framework.py 259.4 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15 16
from __future__ import print_function

Y
Yu Yang 已提交
17
import collections
X
Xin Pan 已提交
18
from collections import defaultdict
H
huzhiqiang 已提交
19
from collections.abc import Iterable
Q
qiaolongfei 已提交
20
import contextlib
21
from .wrapped_decorator import signature_safe_contextmanager, wrap_decorator
P
peizhilin 已提交
22
import os
F
fengjiayi 已提交
23
import re
24
import traceback
25
import six
26
import copy
27
from types import MethodType, FunctionType
28

Y
Yu Yang 已提交
29
import numpy as np
30
import subprocess
S
sneaxiy 已提交
31
import multiprocessing
32
import sys
33
import logging
M
minqiyang 已提交
34
from .. import compat as cpt
35
from .proto import framework_pb2
36 37

from . import core
38
from . import unique_name
39 40
import paddle.version as fluid_version
import warnings
41
import functools
42
from .variable_index import _getitem_impl_, _setitem_impl_
Y
Yu Yang 已提交
43

44
__all__ = [
45 46 47 48
    'Program',
    'default_startup_program',
    'default_main_program',
    'program_guard',
49
    'name_scope',
J
jianghaicheng 已提交
50
    'ipu_shard_guard',
51
    'set_ipu_shard',
S
sneaxiy 已提交
52 53
    'cuda_places',
    'cpu_places',
54
    'xpu_places',
55
    'mlu_places',
S
sneaxiy 已提交
56
    'cuda_pinned_places',
J
Jiabin Yang 已提交
57
    '_non_static_mode',
L
lujun 已提交
58
    'in_dygraph_mode',
59
    'is_compiled_with_cinn',
C
chengduo 已提交
60
    'is_compiled_with_cuda',
61
    'is_compiled_with_rocm',
62
    'is_compiled_with_xpu',
63
    'is_compiled_with_npu',
64
    'Variable',
65
    'require_version',
66
    'device_guard',
G
guofei 已提交
67 68
    'set_flags',
    'get_flags',
69
]
Y
Yu Yang 已提交
70

Q
qiaolongfei 已提交
71 72 73 74
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
75 76
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()

L
lujun 已提交
77
_dygraph_tracer_ = None
78
_in_eager_mode_ = (os.environ.get('FLAGS_enable_eager_mode', '1') == '1')
79
_global_expected_place_ = None
80
_current_device = None
81
global_prog_seed = 0
82
_current_pipeline_stage = None
83
_already_patch_eager_tensor = False
J
Jiabin Yang 已提交
84
_already_patch_varbase = False
85
_current_cuda_graph_mode = None
86
_global_flags_ = core.globals()
L
Leo Chen 已提交
87 88
_enable_standalone_executor_ = (os.environ.get('FLAGS_USE_STANDALONE_EXECUTOR',
                                               None))
89
_dy2st_enable_standalone_executor_ = (os.environ.get(
90
    'FLAGS_DY2ST_USE_STANDALONE_EXECUTOR', 0))
J
Jiabin Yang 已提交
91 92

# Some explanation of our execution system 2022.03
93
# For now we have 3 kinds of execution system, since we refactored dygraph mode to
J
Jiabin Yang 已提交
94
# build a fast execution system for dynamic mode. But we can't just remove all legacy
95
# code once we present the new system for some historical reason. That's why we have
J
Jiabin Yang 已提交
96
# these flags.
97
#
J
Jiabin Yang 已提交
98
# 1. _non_static_mode():
99
# _non_static_mode means  we are now running in legacy dygraph mode or dygraph mode.
J
Jiabin Yang 已提交
100 101 102 103
# 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
104
#
J
Jiabin Yang 已提交
105
# They have a relation ship as below:
106
# Both dygraph_mode and _in_legacy_dygraph are _non_static_mode, but if you are running in
J
Jiabin Yang 已提交
107
# dygraph mode means you are not in _in_legacy_dygraph.
108
#
J
Jiabin Yang 已提交
109 110 111 112 113 114
# 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.


115 116 117 118 119
def _update_monkey_methods(is_eager):
    """
    Update monkey methods of VarBase or eager.Tensor while
    switching eager mode and legacy mode.
    """
120
    from paddle import _C_ops, _legacy_C_ops
121 122 123
    from .dygraph.varbase_patch_methods import monkey_patch_varbase
    from .dygraph import monkey_patch_math_varbase

124 125 126
    global _already_patch_eager_tensor
    global _already_patch_varbase

127
    assert isinstance(is_eager, bool)
128
    # switch into eager mode
129
    if is_eager:
130
        _legacy_C_ops.switch_to_eager_ops()
131 132 133 134 135 136
        if not _already_patch_eager_tensor:
            monkey_patch_varbase()
            monkey_patch_math_varbase()

            _already_patch_eager_tensor = True
    # switch back into legacy mode
137
    else:
138
        _legacy_C_ops.switch_to_core_ops()
139 140 141 142 143
        if not _already_patch_varbase:
            monkey_patch_varbase()
            monkey_patch_math_varbase()

            _already_patch_varbase = True
144

145 146 147 148 149 150 151 152 153 154 155
    # switch Paddle.Tensor bind type
    _switch_tensor_bind_type(is_eager)


def _switch_tensor_bind_type(is_eager):
    import paddle
    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_


175 176 177 178 179 180 181 182 183 184 185
# 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
    # Only enable eager on CPU/GPU
    is_not_support = core.is_compiled_with_xpu() or core.is_compiled_with_npu(
J
Jiabin Yang 已提交
186
    ) or core.is_compiled_with_ipu() or core.is_compiled_with_mlu()
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207

    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
208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243
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
244 245 246


@signature_safe_contextmanager
J
Jiabin Yang 已提交
247
def _test_eager_guard(place=None):
248 249 250 251 252
    # 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.
    already_fallback = _fallback_legacy_dygraph()
    if not already_fallback:
        _disable_legacy_dygraph()
253
    try:
J
Jiabin Yang 已提交
254
        yield
255
    finally:
256 257
        if not already_fallback:
            _enable_legacy_dygraph()
258 259


260 261
global_ipu_index = -1
global_ipu_stage = -1
J
jianghaicheng 已提交
262 263 264 265
ipu_index_attr_name = 'ipu_index'
ipu_stage_attr_name = 'ipu_stage'


L
Leo Chen 已提交
266 267 268 269 270 271 272 273 274 275 276
@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
277
@signature_safe_contextmanager
278
def ipu_shard_guard(index=-1, stage=-1):
J
jianghaicheng 已提交
279 280 281 282
    """
    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
283
        index(int, optional): Specify which ipu the Tensor is computed on, (such as '0, 1, 2, 3').
284
            The default value is -1, which means the Op only run on IPU 0.
W
Weilong Wu 已提交
285
        stage(int, optional): Specify the computation order of the sharded model(such as '0, 1, 2, 3').
286
            The sharded model will be computed from small to large. The default value is -1, 
J
jianghaicheng 已提交
287 288 289
            which means no pipelining computation order and run Ops in terms of graph.
    
    **Note**:
290
    Only if the enable_manual_shard=True, the 'index' is able to be set not -1. Please refer 
J
jianghaicheng 已提交
291
    to :code:`paddle.static.IpuStrategy` . 
292
    Only if the enable_pipelining=True, the 'stage' is able to be set not -1. Please refer 
J
jianghaicheng 已提交
293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329
    to :code:`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.

    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


330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386
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.

    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’).
            The sharded model will be computed from small to large. The default value is -1, 
            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
    if not isinstance(call_func, Layer):
        if callable(call_func):
            return decorate(call_func)
        else:
            raise TypeError(
                "Unsupported type. Only accept paddle.nn.Layer or function.")

    # 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


387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471
def require_version(min_version, max_version=None):
    """
        Check if the installed version of PaddlePaddle is in [min_version, max_version],
        if the installed version is lower than ``min_version`` or higher than ``max_version``,
        an exception will be thrown, NO returns if the installed version is satisfied.

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

        Returns:
            None.

