654
        op_proto = framework_pb2.OpProto.FromString(bytes(pbstr))
F
fengjiayi 已提交
2655 2656 2657 2658 2659
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
2660 2661 2662 2663
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
2664 2665 2666 2667 2668 2669 2670 2671
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
2672 2673
            self.__class__, '_instance'
        ), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
2674 2675 2676 2677 2678 2679
        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):
2680 2681 2682 2683 2684 2685 2686 2687
        """
        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
2688 2689
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
2690 2691
        return self.op_proto_map[type]

2692 2693
    def update_op_proto(self):
        op_protos = get_all_op_protos()
2694
        custom_op_names = []
2695 2696 2697
        for proto in op_protos:
            if proto.type not in self.op_proto_map:
                self.op_proto_map[proto.type] = proto
2698 2699 2700
                custom_op_names.append(proto.type)

        return custom_op_names
2701

2702 2703 2704 2705
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
2706
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
2707
            core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
2708
            core.op_proto_and_checker_maker.kOpCreationCallstackAttrName(),
2709
            core.op_proto_and_checker_maker.kOpDeviceAttrName(),
2710 2711
        }

F
fengjiayi 已提交
2712

X
Xin Pan 已提交
2713
class Operator(object):
2714
    """
2715 2716 2717 2718 2719 2720 2721
    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
2722
        type(str): The type of operator. Default None.
2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742
        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
2743
        Block.append_op or Block._prepend_op instead.
2744 2745 2746 2747

    Examples:
        .. code-block:: python

2748
            import paddle.fluid as fluid
2749
            cur_program = fluid.Program()
2750 2751 2752 2753 2754
            cur_block = cur_program.current_block()
            # var1 += var2 + var3
            cur_block.append_op(type="sum",
                                inputs={"X": [var1, var2, var3]},
                                outputs={"Out": [var1]})
2755
    """
2756

2757
    OP_WITHOUT_KERNEL_SET = {
2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788
        'feed',
        'fetch',
        'recurrent',
        'go',
        'rnn_memory_helper_grad',
        'conditional_block',
        'while',
        'send',
        'recv',
        'listen_and_serv',
        'fl_listen_and_serv',
        'ncclInit',
        'select',
        'checkpoint_notify',
        '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',
        'c_wait_comm',
        'c_wait_compute',
        'c_gen_hccl_id',
        'c_comm_init_hccl',
        'copy_cross_scope',
        'c_gen_cncl_id',
2789
    }
2790

2791 2792 2793
    def __init__(
        self, block, desc, type=None, inputs=None, outputs=None, attrs=None
    ):
2794 2795 2796 2797 2798 2799 2800 2801 2802 2803
        # read attr type index from op proto to avoid unexpected type
        # conversions, e.g. narrowing conversion like double to float
        try:
            proto = OpProtoHolder.instance().get_op_proto(type)
            self._attr_types = {}
            for attr in proto.attrs:
                self._attr_types[attr.name] = attr.type
        except ValueError:
            pass

J
Jiabin Yang 已提交
2804
        if _non_static_mode():
2805 2806
            if type is None:
                raise ValueError(
2807 2808
                    "`type` to initialized an Operator can not be None."
                )
J
Jiabin Yang 已提交
2809
            self._type = type
M
minqiyang 已提交
2810
            self.attrs = attrs if attrs else {}
2811 2812 2813 2814 2815 2816 2817 2818 2819 2820
        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

2821 2822 2823
            # attr for static mode cuda graph
            self._cuda_graph_attr = _current_cuda_graph_mode

2824 2825 2826
            op_maker = core.op_proto_and_checker_maker

            if op_maker.kOpRoleAttrName() not in op_attrs:
2827
                op_attrs[
2828 2829
                    op_maker.kOpRoleAttrName()
                ] = self.block.program._op_role
2830 2831

            role_var_name = op_maker.kOpRoleVarAttrName()
2832 2833 2834 2835
            if (
                len(self.block.program._op_role_var) != 0
                and role_var_name not in op_attrs
            ):
2836
                op_attrs[role_var_name] = self.block.program._op_role_var
2837 2838 2839 2840 2841

            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:
2842 2843 2844 2845 2846
                # NOTE(Aurelius84): prog.clone() will lead that var.op is always None,
                # we add this to fix the problem.
                for arg in self.desc.output_arg_names():
                    if block.has_var(arg) and block.var(arg).op is None:
                        block.var(arg).op = self
2847 2848 2849
                return
            if type is None:
                raise ValueError(
2850 2851
                    "`type` to initialized an Operator can not be None."
                )
2852 2853
            else:
                callstack_var_name = op_maker.kOpCreationCallstackAttrName()
2854 2855 2856
                op_attrs[callstack_var_name] = []
                for frame in traceback.extract_stack():
                    op_attrs[callstack_var_name].append(
2857
                        '  File "{}", line {}, in {}'.format(
2858 2859 2860 2861 2862 2863
                            frame[0], frame[1], frame[2]
                        )
                    )
                    op_attrs[callstack_var_name].append(
                        '    {}'.format(frame[3])
                    )
2864 2865 2866 2867 2868 2869 2870

            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()

2871 2872 2873 2874 2875 2876 2877 2878
            # 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:
2879 2880 2881
                    warnings.warn(
                        "The Op(%s) is not support to set device." % type
                    )
2882
                if 'force_cpu' in op_attrs:
2883 2884 2885
                    if (
                        type == 'less_than' and op_attrs['force_cpu'] != None
                    ) or op_attrs['force_cpu'] != False:
2886 2887 2888
                        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 "
2889 2890
                            "used at the same time." % type
                        )
2891
            if _current_pipeline_stage is not None:
2892 2893 2894 2895 2896
                pipeline_attr_name = (
                    'pipeline_stage' + core.kAutoParallelSuffix()
                )
                self._update_desc_attr(
                    pipeline_attr_name, _current_pipeline_stage
2897
                )
2898

2899 2900 2901 2902 2903 2904 2905 2906 2907
            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)
2908 2909 2910
                    assert (
                        found or in_proto.dispensable
                    ), "Input {} not found".format(in_proto.name)
2911 2912
                    if found:
                        in_args = inputs[in_proto.name]
2913
                        if not isinstance(in_args, (list, tuple)):
2914 2915 2916 2917
                            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."
2918 2919
                                % (in_proto.name, len(in_args))
                            )
2920
                        in_arg_names = []
2921
                        for index, arg in enumerate(in_args):
2922
                            if isinstance(arg, str):
2923
                                in_arg_names.append(arg)
2924
                            elif isinstance(arg, bytes):
2925
                                in_arg_names.append(arg.decode())
2926
                            elif isinstance(arg, (Variable, core.VarBase)):
2927
                                in_arg_names.append(arg.name)
2928
                            else:
2929 2930 2931 2932
                                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."
2933 2934 2935
                                    "but received : %s"
                                    % (in_proto.name, type, arg)
                                )
2936 2937 2938 2939 2940 2941 2942 2943 2944
                        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):
2945
                        raise ValueError(
2946 2947 2948 2949 2950 2951
                            (
                                "Incorrect setting for output(s) of "
                                "operator \"%s\", should set: [%s]."
                            )
                            % (type, m.name)
                        )
2952 2953 2954 2955 2956 2957 2958 2959 2960
                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."
2961 2962
                            % (out_proto.name, len(out_args))
                        )
2963 2964
                    out_arg_names = []
                    for arg in out_args:
2965
                        if isinstance(arg, str):
2966 2967
                            out_arg_names.append(arg)
                        else:
2968
                            out_arg_names.append(arg.name)
2969
                        # TODO(minqiyang): could we remove variable's op in static mode?
J
Jiabin Yang 已提交
2970
                        if not _non_static_mode():
2971
                            if isinstance(arg, str):
2972 2973 2974
                                block.var(arg).op = self
                            else:
                                arg.op = self
2975 2976
                    self.desc.set_output(out_proto.name, out_arg_names)

2977
            extra_attrs_map = core.get_op_extra_attrs(type)
2978 2979 2980 2981 2982
            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
2983 2984 2985
                    if (attr_name not in op_attrs) or (
                        op_attrs[attr_name] is None
                    ):
2986 2987 2988
                        continue
                    attr_val = op_attrs[attr_name]
                    self._update_desc_attr(attr_name, attr_val)
2989
                for attr_name in extra_attrs_map.keys():
2990 2991 2992 2993 2994 2995
                    if (attr_name not in op_attrs) or (
                        op_attrs[attr_name] is None
                    ):
                        self._update_desc_attr(
                            attr_name, extra_attrs_map[attr_name]
                        )
2996 2997
                    else:
                        self._update_desc_attr(attr_name, op_attrs[attr_name])
2998

J
jianghaicheng 已提交
2999 3000
            # proto.attrs doesn't include ipu_index
            if core.is_compiled_with_ipu():
3001
                if global_ipu_index >= 0:
3002 3003 3004
                    self._update_desc_attr(
                        ipu_index_attr_name, global_ipu_index
                    )
3005
                if global_ipu_stage >= 0:
3006 3007 3008
                    self._update_desc_attr(
                        ipu_stage_attr_name, global_ipu_stage
                    )
J
jianghaicheng 已提交
3009

3010 3011 3012 3013 3014
            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
3015
    def _has_kernel(self, op_type):
3016 3017
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
3018
    def to_string(self, throw_on_error):
3019
        """
3020 3021
        Get debug string.

3022
        Args:
3023 3024
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
3025

3026 3027
        Returns:
            str: The debug string.
3028 3029

        """
3030
        protostr = self.desc.serialize_to_string()
3031
        proto = framework_pb2.OpDesc.FromString(bytes(protostr))
Y
Yang Yang(Tony) 已提交
3032 3033
        return _debug_string_(proto, throw_on_error)

3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065
    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
Z
zhangchunle 已提交
3066
        ), "skip_op_callstack parameter's type is error, expect bool, received {}".format(
3067 3068
            type(skip_op_callstack)
        )
3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094
        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

3095 3096 3097
            attr_type = self.desc.attr_type(name, True)
            if attr_type == core.AttrType.VAR:
                attr_var_name = self.desc.attr(name, True).name()
3098 3099 3100
                a = "{name} = Var['{value}']".format(
                    name=name, type=attr_type, value=attr_var_name
                )
3101 3102 3103 3104 3105 3106 3107 3108 3109 3110
                attrs_str += a
                if i != len(attr_names) - 1:
                    attrs_str += ", "
                continue

            if attr_type == core.AttrType.VARS:
                attr_var_names = [
                    "'%s'" % var.name() for var in self.desc.attr(name, True)
                ]
                a = "{name} = Vars[{value}]".format(
3111 3112
                    name=name, type=attr_type, value=','.join(attr_var_names)
                )
3113 3114 3115 3116 3117
                attrs_str += a
                if i != len(attr_names) - 1:
                    attrs_str += ", "
                continue

3118 3119
            if attr_type == core.AttrType.BLOCK:
                a = "{name} = block[{value}]".format(
3120 3121
                    name=name, type=attr_type, value=self._block_attr_id(name)
                )
3122 3123 3124 3125 3126 3127 3128
                attrs_str += a
                if i != len(attr_names) - 1:
                    attrs_str += ", "
                continue

            if attr_type == core.AttrType.BLOCKS:
                a = "{name} = blocks{value}".format(
3129 3130
                    name=name, type=attr_type, value=self._blocks_attr_ids(name)
                )
3131 3132 3133 3134 3135
                attrs_str += a
                if i != len(attr_names) - 1:
                    attrs_str += ", "
                continue

3136
            # it is bytes of serialized protobuf
3137 3138 3139 3140 3141
            if (
                is_compiled_with_cinn()
                and self.type == 'cinn_launch'
                and name == 'compilation_key'
            ):
3142 3143
                key = self.desc.attr(name)
                v = core.get_serialize_comile_key(key)
3144 3145 3146 3147 3148 3149 3150 3151 3152
                prog = Program()
                prog = prog.parse_from_string(v)
                s = prog._to_readable_code()
                lines = s.split('\n')
                value = '\n'.join(['      ' + line for line in lines])
                value = '\n' + value
            else:
                value = self.desc.attr(name)

3153 3154 3155
            a = "{name} = {value}".format(
                name=name, type=attr_type, value=value
            )
3156

3157 3158 3159 3160
            attrs_str += a
            if i != len(attr_names) - 1:
                attrs_str += ", "

3161 3162 3163 3164
        from paddle.distributed.auto_parallel.dist_context import (
            get_default_distributed_context,
        )

3165
        dist_context = get_default_distributed_context()
3166 3167
        dist_op = dist_context.get_dist_op_for_program(self)
        if dist_op is not None:
3168 3169 3170
            attrs_str += ", {name} = {value}".format(
                name="dist_attr", value=dist_op
            )
3171

3172
        if outputs_str != "{}":
3173 3174 3175 3176 3177 3178
            op_str = "{outputs} = {op_type}(inputs={inputs}, {attrs})".format(
                outputs=outputs_str,
                op_type=self.type,
                inputs=inputs_str,
                attrs=attrs_str,
            )
3179
        else:
3180 3181 3182
            op_str = "{op_type}(inputs={inputs}, {attrs})".format(
                op_type=self.type, inputs=inputs_str, attrs=attrs_str
            )
3183 3184
        return op_str

Y
Yang Yang(Tony) 已提交
3185
    def __str__(self):
3186
        return self._to_readable_code()
3187 3188 3189

    __repr__ = __str__

F
fengjiayi 已提交
3190 3191
    @property
    def type(self):
3192
        return self.desc.type()
F
fengjiayi 已提交
3193 3194

    def input(self, name):
3195
        r"""
3196
        Get the input arguments according to the input parameter name.
3197

3198 3199
        Args:
            name(str): The input parameter name.
3200

3201 3202 3203
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
3204
        """
F
fengjiayi 已提交
3205 3206
        return self.desc.input(name)

W
Wu Yi 已提交
3207
    def _rename_input(self, old_name, new_name):
3208 3209 3210 3211 3212 3213 3214 3215 3216 3217
        """
        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
3218
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
3219

W
Wu Yi 已提交
3220
    def _rename_output(self, old_name, new_name):
3221 3222 3223 3224 3225 3226 3227 3228 3229 3230
        """
        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
3231
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
3232

F
fengjiayi 已提交
3233 3234 3235 3236
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
3237 3238 3239 3240 3241 3242 3243 3244
    @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
3245
    def output(self, name):
3246
        r"""
3247
        Get output arguments by the output parameter name.
3248

3249 3250
        Args:
            name(str): The output parameter name.
3251

3252 3253 3254
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
3255
        """
F
fengjiayi 已提交
3256 3257 3258 3259 3260 3261
        return self.desc.output(name)

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

3262 3263 3264 3265 3266 3267
    @property
    def idx(self):
        for i, op in enumerate(self.block.ops):
            if op == self:
                return i
        raise ValueError(
3268 3269
            "Can't find op itself in it's block. It could be a bug of Paddle."
        )
3270

F
fengjiayi 已提交
3271
    def has_attr(self, name):
3272
        """
3273 3274
        Whether this Operator has the attribute with name or not.

3275
        Args:
3276
            name(str): the attribute name.
3277

3278 3279
        Returns:
            bool: True if has this attribute.
3280 3281

        """
F
fengjiayi 已提交
3282 3283 3284
        return self.desc.has_attr(name)

    def attr_type(self, name):
3285
        """
3286
        Get the type of attribute by attribute's name.
3287

3288 3289
        Args:
            name(str): the attribute name.
3290

3291 3292
        Returns:
            core.AttrType: the attribute type.
3293
        """
3294
        return self.desc.attr_type(name, True)
F
fengjiayi 已提交
3295

W
Wu Yi 已提交
3296
    def _set_attr(self, name, val):
3297 3298 3299 3300 3301 3302 3303 3304 3305 3306
        """
        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
3307 3308
        self._update_desc_attr(name, val)

3309 3310 3311
    def _remove_attr(self, name):
        self.desc.remove_attr(name)

G
gongweibao 已提交
3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322
    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).
        """
3323 3324 3325 3326 3327
        if isinstance(val, Variable):
            self.desc.set_var_attr(name, val.desc)
        elif isinstance(val, list) and _all_is_type(val, Variable):
            self.desc.set_vars_attr(name, [v.desc for v in val])
        elif isinstance(val, Block):
Q
Qiyang Min 已提交
3328
            self.desc.set_block_attr(name, val.desc)
3329
        elif isinstance(val, list) and val and _all_is_type(val, Block):
3330
            self.desc.set_blocks_attr(name, [v.desc for v in val])
3331 3332 3333
        elif isinstance(val, core.BlockDesc) or isinstance(
            val, core.ProgramDesc
        ):
Q
Qiyang Min 已提交
3334 3335
            self.desc.set_serialized_attr(name, val.serialize_to_string())
        else:
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
            self._update_desc_plain_attr(name, val)

    def _update_desc_plain_attr(self, name, val):
        desc = self.desc
        if not hasattr(self, "_attr_types") or (name not in self._attr_types):
            desc._set_attr(name, val)
            return

        type_index = self._attr_types[name]
        if type_index == core.AttrType.BOOL:
            desc._set_bool_attr(name, val)
        elif type_index == core.AttrType.INT:
            desc._set_int32_attr(name, val)
        elif type_index == core.AttrType.LONG:
            desc._set_int64_attr(name, val)
        elif type_index == core.AttrType.FLOAT:
            desc._set_float32_attr(name, val)
        # elif type_index == core.AttrType.FLOAT64:
        #     desc._set_float64_attr(name, val)
        elif type_index == core.AttrType.STRING:
            desc._set_str_attr(name, val)
        elif type_index == core.AttrType.BOOLS:
            desc._set_bools_attr(name, val)
        elif type_index == core.AttrType.INTS:
            desc._set_int32s_attr(name, val)
        elif type_index == core.AttrType.LONGS:
            desc._set_int64s_attr(name, val)
        elif type_index == core.AttrType.FLOATS:
            desc._set_float32s_attr(name, val)
        elif type_index == core.AttrType.FLOAT64S:
            desc._set_float64s_attr(name, val)
        elif type_index == core.AttrType.STRINGS:
            desc._set_strs_attr(name, val)
        else:
            # defaults to old methods
            desc._set_attr(name, val)
Y
yuyang18 已提交
3372

F
fengjiayi 已提交
3373 3374
    @property
    def attr_names(self):
3375
        return self.desc.attr_names(True)
F
fengjiayi 已提交
3376 3377

    def attr(self, name):
3378
        """
3379 3380
        Get the attribute by name.