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

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

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

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

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

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

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

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

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

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

    min_version_split = min_version.split('.')
472 473
    min_version_to_check = min_version_split + zero_version[
        len(min_version_split):]
474 475 476

    if max_version is not None:
        max_version_split = max_version.split('.')
477 478
        max_version_to_check = max_version_split + zero_version[
            len(max_version_split):]
479 480 481 482 483 484 485 486 487 488 489 490 491 492 493

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


494
def _dygraph_not_support_(func):
495

496
    def __impl__(*args, **kwargs):
J
Jiabin Yang 已提交
497 498
        assert not _non_static_mode(
        ), "We don't support %s in dynamic graph mode" % func.__name__
499 500 501 502 503 504
        return func(*args, **kwargs)

    return __impl__


def _dygraph_only_(func):
505

506
    def __impl__(*args, **kwargs):
J
Jiabin Yang 已提交
507
        assert _non_static_mode(
508 509 510 511 512 513
        ), "We only support '%s()' in dynamic graph mode, please call 'paddle.disable_static()' to enter dynamic graph mode." % func.__name__
        return func(*args, **kwargs)

    return __impl__


514 515 516 517 518 519 520 521 522 523 524
def _non_static_only_(func):

    def __impl__(*args, **kwargs):
        from .dygraph.base import in_declarative_mode
        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__
        return func(*args, **kwargs)

    return __impl__


525
def _static_only_(func):
526

527
    def __impl__(*args, **kwargs):
J
Jiabin Yang 已提交
528
        assert not _non_static_mode(
529
        ), "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__
530 531 532 533 534
        return func(*args, **kwargs)

    return __impl__


535 536 537 538 539
def _set_pipeline_stage(stage):
    global _current_pipeline_stage
    _current_pipeline_stage = stage


540 541 542 543 544 545
# 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
546
# same base class.
547
def _fake_interface_only_(func):
548

549 550
    def __impl__(*args, **kwargs):
        raise AssertionError(
551 552 553 554 555
            "'%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'."
            % (func.__name__, func.__name__))
556 557 558 559

    return __impl__


T
tangwei12 已提交
560 561
# 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
562 563 564 565
# 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):
566

567 568 569 570 571 572 573 574 575 576 577 578 579
    @functools.wraps(func)
    def wrapper(*args, **kwargs):
        if 'stat_dict' in kwargs:
            warnings.warn(
                "The argument `stat_dict` has deprecated, please change it to `state_dict`.",
                DeprecationWarning)
            kwargs['state_dict'] = kwargs['stat_dict']
            kwargs.pop('stat_dict')
        return func(*args, **kwargs)

    return wrapper


580 581
dygraph_not_support = wrap_decorator(_dygraph_not_support_)
dygraph_only = wrap_decorator(_dygraph_only_)
582
static_only = wrap_decorator(_static_only_)
583
fake_interface_only = wrap_decorator(_fake_interface_only_)
584
non_static_only = wrap_decorator(_non_static_only_)
585 586


L
lujun 已提交
587 588
def _dygraph_tracer():
    return _dygraph_tracer_
589

W
Wu Yi 已提交
590

591 592 593 594
def _global_flags():
    return _global_flags_


M
minqiyang 已提交
595
def _current_expected_place():
596 597 598
    global _global_expected_place_
    if _global_expected_place_ is None:
        if core.is_compiled_with_cuda():
599 600 601 602 603
            try:
                device_count = core.get_cuda_device_count()
            except Exception as e:
                device_count = 0
            if device_count > 0:
604
                _global_expected_place_ = core.CUDAPlace(_cuda_ids()[0])
605 606 607 608 609
            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()
610 611 612 613 614 615
        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:
616
                _global_expected_place_ = core.XPUPlace(_xpu_ids()[0])
617 618 619 620 621
            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()
622 623 624 625 626 627
        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:
628
                _global_expected_place_ = core.MLUPlace(_mlu_ids()[0])
629 630 631 632 633
            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()
634 635 636 637 638 639 640 641 642 643 644 645 646 647 648
        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
649
    _set_dygraph_tracer_expected_place(place)
M
minqiyang 已提交
650 651


L
Leo Chen 已提交
652 653 654 655
# TODO(zhiqiu): remove this function.
def _var_base_to_np(var_base):
    """	
    convert VarBase tp numpy	
T
tangwei12 已提交
656

L
Leo Chen 已提交
657 658 659 660 661 662 663 664 665 666 667 668
    Args:	
        var_base(VarBase) : the VarBase to convert	
    Returns (np.ndarray): the np.ndarray contain the value of VarBase	
    """

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

    return var_base.numpy()


S
sneaxiy 已提交
669
def _cpu_num():
670
    if "CPU_NUM" not in os.environ.keys():
C
chengduo 已提交
671 672 673 674 675 676 677 678
        if multiprocessing.cpu_count() > 1:
            sys.stderr.write(
                '!!! The CPU_NUM is not specified, you should set CPU_NUM in the environment variable list.\n'
                'CPU_NUM indicates that how many CPUPlace are used in the current task.\n'
                'And if this parameter are set as N (equal to the number of physical CPU core) the program may be faster.\n\n'
                'export CPU_NUM={} # for example, set CPU_NUM as number of physical CPU core which is {}.\n\n'
                '!!! The default number of CPU_NUM=1.\n'.format(
                    multiprocessing.cpu_count(), multiprocessing.cpu_count()))
C
chengduo 已提交
679
        os.environ['CPU_NUM'] = str(1)
680
    cpu_num = os.environ.get('CPU_NUM')
C
chengduo 已提交
681 682 683 684 685 686 687 688 689 690
    return int(cpu_num)


def _cuda_ids():
    gpus_env = os.getenv("FLAGS_selected_gpus")
    if gpus_env:
        device_ids = [int(s) for s in gpus_env.split(",")]
    else:
        device_ids = six.moves.range(core.get_cuda_device_count())
    return device_ids
S
sneaxiy 已提交
691 692


693 694 695 696 697 698 699 700 701
def _xpu_ids():
    xpus_env = os.getenv("FLAGS_selected_xpus")
    if xpus_env:
        device_ids = [int(s) for s in xpus_env.split(",")]
    else:
        device_ids = six.moves.range(core.get_xpu_device_count())
    return device_ids


702 703 704 705 706 707 708 709 710
def _npu_ids():
    npus_env = os.getenv("FLAGS_selected_npus")
    if npus_env:
        device_ids = [int(s) for s in npus_env.split(",")]
    else:
        device_ids = six.moves.range(core.get_npu_device_count())
    return device_ids


711 712 713 714 715 716 717 718 719
def _mlu_ids():
    mlus_env = os.getenv("FLAGS_selected_mlus")
    if mlus_env:
        device_ids = [int(s) for s in mlus_env.split(",")]
    else:
        device_ids = six.moves.range(core.get_mlu_device_count())
    return device_ids


720 721 722 723 724 725 726 727 728 729 730 731 732 733 734
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()


735 736 737 738 739 740 741 742 743 744 745 746 747 748 749
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()


750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774
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.
    
    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.

    Returns: None 

    Examples:
        .. code-block:: python

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


775 776 777 778 779 780 781 782 783 784 785 786 787 788 789
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
790 791 792 793
def is_compiled_with_cuda():
    """
    Whether this whl package can be used to run the model on GPU.

794
    Returns (bool): `True` if CUDA is currently available, otherwise `False`.
C
chengduo 已提交
795 796 797 798

    Examples:
        .. code-block:: python

799
            import paddle
800
            support_gpu = paddle.device.is_compiled_with_cuda()
C
chengduo 已提交
801 802 803 804
    """
    return core.is_compiled_with_cuda()


805 806 807 808 809 810 811 812 813 814
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
815
            support_gpu = paddle.device.is_compiled_with_rocm()
816 817 818 819
    """
    return core.is_compiled_with_rocm()


S
sneaxiy 已提交
820
def cuda_places(device_ids=None):
L
lujun 已提交
821
    """
822
    Note:
823 824 825
        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
826
    This function creates a list of :code:`paddle.CUDAPlace` objects.
S
add doc  
sneaxiy 已提交
827 828

    If :code:`device_ids` is None, environment variable of
829
    :code:`FLAGS_selected_gpus` would be checked first. For example, if
S
add doc  
sneaxiy 已提交
830
    :code:`FLAGS_selected_gpus=0,1,2`, the returned list would
C
Chen Weihang 已提交
831
    be [paddle.CUDAPlace(0), paddle.CUDAPlace(1), paddle.CUDAPlace(2)].
S
add doc  
sneaxiy 已提交
832
    If :code:`FLAGS_selected_gpus` is not set, all visible
833
    gpu places would be returned according to the :code:`CUDA_VISIBLE_DEVICES` environment variable.
S
add doc  
sneaxiy 已提交
834 835