3381
        Args:
3382
            name(str): the attribute name.
3383

3384 3385
        Returns:
            bool|int|str|float|list: The attribute value. The return value
3386 3387
            can be any valid attribute type.
        """
F
fengjiayi 已提交
3388
        return self.desc.attr(name)
Y
Yu Yang 已提交
3389

W
Wu Yi 已提交
3390
    def _block_attr_id(self, name):
3391
        """
G
gongweibao 已提交
3392
        Get the block attribute's id by name.
3393

3394 3395
        Args:
            name(str): the attribute name.
3396

3397 3398
        Returns:
            int: the block index.
3399
        """
W
Wu Yi 已提交
3400
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
3401

W
Wu Yi 已提交
3402
    def _block_attr(self, name):
G
gongweibao 已提交
3403 3404 3405 3406 3407 3408 3409 3410 3411 3412
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
3413
        id = self._block_attr_id(name)
3414
        assert id >= 0 and id < len(self.block.program.blocks)
G
gongweibao 已提交
3415 3416
        return self.block.program.blocks[id]

W
Wu Yi 已提交
3417
    def _blocks_attr(self, name):
G
gongweibao 已提交
3418 3419 3420 3421 3422 3423 3424 3425 3426 3427
        """
        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
3428
        for i in self._blocks_attr_ids(name):
3429
            assert i >= 0 and i < len(self.block.program.blocks)
G
gongweibao 已提交
3430 3431 3432 3433
            attrs.append(self.block.program.blocks[i])

        return attrs

W
Wu Yi 已提交
3434
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
3435 3436 3437 3438 3439 3440 3441 3442 3443 3444
        """
        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
3445
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
3446

3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457
    def _var_attr(self, name):
        """
        Get the Variable attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            Variable: the Variable attribute.
        """
        attr_type = self.desc.attr_type(name, True)
3458 3459 3460 3461 3462
        assert (
            attr_type == core.AttrType.VAR
        ), "Required type attr({}) is Variable, but received {}".format(
            name, attr_type
        )
3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476
        attr_var_name = self.desc.attr(name, True).name()
        return self.block._var_recursive(attr_var_name)

    def _vars_attr(self, name):
        """
        Get the Variables attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            Variables: the Variables attribute.
        """
        attr_type = self.desc.attr_type(name, True)
3477 3478 3479 3480 3481
        assert (
            attr_type == core.AttrType.VARS
        ), "Required type attr({}) is list[Variable], but received {}".format(
            name, attr_type
        )
3482 3483 3484 3485 3486 3487
        attr_vars = [
            self.block._var_recursive(var.name())
            for var in self.desc.attr(name, True)
        ]
        return attr_vars

J
JiayiFeng 已提交
3488
    def all_attrs(self):
F
fengjiayi 已提交
3489
        """
3490 3491 3492
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
3493
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
3494 3495 3496 3497
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
3498
            attr_type = self.desc.attr_type(n, True)
G
gongweibao 已提交
3499
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
3500
                attr_map[n] = self._block_attr(n)
3501
            elif attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
3502
                attr_map[n] = self._blocks_attr(n)
3503 3504 3505 3506 3507 3508
            elif attr_type == core.AttrType.VAR:
                attr_map[n] = self._var_attr(n)
            elif attr_type == core.AttrType.VARS:
                attr_map[n] = self._vars_attr(n)
            else:
                attr_map[n] = self.attr(n)
G
gongweibao 已提交
3509

F
fengjiayi 已提交
3510 3511
        return attr_map

3512 3513 3514
    def _is_optimize_op(self):
        op_maker = core.op_proto_and_checker_maker
        OPTIMIZE = core.op_proto_and_checker_maker.OpRole.Optimize
3515 3516 3517 3518

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

3519 3520 3521
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(OPTIMIZE):
            return True
3522 3523 3524 3525 3526 3527 3528 3529

        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()):
3530 3531
            return False

3532 3533 3534 3535 3536 3537
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(BACKWARD):
            return True

        return False

3538
    @property
3539
    def dist_attr(self):
3540
        """
3541
        Get distributed attribute of this Variable.
3542
        """
3543
        return self.desc.dist_attr
3544

3545 3546
    @dist_attr.setter
    def dist_attr(self, dist_attr):
3547
        """
3548
        Set distributed attribute of this Variable.
3549
        """
3550
        self.desc.dist_attr = dist_attr
3551

Y
Yu Yang 已提交
3552

Y
Yu Yang 已提交
3553
class Block(object):
3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567
    """
    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
3568
        use `Program._create_block()` to create a block.
3569 3570 3571 3572

    Examples:
        .. code-block:: python

3573 3574 3575
            import paddle.fluid as fluid

            cur_program = fluid.Program()
3576 3577 3578 3579 3580 3581 3582 3583 3584
            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
3585
    def __init__(self, program, idx):
Y
Yu Yang 已提交
3586
        self.desc = program.desc.block(idx)
3587
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
3588
        self.ops = list()  # operator list
Y
Yu Yang 已提交
3589
        self.program = program
3590
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
3591

3592
    def __str__(self):
3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626
        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
Z
zhangchunle 已提交
3627
        ), "skip_op_callstack parameter's type is error, expect bool, received {}".format(
3628 3629
            type(skip_op_callstack)
        )
3630 3631 3632 3633 3634 3635 3636
        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(
3637 3638
                op._to_readable_code(skip_op_callstack)
            )
3639 3640
        block_str += "}"
        return block_str
Y
Yang Yang(Tony) 已提交
3641

F
fengjiayi 已提交
3642 3643
    def to_string(self, throw_on_error, with_details=False):
        """
3644 3645
        Get debug string.

F
fengjiayi 已提交
3646 3647
        Args:
            throw_on_error(bool): raise exception when self is not initialized
3648
                when throw_on_error is True.
F
update  
fengjiayi 已提交
3649
            with_details(bool): more details about variables and parameters
3650 3651
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
3652

3653 3654
        Returns:
            str: The debug string.
F
fengjiayi 已提交
3655
        """
3656
        assert isinstance(throw_on_error, bool) and isinstance(
3657 3658
            with_details, bool
        )
F
fengjiayi 已提交
3659
        if with_details:
F
fengjiayi 已提交
3660
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
3661
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
3662 3663 3664
                self.idx,
                self.parent_idx,
            )
3665
            for var in list(self.vars.values()):
F
fengjiayi 已提交
3666
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
3667 3668
                    r"\n    \1", var.to_string(throw_on_error, with_details)
                )
F
fengjiayi 已提交
3669
            for op in self.ops:
F
fengjiayi 已提交
3670
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
3671 3672
                    r"\n    \1", op.to_string(throw_on_error)
                )
F
fengjiayi 已提交
3673 3674 3675
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
3676
            proto = framework_pb2.BlockDesc.FromString(bytes(protostr))
F
fengjiayi 已提交
3677 3678
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
3679 3680 3681

    __repr__ = __str__

Y
Yu Yang 已提交
3682 3683
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
3684
        return self.desc.parent
Y
Yu Yang 已提交
3685

Y
Yu Yang 已提交
3686 3687 3688 3689
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
3690
    def _set_forward_block_idx(self, idx):
3691 3692 3693 3694 3695 3696 3697 3698 3699
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
3700
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
3701

3702 3703 3704 3705 3706 3707 3708 3709
    @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
3710 3711
    @property
    def idx(self):
Y
Yu Yang 已提交
3712
        return self.desc.id
Y
Yu Yang 已提交
3713

Q
Qiao Longfei 已提交
3714
    def var(self, name):
3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727
        """
        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.
        """
3728
        if not isinstance(name, str):
M
minqiyang 已提交
3729
            raise TypeError(
3730 3731 3732
                "var require string as parameter, but get %s instead."
                % (type(name))
            )
Y
Yu Yang 已提交
3733 3734
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
3735
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
3736
        return v
Q
Qiao Longfei 已提交
3737

X
Xin Pan 已提交
3738
    def _find_var_recursive(self, name):
3739 3740 3741 3742 3743 3744 3745
        """
        Get a Variable by name from this block recursively.

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

        Returns:
X
Xin Pan 已提交
3746
            Variable: the Variable with the giving name. Or None if not found.
3747
        """
Y
Yu Yang 已提交
3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771
        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
3772
        return None
Y
Yu Yang 已提交
3773

X
Xin Pan 已提交
3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792
    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
3793

Q
Qiao Longfei 已提交
3794
    def all_parameters(self):
3795
        return list(self.iter_parameters())
3796

3797
    def iter_parameters(self):
3798 3799 3800 3801 3802
        return (
            item[1]
            for item in self.vars.items()
            if isinstance(item[1], Parameter)
        )
Q
Qiao Longfei 已提交
3803

Y
Yu Yang 已提交
3804
    def create_var(self, *args, **kwargs):
J
Jiabin Yang 已提交
3805
        if _non_static_mode():
L
Leo Chen 已提交
3806 3807
            var = _varbase_creator(*args, **kwargs)
        else:
3808 3809 3810
            var = Variable(block=self, *args, **kwargs)
            if 'initializer' in kwargs:
                kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
3811
        return var
Y
Yu Yang 已提交
3812

Q
Qiao Longfei 已提交
3813 3814 3815
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
3816
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
3817 3818
        """
        Rename variable in vars and ops' inputs and outputs
3819 3820

        Args:
3821 3822
            name(str|bytes): the name that need to be renamed.
            new_name(str|bytes): the name that need to rename to.
3823 3824 3825 3826 3827 3828 3829 3830

        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
3831
        """
3832 3833
        # Ensure the type of name and new_name is str
        name = name.decode() if isinstance(name, bytes) else name
3834 3835 3836
        new_name = (
            new_name.decode() if isinstance(new_name, bytes) else new_name
        )
M
minqiyang 已提交
3837

T
typhoonzero 已提交
3838
        if not self.has_var(name):
3839
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
3840 3841
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
3842
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
3843 3844 3845 3846 3847 3848
            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
3849
            var_type = "Variable"
T
wip  
typhoonzero 已提交
3850 3851 3852 3853
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
3854
        orig_var_type = v.type
3855
        self.desc._rename_var(name.encode(), new_name.encode())
W
Wu Yi 已提交
3856
        # NOTE: v is destroyed by C++ after calling _rename_var.
3857
        d = self.desc.find_var(new_name.encode())
T
typhoonzero 已提交
3858
        if var_type == "Parameter":
L
Leo Chen 已提交
3859
            if in_dygraph_mode():
3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870
                var = EagerParamBase(
                    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,
                )
3871
            else:
J
Jiabin Yang 已提交
3872
                if _in_legacy_dygraph():
3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883
                    var = ParamBase(
                        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,
                    )
J
Jiabin Yang 已提交
3884
                else:
3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896
                    var = Parameter(
                        self,
                        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
3897
        elif var_type == "Variable":
3898 3899 3900 3901 3902 3903 3904
            var = Variable(
                self,
                type=orig_var_type,
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient,
            )
T
wip  
typhoonzero 已提交
3905

W
Wu Yi 已提交
3906
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
3907 3908 3909
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
3910
        self._sync_with_cpp()
3911
        return var
T
typhoonzero 已提交
3912

3913 3914 3915
    def _remove_var(self, name, sync=True):
        if sync == True:
            self._sync_with_cpp()
3916
        self.desc._remove_var(name.encode())
3917 3918
        del self.vars[name]

Y
Yu Yang 已提交
3919 3920
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
3921
        param = None
L
Leo Chen 已提交
3922
        if in_dygraph_mode():
J
Jiabin Yang 已提交
3923
            param = EagerParamBase(*args, **kwargs)
L
Leo Chen 已提交
3924
        else:
J
Jiabin Yang 已提交
3925 3926 3927 3928
            if _in_legacy_dygraph():
                param = ParamBase(*args, **kwargs)
            else:
                param = Parameter(global_block, *args, **kwargs)
3929

3930
        if 'initializer' in kwargs:
3931 3932 3933 3934 3935

            def _is_inited_by(block, var):
                init_ops = []
                for op in block.ops:
                    if var.name in op.output_arg_names:
3936
                        # In startup_program, "c_broadcast" and "c_sync_comm_stream"
T
tangwei12 已提交
3937
                        # are treated as initialization ops that cause error.
3938
                        # Think of "c_broadcast" and "c_sync_comm_stream" as a special case here.
3939 3940
                        # NOTE: "coalesce_tensor" is a special case for rnn with cudnn support
                        if op.type in [
3941 3942 3943
                            "c_broadcast",
                            "c_sync_comm_stream",
                            "coalesce_tensor",
3944
                        ]:
3945
                            continue
3946 3947 3948 3949 3950 3951 3952
                        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:
3953 3954 3955 3956 3957 3958
                raise RuntimeError(
                    "param "
                    + param.name
                    + " is inited by multiple init ops "
                    + str(init_ops)
                )
3959
            elif init_ops_len == 1:
3960
                # TODO already inited, do nothing, should log a warning
3961 3962 3963
                pass
            else:
                initializer(param, self)
Q
Qiao Longfei 已提交
3964
        return param
Y
Yu Yang 已提交
3965

Y
Yu Yang 已提交
3966
    def append_op(self, *args, **kwargs):
3967 3968 3969 3970 3971 3972
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
J
Jiabin Yang 已提交
3973
        if _non_static_mode():
3974
            attrs = kwargs.get("attrs", {})
Z
zyfncg 已提交
3975
            inplace_map = kwargs.get("inplace_map", None)
J
Jiabin Yang 已提交
3976
            type = kwargs.get("type", None)
3977 3978 3979
            warnings.warn(
                "Op `%s` is executed through `append_op` under the dynamic mode, "
                "the corresponding API implementation needs to be upgraded to "
3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990
                "using `_C_ops` method." % type,
                DeprecationWarning,
            )
            op = Operator(
                block=self,
                desc=None,
                type=type,
                inputs=None,
                outputs=None,
                attrs=attrs,
            )
3991

M
minqiyang 已提交
3992 3993 3994
            # record ops in tracer rather than blocks
            #
            # TODO(minqiyang): add op stop_gradient support in static mode too.
L
lujun 已提交
3995
            # currently, we only support stop_gradient in dygraph mode.
J
Jiabin Yang 已提交
3996

3997 3998 3999 4000 4001 4002 4003 4004
            _dygraph_tracer().trace_op(
                type,
                kwargs.get("inputs", {}),
                kwargs.get("outputs", {}),
                attrs if attrs else {},
                kwargs.get("stop_gradient", False),
                inplace_map,
            )
M
minqiyang 已提交
4005
        else:
4006 4007
            from paddle.fluid.dygraph.base import param_guard

4008
            op_desc = self.desc.append_op()
4009 4010 4011 4012 4013 4014
            # NOTE(Aurelius84): In case of @to_static, all VarBase(s) should
            # be converted into Variable(s) with same name and block location.
            # This is ONE and ONLY logic of type transformation of dy2static.
            inputs = kwargs.get("inputs", None)
            outputs = kwargs.get("outputs", None)
            with param_guard(inputs), param_guard(outputs):
4015 4016 4017 4018 4019 4020 4021 4022
                op = Operator(
                    block=self,
                    desc=op_desc,
                    type=kwargs.get("type", None),
                    inputs=inputs,
                    outputs=outputs,
                    attrs=kwargs.get("attrs", None),
                )
4023

M
minqiyang 已提交
4024
            self.ops.append(op)
M
minqiyang 已提交
4025

4026 4027
        return op

W
Wu Yi 已提交
4028
    def _insert_op(self, index, *args, **kwargs):
4029 4030 4031 4032 4033 4034 4035 4036 4037
        """
        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
4038
        self._sync_with_cpp()
F
fangshuixun007 已提交
4039
        return self._insert_op_without_sync(index, *args, **kwargs)
Q
qiaolongfei 已提交
4040

4041 4042
    def _insert_op_without_sync(self, index, *args, **kwargs):
        """
4043
        Insert an Operator according to the giving arguments,
4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055 4056 4057
        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):
4058 4059 4060 4061 4062 4063 4064 4065 4066
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
4067 4068
        if sync == True:
            self._sync_with_cpp()
W
Wu Yi 已提交
4069
        self.desc._remove_op(index, index + 1)
4070 4071
        del self.ops[index]

W
Wu Yi 已提交
4072
    def _slice_ops(self, start, end):
4073 4074 4075 4076 4077 4078 4079 4080 4081 4082
        """
        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
4083
        return self.ops[start:end]
Y
Yancey1989 已提交
4084

W
Wu Yi 已提交
4085
    def _prepend_op(self, *args, **kwargs):
J
Jiabin Yang 已提交
4086
        if _non_static_mode():
J
Jiabin Yang 已提交
4087 4088
            type = kwargs.get("type", None)
            attrs = kwargs.get("attrs", {})
4089 4090 4091 4092 4093 4094 4095 4096 4097 4098 4099
            op = Operator(
                self, None, type=type, inputs=None, outputs=None, attrs=attrs
            )

            _dygraph_tracer().trace_op(
                type,
                kwargs.get("inputs", {}),
                kwargs.get("outputs", {}),
                attrs if attrs else {},
                kwargs.get("stop_gradient", False),
            )
M
minqiyang 已提交
4100
        else:
4101
            op_desc = self.desc._prepend_op()
4102 4103 4104 4105 4106 4107 4108 4109
            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
4110
            self.ops.insert(0, op)
4111

Y
Yu Yang 已提交
4112 4113
        return op

W
Wu Yi 已提交
4114
    def _sync_with_cpp(self):
4115
        """
4116 4117
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
4118
        """
Q
Qiao Longfei 已提交
4119 4120 4121
        # sync variables from cpp
        for var in self.desc.all_vars():
            if not self.has_var(var.name()):
4122 4123 4124 4125
                is_stop_gradient = False
                if var.has_stop_gradient():
                    is_stop_gradient = var.stop_gradient()
                if var.has_is_parameter() and var.is_parameter():
4126 4127 4128 4129 4130 4131 4132 4133
                    self.create_parameter(
                        name=var.name(),
                        desc=var,
                        type=var.type(),
                        shape=var.shape(),
                        dtype=var.dtype(),
                        stop_gradient=is_stop_gradient,
                    )
4134
                else:
4135 4136 4137 4138 4139 4140
                    self.create_var(
                        name=var.name(),
                        desc=var,
                        type=var.type(),
                        stop_gradient=is_stop_gradient,
                    )
Q
Qiao Longfei 已提交
4141

4142
        # sync variables removed from c++ end
4143
        for var in list(self.vars.keys()):
4144
            if not self.desc.find_var(var.encode()):
4145 4146
                self.vars.pop(var)

Q
Qiao Longfei 已提交
4147
        # sync operators from cpp
4148 4149 4150 4151
        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
4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167
        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
4168 4169 4170 4171 4172

        # 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
4173
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
4174 4175 4176 4177 4178 4179 4180

        # 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)

4181 4182 4183 4184 4185
        # 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(
4186 4187 4188 4189 4190 4191
                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]
                ):
4192 4193 4194 4195 4196
                    del self.ops[ops_in_python_index]
                else:
                    ops_in_cpp_index += 1
                    ops_in_python_index += 1

Q
Qiao Longfei 已提交
4197 4198 4199 4200
        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
4201
    def _copy_param_info_from(self, other):
4202
        """
4203 4204
        Copy the information of parameters from the other block.

4205
        Args:
4206 4207 4208 4209 4210
            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.
4211 4212 4213 4214 4215

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
4216
            raise TypeError(
4217 4218
                "_copy_param_info_from should be invoked with Block"
            )
4219
        for p in other.iter_parameters():
4220 4221 4222
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
4223 4224
                # if the Parameter is pruned, v may be None
                continue
4225
            assert isinstance(v, Variable)
4226
            new_p = None
L
Leo Chen 已提交
4227
            if in_dygraph_mode():
4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239
                new_p = EagerParamBase(
                    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,
                )
4240
            else:
J
Jiabin Yang 已提交
4241
                if _in_legacy_dygraph():
4242 4243 4244 4245 4246 4247 4248 4249 4250 4251 4252 4253
                    new_p = ParamBase(
                        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,
                    )
J
Jiabin Yang 已提交
4254 4255 4256 4257 4258 4259 4260
                else:
                    new_p = Parameter(
                        block=self,
                        shape=v.shape,
                        dtype=v.dtype,
                        type=v.type,
                        lod_level=v.lod_level
4261 4262
                        if v.type == core.VarDesc.VarType.LOD_TENSOR
                        else None,
J
Jiabin Yang 已提交
4263 4264 4265 4266 4267
                        stop_gradient=p.stop_gradient,
                        trainable=p.trainable,
                        optimize_attr=p.optimize_attr,
                        regularizer=p.regularizer,
                        error_clip=p.error_clip,
4268 4269
                        name=v.name,
                    )
4270 4271
            self.vars[new_p.name] = new_p

4272
    def _clone_variable(self, var, force_persistable=True):
4273 4274
        """
        Clone a variable into current block.
4275

4276 4277
        Args:
            var: the variable to be cloned.
4278 4279 4280
            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.
4281 4282

        Returns:
4283
            Variable: the new  variable cloned from 'var' in current block.
4284 4285
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
4286 4287 4288
        ret_var = None
        # make STEP_SCOPES var can be safely cloned.
        if var.type == core.VarDesc.VarType.STEP_SCOPES:
4289 4290 4291
            ret_var = self.create_var(
                name=var.name, persistable=var.persistable, type=var.type
            )
T
tangwei12 已提交
4292
        elif var.type == core.VarDesc.VarType.RAW:
4293 4294 4295
            ret_var = self.create_var(
                name=var.name, persistable=var.persistable, type=var.type
            )
T
typhoonzero 已提交
4296 4297 4298 4299 4300 4301
        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,
4302
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
4303
                is_data=var.is_data,
4304 4305
                need_check_feed=var.desc.need_check_feed(),
            )
T
update  
typhoonzero 已提交
4306 4307 4308 4309 4310 4311 4312
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
4313
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
4314
                is_data=var.is_data,
4315 4316
                need_check_feed=var.desc.need_check_feed(),
            )
T
update  
typhoonzero 已提交
4317
        return ret_var
4318

Y
Yu Yang 已提交
4319

4320 4321 4322 4323
# NOTE(zjl): you should be careful that after you call this method,
# some Python Variable and all Python Operators should not be used
# again. Because all Python Variables and all Python Operators are
# re-constructed inside this method. The underlying VarDesc(OpDesc)
4324
# of some old Python Variables(all old Python Operators) may have
4325
# been destructed.
4326 4327 4328
def _apply_pass(
    main_program, startup_program, pass_name, pass_attrs={}, pass_attr_types={}
):
4329 4330 4331 4332
    assert isinstance(pass_attrs, dict), "pass_attrs must be dict"
    assert isinstance(pass_attr_types, dict), "pass_attr_types must be dict"
    tmp_main_program = core.ProgramDesc(main_program.desc)
    tmp_startup_program = core.ProgramDesc(startup_program.desc)
4333 4334 4335 4336 4337 4338 4339
    attrs = core.apply_pass(
        tmp_main_program,
        tmp_startup_program,
        pass_name,
        pass_attrs,
        pass_attr_types,
    )
4340 4341 4342 4343 4344
    main_program._rebuild_from_desc(tmp_main_program)
    startup_program._rebuild_from_desc(tmp_startup_program)
    return attrs


4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356
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.
        """
4357 4358 4359
        assert isinstance(
            node, core.Node
        ), 'node must be the instance of core.Node.'
4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430 4431 4432 4433 4434 4435 4436 4437 4438 4439 4440
        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()

4441
    def remove_input_by_id(self, node_id):
4442 4443 4444 4445 4446 4447
        """
        Remove a node from inputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
4448
        self.node.remove_input(node_id)
4449

4450
    def remove_input(self, node):
4451 4452 4453 4454
        """
        Remove a node from inputs.

        Args:
4455
            node(IrNode): the node being removed.
4456
        """
4457
        self.node.remove_input(node.node)
4458

4459
    def append_input(self, node):
4460 4461 4462 4463
        """
        Append a node in inputs.

        Args:
4464
            node(IrNode): the node being appended.
4465
        """
4466
        self.node.append_input(node.node)
4467 4468 4469 4470 4471 4472 4473 4474

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

4475
    def remove_output_by_id(self, node_id):
4476 4477 4478 4479 4480 4481
        """
        Remove a node from outputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
4482
        self.node.remove_output(node_id)
4483

4484
    def remove_output(self, node):
4485 4486 4487 4488
        """
        Remove a node from outputs.

        Args:
4489
            node(IrNode): the node being removed.
4490
        """
4491
        self.node.remove_output(node.node)
4492

4493
    def append_output(self, node):
4494 4495 4496 4497
        """
        Append a node in outputs.

        Args:
4498
            node(IrNode): the node being appended.
4499
        """
4500
        self.node.append_output(node.node)
4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512 4513 4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525 4526 4527 4528 4529 4530 4531 4532 4533 4534

    @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.
        """
4535 4536 4537
        assert (
            isinstance(node, core.Node) and node.is_var()
        ), 'node must be the instance of core.Node and it must be a variable node.'
4538 4539 4540 4541 4542 4543 4544 4545 4546 4547
        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.
        """
4548 4549 4550
        assert (
            self.node.var() is not None
        ), "The node variable description can not be None."
4551 4552 4553 4554 4555 4556 4557 4558 4559
        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.
        """
4560 4561 4562
        assert (
            self.node.var() is not None
        ), "The node variable description can not be None."
4563 4564
        return self.node.var().persistable()

4565 4566 4567 4568 4569 4570 4571
    def type(self):
        """
        Return the variable type.

        Returns:
            core.VarDesc.VarType: the variable type.
        """
4572 4573 4574
        assert (
            self.node.var() is not None
        ), "The node variable description can not be None."
4575 4576 4577 4578 4579 4580 4581 4582 4583
        return self.node.var().type()

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

        Returns:
            core.VarDesc.VarType: the variable data type.
        """
4584 4585 4586
        assert (
            self.node.var() is not None
        ), "The node variable description can not be None."
4587 4588 4589 4590 4591 4592 4593 4594 4595
        return self.node.var().dtype()

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

        Returns:
            list: the variable shape.
        """
4596 4597 4598
        assert (
            self.node.var() is not None
        ), "The node variable description can not be None."
4599 4600
        return self.node.var().shape()

4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617 4618 4619 4620 4621 4622 4623 4624 4625 4626 4627 4628 4629 4630 4631 4632 4633
    @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.
        """
4634 4635 4636
        assert (
            isinstance(node, core.Node) and node.is_op()
        ), 'node must be the instance of core.Node and it must be a operator node.'
4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647
        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.
        """
4648 4649 4650
        assert (
            self.node.op() is not None
        ), "The node operator description can not be None."
4651 4652
        self.node.op()._rename_input(old_input_name, new_input_name)

4653 4654 4655 4656 4657 4658 4659 4660
    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.
        """
4661 4662 4663
        assert (
            self.node.op() is not None
        ), "The node operator description can not be None."
4664 4665
        self.node.op()._rename_output(old_output_name, new_output_name)

4666 4667 4668 4669 4670 4671 4672 4673 4674 4675
    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.
        """
4676 4677 4678
        assert (
            self.node.op() is not None
        ), "The node operator description can not be None."
4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690
        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.
        """
4691 4692 4693
        assert (
            self.node.op() is not None
        ), "The node operator description can not be None."
4694 4695 4696 4697 4698 4699 4700 4701 4702
        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.
        """
4703 4704 4705
        assert (
            self.node.op() is not None
        ), "The node operator description can not be None."
4706 4707
        return self.node.op().set_type(new_type)

4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721
    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.
        """
4722 4723 4724
        assert (
            self.node.op() is not None
        ), "The node operator description can not be None."
4725
        desc = self.node.op()
4726 4727 4728 4729 4730
        if isinstance(val, Variable):
            desc.set_var_attr(name, val.desc)
        elif isinstance(val, list) and _all_is_type(val, Variable):
            desc.set_vars_attr(name, [v.desc for v in val])
        elif isinstance(val, Block):
4731
            desc.set_block_attr(name, val.desc)
4732
        elif isinstance(val, list) and val and _all_is_type(val, Block):
4733
            desc.set_blocks_attr(name, [v.desc for v in val])
4734 4735 4736
        elif isinstance(val, core.BlockDesc) or isinstance(
            val, core.ProgramDesc
        ):
4737 4738 4739 4740
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)

4741 4742 4743 4744 4745 4746 4747
    def input_arg_names(self):
        """
        Return input arguments' names of this op node.

        Returns:
            list(str): input arguments' names of this op node.
        """
4748 4749 4750
        assert (
            self.node.op() is not None
        ), "The node operator description can not be None."
4751 4752 4753 4754 4755 4756 4757 4758 4759
        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.
        """
4760 4761 4762
        assert (
            self.node.op() is not None
        ), "The node operator description can not be None."
4763 4764
        return self.node.op().output_arg_names()

4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785
    @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]


4786 4787
class IrGraph(object):
    """
4788
    Python IrGraph. Beneath it is a core.Graph, which is used for
4789
    creating a c++ Ir Pass Graph. An IrGraph is just a graph view of
4790 4791
    a Program. In an IrGraph, both Variables and Operators are graph
    nodes.
4792 4793 4794 4795
    """

    def __init__(self, graph, for_test=False):
        """
4796 4797
        Construct an IrGraph using core.Graph.

4798 4799 4800 4801 4802
        Args:
            graph(core.Graph): C++ Graph.
            for_test(bool): True for the test graph and false for the train graph.
        """
        assert isinstance(
4803 4804
            graph, core.Graph
        ), 'graph must be the instance of core.Graph.'
4805 4806 4807
        self.graph = graph
        self._for_test = for_test

4808 4809 4810 4811
    def clone(self):
        """
        Create a new and duplicated IrGraph.

4812 4813 4814
        Warns:
            The method only clones the graph structure, not its attributes.

4815 4816 4817
        Returns:
            IrGraph: A new and duplicated graph.
        """
4818
        g = self.graph.clone()
4819 4820
        return IrGraph(g, self._for_test)

4821
    def is_test(self):
4822 4823 4824
        """
        If the graph is used for testing, the function returns true. Otherwise, returns false.
        """
4825 4826
        return self._for_test

W
WangZhen 已提交
4827
    def all_nodes(self):
4828 4829 4830
        """
        Return all nodes included in the graph as a set.
        """
4831
        return {IrNode(node) for node in self.graph.nodes()}
4832

4833
    def all_var_nodes(self):
4834 4835 4836
        """
        Return all variable nodes included in the graph as a set.
        """
4837
        return {IrVarNode(node) for node in self.graph.nodes() if node.is_var()}
4838

4839
    def all_persistable_nodes(self):
4840 4841 4842
        """
        Return all persistable variable nodes included in the graph as a set.
        """
W
WangZhen 已提交
4843 4844
        persistable_nodes = set()
        for node in self.graph.nodes():
4845 4846 4847 4848 4849
            if (
                node.is_var()
                and node.var() is not None
                and node.var().persistable()
            ):
W
WangZhen 已提交
4850
                persistable_nodes.add(node)
4851
        return {IrVarNode(p) for p in persistable_nodes}
W
WangZhen 已提交
4852

4853
    def all_op_nodes(self):
4854 4855 4856
        """
        Return all operator nodes included in the graph as a set.
        """
4857
        return {IrOpNode(node) for node in self.graph.nodes() if node.is_op()}
4858

4859 4860 4861 4862 4863 4864
    def all_sub_graphs(self, for_test=False):
        """
        Return all sub_graphs included in the main graph as a set.
        """

        return [
4865
            IrGraph(self.graph.get_sub_graph(i), for_test=for_test)
4866 4867 4868 4869 4870 4871 4872 4873 4874
            for i in range(self.graph.sub_graph_size())
        ]

    def get_sub_graph(self, i, for_test=False):
        """
        Return i-th sub_graph in the main graph.
        """
        return IrGraph(self.graph.get_sub_graph(i), for_test=for_test)

4875
    def create_persistable_node(self, name, var_type, shape, var_dtype):
4876 4877 4878 4879 4880 4881 4882 4883 4884 4885 4886
        """
        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:
4887
            IrVarNode: the created persistable variable node.
4888
        """
4889 4890 4891 4892 4893
        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)
4894
        return IrVarNode(self.graph.create_var_node(var_desc))
4895 4896

    def create_var_node(self, name, var_type, shape, var_dtype):
4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907
        """
        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:
4908
            IrVarNode: the created variable node.
4909 4910
        """

4911 4912 4913 4914
        var_desc = core.VarDesc(name)
        var_desc.set_type(var_type)
        var_desc.set_shape(shape)
        var_desc.set_dtype(var_dtype)
4915
        return IrVarNode(self.graph.create_var_node(var_desc))
4916

4917 4918 4919 4920 4921 4922
    def create_control_dep_var(self):
        """
        create a control var
        """
        return IrVarNode(self.graph.create_control_dep_var())

4923
    def create_var_node_from_desc(self, var_desc):
4924 4925 4926 4927 4928 4929 4930 4931
        """
        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:
4932
            IrVarNode: the created variable node.
4933
        """
4934
        return IrVarNode(self.graph.create_var_node(var_desc))
4935 4936

    def create_op_node(self, op_type, attrs, inputs, outputs):
4937 4938 4939 4940 4941 4942 4943
        """
        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
4944
            outputs(dict): the outputs of the operator node.
4945 4946

        Returns:
4947
            IrOpNode: the created operator node.
4948
        """
4949 4950
        op_desc = core.OpDesc()
        op_desc.set_type(op_type)
4951
        for attr, value in attrs.items():
4952
            self._update_desc_attr(op_desc, attr, value)
4953
        for input_name, var_nodes in inputs.items():
4954 4955
            if not isinstance(var_nodes, list):
                var_nodes = [var_nodes]
4956 4957 4958
            op_desc.set_input(
                input_name, [var_node.name() for var_node in var_nodes]
            )
4959
        for output_name, var_nodes in outputs.items():
4960 4961
            if not isinstance(var_nodes, list):
                var_nodes = [var_nodes]
4962 4963 4964
            op_desc.set_output(
                output_name, [var_node.name() for var_node in var_nodes]
            )
4965
        return IrOpNode(self.graph.create_op_node(op_desc))
4966 4967

    def create_op_node_from_desc(self, op_desc):
4968 4969 4970 4971 4972 4973 4974
        """
        Create a operator node by using an existing OpDesc in the graph.