    If :code:`device_ids` is not None, it should be the device
836
    ids of GPUs. For example, if :code:`device_ids=[0,1,2]`,
S
add doc  
sneaxiy 已提交
837
    the returned list would be 
C
Chen Weihang 已提交
838
    [paddle.CUDAPlace(0), paddle.CUDAPlace(1), paddle.CUDAPlace(2)].
T
tangwei12 已提交
839

840
    Parameters:
841
        device_ids (list|tuple, optional): A list/tuple of int of GPU device ids.
S
add doc  
sneaxiy 已提交
842 843

    Returns:
C
Chen Weihang 已提交
844
        list of paddle.CUDAPlace: Created GPU place list.
L
lujun 已提交
845 846

    Examples:
847
    
L
lujun 已提交
848 849
        .. code-block:: python

C
Chen Weihang 已提交
850 851
            import paddle
            import paddle.static as static
T
tangwei12 已提交
852

853 854
            # required: gpu
            
C
Chen Weihang 已提交
855 856 857
            paddle.enable_static()

            cuda_places = static.cuda_places()
L
lujun 已提交
858 859

    """
S
sneaxiy 已提交
860 861 862
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_ids is None:
C
chengduo 已提交
863
        device_ids = _cuda_ids()
S
sneaxiy 已提交
864 865 866 867 868
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.CUDAPlace(dev_id) for dev_id in device_ids]


869 870 871 872
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
873 874 875 876 877 878 879 880 881 882 883
        This function creates a list of :code:`paddle.XPUPlace` objects.
        If :code:`device_ids` is None, environment variable of
        :code:`FLAGS_selected_xpus` would be checked first. For example, if
        :code:`FLAGS_selected_xpus=0,1,2`, the returned list would
        be [paddle.XPUPlace(0), paddle.XPUPlace(1), paddle.XPUPlace(2)].
        If :code:`FLAGS_selected_xpus` is not set, all visible
        xpu places would be returned.
        If :code:`device_ids` is not None, it should be the device
        ids of XPUs. For example, if :code:`device_ids=[0,1,2]`,
        the returned list would be 
        [paddle.XPUPlace(0), paddle.XPUPlace(1), paddle.XPUPlace(2)].
884 885 886 887 888 889 890
    
    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
891

892 893
            # required: xpu

894 895 896 897 898 899 900 901 902 903 904 905 906 907 908
            import paddle
            import paddle.static as static
            
            paddle.enable_static()
            xpu_places = static.xpu_places()
    """
    assert core.is_compiled_with_xpu(), \
        "Not compiled with XPU"
    if device_ids is None:
        device_ids = _xpu_ids()
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.XPUPlace(dev_id) for dev_id in device_ids]


909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949
def npu_places(device_ids=None):
    """
    **Note**:
        For multi-card tasks, please use `FLAGS_selected_npus` environment variable to set the visible NPU device.
    
    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]`,
    the returned list would be 
    [paddle.NPUPlace(0), paddle.NPUPlace(1), paddle.NPUPlace(2)].
    
    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
            
            paddle.enable_static()
            npu_places = static.npu_places()
    """
    assert core.is_compiled_with_npu(), \
        "Not compiled with NPU"
    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
950
def cpu_places(device_count=None):
L
lujun 已提交
951
    """
C
Chen Weihang 已提交
952
    This function creates a list of :code:`paddle.CPUPlace` objects, and returns the created list.
T
tangwei12 已提交
953

S
add doc  
sneaxiy 已提交
954 955
    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
C
chengduo 已提交
956 957
    If :code:`CPU_NUM` is not set, the default value is 1,
    i.e. CPU_NUM=1.
958 959
    :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
960

961 962
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
963 964

    Returns:
C
Chen Weihang 已提交
965
        list of paddle.CPUPlace: Created list of CPU places.
L
lujun 已提交
966 967

    Examples:
968
    
L
lujun 已提交
969 970
        .. code-block:: python

C
Chen Weihang 已提交
971 972
            import paddle
            import paddle.static as static
T
tangwei12 已提交
973

C
Chen Weihang 已提交
974 975 976
            paddle.enable_static()

            cpu_places = static.cpu_places()
L
lujun 已提交
977 978
    """

S
sneaxiy 已提交
979 980 981 982 983 984
    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
985
    """
986
    This function creates a list of :code:`fluid.CUDAPinnedPlace` objects.
S
add doc  
sneaxiy 已提交
987 988 989

    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
990 991 992 993
    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
994

995 996
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
997 998

    Returns:
999
        list of fluid.CUDAPinnedPlace: Created list of CUDA pinned places.
L
lujun 已提交
1000 1001 1002 1003

    Examples:
        .. code-block:: python

1004
            import paddle.fluid as fluid
L
lujun 已提交
1005 1006 1007 1008 1009
            cuda_pinned_places_cpu_num = fluid.cuda_pinned_places()
            # or
            cuda_pinned_places = fluid.cuda_pinned_places(1)

    """
S
sneaxiy 已提交
1010 1011 1012
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_count is None:
1013 1014
        device_count = len(_cuda_ids())
    return [core.CUDAPinnedPlace()] * device_count
S
sneaxiy 已提交
1015 1016


1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058
def mlu_places(device_ids=None):
    """
    **Note**:
        For multi-card tasks, please use `FLAGS_selected_mlus` environment variable to set the visible MLU device.
        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)].

    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()
    """
    assert core.is_compiled_with_mlu(), \
        "Not compiled with MLU"
    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]


1059
class NameScope(object):
1060

1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085
    def __init__(self, name="", parent=None):
        self._children = dict()
        self._name = name
        self._parent = parent

    def child(self, prefix):
        if prefix not in self._children:
            new_child = NameScope(prefix, self)
            self._children[prefix] = [new_child]
        else:
            new_child = NameScope(prefix + "_%d" % len(self._children[prefix]),
                                  self)
            self._children[prefix].append(new_child)
        return new_child

    def parent(self):
        return self._parent

    def name(self):
        return self._name


_name_scope = NameScope()


S
rename  
sneaxiy 已提交
1086
@signature_safe_contextmanager
1087 1088
def name_scope(prefix=None):
    """
1089

1090
    Generate hierarchical name prefix for the operators in Static Graph.
1091

T
Tao Luo 已提交
1092 1093 1094
    Note: 
        This should only used for debugging and visualization purpose.
        Don't use it for serious analysis such as graph/program transformations.
1095
        Don't use it in dygraph, since it will cause memory leak.
1096 1097

    Args:
T
Tao Luo 已提交
1098
        prefix(str, optional): prefix. Default is none.
1099 1100

    Examples:
1101
    
1102
        .. code-block:: python
T
Tink_Y 已提交
1103

1104 1105 1106
          import paddle
          paddle.enable_static()
          with paddle.static.name_scope("s1"):
1107
             a = paddle.static.data(name='data', shape=[None, 1], dtype='int32')
T
Tao Luo 已提交
1108
             b = a + 1
1109
             with paddle.static.name_scope("s2"):
T
Tao Luo 已提交
1110
                c = b * 1
1111
             with paddle.static.name_scope("s3"):
T
Tao Luo 已提交
1112
                d = c / 1
1113 1114 1115
          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
1116 1117 1118
                g = f - 1

          # Op are created in the default main program.  
1119
          for op in paddle.static.default_main_program().block(0).ops:
T
Tao Luo 已提交
1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134
              # 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/'
1135 1136
    """
    # TODO(panyx0718): Only [0-9a-z].
1137
    # in dygraph we don't need namescope since it will cause mem leak
J
Jiabin Yang 已提交
1138
    if _non_static_mode():
L
Leo Chen 已提交
1139 1140
        yield
    else:
T
tianshuo78520a 已提交
1141
        assert prefix, "namescope prefix can not be empty."
1142 1143
        global _name_scope
        _name_scope = _name_scope.child(prefix)
1144 1145 1146 1147
        try:
            yield
        finally:
            _name_scope = _name_scope.parent()
1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159


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
1160 1161 1162
def generate_control_dev_var_name():
    import random
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
1163 1164 1165 1166


def grad_var_name(var_name):
    """
1167 1168
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
1169 1170 1171
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
1172

1173
def convert_np_dtype_to_dtype_(np_dtype):
1174 1175
    """
    Convert the data type in numpy to the data type in Paddle
1176