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

        Returns:
4975
            IrOpNode: the created operator node.
4976
        """
4977
        return IrOpNode(self.graph.create_op_node(op_desc))
4978 4979

    def update_input_link(self, old_input_node, new_input_node, op_node):
4980 4981 4982 4983
        """
        Update the input's link of a operator node.

        Args:
4984 4985 4986
            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.
4987
        """
4988 4989 4990 4991 4992
        assert (
            old_input_node.node in self.graph.nodes()
            and new_input_node.node in self.graph.nodes()
            and op_node.node in self.graph.nodes()
        ), 'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.'
4993 4994 4995 4996
        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)
4997
        op_node.rename_input(old_input_node.name(), new_input_node.name())
4998

4999 5000 5001 5002 5003 5004 5005 5006 5007
    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.
        """
5008 5009 5010 5011 5012
        assert (
            old_output_node.node in self.graph.nodes()
            and new_output_node.node in self.graph.nodes()
            and op_node.node in self.graph.nodes()
        ), 'The three arguments(old_output_node &new_output_node &op_node) must be in the graph nodes.'
5013 5014 5015 5016 5017 5018
        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())

5019
    def link_to(self, node_in, node_out):
5020 5021 5022 5023
        """
        Connect two nodes.

        Args:
5024 5025
            node_in(IrNode): the input node.
            node_out(IrNode): the output node.
5026
        """
5027
        assert node_in.node in self.graph.nodes(), (
5028 5029
            'node_in(%s) must be in the graph nodes.' % node_in.node.name()
        )
5030
        assert node_out.node in self.graph.nodes(), (
5031 5032
            'node_out(%s) must be in the graph nodes.' % node_out.node.name()
        )
5033 5034
        node_in.append_output(node_out)
        node_out.append_input(node_in)
5035 5036

    def safe_remove_nodes(self, remove_nodes):
5037 5038 5039 5040 5041 5042 5043
        """
        Remove nodes safely since links connected to these removed nodes are
        also removed.

        Args:
            remove_nodes(set): the nodes prepared to be removed.
        """
5044
        if not isinstance(remove_nodes, set):
W
WangZhen 已提交
5045 5046 5047 5048
            if isinstance(remove_nodes, Iterable):
                remove_nodes = set(remove_nodes)
            else:
                remove_nodes = {remove_nodes}
5049 5050
        original_nodes = {n.node for n in remove_nodes}
        core.graph_safe_remove_nodes(self.graph, original_nodes)
5051

Z
Zhen Wang 已提交
5052 5053 5054 5055 5056 5057 5058 5059
    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] = [
5060
                            self._find_node_by_name(node.inputs, each_var_name)
Z
Zhen Wang 已提交
5061 5062 5063 5064
                        ]
                for each_var_name in node.op().output_arg_names():
                    if each_var_name not in var_nodes:
                        var_nodes[each_var_name] = [
5065
                            self._find_node_by_name(node.outputs, each_var_name)
Z
Zhen Wang 已提交
5066 5067 5068
                        ]
                    else:
                        var_nodes[each_var_name].append(
5069 5070
                            self._find_node_by_name(node.outputs, each_var_name)
                        )
Z
Zhen Wang 已提交
5071 5072
        self.graph.resolve_hazard(var_nodes)

W
WangZhen 已提交
5073
    def has_circle(self):
5074 5075 5076 5077 5078 5079
        """
        Check if the graph has a circle.

        Returns:
            bool: True if the graph has a circle else False.
        """
W
WangZhen 已提交
5080 5081 5082
        return core.has_circle(self.graph)

    def graph_num(self):
5083 5084 5085 5086 5087 5088
        """
        Count the number of unconnected graphs in this graph.

        Returns:
            int: the number of unconnected graphs.
        """
W
WangZhen 已提交
5089 5090 5091
        return core.graph_num(self.graph)

    def topology_sort(self):
5092 5093 5094
        """
        Perform the topology sort operation on the graph.

T
tianshuo78520a 已提交
5095
        Notes: the `graph` can not contain a circle.
5096 5097

        Returns:
Z
Zhen Wang 已提交
5098
            list(IrNode): nodes in topology order.
5099
        """
5100
        ordered_nodes = core.topology_sort(self.graph)
Z
Zhen Wang 已提交
5101
        return [IrNode(n) for n in ordered_nodes]
W
WangZhen 已提交
5102 5103

    def build_adjacency_list(self):
5104 5105 5106 5107
        """
        Build an adjacency list of operations for the `graph`.

        Returns:
5108
            dict{IrNode: set(IrNode)}: the adjacency list.
5109
        """
5110 5111
        adj_list = core.build_adjacency_list(self.graph)
        wrapped_adj_list = dict()
5112
        for k, v in adj_list.items():
5113 5114
            wrapped_adj_list[IrNode(k)] = {IrNode(n) for n in v}
        return wrapped_adj_list
W
WangZhen 已提交
5115

5116 5117 5118 5119 5120 5121 5122 5123
    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.
5124
            marked_nodes(set(IrNode)): nodes that are needed to be marked.
5125 5126 5127 5128 5129
            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.
        """

5130 5131
        def _convert_to_pdf(dot_file_path):
            pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf'
5132 5133 5134 5135
            exited_code = subprocess.call(
                'dot -Tpdf ' + dot_file_path + ' -o ' + pdf_save_path,
                shell=True,
            )
5136 5137
            if exited_code != 0:
                print('The dot command is needed for creating pdf files.')
5138 5139 5140
                print(
                    'The {} is saved as the dot filetype.'.format(dot_file_path)
                )
5141

5142
        remove_ctr_vars = set()
5143
        if remove_ctr_var:
5144
            for node in self.all_var_nodes():
5145 5146 5147
                if node.is_ctrl_var():
                    remove_ctr_vars.add(node)
            self.safe_remove_nodes(remove_ctr_vars)
5148 5149
        print('Total ops num = {}.'.format(len(self.all_op_nodes())))

5150 5151
        if marked_nodes is not None:
            if not isinstance(marked_nodes, set):
5152 5153 5154 5155 5156 5157
                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}
5158 5159 5160 5161
            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)
5162 5163
        if not os.path.exists(save_path):
            os.makedirs(save_path)
5164 5165 5166 5167 5168 5169 5170
        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):
5171 5172 5173
        """
        Convert the graph into a Program.

Z
Zhen Wang 已提交
5174
        WARN: When the graph includes backward operator nodes, the
5175 5176 5177 5178 5179 5180
        conversion process may be failed. Usually, this function is
        only used to convert a test graph.

        Returns:
            Program: a program converted from the graph.
        """
5181
        convert_pass = core.get_pass('graph_to_program_pass')
5182 5183
        desc = core.ProgramDesc()
        convert_pass.set_not_owned('program', desc)
5184 5185 5186 5187
        convert_pass.apply(self.graph)
        program = Program._construct_from_desc(desc)
        return program

5188 5189 5190 5191 5192 5193 5194 5195
    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
5196
        assert target_node is not None, (
5197 5198
            "Cannot find the target node (%s)in the giving set." % node_name
        )
5199 5200
        return target_node

5201 5202 5203 5204
    def _update_desc_attr(self, desc, name, val):
        """
        Update the value of desc's attribute by attribute's name.
        """
5205 5206 5207 5208 5209
        if isinstance(val, Variable):
            desc.set_var_attr(name, val.desc)
        elif isinstance(val, list) and _all_is_type(val, Variable):
            desc.set_vars_attr(name, [v.desc for v in val])
        elif isinstance(val, Block):
5210
            desc.set_block_attr(name, val.desc)
5211
        elif isinstance(val, list) and val and _all_is_type(val, Block):
5212
            desc.set_blocks_attr(name, [v.desc for v in val])
5213 5214 5215
        elif isinstance(val, core.BlockDesc) or isinstance(
            val, core.ProgramDesc
        ):
5216 5217 5218 5219 5220
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)


Y
Yu Yang 已提交
5221
class Program(object):
D
dzhwinter 已提交
5222
    """
5223
    Create Python Program.  It has at least one :ref:`api_guide_Block_en`, when the
5224
    control flow op like conditional_block, while :ref:`api_paddle_fluid_layers_While` is included,
J
Jiabin Yang 已提交
5225
    it will contain nested block.
5226

J
Jiabin Yang 已提交
5227 5228 5229
    Please reference the
    `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_
    for details.
D
dzhwinter 已提交
5230

J
Jiabin Yang 已提交
5231
    A set of Program usually contains startup program and main program.
J
Jiabin Yang 已提交
5232
    A startup program is set to contain some initial work, eg. initialize the ``Parameter``, and the main
J
Jiabin Yang 已提交
5233 5234 5235 5236 5237 5238 5239
    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
5240
    **Notes**:
5241 5242 5243
        **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
5244 5245

    Returns:
J
Jiabin Yang 已提交
5246
        Program: An empty Program.
D
dzhwinter 已提交
5247 5248

    Examples:
5249 5250
        .. code-block:: python

5251 5252 5253 5254
            import paddle
            import paddle.static as static

            paddle.enable_static()
5255

5256 5257 5258 5259 5260
            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')
5261
                z = static.nn.fc(name="fc", x=x, size=10, activation="relu")
5262 5263 5264

            print("main program is: {}".format(main_program))
            print("start up program is: {}".format(startup_program))
D
dzhwinter 已提交
5265 5266 5267

    """

5268 5269
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
5270 5271
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
5272 5273
        global global_prog_seed
        self._seed = global_prog_seed
Y
yuyang18 已提交
5274
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
5275
        self.__op_role_var = []
T
tangwei12 已提交
5276

5277 5278
        # for distribute training
        # _is_distributed = True if under distributed training
T
tangwei12 已提交
5279
        self._is_distributed = False
5280
        # _is_chief = True if the trainer is the first one, usually No.0
T
tangwei12 已提交
5281
        self._is_chief = False
5282 5283 5284
        # _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
5285
        self._endpoints = []
5286 5287 5288
        # if current role is parameter server, the _ps_endpoint is its "ip:port"
        self._ps_endpoint = None
        # trainers_endpoints, it is used for distribution.
5289
        self._trainers_endpoints = []
5290
        # the distributed lookup table names
T
tangwei12 已提交
5291
        self._distributed_lookup_table = None
5292 5293 5294

        # use Deep gradient comrepssion or not
        self._enable_dgc = False
5295 5296
        self._use_lamb = False

5297 5298 5299
        self._nccl_comm_num = 1
        self._use_hierarchical_allreduce = False
        self._hierarchical_allreduce_inter_nranks = 0
5300

5301 5302 5303
        # 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
5304
        self._program_config = None
5305

H
hutuxian 已提交
5306 5307 5308
        # assigned if this program has been parsed by a pipeline optimizer
        self._pipeline_opt = None

5309 5310 5311
        # assigned if this program has been parsed by a heter pipeline parameter server optimizer
        self._heter_pipeline_opt = None

5312 5313 5314
        # appending gradients times
        self._appending_grad_times = 0

5315 5316
        # identifier for auto checkpoint
        self._auto_checkpoint_name = unique_name.generate(
5317 5318
            "__auto_checkpoint_program__"
        )
5319

5320 5321
        # compiled program, i.e. Graph
        self._graph = None
5322 5323
        # to tag whether is startup_program
        self._is_start_up_program_ = False
5324

5325
    def _find_var_class_kwargs(self, new_desc):
5326 5327 5328 5329 5330 5331 5332 5333
        # NOTE: not all variables support shape/dtype/lod_level methods.
        # For example: RAW, STEP_SCOPES, etc.
        def get_var_desc_attr_or_none(var_desc, attr_name, allowed_types):
            if var_desc.type() in allowed_types:
                return getattr(var_desc, attr_name)()
            else:
                return None

5334 5335 5336 5337
        old_desc = self.desc
        all_new_vars = []
        block_num = new_desc.num_blocks()
        for idx in range(block_num):
5338
            if idx > (len(self.blocks) - 1):
5339
                self._create_block()
5340 5341 5342 5343 5344 5345 5346 5347 5348 5349
            new_block_desc = new_desc.block(idx)
            all_new_vars.append([])
            block_new_vars = all_new_vars[-1]
            for new_var_desc in new_block_desc.all_vars():
                if self.blocks[idx].has_var(new_var_desc.name()):
                    old_var = self.blocks[idx].var(new_var_desc.name())
                else:
                    old_var = None

                kwargs = {
5350 5351 5352 5353 5354 5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382 5383 5384 5385 5386 5387 5388 5389 5390
                    'type': new_var_desc.type(),
                    'name': new_var_desc.name(),
                    'shape': get_var_desc_attr_or_none(
                        new_var_desc,
                        "shape",
                        [
                            core.VarDesc.VarType.LOD_TENSOR,
                            core.VarDesc.VarType.SELECTED_ROWS,
                            core.VarDesc.VarType.LOD_TENSOR_ARRAY,
                        ],
                    ),
                    'dtype': get_var_desc_attr_or_none(
                        new_var_desc,
                        "dtype",
                        [
                            core.VarDesc.VarType.LOD_TENSOR,
                            core.VarDesc.VarType.SELECTED_ROWS,
                            core.VarDesc.VarType.LOD_TENSOR_ARRAY,
                        ],
                    ),
                    'lod_level': get_var_desc_attr_or_none(
                        new_var_desc,
                        "lod_level",
                        [
                            core.VarDesc.VarType.LOD_TENSOR,
                            core.VarDesc.VarType.LOD_TENSOR_ARRAY,
                        ],
                    ),
                    'error_clip': old_var.error_clip
                    if old_var is not None
                    else None,
                    'stop_gradient': old_var.stop_gradient
                    if old_var is not None
                    else False,
                    'is_data': old_var.is_data
                    if old_var is not None
                    else False,
                    'need_check_feed': new_var_desc.need_check_feed(),
                    'belong_to_optimizer': old_var.belong_to_optimizer
                    if old_var is not None
                    else False,
5391 5392 5393
                }

                if isinstance(old_var, Parameter):
5394 5395 5396 5397 5398 5399 5400 5401 5402 5403 5404 5405 5406 5407 5408 5409 5410
                    kwargs.update(
                        {
                            'trainable': old_var.trainable,
                            'optimize_attr': old_var.optimize_attr,
                            'regularizer': old_var.regularizer,
                            'do_model_average': old_var.do_model_average,
                            'need_clip': old_var.need_clip,
                            'is_distributed': old_var.is_distributed,
                            'is_parameter': old_var.is_parameter,
                        }
                    )
                    block_new_vars.append(
                        {
                            'class': Parameter,
                            'kwargs': copy.deepcopy(kwargs),
                        }
                    )
5411 5412
                else:
                    kwargs['persistable'] = new_var_desc.persistable()
5413 5414 5415 5416 5417 5418
                    block_new_vars.append(
                        {
                            'class': Variable,
                            'kwargs': copy.deepcopy(kwargs),
                        }
                    )
5419 5420 5421 5422 5423 5424 5425

        return all_new_vars

    def _rebuild_from_desc(self, desc):
        all_new_vars = self._find_var_class_kwargs(desc)
        block_num = desc.num_blocks()
        assert block_num == len(all_new_vars)
5426
        assert block_num == self.desc.num_blocks()
5427 5428

        # clear old blocks and desc
5429 5430 5431 5432 5433 5434 5435 5436 5437
        for idx in range(block_num):
            block = self.blocks[idx]
            block.vars.clear()
            block.ops.clear()

        for idx in range(block_num):
            block_desc = self.blocks[idx].desc
            new_block_desc = desc.block(idx)
            block_desc._move_from(new_block_desc)
5438

5439
        del desc
5440 5441 5442 5443 5444 5445 5446 5447 5448 5449 5450 5451 5452 5453 5454 5455 5456 5457 5458

        # add new vars first
        for idx in range(block_num):
            block = self.blocks[idx]
            for new_var in all_new_vars[idx]:
                clazz = new_var['class']
                kwargs = new_var['kwargs']
                kwargs['block'] = block
                clazz(**kwargs)

        # then append op
        for idx in range(block_num):
            block = self.blocks[idx]
            block_desc = self.desc.block(idx)
            for op_idx in range(block_desc.op_size()):
                op_desc = block_desc.op(op_idx)
                op = Operator(block=block, desc=op_desc)
                block.ops.append(op)

5459 5460 5461 5462 5463 5464 5465 5466 5467 5468
    def global_seed(self, seed=0):
        """
        Set global seed for Program

        Returns:
            None.