1177
    Args:
1178
        np_dtype(np.dtype): the data type in numpy.
1179

1180 1181
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
1182 1183

    """
1184 1185
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
1186
        return core.VarDesc.VarType.FP32
1187
    elif dtype == np.float64:
1188
        return core.VarDesc.VarType.FP64
1189
    elif dtype == np.float16:
1190
        return core.VarDesc.VarType.FP16
1191
    elif dtype == np.int32:
1192
        return core.VarDesc.VarType.INT32
1193
    elif dtype == np.int16:
1194
        return core.VarDesc.VarType.INT16
1195
    elif dtype == np.int64:
1196
        return core.VarDesc.VarType.INT64
1197
    elif dtype == np.bool_:
1198
        return core.VarDesc.VarType.BOOL
1199
    elif dtype == np.uint16:
1200 1201 1202
        # since there is still no support for bfloat16 in NumPy,
        # uint16 is used for casting bfloat16
        return core.VarDesc.VarType.BF16
1203 1204
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
1205 1206
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
1207 1208 1209 1210
    elif dtype == np.complex64:
        return core.VarDesc.VarType.COMPLEX64
    elif dtype == np.complex128:
        return core.VarDesc.VarType.COMPLEX128
1211
    else:
M
minqiyang 已提交
1212
        raise ValueError("Not supported numpy dtype %s" % dtype)
1213 1214 1215


def dtype_is_floating(dtype):
1216 1217 1218
    """
    Check the data type is floating or not.
    Args:
1219
        dtype(np.dtype|core.VarDesc.VarType): data type.
1220 1221 1222 1223 1224
            Could be numpy format or Paddle format

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

    """
1225
    if not isinstance(dtype, core.VarDesc.VarType):
1226 1227
        dtype = convert_np_dtype_to_dtype_(dtype)

1228 1229 1230 1231
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
1232 1233


Y
Yang Yang(Tony) 已提交
1234
def _debug_string_(proto, throw_on_error=True):
1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245
    """
    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
1246
    error_fields = list()
Y
Yang Yang(Tony) 已提交
1247
    if not proto.IsInitialized(error_fields) and throw_on_error:
1248 1249 1250
        raise ValueError(
            "{0} are not initialized.\nThe message is {1}:\n".format(
                error_fields, proto))
Y
Yu Yang 已提交
1251 1252 1253
    return proto.__str__()


1254 1255 1256 1257 1258 1259 1260 1261 1262 1263
def _varbase_creator(type=core.VarDesc.VarType.LOD_TENSOR,
                     name=None,
                     shape=None,
                     dtype=None,
                     persistable=None,
                     **kwargs):
    if dtype is not None:
        if not isinstance(dtype, core.VarDesc.VarType):
            dtype = convert_np_dtype_to_dtype_(dtype)

J
Jiabin Yang 已提交
1264
    if _in_eager_mode_:
1265
        eager_tensor = core.eager.Tensor(
1266
            dtype if dtype else core.VarDesc.VarType.FP32,
1267 1268 1269
            list(shape) if shape else [], name,
            type if type else core.VarDesc.VarType.LOD_TENSOR,
            True if persistable else False)
1270 1271
        eager_tensor.retain_grads()
        return eager_tensor
J
Jiabin Yang 已提交
1272 1273
    else:
        return core.VarBase(dtype if dtype else core.VarDesc.VarType.FP32,
1274 1275 1276
                            list(shape) if shape else [], name,
                            type if type else core.VarDesc.VarType.LOD_TENSOR,
                            True if persistable else False)
1277 1278


1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289
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))
    if not vals: return False
    return all(isinstance(v, expected_type) for v in vals)


1290
class VariableMetaClass(type):
1291

1292 1293 1294 1295
    @classmethod
    def __instancecheck__(cls, instance):
        t = type(instance)
        if in_dygraph_mode():
J
Jiabin Yang 已提交
1296
            return issubclass(t, core.eager.Tensor)
1297
        else:
J
Jiabin Yang 已提交
1298 1299
            if _in_legacy_dygraph():
                return issubclass(t, core.VarBase)
1300 1301 1302 1303
            return issubclass(t, Variable)


class ParameterMetaClass(VariableMetaClass):
1304

1305 1306 1307 1308
    @classmethod
    def __instancecheck__(cls, instance):
        t = type(instance)
        if in_dygraph_mode():
J
Jiabin Yang 已提交
1309
            return issubclass(t, EagerParamBase)
1310
        else:
J
Jiabin Yang 已提交
1311 1312
            if _in_legacy_dygraph():
                return issubclass(t, ParamBase)
1313 1314 1315 1316
            return issubclass(t, Parameter)


@six.add_metaclass(VariableMetaClass)
X
Xin Pan 已提交
1317
class Variable(object):
1318
    """
J
Jiabin Yang 已提交
1319
    **Notes**:
1320
        **The constructor of Variable should not be invoked directly.**
J
Jiabin Yang 已提交
1321

1322 1323
        **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
1324 1325 1326
        **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
1327
    cases, variables are used for holding different kinds of data or training
J
Jiabin Yang 已提交
1328 1329
    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.
1330

1331
    There are many kinds of variables. Each kind of them has its own attributes
J
Jiabin Yang 已提交
1332
    and usages. Please refer to the `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_ for details.
1333

T
tianshuo78520a 已提交
1334
    Most of a Variable's member variables can be set to be None. It mean
1335
    it is not available or will be specified later.
1336

1337
    Examples:
1338 1339
        In Static Graph Mode:

1340 1341
        .. code-block:: python

1342
            import paddle.fluid as fluid
1343
            cur_program = fluid.Program()
1344 1345 1346 1347
            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
1348

J
Jiabin Yang 已提交
1349
        In `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_  Mode:
1350 1351 1352 1353 1354 1355 1356 1357 1358

        .. 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))

1359 1360
    """

Y
Yu Yang 已提交
1361 1362
    def __init__(self,
                 block,
Y
Yu Yang 已提交
1363
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
1364 1365 1366 1367
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
1368
                 capacity=None,
Q
QI JUN 已提交
1369
                 persistable=None,
F
fengjiayi 已提交
1370
                 error_clip=None,
Y
Yu Yang 已提交
1371
                 stop_gradient=False,
F
fengjiayi 已提交
1372
                 is_data=False,
H
Huihuang Zheng 已提交
1373
                 need_check_feed=False,
H
hong 已提交
1374
                 belong_to_optimizer=False,
Y
Yu Yang 已提交
1375
                 **kwargs):
Y
Yu Yang 已提交
1376 1377
        self.block = block
        if name is None:
Y
Yu Yang 已提交
1378
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
1379

Y
Yu Yang 已提交
1380
        if dtype is not None:
1381
            if not isinstance(dtype, core.VarDesc.VarType):
1382
                dtype = convert_np_dtype_to_dtype_(dtype)
1383

S
Steffy-zxf 已提交
1384 1385 1386 1387
        if dtype == core.VarDesc.VarType.STRINGS:
            type = core.VarDesc.VarType.STRINGS
            lod_level = None

1388 1389 1390
        if type == core.VarDesc.VarType.SPARSE_COO:
            lod_level = None

H
hong 已提交
1391 1392
        self.belong_to_optimizer = belong_to_optimizer

1393 1394 1395 1396 1397
        self.error_clip = error_clip

        is_new_var = False
        name = cpt.to_text(name)
        self.desc = self.block.desc.find_var(cpt.to_bytes(name))
1398

1399 1400 1401
        if self.desc is None:
            self.desc = self.block.desc.var(cpt.to_bytes(name))
            is_new_var = True
1402

1403 1404 1405
        if is_new_var:
            self.desc.set_type(type)
        elif self.desc.type() != type:
L
Leo Chen 已提交
1406 1407
            raise ValueError("Variable '{0}' has been created before. The "
                             "previous type is {1}, the new type is {2}. They"
1408 1409
                             " are not matched".format(self.name,
                                                       self.desc.type(), type))
1410

1411
        if shape is not None:
1412
            if is_new_var:
1413 1414 1415 1416 1417 1418
                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
1419 1420
                        "Variable '{0}' has been created before. The previous "
                        "shape is {1}, the new shape is {2}. They are not "
1421 1422 1423 1424 1425 1426 1427
                        "matched.".format(self.name, old_shape, shape))
        if dtype is not None:
            if is_new_var:
                self.desc.set_dtype(dtype)
            else:
                old_dtype = self.dtype
                if dtype != old_dtype:
L
Leo Chen 已提交
1428 1429
                    raise ValueError("Variable '{0}' has been created before. "
                                     "The previous data type is {1}, the new "
1430 1431 1432 1433 1434 1435 1436 1437 1438
                                     "data type is {2}. They are not "
                                     "matched.".format(self.name, old_dtype,
                                                       dtype))

        if lod_level is not None:
            if is_new_var:
                self.desc.set_lod_level(lod_level)
            else:
                if lod_level != self.lod_level:
L
Leo Chen 已提交
1439 1440
                    raise ValueError("Variable '{0}' has been created before. "
                                     "The previous lod_level is {1}, the new "
1441 1442 1443 1444 1445 1446 1447 1448 1449
                                     "lod_level is {2}. They are not "
                                     "matched".format(self.name, self.lod_level,
                                                      lod_level))
        if persistable is not None:
            if is_new_var:
                self.desc.set_persistable(persistable)
            else:
                if persistable != self.persistable:
                    raise ValueError(
L
Leo Chen 已提交
1450 1451
                        "Variable '{0}' has been created before."
                        "The previous persistable is {1}, the new "
1452 1453
                        "persistable is {2}. They are not matched".format(
                            self.name, self.persistable, persistable))
1454

1455 1456
        if need_check_feed and is_new_var:
            self.desc.set_need_check_feed(need_check_feed)
H
Huihuang Zheng 已提交
1457

1458 1459 1460 1461 1462 1463 1464
        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
1465

1466 1467
        self.block.vars[name] = self
        self.op = None
1468
        self.stop_gradient = stop_gradient
1469
        self.is_data = is_data
Y
Yu Yang 已提交
1470

1471 1472 1473
    def detach(self):
        """
        Returns a new Variable, detached from the current graph.
1474 1475
        It will share data with origin Variable and without tensor copy.
        In addition, the detached Variable doesn't provide gradient propagation.
1476

1477
        Returns:
J
Jiabin Yang 已提交
1478
             ( :ref:`api_guide_Variable_en` | dtype is same as current Variable): The detached Variable.
1479 1480 1481 1482

        Examples:
            .. code-block:: python

1483
                import paddle
1484

1485 1486 1487 1488
                paddle.enable_static()

                # create a static Variable
                x = paddle.static.data(name='x', shape=[3, 2, 1])
1489

1490 1491
                # create a detached Variable
                y = x.detach()
1492
        """
1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504

        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"

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

1505 1506 1507
        self.block.append_op(type='share_data',
                             inputs={'X': [self]},
                             outputs={'Out': [output]})
1508
        return output
1509

1510
    @fake_interface_only
1511
    def numpy(self):
1512
        """
J
Jiabin Yang 已提交
1513
        **Notes**:
T
tianshuo78520a 已提交
1514
            **This API is ONLY available in Dygraph mode**
1515

J
Jiabin Yang 已提交
1516
        Returns a numpy array shows the value of current :ref:`api_guide_Variable_en`
1517 1518 1519 1520 1521

        Returns:
            ndarray: The numpy value of current Variable.

        Returns type:
J
Jiabin Yang 已提交
1522
            ndarray: dtype is same as current Variable
1523 1524 1525 1526 1527 1528

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1529
                from paddle.fluid.dygraph import Linear
1530 1531 1532 1533
                import numpy as np

                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
                with fluid.dygraph.guard():
1534
                    linear = Linear(32, 64)
1535
                    data = to_variable(data)
1536
                    x = linear(data)
1537 1538 1539
                    print(x.numpy())

        """
1540
        pass
1541

1542
    @fake_interface_only
1543
    def backward(self, retain_graph=False):
1544
        """
J
Jiabin Yang 已提交
1545
        **Notes**:
T
tianshuo78520a 已提交
1546
            **This API is ONLY available in Dygraph mode**
1547

1548
        Run backward of current Graph which starts from current Tensor.
1549

J
Jiabin Yang 已提交
1550
        Args:
1551 1552 1553 1554
            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.
1555

J
Jiabin Yang 已提交
1556 1557
        Returns:
            NoneType: None
1558 1559 1560 1561 1562

        Examples:
            .. code-block:: python

                import numpy as np
1563 1564
                import paddle
                paddle.disable_static()
1565 1566

                x = np.ones([2, 2], np.float32)
1567 1568 1569 1570 1571 1572 1573
                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)
1574 1575
                ret = paddle.add_n(inputs)
                loss = paddle.sum(ret)
1576
                loss.backward()
1577 1578

        """
1579
        pass
1580

1581
    @fake_interface_only
1582
    def gradient(self):
1583
        """
J
Jiabin Yang 已提交
1584
        **Notes**:
T
tianshuo78520a 已提交
1585
            **This API is ONLY available in Dygraph mode**
1586 1587 1588

        Get the Gradient of Current Variable

J
Jiabin Yang 已提交
1589
        Returns:
1590
            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.
1591 1592 1593 1594 1595 1596 1597

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

1598
                # example1: return ndarray
1599 1600 1601 1602 1603 1604 1605 1606 1607
                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)
1608
                    loss2.backward()
1609 1610
                    print(loss2.gradient())

1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623
                # 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())

1624
        """
1625
        pass
1626

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

            **2. Use it only Variable has gradient, normally we use this for Parameters since other temporal Variable will be deleted by Python's GC**
1634

J
Jiabin Yang 已提交
1635
        Clear  (set to ``0`` ) the Gradient of Current Variable
1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653

        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)
1654
                    loss2.backward()
1655 1656 1657 1658 1659
                    print(loss2.gradient())
                    loss2.clear_gradient()
                    print("After clear {}".format(loss2.gradient()))

        """
1660
        pass
X
Xin Pan 已提交
1661

1662 1663 1664 1665
    @fake_interface_only
    def register_hook(self, hook):
        pass

1666
    def __str__(self):
1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682
        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

1683 1684
                import paddle
                import paddle.static as static
1685

1686 1687 1688
                paddle.enable_static()

                cur_program = static.Program()
1689 1690 1691 1692 1693 1694
                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())
        """
1695 1696
        # VarType.LOD_TENSOR -> LOD_TENSOR
        type_str = str(self.type).split('.')[1]
1697
        if self.type == core.VarDesc.VarType.SELECTED_ROWS or self.type == core.VarDesc.VarType.LOD_TENSOR:
1698 1699
            dtype_str = str(self.dtype).split('.')[1]
            var_str = "{name} : {type}.shape{shape}.dtype({dtype}).stop_gradient({stop_gradient})".\
T
tangwei12 已提交
1700 1701
                format(name=self.name, type=type_str, shape=self.shape,
                       dtype=dtype_str, stop_gradient=self.stop_gradient)
1702
        else:
1703 1704
            var_str = "{name} : {type})".\
                format(name=self.name, type=type_str)
1705

1706
        if self.is_parameter:
1707 1708 1709 1710 1711 1712 1713 1714 1715 1716
            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

1717
        from paddle.distributed.auto_parallel.dist_context import get_default_distributed_context
1718
        dist_context = get_default_distributed_context()
1719 1720
        dist_tensor = dist_context.get_dist_tensor_for_program(self)
        if dist_tensor is not None:
1721 1722
            var_str += ", {name} = {value}".format(name="dist_attr",
                                                   value=dist_tensor)
1723

1724
        return var_str
Y
Yang Yang(Tony) 已提交
1725

F
update  
fengjiayi 已提交
1726
    def to_string(self, throw_on_error, with_details=False):
1727 1728 1729
        """
        Get debug string.

J
Jiabin Yang 已提交
1730 1731 1732 1733 1734
        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;
1735

1736 1737
        Returns:
            str: The debug string.
1738 1739 1740 1741 1742

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
1743
                import paddle
1744

1745
                paddle.enable_static()
1746 1747 1748 1749 1750
                cur_program = fluid.Program()
                cur_block = cur_program.current_block()
                new_variable = cur_block.create_var(name="X",
                                                    shape=[-1, 23, 48],
                                                    dtype='float32')
1751
                print(new_variable.to_string(True))
J
Jiabin Yang 已提交
1752
                print("=============with detail===============")
1753
                print(new_variable.to_string(True, True))
1754
        """
1755 1756
        assert isinstance(throw_on_error, bool) and isinstance(
            with_details, bool)
1757
        protostr = self.desc.serialize_to_string()
1758
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
1759 1760
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
1761
            additional_attr = ("error_clip", )
F
update  
fengjiayi 已提交
1762
            for attr_name in additional_attr:
1763 1764 1765
                res_str += "%s: %s\n" % (attr_name,
                                         cpt.to_text(getattr(self, attr_name)))

F
update  
fengjiayi 已提交
1766
        return res_str
1767 1768 1769

    __repr__ = __str__

1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796
    def element_size(self):
        """
        Returns the size in bytes of an element in the Tensor.
        