        Examples:
            .. code-block:: python

5469 5470
                import paddle
                import paddle.static as static
5471

5472 5473 5474
                paddle.enable_static()

                prog = static.default_main_program()
5475 5476 5477 5478 5479
                print(prog.random_seed)
                ## 0
                ## the default random seed is 0

                prog.global_seed(102)
5480
                prog1 = static.default_main_program()
5481 5482 5483 5484 5485 5486 5487 5488
                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
5489
    @property
5490
    def _op_role(self):
Y
yuyang18 已提交
5491 5492 5493 5494 5495 5496 5497 5498
        """
        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
5499
        parameter gradient of backward (use :code:`_op_role_var` to get this
Y
yuyang18 已提交
5500 5501 5502 5503
        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
5504 5505
        return self._current_role

5506 5507
    @_op_role.setter
    def _op_role(self, role):
Y
yuyang18 已提交
5508 5509 5510
        self._current_role = role

    @property
5511
    def _op_role_var(self):
Y
yuyang18 已提交
5512
        """
5513
        The auxiliary variables for :code:`_op_role` property.
Y
yuyang18 已提交
5514

5515
        See Also: :code:`Program._op_role`'s documentation for details.
Y
yuyang18 已提交
5516 5517 5518

        Notes: This is a very low-level API. Users should not use it directly.
        """
5519
        return self.__op_role_var
Y
yuyang18 已提交
5520

5521
    @signature_safe_contextmanager
5522 5523 5524 5525 5526
    def _backward_role_guard(self):
        tmp_role = self._current_role

        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Backward
5527 5528 5529 5530
        try:
            yield
        finally:
            self._current_role = tmp_role
5531

S
rename  
sneaxiy 已提交
5532
    @signature_safe_contextmanager
W
Wu Yi 已提交
5533
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
5534 5535 5536 5537 5538 5539 5540
        """
        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:
5541
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
5542 5543 5544

        Examples:

5545
            >>> import paddle.fluid as fluid
Y
yuyang18 已提交
5546
            >>> p, g = backward(...)
W
Wu Yi 已提交
5547
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
5548 5549
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
5550
        tmp_role = self._current_role
5551
        tmp_var = self.__op_role_var
X
Xin Pan 已提交
5552

Y
yuyang18 已提交
5553 5554
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
5555
        self.__op_role_var = [
5556 5557 5558
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
5559 5560 5561 5562 5563
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
Y
Yu Yang 已提交
5564

S
rename  
sneaxiy 已提交
5565
    @signature_safe_contextmanager
X
Xin Pan 已提交
5566
    def _lr_schedule_guard(self, is_with_opt=False):
5567 5568 5569 5570 5571 5572 5573
        """
        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
5574 5575 5576 5577
        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.
5578 5579 5580

        Examples:

5581
            >>> import paddle.fluid as fluid
5582 5583 5584 5585
            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
5586 5587

        tmp_role = self._current_role
5588
        tmp_var = self.__op_role_var
5589

5590 5591
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
5592 5593
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
5594
        # TODO(typhoonzero): how to set target learning rate var
5595
        self.__op_role_var = []
5596 5597 5598 5599 5600
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
5601

5602
    def __str__(self):
Y
yuyang18 已提交
5603 5604 5605 5606 5607 5608 5609 5610 5611
        """
        Get the protobuf debug string of this Program.

        Returns:
            (str): The protobuf debug string.

        Raises:
            ValueError: If any of required fields is not set.
        """
5612 5613 5614 5615 5616 5617 5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631
        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

5632 5633
            import paddle
            import paddle.static as static
5634

5635 5636 5637
            paddle.enable_static()

            cur_program = static.Program()
5638 5639 5640 5641 5642 5643 5644 5645 5646 5647 5648
            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
Z
zhangchunle 已提交
5649
        ), "skip_op_callstack parameter's type is error, expect bool, received {}".format(
5650 5651
            type(skip_op_callstack)
        )
5652 5653 5654
        program_str = ""
        for block in self.blocks:
            program_str += block._to_readable_code(skip_op_callstack)
5655
            program_str += '\n'
5656
        return program_str
Y
Yang Yang(Tony) 已提交
5657

F
fengjiayi 已提交
5658 5659 5660
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
5661

J
Jiabin Yang 已提交
5662 5663 5664
        Args:

            throw_on_error (bool): raise Value error when any of required fields is not set.
F
fengjiayi 已提交
5665

J
Jiabin Yang 已提交
5666
            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
5667

H
haowang101779990 已提交
5668
        Returns:
J
Jiabin Yang 已提交
5669
            str: The debug string describe current Program.
Y
yuyang18 已提交
5670 5671

        Raises:
J
Jiabin Yang 已提交
5672
            ValueError: If any of required fields is not set and throw_on_error is True.
F
fengjiayi 已提交
5673

5674 5675 5676
        Examples:
            .. code-block:: python

5677 5678 5679 5680
                import paddle
                import paddle.static as static

                paddle.enable_static()
5681

5682 5683 5684
                prog = static.default_main_program()
                x = static.data(name="X", shape=[2,3], dtype="float32")
                pred = static.nn.fc(x, size=3)
5685
                prog_string = prog.to_string(throw_on_error=True, with_details=False)
5686
                prog_string_with_details = prog.to_string(throw_on_error=False, with_details=True)
T
tianshuo78520a 已提交
5687
                print("program string without detail: {}".format(prog_string))
5688
                print("program string with detail: {}".format(prog_string_with_details))
F
fengjiayi 已提交
5689
        """
5690 5691 5692
        assert isinstance(
            throw_on_error, bool
        ), "The type of throw_on_error parameter is wrong, expected bool, but received {}.".format(
5693 5694
            type(throw_on_error)
        )
5695 5696 5697
        assert isinstance(
            with_details, bool
        ), "The type of with_details parameter is wrong, expected bool, but received {}.".format(
5698 5699
            type(with_details)
        )
5700

F
fengjiayi 已提交
5701 5702 5703 5704 5705 5706
        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()
5707
            proto = framework_pb2.ProgramDesc.FromString(bytes(protostr))
F
fengjiayi 已提交
5708 5709
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
5710

W
Wu Yi 已提交
5711
    def _get_desc(self):
Y
yuyang18 已提交
5712 5713 5714 5715 5716 5717 5718
        """
        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.
        """
5719 5720
        return self.desc

X
version  
Xin Pan 已提交
5721 5722 5723
    def _version(self):
        return self.desc._version()

5724
    def clone(self, for_test=False):
Y
yuyang18 已提交
5725
        """
5726
        .. note:::
5727 5728
            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` .
5729
            3. This API has no effect in Dygraph Mode.
Y
yuyang18 已提交
5730

5731
        Create a new Program with forward content of original one when ``for_test=True``.
5732
        Create a new Program as same as the original one when ``for_test=False``.
5733

5734
        Some operators, e.g., :ref:`api_paddle_fluid_layers_batch_norm` , behave differently between
Y
yuyang18 已提交
5735 5736 5737
        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`.
5738

5739 5740
        * 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.
5741 5742
          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
5743
          recommend you to use :code:`clone` before using :code:`Opimizer.minimize`.
Y
yuyang18 已提交
5744

J
Jiabin Yang 已提交
5745
        For Example:
5746
          ::
L
Luo Tao 已提交
5747

5748 5749 5750 5751 5752 5753
            import paddle
            import paddle.static as static

            paddle.enable_static()

            img = static.data(name='image', shape=[None, 784])
5754
            pred = static.nn.fc(x=img, size=10, actvation='relu')
5755
            loss = paddle.mean(pred)
5756
            # Here we use clone before Momentum
5757 5758
            test_program = static.default_main_program().clone(for_test=True)
            optimizer = paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
5759
            optimizer.minimize(loss)
5760

J
Jiabin Yang 已提交
5761
        Args:
5762

5763 5764
            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` .
5765

J
Jiabin Yang 已提交
5766
        Returns:
5767
            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``
5768

Y
yuyang18 已提交
5769 5770 5771

        Examples:

5772 5773 5774 5775 5776 5777 5778
            .. 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`:

5779 5780 5781 5782 5783 5784 5785 5786 5787 5788 5789
            .. 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))
5790
                        for key, value in sorted(op.all_attrs().items()):
5791 5792 5793 5794
                            if key not in ['op_callstack', 'op_role_var']:
                                print(" [ attrs: {}:   {} ]".format(key, value))


5795
            1. To clone a test program, the sample code is:
5796 5797
                .. code-block:: python

5798 5799 5800 5801 5802 5803
                    import paddle
                    import paddle.static as static
                    import paddle.utils as utils
                    import paddle.nn.functional as F

                    paddle.enable_static()
5804 5805

                    def print_prog(prog):
5806
                        for name, value in sorted(prog.block(0).vars.items()):
5807 5808 5809 5810 5811
                            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))
5812
                            for key, value in sorted(op.all_attrs().items()):
5813 5814 5815
                                if key not in ['op_callstack', 'op_role_var']:
                                    print(" [ attrs: {}:   {} ]".format(key, value))

5816 5817
                    train_program = static.Program()
                    startup_program = static.Program()
J
Jiabin Yang 已提交
5818 5819 5820

                    # startup_program is used to do some parameter init work,
                    # and main program is used to hold the network
5821 5822 5823
                    with static.program_guard(train_program, startup_program):
                        with utils.unique_name.guard():
                            img = static.data(name='image', shape=[None, 784])
5824
                            hidden = static.nn.fc(x=img, size=200, activation='relu')
5825 5826
                            hidden = F.dropout(hidden, p=0.5)
                            loss = F.cross_entropy(
5827
                                input=static.nn.fc(x=hidden, size=10, activation='softmax'),
5828 5829
                                label=static.data(name='label', shape=[1], dtype='int64'))
                            avg_loss = paddle.mean(loss)
5830
                            test_program = train_program.clone(for_test=True)
5831
                    print_prog(test_program)
J
Jiabin Yang 已提交
5832 5833 5834 5835

                    # 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

5836
                    # In Paddle we will share weights by using the same Tensor name. In train and test program
J
Jiabin Yang 已提交
5837 5838 5839 5840
                    # 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.

5841 5842 5843
                    with static.program_guard(train_program, startup_program):
                        with utils.unique_name.guard():
                            sgd = paddle.optimizer.SGD(learning_rate=1e-3)
5844 5845 5846
                            sgd.minimize(avg_loss)


5847
            2. The clone method can be avoid if you create program for training and program for testing individually.
5848 5849
                .. code-block:: python

5850 5851 5852 5853 5854 5855
                    import paddle
                    import paddle.static as static
                    import paddle.utils as utils
                    import paddle.nn.functional as F

                    paddle.enable_static()
5856 5857

                    def print_prog(prog):
5858
                        for name, value in sorted(prog.block(0).vars.items()):
5859 5860 5861 5862 5863
                            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))
5864
                            for key, value in sorted(op.all_attrs().items()):
5865 5866
                                if key not in ['op_callstack', 'op_role_var']:
                                    print(" [ attrs: {}:   {} ]".format(key, value))
5867

5868
                    def network():
5869
                        img = static.data(name='image', shape=[None, 784])
5870
                        hidden = static.nn.fc(x=img, size=200, activation='relu')
5871 5872
                        hidden = F.dropout(hidden, p=0.5)
                        loss = F.cross_entropy(
5873
                            input=static.nn.fc(x=hidden, size=10, activation='softmax'),
5874 5875
                            label=static.data(name='label', shape=[1], dtype='int64'))
                        avg_loss = paddle.mean(loss)
5876 5877
                        return avg_loss

5878 5879 5880 5881 5882
                    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():
5883
                            avg_loss = network()
5884
                            sgd = paddle.optimizer.SGD(learning_rate=1e-3)
5885
                            sgd.minimize(avg_loss)
5886
                    # the test startup program is not used.
5887 5888
                    with static.program_guard(test_program_2, startup_program_2):
                        with utils.unique_name.guard():
5889 5890
                            avg_loss = network()
                    print_prog(test_program_2)
5891

5892
            The two code snippets above will generate and print same programs.
5893
        """
5894

T
tangwei12 已提交
5895
        # NOTE(zhiqiu): we sync the original program first, since its program may diff with
5896 5897 5898
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

5899
        pruned_origin_block_id_map = None
5900
        if for_test:
5901 5902
            forward_prog = Program()
            forward_prog.desc, pruned_origin_block_id_map = core.prune_backward(
5903 5904
                self.desc
            )
5905 5906
            forward_prog.blocks = [
                Block(forward_prog, i)
5907
                for i in range(forward_prog.desc.num_blocks())
5908 5909 5910
            ]
            forward_prog._sync_with_cpp()
            p = forward_prog._inference_optimize(prune_read_op=False)
5911
        else:
5912
            p = Program()
G
gongweibao 已提交
5913 5914
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
5915
            p.desc = core.ProgramDesc(self.desc)
5916
            p.blocks = [Block(p, i) for i in range(self.desc.num_blocks())]
G
gongweibao 已提交
5917 5918

            p._current_role = self._current_role
5919
            p.__op_role_var = self.__op_role_var
5920
            p._appending_grad_times = self._appending_grad_times
5921 5922
            if hasattr(self, 'lr_sheduler'):
                p.lr_sheduler = self.lr_sheduler
G
gongweibao 已提交
5923

T
tangwei12 已提交
5924
            # NOTE(zhiqiu): we sync the cloned program, to update its program by
5925
            # its desc.
W
Wu Yi 已提交
5926
            p._sync_with_cpp()
5927

W
Wu Yi 已提交
5928
        p._copy_param_info_from(self)
5929
        p._copy_data_info_from(self, pruned_origin_block_id_map)
5930
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
5931
        return p
5932

5933
    def _prune(self, targets):
Y
yuyang18 已提交
5934 5935 5936 5937 5938 5939 5940 5941
        """
        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:
5942
            targets(list|Variable|Operator): A list of variables, operators, or variable names
Y
yuyang18 已提交
5943 5944 5945 5946
                need to be pruned

        Returns:
            Program:  A new, pruned program.
5947
        """
5948
        return self._prune_with_input([], targets)
5949 5950

    def _prune_with_input(self, feeded_var_names, targets):
Y
yuyang18 已提交
5951
        """
5952
        Prune operators and variables which are not needed to generate
5953 5954
        :code:`targets`. Prune operators and variables which are needed
        to generate feeded_var
5955 5956 5957 5958 5959 5960 5961

        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()
5962
            targets(list|Variable|Operator): A list of variables, operators, or variable names
5963 5964 5965 5966 5967 5968
                need to be pruned

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

T
tangwei12 已提交
5969
        # NOTE(zhiqiu): we sync the original program first, since its program may diff with
5970 5971 5972
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

5973 5974
        if not isinstance(feeded_var_names, list):
            feeded_var_names = [feeded_var_names]
5975 5976
        if not isinstance(targets, list):
            targets = [targets]
5977 5978

        for var in feeded_var_names:
5979
            if not isinstance(var, str):
5980 5981
                raise ValueError(
                    "All feeded_var_names of Program._prune_with_input() can only be "
5982 5983
                    "str, but received %s." % type(var)
                )
5984

5985 5986 5987 5988 5989 5990 5991 5992 5993 5994 5995 5996 5997 5998 5999 6000
        # find out all variables that can be generated or updated with given feed
        generatable_vars = set()

        for idx, op in enumerate(self.global_block().ops):
            runnable_op = True
            for name in op.input_arg_names:
                if not self.global_block().has_var(name):
                    continue
                if self.global_block().var(name).persistable:
                    continue
                if name not in generatable_vars.union(feeded_var_names):
                    runnable_op = False
                    break
            if runnable_op:
                generatable_vars = generatable_vars.union(op.output_arg_names)

6001 6002 6003 6004
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
6005
                    name = t.name
6006
                elif isinstance(t, str):
6007
                    name = str(t)
6008
                else:
6009 6010
                    raise ValueError(
                        "All targets of Program._prune_with_input() can only be "
6011 6012
                        "Variable or Operator, but received %s." % type(t)
                    )
6013 6014 6015 6016 6017 6018

                # 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:
6019 6020 6021
                    # however if the var is also updated by a runnable op, will shall keep it
                    if name not in generatable_vars:
                        continue
6022

6023 6024 6025 6026 6027 6028 6029 6030 6031
                # 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
6032
                        # Skip optimize op except for optimize op in targets,
6033 6034 6035 6036 6037
                        # since optimize op generates parameters.
                        if op._is_optimize_op() and op not in targets:
                            continue
                        else:
                            target_op = op
6038

6039
                if target_op is not None:
6040 6041 6042
                    targets_idx.append([target_op.block.idx, target_op.idx])
            else:
                targets_idx.append([t.block.idx, t.idx])
6043

6044
        res = Program()
6045
        res.desc, pruned_origin_block_id_map = core.prune(
6046 6047
            self.desc, set(feeded_var_names), targets_idx
        )
6048
        res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())]
W
Wu Yi 已提交
6049
        res._sync_with_cpp()
6050 6051 6052 6053 6054

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

6055 6056
        return res

X
Xin Pan 已提交
6057
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
6058
        """
F
fengjiayi 已提交
6059 6060 6061 6062 6063
        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.