        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()

1797
    @property
1798
    def stop_gradient(self):
J
Jiabin Yang 已提交
1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813
        """
        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")
1814 1815
                linear = fluid.Linear(13, 5, dtype="float32")
                linear2 = fluid.Linear(3, 3, dtype="float32")
J
Jiabin Yang 已提交
1816 1817 1818
                a = fluid.dygraph.to_variable(value0)
                b = fluid.dygraph.to_variable(value1)
                c = fluid.dygraph.to_variable(value2)
1819 1820
                out1 = linear(a)
                out2 = linear2(b)
J
Jiabin Yang 已提交
1821 1822 1823 1824
                out1.stop_gradient = True
                out = fluid.layers.concat(input=[out1, out2, c], axis=1)
                out.backward()

1825
                assert linear.weight.gradient() is None
J
Jiabin Yang 已提交
1826 1827
                assert (out1.gradient() == 0).all()
        """
1828
        return self.desc.stop_gradient()
1829

1830 1831
    @stop_gradient.setter
    def stop_gradient(self, s):
1832
        self.desc.set_stop_gradient(s)
1833

1834 1835
    @property
    def persistable(self):
J
Jiabin Yang 已提交
1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856
        """
        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))
        """
1857
        return self.desc.persistable()
1858

Y
Yu Yang 已提交
1859 1860
    @persistable.setter
    def persistable(self, p):
1861
        self.desc.set_persistable(p)
Y
Yu Yang 已提交
1862

1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887
    @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
1888 1889
    @property
    def name(self):
J
Jiabin Yang 已提交
1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905
        """
        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))
        """
1906
        return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
1907

1908 1909 1910 1911 1912 1913
    @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
1914 1915
        gradient Variable from a naming convention but doesn't guarantee
        the gradient exists.**
T
tangwei12 已提交
1916

1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927
        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
1928 1929
    @name.setter
    def name(self, new_name):
1930
        self.desc.set_name(new_name)
T
typhoonzero 已提交
1931

Y
Yu Yang 已提交
1932 1933
    @property
    def shape(self):
J
Jiabin Yang 已提交
1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950
        """
        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
1951
        # convert to tuple, make it as same as numpy API.
1952
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
1953 1954

    @property
F
fengjiayi 已提交
1955
    def dtype(self):
J
Jiabin Yang 已提交
1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971
        """
        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))
        """
1972
        return self.desc.dtype()
Y
Yu Yang 已提交
1973 1974 1975

    @property
    def lod_level(self):
J
Jiabin Yang 已提交
1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988
        """
        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

1989
            import paddle
J
Jiabin Yang 已提交
1990
            import paddle.fluid as fluid
1991 1992

            paddle.enable_static()
J
Jiabin Yang 已提交
1993 1994 1995 1996 1997 1998 1999
            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))
        """
2000 2001
        if self.type == core.VarDesc.VarType.SELECTED_ROWS:
            raise Exception("SelectedRows DO NOT supprt lod")
2002 2003
        if self.type == core.VarDesc.VarType.STRINGS:
            return None
2004
        return self.desc.lod_level()
Y
Yu Yang 已提交
2005

Y
Yu Yang 已提交
2006 2007
    @property
    def type(self):
J
Jiabin Yang 已提交
2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023
        """
        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))
        """
2024
        return self.desc.type()
Y
Yu Yang 已提交
2025

2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066
    @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,
            stop_gradient=False)
        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,
            stop_gradient=False)

2067 2068 2069 2070 2071 2072 2073
        self.block.append_op(type='transpose2',
                             inputs={'X': [self]},
                             outputs={
                                 'Out': [out],
                                 'XShape': [input_shape]
                             },
                             attrs={'axis': perm})
2074 2075
        return out

2076 2077 2078 2079 2080 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
    def clone(self):
        """
        Returns a new static Variable, which is the clone of the original static
        Variable. It remains in the current graph, that is, the cloned Variable 
        provides gradient propagation. Calling ``out = tensor.clone()`` is same
        as ``out = assign(tensor)`` .

        Returns:
            Variable: The cloned Variable.

        Examples:
            .. code-block:: python

                import paddle

                paddle.enable_static()

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

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

2106 2107 2108
        self.block.append_op(type='assign',
                             inputs={'X': [self]},
                             outputs={'Out': [output]})
2109 2110
        return output

W
Wu Yi 已提交
2111
    def _set_error_clip(self, error_clip):
2112 2113 2114 2115 2116 2117 2118 2119 2120
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
2121 2122
        self.error_clip = error_clip

2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151
    def _set_info(self, key, value):
        """
        Set key-value information for this variable.

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

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

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

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

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

2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162
    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
2163
            raise ValueError("slice step can not be zero")
2164 2165 2166 2167 2168 2169 2170 2171 2172 2173

        # 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
2174 2175
            start = max(start +
                        length, lower) if start < 0 else min(start, upper)
2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238

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

        return start, stop, step

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

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

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

L
lujun 已提交
2239
    def _cloneVar(self, copy=False):
2240 2241
        if not copy:
            return self.block.create_var(
H
Hongyu Liu 已提交
2242 2243
                name=unique_name.generate_with_ignorable_key(self.name),
                dtype=self.dtype)
2244 2245 2246 2247
        else:
            return self

    def _sliceVar(self, axes, starts, ends):
L
lujun 已提交
2248
        new_var = self._cloneVar()
2249 2250 2251 2252 2253 2254 2255 2256
        self.block.append_op(type="slice",
                             inputs={'Input': [self]},
                             outputs={'Out': [new_var]},
                             attrs={
                                 'axes': axes,
                                 'starts': starts,
                                 'ends': ends
                             })
2257 2258 2259
        return new_var

    def _concatVar(self, inputs, axis):
L
lujun 已提交
2260
        new_var = self._cloneVar()
2261 2262 2263 2264 2265 2266
        self.block.append_op(type="concat",
                             inputs={'X': inputs},
                             outputs={'Out': [new_var]},
                             attrs={
                                 'axis': axis,
                             })
2267 2268 2269 2270 2271
        return new_var

    def _sliceAndConcatVar(self, item, axis):
        if isinstance(item, slice):
            if self.shape[axis] < 0:
L
lujun 已提交
2272
                return self._cloneVar(True)
2273 2274 2275 2276 2277 2278 2279
            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:
2280 2281
                        vars.append(self._sliceVar([axis], [start],
                                                   [start + 1]))
2282 2283 2284
                        start += step
                else:
                    while start > stop:
2285 2286
                        vars.append(self._sliceVar([axis], [start],
                                                   [start + 1]))
2287 2288 2289 2290
                        start += step
                return self._concatVar(vars, axis)
        elif isinstance(item, int):
            if self.shape[axis] < 0:
L
lujun 已提交
2291
                return self._cloneVar(True)
2292
            index = int(item)
2293
            if (index > 0 and index >= self.shape[axis]) \
2294 2295 2296 2297 2298 2299 2300
                    or (index < 0 and (index + self.shape[axis]) < 0):
                raise IndexError("invalid index")
            return self._sliceVar([axis], [index], [index + 1])
        else:
            raise IndexError("Valid index accept int or slice or tuple")

    def __getitem__(self, item):
2301
        return _getitem_impl_(self, item)
2302

2303
    def __setitem__(self, item, value):
2304
        return _setitem_impl_(self, item, value)
2305

2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346
    def get_value(self, scope=None):
        """
        Get the value of variable in given scope. 

        Args:
            scope(Scope, optional) : If `scope` is None, it will be set to global scope 
                obtained through 'paddle.static.global_scope()'. Otherwise, use `scope`.
                Default: None

        Returns:
            Tensor: the value in given scope.

        Examples:
            .. code-block:: python

                import paddle
                import paddle.static as static 
                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)
        """
2347 2348
        # The 'framework' is a low-level module, and 'executor'
        # can not be imported at the begainning of this file.
2349 2350 2351 2352
        # Therefore, the above two modules are dynamically imported.
        from .executor import global_scope
        if scope is not None and not isinstance(scope, core._Scope):
            raise TypeError(
2353 2354
                "`scope` should be None or `paddle.static.Scope` type, but received {}."
                .format(type(scope)))
2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408

        if scope is None:
            scope = global_scope()
        var_temp = scope.find_var(self.name)
        if var_temp is None:
            raise ValueError("Can not find Variable '{}' in the Scope.".format(
                self.name))
        t = var_temp.get_tensor()
        return t

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

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

        Returns:
            None
        
        Examples:
            .. code-block:: python

                import paddle
                import paddle.static as static 
                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'
2409
        # can not be imported at the begainning of this file.
2410 2411 2412 2413 2414
        # 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(
2415 2416
                "`value` should be `numpy.ndarray` or `LoDTensor`, but received {}."
                .format(type(value)))
2417 2418 2419

        if scope is not None and not isinstance(scope, core._Scope):
            raise TypeError(
2420 2421
                "`scope` should be None or `paddle.static.Scope` type, but received {}."
                .format(type(scope)))
2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451

        if scope is None:
            scope = global_scope()

        var_temp = scope.find_var(self.name)
        if var_temp is None:
            raise ValueError("Can not find Variable '{}' in the Scope.".format(
                self.name))

        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(
                    "{} expected a shape {}, but the received shape is {}.".
                    format(self.name, list(t.shape()), list(value_shape)))

        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())
2452 2453 2454 2455
        elif p.is_npu_place():
            p = core.Place()
            p.set_place(t._place())
            place = core.NPUPlace(p.npu_device_id())
2456 2457 2458 2459
        elif p.is_mlu_place():
            p = core.Place()
            p.set_place(t._place())
            place = core.MLUPlace(p.mlu_device_id())
2460 2461 2462 2463 2464 2465 2466
        else:
            p = core.Place()
            p.set_place(t._place())
            place = core.CUDAPlace(p.gpu_device_id())

        t.set(value, place)

2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491
    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"),
            dtype=core.VarDesc.VarType.INT64)

2492 2493 2494
        self.block.append_op(type='size',
                             inputs={'Input': [self]},
                             outputs={'Out': [output]})
2495 2496
        return output

2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550
    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
2551
    def dist_attr(self):
2552
        """
2553
        Get distributed attribute of this Variable.
2554
        """
2555
        return self.desc.dist_attr
2556

2557 2558
    @dist_attr.setter
    def dist_attr(self, dist_attr):
2559
        """
2560
        Set distributed attribute of this Variable.
2561
        """
2562
        self.desc.dist_attr = dist_attr
2563

Y
Yu Yang 已提交
2564

F
fengjiayi 已提交
2565 2566 2567
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
2568

2569 2570
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
2571 2572 2573 2574
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
2575
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
2576 2577 2578 2579 2580
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
2581 2582 2583 2584
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
2585 2586 2587 2588 2589 2590 2591 2592 2593
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
2594
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
2595 2596 2597 2598 2599 2600
        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):
2601 2602 2603 2604 2605 2606 2607 2608
        """
        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
2609 2610
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
2611 2612
        return self.op_proto_map[type]

2613 2614
    def update_op_proto(self):
        op_protos = get_all_op_protos()
2615
        custom_op_names = []
2616 2617 2618
        for proto in op_protos:
            if proto.type not in self.op_proto_map:
                self.op_proto_map[proto.type] = proto
2619 2620 2621
                custom_op_names.append(proto.type)

        return custom_op_names
2622

2623 2624 2625 2626
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
2627
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
2628
            core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
2629 2630
            core.op_proto_and_checker_maker.kOpCreationCallstackAttrName(),
            core.op_proto_and_checker_maker.kOpDeviceAttrName()
2631 2632
        }

F
fengjiayi 已提交
2633

X
Xin Pan 已提交
2634
class Operator(object):
2635
    """
2636 2637 2638 2639 2640 2641 2642
    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
2643
        type(str): The type of operator. Default None.
2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663
        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
2664
        Block.append_op or Block._prepend_op instead.
2665 2666 2667 2668

    Examples:
        .. code-block:: python

2669
            import paddle.fluid as fluid
2670
            cur_program = fluid.Program()
2671 2672 2673 2674 2675
            cur_block = cur_program.current_block()
            # var1 += var2 + var3
            cur_block.append_op(type="sum",
                                inputs={"X": [var1, var2, var3]},
                                outputs={"Out": [var1]})
2676
    """
2677
    OP_WITHOUT_KERNEL_SET = {
2678 2679
        'feed', 'fetch', 'recurrent', 'go', 'rnn_memory_helper_grad',
        'conditional_block', 'while', 'send', 'recv', 'listen_and_serv',
2680
        'fl_listen_and_serv', 'ncclInit', 'select', 'checkpoint_notify',
2681 2682
        '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',
W
WangXi 已提交
2683
        'queue_generator', 'dequeue', 'enqueue', 'heter_listen_and_serv',
B
Baibaifan 已提交
2684
        'c_wait_comm', 'c_wait_compute', 'c_gen_hccl_id', 'c_comm_init_hccl',
2685
        'copy_cross_scope', 'c_gen_cncl_id'
2686
    }
2687

Y
Yu Yang 已提交
2688 2689
    def __init__(self,
                 block,
Y
Yu Yang 已提交
2690
                 desc,
Y
Yu Yang 已提交
2691 2692 2693
                 type=None,
                 inputs=None,
                 outputs=None,
M
minqiyang 已提交
2694
                 attrs=None):
2695 2696 2697 2698 2699 2700 2701 2702 2703 2704
        # 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
2705
        if _non_static_mode():
2706 2707
            if type is None:
                raise ValueError(
2708
                    "`type` to initialized an Operator can not be None.")
J
Jiabin Yang 已提交
2709
            self._type = type
M
minqiyang 已提交
2710
            self.attrs = attrs if attrs else {}
2711 2712 2713 2714 2715 2716 2717 2718 2719 2720
        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

2721 2722 2723
            # attr for static mode cuda graph
            self._cuda_graph_attr = _current_cuda_graph_mode

2724 2725 2726
            op_maker = core.op_proto_and_checker_maker

            if op_maker.kOpRoleAttrName() not in op_attrs:
2727 2728
                op_attrs[
                    op_maker.kOpRoleAttrName()] = self.block.program._op_role
2729 2730

            role_var_name = op_maker.kOpRoleVarAttrName()
2731 2732
            if len(self.block.program._op_role_var
                   ) != 0 and role_var_name not in op_attrs:
2733
                op_attrs[role_var_name] = self.block.program._op_role_var
2734 2735 2736 2737 2738

            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:
2739 2740 2741 2742 2743
                # 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
2744 2745 2746
                return
            if type is None:
                raise ValueError(
2747
                    "`type` to initialized an Operator can not be None.")
2748 2749
            else:
                callstack_var_name = op_maker.kOpCreationCallstackAttrName()
2750 2751 2752
                op_attrs[callstack_var_name] = []
                for frame in traceback.extract_stack():
                    op_attrs[callstack_var_name].append(
2753 2754 2755 2756
                        '  File "{}", line {}, in {}'.format(
                            frame[0], frame[1], frame[2]))
                    op_attrs[callstack_var_name].append('    {}'.format(
                        frame[3]))
2757 2758 2759 2760 2761 2762 2763

            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()

2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774
            # set device for op with kernels, give warning for op without kernels
            # when force_cpu and device_guard are used at the same time, a warning will be given.
            # TODO(zhangting2020): when force_cpu is removed, clear warning below.
            if _current_device is not None:
                if self._has_kernel(type):
                    op_device = op_maker.kOpDeviceAttrName()
                    op_attrs[op_device] = _current_device
                else:
                    warnings.warn("The Op(%s) is not support to set device." %
                                  type)
                if 'force_cpu' in op_attrs:
2775
                    if (type == 'less_than' and op_attrs['force_cpu'] != None
2776 2777 2778 2779 2780
                        ) or op_attrs['force_cpu'] != False:
                        warnings.warn(
                            "The Attr(force_cpu) of Op(%s) will be deprecated in the future, "
                            "please use 'device_guard' instead. 'device_guard' has higher priority when they are "
                            "used at the same time." % type)
2781 2782 2783 2784 2785
            if _current_pipeline_stage is not None:
                pipeline_attr_name = 'pipeline_stage' + core.kAutoParallelSuffix(
                )
                self._update_desc_attr(pipeline_attr_name,
                                       _current_pipeline_stage)
2786