6064
        3. change the :code:`is_test`
Y
yuyang18 已提交
6065 6066 6067
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

6068
        Args:
X
Xin Pan 已提交
6069 6070
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
6071

Y
yuyang18 已提交
6072 6073 6074 6075 6076 6077
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
6078
        res = Program()
6079
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
6080 6081 6082 6083

        # 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
6084
        if prune_read_op:
6085
            while True:
6086 6087 6088 6089
                if (
                    read_op_idx >= root_block.op_size()
                    or root_block.op(read_op_idx).type() == 'read'
                ):
6090 6091 6092 6093 6094 6095
                    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:
6096
                    root_block._remove_var(var.name().encode())
F
fengjiayi 已提交
6097 6098

        # change all `is_test` attributes to True
6099
        for i in range(res.desc.num_blocks()):
6100
            block = res.desc.block(i)
6101
            for j in range(block.op_size()):
6102 6103
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
6104
                    op._set_attr('is_test', True)
6105 6106 6107
                if op.type() == "batch_norm":
                    # Remove the output ReserveSpace of batch_norm if exists.
                    op.remove_output("ReserveSpace")
6108
        res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())]
W
Wu Yi 已提交
6109
        res._sync_with_cpp()
6110 6111
        return res

6112
    def _remove_training_info(self, clip_extra=True):
6113 6114 6115 6116 6117 6118 6119 6120 6121 6122 6123 6124 6125 6126
        """
        This method will create a new program and do following adjustments on it:
        1. Remove all variable's `is_parameter` attribute if exist.

        2. Remove all variable's `stop_gradient` attribute if exist.

        Notes: This API is a very low level API.

        Returns:
            Program: The new program.
        """
        res = Program()
        res.desc = core.ProgramDesc(self.desc)

6127
        res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())]
6128 6129
        res._sync_with_cpp()

6130 6131
        # Note: The op_role and op_role_var cann't be deleted currently,
        # and we will try to remove them in the future.
6132
        common_clipped_attrs_list = ['op_callstack', 'with_quant_attr']
6133

6134
        for i in range(res.desc.num_blocks()):
6135 6136 6137 6138
            block = res.desc.block(i)
            for var in block.all_vars():
                var.clear_is_parameter()
                var.clear_stop_gradient()
6139 6140
            if not clip_extra:
                continue
6141 6142 6143 6144
            for op_idx in range(0, block.op_size()):
                op = block.op(op_idx)
                if op.type() not in OpProtoHolder.instance().op_proto_map:
                    continue
6145 6146 6147

                extra_attrs_map = core.get_op_extra_attrs(op.type())

6148 6149 6150 6151 6152 6153 6154 6155 6156 6157 6158 6159 6160
                proto = OpProtoHolder.instance().get_op_proto(op.type())
                remove_input_list = []
                for name in op.input_names():
                    find = False
                    for input_proto in proto.inputs:
                        if input_proto.name != name:
                            continue
                        if input_proto.extra:
                            remove_input_list.append(name)
                        find = True
                        break
                    if not find:
                        remove_input_list.append(name)
6161 6162 6163
                # The extra input of op will be removed in the future
                # for name in remove_input_list:
                #     op.remove_input(name)
6164 6165 6166 6167 6168 6169 6170 6171 6172 6173 6174 6175 6176

                remove_output_list = []
                for name in op.output_names():
                    find = False
                    for output_proto in proto.outputs:
                        if output_proto.name != name:
                            continue
                        if output_proto.extra:
                            remove_output_list.append(name)
                        find = True
                        break
                    if not find:
                        remove_output_list.append(name)
6177 6178 6179
                # The extra output of op will be removed in the future
                # for name in remove_output_list:
                #     op.remove_output(name)
6180

6181 6182 6183 6184 6185 6186 6187
                op_quant_name = (
                    core.op_proto_and_checker_maker.kOpWithQuantAttrName()
                )
                quant = (
                    bool(op.attr(op_quant_name))
                    if op_quant_name in op.attr_names()
                    else False
6188 6189
                )
                quant_attrs = [
6190 6191 6192 6193 6194 6195 6196
                    op_quant_name,
                    "quantization_type",
                    "skip_quant",
                    "activation_bits",
                    "bit_length",
                    "quantize_weight_bits",
                    "weight_quant_scale",
6197
                ]
6198 6199
                for extra_attr_name in extra_attrs_map.keys():
                    op.remove_attr(extra_attr_name)
6200
                remove_attr_list = []
6201 6202 6203 6204 6205 6206
                for name in op.attr_names():
                    if quant:
                        if name in quant_attrs:
                            continue
                        if name.endswith("_threshold"):
                            continue
6207
                    if len(extra_attrs_map) > 0:
6208
                        if name in common_clipped_attrs_list:
6209
                            op.remove_attr(name)
6210
                        continue
6211 6212 6213 6214 6215 6216 6217 6218 6219 6220
                    find = False
                    for attr_proto in proto.attrs:
                        if attr_proto.name != name:
                            continue
                        find = True
                        break
                    if not find:
                        remove_attr_list.append(name)
                for name in remove_attr_list:
                    op.remove_attr(name)
6221 6222
        return res

6223 6224
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
6225
        """
6226
        .. note::
6227
            1. All information about parameters will be lost after serialization;
6228
            2. This API has no effect in Dygraph mode.
Y
yuyang18 已提交
6229

6230 6231
        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
6232

J
Jiabin Yang 已提交
6233
        Args:
Y
yuyang18 已提交
6234

J
Jiabin Yang 已提交
6235
            binary_str_type (str): the binary prootbuf string.
6236

J
Jiabin Yang 已提交
6237 6238
        Returns:
            Program: A deserialized Program.
6239 6240 6241 6242

        Examples:
            .. code-block:: python

6243 6244 6245 6246
                import paddle
                import paddle.static as static

                paddle.enable_static()
6247

6248 6249 6250 6251
                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')
6252

6253
                    y = static.data(name='Y', shape=[784, 100], dtype='float32')
6254

6255
                    z = paddle.matmul(x=x, y=y)
6256

6257 6258
                    binary_str = static.default_main_program().desc.serialize_to_string()
                    prog_restored = static.default_main_program().parse_from_string(binary_str)
6259

6260
                    print(static.default_main_program())
6261
                    print(prog_restored)
Y
yuyang18 已提交
6262
        """
6263 6264
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
6265
        p.blocks = [Block(p, i) for i in range(p.desc.num_blocks())]
W
Wu Yi 已提交
6266
        p._sync_with_cpp()
6267
        return p
Y
Yu Yang 已提交
6268

6269
    @staticmethod
6270
    def _construct_from_desc(desc):
6271 6272 6273 6274 6275 6276 6277 6278 6279 6280 6281
        """
        Construct a program from program desc.

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

        Returns:
            Program: A program.
        """
        p = Program()
        p.desc = desc
6282
        p.blocks = [Block(p, i) for i in range(p.desc.num_blocks())]
6283 6284 6285
        p._sync_with_cpp()
        return p

D
dzhwinter 已提交
6286 6287
    @property
    def random_seed(self):
Y
yuyang18 已提交
6288
        """
J
Jiabin Yang 已提交
6289
        The default random seed for random operators in Program. ``0`` means get
Y
yuyang18 已提交
6290 6291
        the random seed from random device.

6292
        .. note::
6293
            It must be set before the operators have been added.
J
Jiabin Yang 已提交
6294 6295 6296

        Returns:
            int64: Random seed in current Program
6297

6298 6299 6300 6301

        Examples:
            .. code-block:: python

6302 6303 6304
                import paddle
                import paddle.static as static
                import paddle.nn.functional as F
6305

6306 6307 6308
                paddle.enable_static()

                prog = static.default_main_program()
6309
                random_seed = prog.random_seed
6310
                x_var = static.data(name="X", shape=[3,3], dtype="float32")
6311 6312 6313
                print(random_seed)
                ## 0
                ## the default random seed is 0
6314

6315
                # Here we need to set random seed before we use paddle.nn.functional.dropout
6316
                prog.random_seed = 1
6317
                z_var = F.dropout(x_var, 0.7)
6318

6319
                print(prog.random_seed)
6320 6321
                ## 1
                ## the random seed is change to 1
Y
yuyang18 已提交
6322
        """
D
dzhwinter 已提交
6323 6324
        return self._seed

Q
qiaolongfei 已提交
6325 6326
    @property
    def num_blocks(self):
Y
yuyang18 已提交
6327
        """
6328 6329
        The number of :ref:`api_guide_Block_en`  in this Program.

6330
        .. note::
6331
            This API has no effect in Dygraph mode.
J
Jiabin Yang 已提交
6332 6333 6334

        Returns:
            int(Platform-dependent size): num of :ref:`api_guide_Block_en`  in current Program
6335

6336 6337 6338 6339

        Examples:
            .. code-block:: python

6340 6341 6342 6343
                import paddle
                import paddle.static as static

                paddle.enable_static()
6344

6345
                prog = static.default_main_program()
6346 6347
                num_blocks = prog.num_blocks
                print(num_blocks)
6348

6349 6350
                # print result:
                # 1
Y
yuyang18 已提交
6351
        """
Q
qiaolongfei 已提交
6352 6353
        return self.desc.num_blocks()

D
dzhwinter 已提交
6354 6355 6356
    @random_seed.setter
    def random_seed(self, seed):
        if not isinstance(seed, int):
6357 6358
            raise ValueError(
                "Program.random_seed's input seed must be an integer, but received %s."
6359 6360
                % type(seed)
            )
D
dzhwinter 已提交
6361 6362
        self._seed = seed

Y
Yu Yang 已提交
6363
    def __repr__(self):
6364
        return self.__str__()
6365

Y
Yu Yang 已提交
6366
    def global_block(self):
Y
yuyang18 已提交
6367
        """
6368 6369
        .. note::
            This API has no effect in Dygraph mode.
6370 6371 6372

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

J
Jiabin Yang 已提交
6373 6374
        Returns:
            :ref:`api_guide_Block_en`: The first :ref:`api_guide_Block_en`  of this Program.
6375

6376 6377 6378 6379

        Examples:
            .. code-block:: python

6380 6381 6382 6383
                import paddle
                import paddle.static as static

                paddle.enable_static()
6384

6385
                prog = static.default_main_program()
6386 6387
                gb_block = prog.global_block()
                print(gb_block)
6388

Y
yuyang18 已提交
6389
        """
Y
Yu Yang 已提交
6390 6391
        return self.blocks[0]

Q
Qiao Longfei 已提交
6392
    def block(self, index):
Y
yuyang18 已提交
6393
        """
6394 6395
        .. note::
            This API has no effect in Dygraph mode.
Y
yuyang18 已提交
6396

6397 6398
        Get the :code:`index`  :ref:`api_guide_Block_en`  of this Program

J
Jiabin Yang 已提交
6399 6400
        Args:
            index (int) - The index of  :ref:`api_guide_Block_en`  to get
6401

J
Jiabin Yang 已提交
6402 6403
        Returns:
            :ref:`api_guide_Block_en`: The :code:`index` block
6404 6405 6406 6407

        Examples:
            .. code-block:: python

6408 6409 6410 6411
                import paddle
                import paddle.static as static

                paddle.enable_static()
6412

6413
                prog = static.default_main_program()
6414 6415
                block_0 = prog.block(0)
                print(block_0)
Y
yuyang18 已提交
6416
        """
Q
Qiao Longfei 已提交
6417 6418
        return self.blocks[index]

Y
Yu Yang 已提交
6419
    def current_block(self):
Y
yuyang18 已提交
6420
        """
6421 6422
        .. note::
            This API has no effect in Dygraph mode.
6423

J
Jiabin Yang 已提交
6424 6425
        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.
6426

J
Jiabin Yang 已提交
6427 6428
        Returns:
             :ref:`api_guide_Block_en`: The :code:`index`  :ref:`api_guide_Block_en`
6429

6430 6431 6432
        Examples:
            .. code-block:: python

6433 6434 6435 6436
                import paddle
                import paddle.static as static

                paddle.enable_static()
6437

6438
                prog = static.default_main_program()
6439 6440
                current_blk = prog.current_block()
                print(current_blk)
Y
yuyang18 已提交
6441
        """
Y
Yu Yang 已提交
6442 6443
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
6444
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
6445 6446 6447 6448 6449
        """
        Create a new block with the :code:`parent_idx` and change the current block
        to new block.

        Args:
J
Jiabin Yang 已提交
6450

Y
yuyang18 已提交
6451 6452 6453 6454 6455
            parent_idx(int): The parent block index.

        Returns:
            Block: The new block.
        """
Y
Yu Yang 已提交
6456
        new_block_idx = len(self.blocks)
6457 6458 6459 6460 6461
        parent = (
            self.current_block()
            if parent_idx is None
            else self.block(parent_idx)
        )
F
update  
fengjiayi 已提交
6462
        self.desc.append_block(parent.desc)
Y
Yu Yang 已提交
6463 6464 6465 6466
        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
6467
    def _rollback(self):
Y
yuyang18 已提交
6468 6469 6470 6471 6472
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
6473 6474
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
6475
    def _sync_with_cpp(self):
Y
yuyang18 已提交
6476 6477 6478 6479 6480 6481 6482 6483 6484 6485
        """
        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
6486 6487 6488
        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
6489
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
6490

W
Wu Yi 已提交
6491
    def _copy_param_info_from(self, other):
6492
        """
6493
        Copy the information of parameters from other program.
D
dzhwinter 已提交
6494

Y
yuyang18 已提交
6495 6496 6497
        Notes: This is a very low level API. Users should not invoke it
        directly.

6498 6499 6500 6501 6502 6503 6504
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
6505 6506
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
6507 6508
                % type(other)
            )
6509

W
Wu Yi 已提交
6510
        self.global_block()._copy_param_info_from(other.global_block())
6511

6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522
    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):
6523 6524
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
6525 6526
                % type(other)
            )
6527 6528
        self._is_distributed = other._is_distributed
        self._is_chief = other._is_chief
6529
        self._parameters_on_pservers = other._parameters_on_pservers
6530
        self._endpoints = other._endpoints
6531
        self._ps_endpoint = other._ps_endpoint
6532 6533
        self._distributed_lookup_table = other._distributed_lookup_table

6534
    def _copy_data_info_from(self, other, pruned_origin_block_id_map=None):
F
fengjiayi 已提交
6535 6536
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
6537

Y
yuyang18 已提交
6538 6539 6540
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
6541 6542
        Args:
            other(Program): Other program
6543
            pruned_origin_block_id_map(dict{int:int}): A dict which maps the block id in program
6544 6545
            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,
6546
            {0:0, 1:1,..., n:n}.
F
fengjiayi 已提交
6547 6548 6549 6550 6551

        Returns:
            None
        """
        if not isinstance(other, Program):
6552 6553
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
6554 6555
                % type(other)
            )
F
fengjiayi 已提交
6556

6557 6558
        if not pruned_origin_block_id_map:
            pruned_origin_block_id_map = {
6559
                i: i for i in range(self.desc.num_blocks())
6560
            }
6561 6562 6563

        # NOTE(zhiqiu): All vars in cloned program exist in original program.
        # The reverse is not true, due to backward pruning.
6564 6565
        for i, block in enumerate(self.blocks):
            other_block = other.blocks[pruned_origin_block_id_map[i]]
6566
            for var in list(block.vars.values()):
6567 6568 6569 6570 6571 6572 6573
                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
6574

6575
    def list_vars(self):
Y
yuyang18 已提交
6576
        """
6577
        Get all Tensors from this Program. A iterable object is returned.
Y
yuyang18 已提交
6578

J
Jiabin Yang 已提交
6579
        Returns:
6580
            iterable Tensors: The Generator will yield every Tensor in this program.
6581 6582 6583 6584

        Examples:
            .. code-block:: python

6585 6586
                import paddle
                import paddle.static as static
6587

6588 6589 6590 6591 6592
                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')
6593 6594
                for var in prog.list_vars():
                    print(var)
T
tangwei12 已提交
6595

6596 6597
                # var img : LOD_TENSOR.shape(-1, 1, 28, 28).dtype(float32).stop_gradient(True)
                # var label : LOD_TENSOR.shape(-1, 1).dtype(int64).stop_gradient(True)
Y
yuyang18 已提交
6598
        """
6599
        for each_block in self.blocks:
6600
            for each_var in list(each_block.vars.values()):
6601 6602
                yield each_var

6603 6604 6605 6606 6607 6608 6609 6610 6611 6612
    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

6613 6614 6615 6616
                import paddle
                import paddle.static as static

                paddle.enable_static()
6617

6618 6619
                program = static.default_main_program()
                data = static.data(name='x', shape=[None, 13], dtype='float32')
6620
                hidden = static.nn.fc(x=data, size=10)
6621 6622
                loss = paddle.mean(hidden)
                paddle.optimizer.SGD(learning_rate=0.01).minimize(loss)
6623 6624 6625 6626 6627 6628 6629

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

                # Here will print all parameters in current program, in this example,
                # the result is like:
                #
6630 6631
                # persist trainable param fc_0.w_0 : LOD_TENSOR.shape(13, 10).dtype(float32).stop_gradient(False)
                # persist trainable param fc_0.b_0 : LOD_TENSOR.shape(10,).dtype(float32).stop_gradient(False)
6632 6633 6634 6635 6636 6637 6638 6639 6640 6641
                #
                # 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

6642 6643 6644 6645 6646 6647 6648 6649 6650
    def state_dict(self, mode='all', scope=None):
        """
        Get parameters and persistable buffers of program as a dict. The key is the name of the parameter or the name of the buffer.
        The value is the tensor of this variable in the given scope.

        .. note::
            This function MUST called after run start_up_program

        Args:
6651 6652 6653
            mode(str, optional): Source of the obtained parameters and buffers.
                    'opt' :  The return value only contains the variable in the optimizer.
                    'param' : The return value only contains the variable in the network, not the variable in the optimizer.
6654 6655
                    'all' : The return value contains the variable in the network and optimizer.
                    Default: 'all'
6656
            scope(Scope, optional) : If scope is None, state_dict will be set to global scope
6657 6658 6659 6660 6661 6662 6663 6664 6665 6666 6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678 6679 6680 6681 6682 6683
                obtained through 'paddle.static.global_scope()'. Otherwise, value will be set to scope.
                Default: None

        Retruns:
            dict: a dict contains the parameters and persistable buffers.

        Examples:
            .. code-block:: python

                import paddle
                import paddle.static as static

                paddle.enable_static()

                x = static.data(name="x", shape=[10, 10], dtype='float32')
                y = static.nn.fc(x, 10)
                z = static.nn.fc(y, 10)

                place = paddle.CPUPlace()
                exe = static.Executor(place)
                exe.run(static.default_startup_program())
                prog = static.default_main_program()

                path = "./temp/model.pdparams"
                paddle.save(prog.state_dict(), path)
        """
        # The 'framework' is a low-level module, and 'executor'
6684
        # can not be imported at the begainning of this file.
6685 6686
        # Therefore, the above two modules are dynamically imported.
        from .executor import global_scope
6687

6688 6689
        if scope is not None and not isinstance(scope, core._Scope):
            raise TypeError(
6690 6691 6692 6693
                "`scope` should be None or `paddle.static.Scope'` type, but received {}.".format(
                    type(scope)
                )
            )
6694 6695 6696 6697 6698

        if scope is None:
            scope = global_scope()

        if not isinstance(mode, str):
6699 6700
            raise TypeError(
                "Type of `mode` should be string, but received {}.".format(
6701 6702 6703
                    type(mode)
                )
            )
6704 6705 6706 6707 6708

        def is_parameter(var):
            return isinstance(var, Parameter)

        def is_persistable(var):
6709 6710 6711 6712 6713
            if (
                var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH
                or var.desc.type() == core.VarDesc.VarType.FETCH_LIST
                or var.desc.type() == core.VarDesc.VarType.READER
            ):
6714 6715 6716 6717 6718 6719 6720 6721 6722 6723 6724 6725 6726 6727 6728 6729 6730 6731
                return False
            return var.persistable

        def is_belong_to_optimizer(var):
            if not (isinstance(var, Parameter) or var.desc.need_check_feed()):
                return is_persistable(var)
            return False

        def condition(var):

            if mode == 'param':
                return is_parameter(var)
            elif mode == 'opt':
                return is_belong_to_optimizer(var)
            elif mode == 'all':
                return is_parameter(var) or is_belong_to_optimizer(var)
            else:
                raise ValueError(
6732 6733 6734 6735
                    "`mode` string should be 'param', 'opt' or 'all', but received {}.".format(
                        mode
                    )
                )
6736 6737 6738 6739 6740 6741 6742 6743

        var_list = filter(condition, self.list_vars())

        state_dict = dict()
        for var in var_list:
            var_temp = scope.find_var(var.name)
            if var_temp is None:
                raise ValueError(
6744 6745 6746 6747
                    "Can not find Variable '{}' in the scope. Make sure it is initialized".format(
                        var.name
                    )
                )
6748 6749 6750 6751 6752 6753
            state_dict[var.name] = var_temp.get_tensor()

        return state_dict

    def set_state_dict(self, state_dict, scope=None):
        """
6754
        Set parameters and persistable buffers in state_dict to program.
6755
        An exception will throw if shape or dtype of the parameters is not match.
6756

6757 6758 6759 6760
        .. note::
            This function MUST called after run start_up_program

        Args:
6761
            state_dict(dict): the dict store parameters and persistable buffers.
6762 6763
                The key is the name of the parameter or the name of the buffer.
                The value is the tensor of this variable in the given scope.
6764
            scope(Scope, optional) : If scope is None, state_dict will be set to global scope
6765 6766
                obtained through 'paddle.static.global_scope()'. Otherwise, value will be set to scope.
                Default: None
6767

6768 6769 6770 6771 6772 6773 6774 6775 6776 6777 6778 6779 6780 6781 6782 6783 6784 6785 6786 6787 6788 6789 6790 6791 6792 6793 6794 6795 6796
        Returns:
            None

        Examples:
            .. code-block:: python

                import paddle
                import paddle.static as static

                paddle.enable_static()

                x = static.data(name="x", shape=[10, 10], dtype='float32')
                y = static.nn.fc(x, 10)
                z = static.nn.fc(y, 10)

                place = paddle.CPUPlace()
                exe = static.Executor(place)
                exe.run(static.default_startup_program())
                prog = static.default_main_program()

                path = "./temp/model.pdparams"
                paddle.save(prog.state_dict(), path)
                state_dict_load = paddle.load(path)
                prog.set_state_dict(state_dict_load)
        """

        if not isinstance(state_dict, dict):
            raise TypeError(
                "Type of `state_dict` should be dict, but received {}.".format(
6797 6798 6799
                    type(state_dict)
                )
            )
6800 6801

        vars_dict = {var.name: var for var in self.list_vars()}
6802 6803 6804
        condition = (
            True if 'StructuredToParameterName@@' in state_dict else False
        )
6805 6806 6807 6808 6809 6810 6811 6812 6813 6814 6815
        for name, value in state_dict.items():
            if condition:
                if name == "StructuredToParameterName@@":
                    continue
                if name in state_dict['StructuredToParameterName@@']:
                    name = state_dict['StructuredToParameterName@@'][name]
            if name in vars_dict:
                try:
                    vars_dict[name].set_value(value, scope)
                except ValueError as err:
                    warnings.warn(
6816 6817
                        ("Skip loading for '{}'. ".format(name) + str(err))
                    )
6818 6819
                except TypeError as err:
                    warnings.warn(
6820 6821
                        ("Skip loading for '{}'. ".format(name) + str(err))
                    )
6822
            else:
6823
                warnings.warn(
6824 6825 6826 6827 6828 6829
                    (
                        "Skip loading for '{0}'. Because '{0}' not in the program.".format(
                            name
                        )
                    )
                )
6830

Y
Yu Yang 已提交
6831

6832
class Parameter(Variable, metaclass=ParameterMetaClass):
6833
    """
6834
    Parameter is derived from Variable. A parameter is a persistable
6835
    Variable, and will be updated by optimizers after each iteration.
6836
    The training of a neural network is essentially the updating of
6837 6838
    its parameters.

6839
    Relative to a general Variable, a Parameter has several its own
6840 6841
    member variables:

6842 6843 6844 6845 6846 6847 6848 6849 6850 6851
    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.
6852
        need_clip (bool): Whether the parameter gradient need to be cliped
6853
            in optimizer. Default is True.
6854 6855
    """

6856 6857 6858 6859 6860 6861 6862 6863
    def __init__(
        self,
        block,
        shape,
        dtype,
        type=core.VarDesc.VarType.LOD_TENSOR,
        **kwargs
    ):
6864 6865 6866 6867 6868
        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
6869 6870
        for each in shape:
            if each < 0:
6871 6872
                raise ValueError(
                    "Each dimension of shape for Parameter must be greater than 0, but received %s"
6873 6874 6875 6876 6877 6878 6879 6880 6881 6882 6883 6884
                    % list(shape)
                )

        Variable.__init__(
            self,
            block,
            persistable=True,
            shape=shape,
            dtype=dtype,
            type=type,
            **kwargs
        )
Y
Yu Yang 已提交
6885 6886 6887 6888
        self.trainable = kwargs.get('trainable', True)

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

6889 6890
        self.regularizer = kwargs.get('regularizer', None)

W
wanghaoshuang 已提交
6891
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
6892

6893 6894
        self.need_clip = kwargs.get('need_clip', True)

6895 6896
        self.is_distributed = False

6897 6898
        self.is_parameter = True

F
fengjiayi 已提交
6899
    def __str__(self):
6900
        return self._to_readable_code()
F
fengjiayi 已提交
6901

F
update  
fengjiayi 已提交
6902 6903 6904
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
6905

F
update  
fengjiayi 已提交
6906 6907 6908 6909 6910 6911 6912 6913
        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.

6914 6915 6916 6917 6918 6919 6920 6921 6922
        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
6923
        """
6924
        assert isinstance(throw_on_error, bool) and isinstance(
6925 6926
            with_details, bool
        )
F
update  
fengjiayi 已提交
6927 6928
        if with_details:
            res_str = Variable.to_string(self, throw_on_error, True)
6929 6930 6931 6932 6933 6934 6935
            additional_attr = (
                "trainable",
                "optimize_attr",
                "regularizer",
                "do_model_average",
                "need_clip",
            )
F
update  
fengjiayi 已提交
6936
            for attr_name in additional_attr:
6937
                res_str += "%s: %s\n" % (attr_name, getattr(self, attr_name))
F
update  
fengjiayi 已提交
6938 6939
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
6940 6941 6942 6943
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
6944

6945 6946
class ParamBase(core.VarBase):
    """
6947 6948
    ParamBase is derived from Tensor( Which is the concept in Dygraph Mode).
    A ParamBase is a persistable Tensor, and will be updated by optimizers
6949
    after each iteration.
6950 6951 6952
    The training of a neural network is essentially the updating of
    its ParamBase.

6953
    Relative to a general Tensor, a ParamBase has several its own
6954 6955 6956 6957 6958 6959 6960 6961 6962 6963 6964 6965
    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.
6966
        need_clip (bool): Whether the parameter gradient need to be cliped
6967
            in optimizer. Default is True.
6968 6969 6970 6971 6972 6973 6974 6975 6976 6977 6978 6979 6980
    """

    @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")

        for each in shape:
            if each < 0:
                raise ValueError(
                    "Each dimension of shape for Parameter must be greater than 0, but received %s"
6981 6982
                    % list(shape)
                )
6983 6984 6985 6986 6987 6988 6989

        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'))

6990 6991 6992 6993 6994 6995 6996
        super(ParamBase, self).__init__(
            dtype if dtype else core.VarDesc.VarType.FP32,
            list(shape) if shape else [],
            name,
            core.VarDesc.VarType.LOD_TENSOR,
            True,
        )
6997

6998 6999
        trainable = kwargs.get('trainable', True)
        self.stop_gradient = not trainable
7000 7001 7002 7003 7004 7005 7006

        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)

7007 7008
        self.need_clip = kwargs.get('need_clip', True)

7009
        self.is_distributed = kwargs.get('is_distributed', False)
7010
        # self.block = default_main_program().global_block()
7011

7012 7013 7014 7015 7016 7017 7018 7019 7020 7021 7022
    @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 ",
7023 7024
                type(trainable),
            )
7025

7026
    def __str__(self):
7027
        """
7028
        Convert a ParamBase object to a readable string.
7029

7030
        Returns(str): A readable string.
7031 7032 7033 7034

        Examples:
            .. code-block:: python

7035
                import paddle
7036 7037 7038 7039 7040 7041 7042
                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]])
7043
        """
7044
        return "Parameter containing:\n{tensor}".format(
7045 7046
            tensor=super(ParamBase, self).__str__()
        )
7047

7048 7049 7050 7051 7052 7053 7054 7055 7056 7057 7058
    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
7059

7060 7061 7062 7063 7064 7065 7066 7067 7068 7069 7070 7071 7072 7073 7074 7075 7076 7077
                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

7078 7079 7080 7081
    def _copy_to(self, device, blocking):
        state = copy.deepcopy(self.__dict__)
        new_param = ParamBase(self.shape, self.dtype, **state)
        core.varbase_copy(self, new_param, device, blocking)
7082 7083 7084 7085 7086 7087
        return new_param

    __repr__ = __str__


if hasattr(core, "eager"):
7088
    _core_eager_eagertensor = core.eager.Tensor
7089 7090 7091 7092 7093 7094
else:
    _core_eager_eagertensor = object


class EagerParamBase(_core_eager_eagertensor):
    """
7095 7096
    EagerParamBase is derived from Tensor( Which is the concept in Eager-Dygraph Mode).
    A EagerParamBase is a persistable Tensor, and will be updated by optimizers
7097 7098 7099 7100 7101 7102 7103 7104 7105 7106 7107 7108 7109 7110 7111 7112 7113
    after each iteration.
    The training of a neural network is essentially the updating of
    its EagerParamBase.

    Relative to a general Tensor, a EagerParamBase has several its own
    member variables:

    Args:
        trainable(bool): True if the EagerParamBase need to be updated after
            iterations.
        optimize_attr(map): EagerParamBase 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 EagerParamBase. Default: None
        do_model_average(bool): True if the model average strategy will
            be applied on this EagerParamBase.
7114
        need_clip (bool): Whether the parameter gradient need to be cliped
7115 7116 7117 7118 7119 7120 7121 7122 7123 7124 7125 7126 7127 7128
            in optimizer. Default is True.
    """

    @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")

        for each in shape:
            if each < 0:
                raise ValueError(
                    "Each dimension of shape for Parameter must be greater than 0, but received %s"
7129 7130
                    % list(shape)
                )
7131 7132 7133 7134 7135 7136 7137

        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('_eager_param_base'))

7138 7139 7140
        if isinstance(shape, core.eager.Tensor):
            shape = shape.numpy()

7141 7142 7143 7144 7145 7146 7147
        super(EagerParamBase, self).__init__(
            dtype if dtype else core.VarDesc.VarType.FP32,
            list(shape) if shape else [],
            name,
            core.VarDesc.VarType.LOD_TENSOR,
            True,
        )
7148 7149 7150 7151 7152 7153 7154 7155 7156 7157 7158 7159 7160 7161
        self.retain_grads()

        trainable = kwargs.get('trainable', True)
        self.stop_gradient = not trainable

        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)

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

        self.is_distributed = kwargs.get('is_distributed', False)
7162 7163 7164
        # hook functions for lazy initialization
        self._init_func = None
        self._init_op_creator = None
7165 7166

    def set_init_func(self, obj):
7167
        self._init_func = obj
7168 7169 7170

    @dygraph_only
    def initialize(self):
7171 7172 7173
        assert (
            self._init_func is not None
        ), "Required self._init_func is not None, but received None."
7174
        self._init_func()
7175
        # clear function handle to release resource
7176
        self._init_func = None
7177 7178 7179 7180 7181 7182 7183 7184 7185 7186 7187 7188

    @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 ",
7189 7190
                type(trainable),
            )
7191

7192 7193 7194 7195
    def _create_init_op(self, block):
        """
        Call init_op_creator function to create initializer operation in block.
        """
7196 7197 7198
        assert (
            self._init_op_creator is not None
        ), "Required self._init_op_creator is not None, but received None."
7199 7200
        self._init_op_creator(block)

7201 7202 7203 7204 7205 7206 7207 7208 7209 7210 7211 7212 7213 7214 7215 7216 7217 7218 7219
    def __str__(self):
        """
        Convert a EagerParamBase object to a readable string.

        Returns(str): A readable string.

        Examples:
            .. code-block:: python

                import paddle
                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]])
        """
        return "Parameter containing:\n{tensor}".format(
7220 7221
            tensor=super(EagerParamBase, self).__str__()
        )
7222 7223 7224 7225 7226 7227 7228 7229 7230 7231 7232 7233 7234 7235 7236 7237 7238 7239 7240 7241 7242 7243 7244 7245 7246 7247 7248 7249 7250 7251 7252 7253 7254 7255 7256

    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)

                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 = EagerParamBase(self.shape, self.dtype, **state)
        memo[id(self)] = new_param
        new_param.copy_(self, True)
        return new_param

    def _copy_to(self, device, blocking):
        state = copy.deepcopy(self.__dict__)
        new_param = EagerParamBase(self.shape, self.dtype, **state)
        core.eager.tensor_copy(self, new_param, device, blocking)
7257 7258
        return new_param

7259 7260 7261
    __repr__ = __str__


Y
Yu Yang 已提交
7262
# program is a global instance.
Y
Yu Yang 已提交
7263 7264
_main_program_ = Program()
_startup_program_ = Program()
7265
_startup_program_._is_start_up_program_ = True
7266

7267

7268
def default_startup_program():
Y
Yu Yang 已提交
7269
    """
Y
yuyang18 已提交
7270 7271
    Get default/global startup program.

7272
    The :code:`paddle.nn` function will append the initialization operators into startup program.
7273
    The :code:`startup_program` will initialize the parameters by the OPs.
T
tangwei12 已提交
7274

7275 7276
    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
7277

7278 7279
    Returns:
        Program: current default startup program.
7280

7281
    Returns type:
7282 7283 7284 7285

    Examples:
        .. code-block:: python

7286
            import paddle
7287

7288
            paddle.enable_static()
7289 7290 7291 7292
            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
7293
    """
Y
Yu Yang 已提交
7294
    return _startup_program_
7295

7296

7297
def default_main_program():
Y
Yu Yang 已提交
7298
    """
7299
    This API can be used to get ``default main program`` which store the
7300
    descriptions of Ops and tensors.
T
tangwei12 已提交
7301

7302 7303
    For example ``z = paddle.add(x, y)`` will create a new ``add``
    Op and a new ``z`` tensor, and they will be recorded in ``default main program`` .
Y
yuyang18 已提交
7304

7305
    The ``default main program`` is the default value for ``Program`` parameter in
7306
    a lot of APIs. For example, the :code:`Executor.run()` will execute the
Y
yuyang18 已提交
7307
    :code:`default_main_program` when the program is not specified.
7308

7309
    If you want to switch the ``default main program``, you can use :ref:`api_paddle_fluid_framework_program_guard` .
T
tangwei12 已提交
7310

Y
Yu Yang 已提交
7311
    Returns:
7312
        Program: A ``Program`` which holding the descriptions of OPs and tensors in the network.
7313 7314 7315 7316

    Examples:
        ..  code-block:: python

7317
            import paddle
7318

7319
            paddle.enable_static()
7320
            # Sample Network:
7321 7322 7323
            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)
7324

7325 7326 7327
            #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
7328
            print(paddle.static.default_main_program())
Y
Yu Yang 已提交
7329
    """
Y
Yu Yang 已提交
7330
    return _main_program_
Y
Yu Yang 已提交
7331 7332 7333 7334 7335


def switch_main_program(program):
    """
    Switch the main program to a new program.
7336

Y
Yu Yang 已提交
7337 7338 7339 7340 7341 7342 7343 7344 7345 7346 7347 7348 7349 7350
    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):
    """
7351
    Switch the startup program to a new program
Y
Yu Yang 已提交
7352 7353 7354 7355 7356 7357 7358 7359 7360 7361 7362 7363
    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
7364
@signature_safe_contextmanager
Y
Yu Yang 已提交
7365 7366
def program_guard(main_program, startup_program=None):
    """
7367 7368
    :api_attr: Static Graph

7369 7370 7371
    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.
7372

G
guofei 已提交
7373
    Args:
7374
        main_program(Program): New main program inside ``with`` statement.
7375 7376
        startup_program(Program, optional): New startup program inside ``with``
            statement. :code:`None` means not changing startup program,
G
guofei 已提交
7377 7378 7379
            default_startup_program is still used.
            Default: None.

Y
Yu Yang 已提交
7380
    Examples:
7381
       .. code-block:: python
T
tangwei12 已提交
7382

7383
          import paddle
Y
yuyang18 已提交
7384

7385 7386 7387 7388 7389
          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')
7390
              hidden = paddle.static.nn.fc(x=data, size=10, activation='relu')
Y
yuyang18 已提交
7391 7392 7393

    Notes: The temporary :code:`Program` can be used if the user does not need
    to construct either of startup program or main program.
7394

Y
Yu Yang 已提交
7395
    Examples:
7396
       .. code-block:: python
Y
yuyang18 已提交
7397

7398
          import paddle
7399

7400 7401 7402 7403 7404
          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
7405

Y
Yu Yang 已提交
7406
    """
7407
    from .data_feeder import check_type
7408 7409 7410 7411

    check_type(
        main_program, 'main_program', Program, 'paddle.static.program_guard'
    )
Y
Yu Yang 已提交
7412 7413
    main_program = switch_main_program(main_program)
    if startup_program is not None:
7414 7415 7416 7417 7418 7419
        check_type(
            startup_program,
            'startup_program',
            Program,
            'paddle.static.program_guard',
        )
7420 7421
        # Tag the program __is_start_up as True
        startup_program._is_start_up_program_ = True
Y
Yu Yang 已提交
7422
        startup_program = switch_startup_program(startup_program)
7423 7424 7425 7426 7427 7428
    try:
        yield
    finally:
        switch_main_program(main_program)
        if startup_program is not None:
            switch_startup_program(startup_program)
X
xuwei06 已提交
7429 7430


W
Wu Yi 已提交
7431
def _get_var(name, program=None):
X
xuwei06 已提交
7432
    """
Y
yuyang18 已提交
7433
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
7434

X
xuwei06 已提交
7435 7436 7437
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
7438
        If None, default_global_program() will be used.
X
xuwei06 已提交
7439 7440 7441 7442 7443 7444 7445

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
7446
    assert isinstance(program, Program)
X
xuwei06 已提交
7447 7448

    return program.global_block().var(name)
7449 7450


S
rename  
sneaxiy 已提交
7451
@signature_safe_contextmanager
L
lujun 已提交
7452 7453
def _dygraph_guard(tracer):
    global _dygraph_tracer_
7454
    tmp_tracer = _dygraph_tracer_
L
lujun 已提交
7455
    _dygraph_tracer_ = tracer
7456
    core._switch_tracer(tracer)
M
minqiyang 已提交
7457

7458 7459 7460
    try:
        yield
    finally:
7461 7462
        core._switch_tracer(tmp_tracer)
        _dygraph_tracer_ = tmp_tracer
P
Paddle CI 已提交
7463 7464


S
rename  
sneaxiy 已提交
7465
@signature_safe_contextmanager
L
lujun 已提交
7466
def _dygraph_place_guard(place):
7467 7468 7469
    global _global_expected_place_
    tmp_place = _global_expected_place_
    _global_expected_place_ = place
7470 7471
    _set_dygraph_tracer_expected_place(place)

7472 7473 7474
    try:
        yield
    finally:
7475
        _global_expected_place_ = tmp_place
J
Jiabin Yang 已提交
7476
        _set_dygraph_tracer_expected_place(_global_expected_place_)
7477 7478


7479 7480 7481 7482 7483 7484 7485 7486 7487 7488
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):
    """
7489

7490 7491
    Note:
        The API only supports static mode.
7492 7493 7494 7495

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

    Args:
7496
        device(str|None): Specify the device to use in the context. It should be ``cpu``,
7497
            ``gpu`` or ``gpu:x``, where ``x`` is the index of the GPUs.
7498 7499 7500 7501 7502 7503 7504
            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:
7505

7506
        .. code-block:: python
7507

7508
            # required: gpu
Z
Zhang Ting 已提交
7509
            import paddle
7510

Z
Zhang Ting 已提交
7511 7512 7513
            paddle.enable_static()
            support_gpu = paddle.is_compiled_with_cuda()
            place = paddle.CPUPlace()
7514
            if support_gpu:
Z
Zhang Ting 已提交
7515
                place = paddle.CUDAPlace(0)
7516 7517

            # if GPU is supported, the three OPs below will be automatically assigned to CUDAPlace(0)
Z
Zhang Ting 已提交
7518 7519 7520
            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)
7521

Z
Zhang Ting 已提交
7522
            with paddle.static.device_guard("cpu"):
7523
                # Ops created here will be placed on CPUPlace
Z
Zhang Ting 已提交
7524 7525
                shape = paddle.slice(shape, axes=[0], starts=[0], ends=[4])
            with paddle.static.device_guard('gpu'):
7526
                # if GPU is supported, OPs created here will be placed on CUDAPlace(0), otherwise on CPUPlace
Z
Zhang Ting 已提交
7527
                out = paddle.reshape(data1, shape=shape)
7528

Z
Zhang Ting 已提交
7529 7530
            exe = paddle.static.Executor(place)
            exe.run(paddle.static.default_startup_program())
7531 7532 7533
            result = exe.run(fetch_list=[out])
    """

7534 7535 7536 7537 7538
    index = None
    if device and ':' in device:
        device, index = device.split(':')
        if device == 'cpu':
            raise ValueError("Should not set device id for cpu.")
7539
    if device not in ['cpu', 'gpu', 'npu', 'xpu', 'mlu', '', None]:
7540
        raise ValueError(
7541
            "The Attr(device) should be 'cpu' 'npu' 'xpu' 'mlu' or 'gpu', and it can also be empty string or None "
7542 7543
            "when there is no need to specify device. But received %s" % device
        )
7544 7545
    if index:
        device = ":".join([device, index])
7546
    pre_device = switch_device(device)
7547 7548 7549 7550
    try:
        yield
    finally:
        switch_device(pre_device)
G
guofei 已提交
7551 7552


7553 7554 7555 7556 7557 7558 7559 7560 7561 7562 7563 7564 7565 7566 7567 7568 7569 7570 7571 7572
def _switch_cuda_graph_mode(cuda_graph_attr):
    global _current_cuda_graph_mode
    pre_mode = _current_cuda_graph_mode
    _current_cuda_graph_mode = cuda_graph_attr
    return pre_mode


@signature_safe_contextmanager
def _cuda_graph_guard(cuda_graph_attr=None):
    """

    Note:
        The API only supports static mode.

    A context manager that specifies the cuda_graph_mode which indicating the cuda graph capture under static mode.

    Args:
        cuda_graph_attr(str|None): The cuda graph attr with the format of:
                                   cuda_graph_capture_mode;memory_pool_id;cuda_graph_id
    """
7573 7574
    assert (
        not _non_static_mode()
7575
    ), "cuda_graph_guard only works under static mode"
7576 7577
    assert (
        core.is_compiled_with_cuda()
7578 7579 7580 7581 7582 7583 7584 7585
    ), "cuda_graph_guard context can be only used when Paddle is compiled with cuda"
    pre_mode = _switch_cuda_graph_mode(cuda_graph_attr)
    try:
        yield
    finally:
        _switch_cuda_graph_mode(pre_mode)


G
guofei 已提交
7586 7587 7588
def set_flags(flags):
    """
    This function sets the GFlags value in Paddle.
7589
    For FLAGS please refer to :ref:`en_guides_flags_flags`
G
guofei 已提交
7590 7591 7592 7593 7594 7595 7596

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

    Examples:
            .. code-block:: python

7597 7598
                import paddle
                paddle.set_flags({'FLAGS_eager_delete_tensor_gb': 1.0})
G
guofei 已提交
7599 7600 7601 7602
    """
    if not isinstance(flags, dict):
        raise TypeError('flags in set_flags should be a dict')
    for key, value in flags.items():
7603 7604
        if _global_flags().is_public(key):
            _global_flags()[key] = value
G
guofei 已提交
7605 7606
        else:
            raise ValueError(
7607 7608
                "Flag %s cannot set its value through this function." % (key)
            )
G
guofei 已提交
7609 7610 7611 7612 7613


def get_flags(flags):
    """
    This function gets the GFlags value in Paddle.
7614
    For FLAGS please refer to :ref:`en_guides_flags_flags`
G
guofei 已提交
7615 7616 7617 7618 7619 7620 7621 7622 7623 7624

    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

7625
            import paddle
G
guofei 已提交
7626 7627

            flags = ['FLAGS_eager_delete_tensor_gb', 'FLAGS_check_nan_inf']
7628
            res = paddle.get_flags(flags)
G
guofei 已提交
7629 7630 7631 7632 7633 7634
            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:
7635
            if _global_flags().is_public(key):
7636
                value = _global_flags()[key]
G
guofei 已提交
7637 7638 7639 7640
                temp = {key: value}
                flags_value.update(temp)
            else:
                raise ValueError(
7641 7642 7643
                    'Flag %s cannot get its value through this function.'
                    % (key)
                )
G
guofei 已提交
7644
    elif isinstance(flags, str):
7645
        if _global_flags().is_public(flags):
7646
            value = _global_flags()[flags]
G
guofei 已提交
7647 7648 7649 7650
            temp = {flags: value}
            flags_value.update(temp)
        else:
            raise ValueError(
7651 7652
                'Flag %s cannot get its value through this function.' % (flags)
            )
G
guofei 已提交
7653 7654 7655
    else:
        raise TypeError('Flags in get_flags should be a list, tuple or string.')
    return flags_value
7656 7657 7658 7659 7660 7661


def _get_paddle_place(place):
    "convert the string to paddle Place"
    if place is None:
        return place
7662 7663 7664 7665 7666 7667 7668 7669 7670 7671 7672 7673 7674 7675
    if isinstance(
        place,
        (
            core.Place,
            core.XPUPlace,
            core.CPUPlace,
            core.CUDAPinnedPlace,
            core.CUDAPlace,
            core.NPUPlace,
            core.IPUPlace,
            core.MLUPlace,
            core.CustomPlace,
        ),
    ):
7676 7677 7678 7679
        return place

    if not isinstance(place, str):
        raise ValueError(
7680 7681
            "place only support string which is 'Place' and so on."
        )
7682 7683

    place = place.lower()
7684
    if place == "cpu":
7685
        return core.CPUPlace()
7686

7687
    if place == "device":
7688 7689
        return core.Place()

7690
    # GPU
7691 7692 7693 7694
    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(
7695 7696 7697
                "The device should not be {}, since PaddlePaddle is "
                "not compiled with CUDA".format(avaliable_gpu_place)
            )
7698 7699 7700 7701 7702 7703 7704 7705 7706
        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)
7707 7708

    # XPU
7709 7710 7711 7712
    avaliable_xpu_place = re.match(r'xpu:\d+', place)
    if avaliable_xpu_place:
        if not core.is_compiled_with_xpu():
            raise ValueError(
7713 7714 7715
                "The device should not be {}, since PaddlePaddle is "
                "not compiled with XPU".format(avaliable_xpu_place)
            )
7716 7717 7718 7719
        place_info_list = place.split(':', 1)
        device_id = place_info_list[1]
        device_id = int(device_id)
        return core.XPUPlace(device_id)
7720 7721 7722 7723 7724 7725

    # NPU
    avaliable_npu_place = re.match(r'npu:\d+', place)
    if avaliable_npu_place:
        if not core.is_compiled_with_npu():
            raise ValueError(
7726 7727 7728
                "The device should not be {}, since PaddlePaddle is "
                "not compiled with NPU".format(avaliable_npu_place)
            )
7729 7730 7731 7732 7733
        place_info_list = place.split(':', 1)
        device_id = place_info_list[1]
        device_id = int(device_id)
        return core.NPUPlace(device_id)

J
jianghaicheng 已提交
7734 7735 7736 7737 7738
    # IPU
    avaliable_ipu_place = re.match(r'ipu:\d+', place)
    if avaliable_ipu_place:
        if not core.is_compiled_with_ipu():
            raise ValueError(
7739 7740 7741
                "The device should not be {}, since PaddlePaddle is "
                "not compiled with IPU".format(avaliable_ipu_place)
            )
J
jianghaicheng 已提交
7742 7743 7744 7745 7746
        place_info_list = place.split(':', 1)
        device_id = place_info_list[1]
        device_id = int(device_id)
        return core.IPUPlace(device_id)

7747 7748 7749 7750 7751
    # MLU
    avaliable_mlu_place = re.match(r'mlu:\d+', place)
    if avaliable_mlu_place:
        if not core.is_compiled_with_mlu():
            raise ValueError(
7752 7753 7754
                "The device should not be {}, since PaddlePaddle is "
                "not compiled with MLU".format(avaliable_mlu_place)
            )
7755 7756 7757 7758 7759
        place_info_list = place.split(':', 1)
        device_id = place_info_list[1]
        device_id = int(device_id)
        return core.MLUPlace(device_id)

7760
    raise ValueError(
7761 7762 7763 7764
        "Paddle supports CPUPlace, CUDAPlace,CUDAPinnedPlace, XPUPlace, IPUPlace, MLUPlace and NPUPlace, but received {}.".format(
            place
        )
    )
7765 7766 7767 7768 7769 7770 7771 7772 7773 7774 7775 7776 7777


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