2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799
            def find_name(var_list, name):
                for var_name in var_list:
                    if var_list[var_name] is not None and var_name == name:
                        return True
                return False

            if inputs is not None:
                for in_proto in proto.inputs:
                    found = find_name(inputs, in_proto.name)
                    assert found or in_proto.dispensable, "Input {} not found".format(
                        in_proto.name)
                    if found:
                        in_args = inputs[in_proto.name]
2800
                        if not isinstance(in_args, (list, tuple)):
2801 2802 2803 2804 2805 2806
                            in_args = [in_args]
                        if not in_proto.duplicable and len(in_args) > 1:
                            raise ValueError(
                                "Input %s expects only one input, but %d are given."
                                % (in_proto.name, len(in_args)))
                        in_arg_names = []
2807
                        for index, arg in enumerate(in_args):
2808 2809 2810 2811
                            if isinstance(arg, six.string_types):
                                in_arg_names.append(arg)
                            elif isinstance(arg, six.binary_type):
                                in_arg_names.append(arg.decode())
2812
                            elif isinstance(arg, (Variable, core.VarBase)):
2813
                                in_arg_names.append(cpt.to_text(arg.name))
2814
                            else:
2815 2816 2817 2818
                                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."
2819 2820
                                    "but received : %s" %
                                    (in_proto.name, type, arg))
2821 2822 2823 2824 2825 2826 2827 2828 2829
                        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):
2830 2831 2832 2833
                        raise ValueError(
                            ("Incorrect setting for output(s) of "
                             "operator \"%s\", should set: [%s].") %
                            (type, m.name))
2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845
                for out_proto in proto.outputs:
                    if out_proto.name not in outputs:
                        continue
                    out_args = outputs[out_proto.name]
                    if not isinstance(out_args, list):
                        out_args = [out_args]
                    if not out_proto.duplicable and len(out_args) > 1:
                        raise ValueError(
                            "Output %s expects only one output, but %d are given."
                            % (out_proto.name, len(out_args)))
                    out_arg_names = []
                    for arg in out_args:
2846 2847 2848 2849
                        if isinstance(arg, six.string_types):
                            out_arg_names.append(arg)
                        else:
                            out_arg_names.append(cpt.to_text(arg.name))
2850
                        # TODO(minqiyang): could we remove variable's op in static mode?
J
Jiabin Yang 已提交
2851
                        if not _non_static_mode():
2852 2853 2854 2855
                            if isinstance(arg, six.string_types):
                                block.var(arg).op = self
                            else:
                                arg.op = self
2856 2857
                    self.desc.set_output(out_proto.name, out_arg_names)

2858
            extra_attrs_map = core.get_op_extra_attrs(type)
2859 2860 2861 2862 2863
            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
2864 2865
                    if (attr_name
                            not in op_attrs) or (op_attrs[attr_name] is None):
2866 2867 2868
                        continue
                    attr_val = op_attrs[attr_name]
                    self._update_desc_attr(attr_name, attr_val)
2869 2870 2871 2872 2873 2874 2875
                for attr_name in extra_attrs_map.keys():
                    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])
                    else:
                        self._update_desc_attr(attr_name, op_attrs[attr_name])
2876

J
jianghaicheng 已提交
2877 2878
            # proto.attrs doesn't include ipu_index
            if core.is_compiled_with_ipu():
2879
                if global_ipu_index >= 0:
J
jianghaicheng 已提交
2880 2881
                    self._update_desc_attr(ipu_index_attr_name,
                                           global_ipu_index)
2882
                if global_ipu_stage >= 0:
J
jianghaicheng 已提交
2883 2884 2885
                    self._update_desc_attr(ipu_stage_attr_name,
                                           global_ipu_stage)

2886 2887 2888 2889 2890
            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
2891
    def _has_kernel(self, op_type):
2892 2893
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
2894
    def to_string(self, throw_on_error):
2895
        """
2896 2897
        Get debug string.

2898
        Args:
2899 2900
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
2901

2902 2903
        Returns:
            str: The debug string.
2904 2905

        """
2906
        protostr = self.desc.serialize_to_string()
2907
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
2908 2909
        return _debug_string_(proto, throw_on_error)

2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941
    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
2942
        ), "skip_op_callstack parameter's type is error, expect bool, received {}".format(
2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969
            type(skip_op_callstack))
        outputs_str = "{"
        for i in range(0, len(self.output_names)):
            outputs_str += "{name}=".format(name=self.output_names[i])
            o = self.output(self.output_names[i])
            outputs_str += "{value}".format(value=o)
            if i != len(self.output_names) - 1:
                outputs_str += ", "
        outputs_str += "}"

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

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

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

2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991
            attr_type = self.desc.attr_type(name, True)
            if attr_type == core.AttrType.VAR:
                attr_var_name = self.desc.attr(name, True).name()
                a = "{name} = Var['{value}']".format(name=name,
                                                     type=attr_type,
                                                     value=attr_var_name)
                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(
                    name=name, type=attr_type, value=','.join(attr_var_names))
                attrs_str += a
                if i != len(attr_names) - 1:
                    attrs_str += ", "
                continue

2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009
            if attr_type == core.AttrType.BLOCK:
                a = "{name} = block[{value}]".format(
                    name=name, type=attr_type, value=self._block_attr_id(name))
                attrs_str += a
                if i != len(attr_names) - 1:
                    attrs_str += ", "
                continue

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

3010
            # it is bytes of serialized protobuf
3011 3012 3013 3014
            if is_compiled_with_cinn(
            ) and self.type == 'cinn_launch' and name == 'compilation_key':
                key = self.desc.attr(name)
                v = core.get_serialize_comile_key(key)
3015 3016 3017 3018 3019 3020 3021 3022 3023
                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)

3024 3025 3026
            a = "{name} = {value}".format(name=name,
                                          type=attr_type,
                                          value=value)
3027

3028 3029 3030 3031
            attrs_str += a
            if i != len(attr_names) - 1:
                attrs_str += ", "

3032
        from paddle.distributed.auto_parallel.dist_context import get_default_distributed_context
3033
        dist_context = get_default_distributed_context()
3034 3035
        dist_op = dist_context.get_dist_op_for_program(self)
        if dist_op is not None:
3036 3037
            attrs_str += ", {name} = {value}".format(name="dist_attr",
                                                     value=dist_op)
3038

3039 3040
        if outputs_str != "{}":
            op_str = "{outputs} = {op_type}(inputs={inputs}, {attrs})".\
T
tangwei12 已提交
3041 3042
                format(outputs=outputs_str, op_type=self.type,
                       inputs=inputs_str, attrs=attrs_str)
3043 3044 3045 3046 3047
        else:
            op_str = "{op_type}(inputs={inputs}, {attrs})".\
                format(op_type=self.type, inputs=inputs_str, attrs=attrs_str)
        return op_str

Y
Yang Yang(Tony) 已提交
3048
    def __str__(self):
3049
        return self._to_readable_code()
3050 3051 3052

    __repr__ = __str__

F
fengjiayi 已提交
3053 3054
    @property
    def type(self):
3055
        return self.desc.type()
F
fengjiayi 已提交
3056 3057

    def input(self, name):
3058
        r"""
3059
        Get the input arguments according to the input parameter name.
3060

3061 3062
        Args:
            name(str): The input parameter name.
3063

3064 3065 3066
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
3067
        """
F
fengjiayi 已提交
3068 3069
        return self.desc.input(name)

W
Wu Yi 已提交
3070
    def _rename_input(self, old_name, new_name):
3071 3072 3073 3074 3075 3076 3077 3078 3079 3080
        """
        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
3081
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
3082

W
Wu Yi 已提交
3083
    def _rename_output(self, old_name, new_name):
3084 3085 3086 3087 3088 3089 3090 3091 3092 3093
        """
        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
3094
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
3095

F
fengjiayi 已提交
3096 3097 3098 3099
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
3100 3101 3102 3103 3104 3105 3106 3107
    @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
3108
    def output(self, name):
3109
        r"""
3110
        Get output arguments by the output parameter name.
3111

3112 3113
        Args:
            name(str): The output parameter name.
3114

3115 3116 3117
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
3118
        """
F
fengjiayi 已提交
3119 3120 3121 3122 3123 3124
        return self.desc.output(name)

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

3125 3126 3127 3128 3129 3130 3131 3132
    @property
    def idx(self):
        for i, op in enumerate(self.block.ops):
            if op == self:
                return i
        raise ValueError(
            "Can't find op itself in it's block. It could be a bug of Paddle.")

F
fengjiayi 已提交
3133
    def has_attr(self, name):
3134
        """
3135 3136
        Whether this Operator has the attribute with name or not.

3137
        Args:
3138
            name(str): the attribute name.
3139

3140 3141
        Returns:
            bool: True if has this attribute.
3142 3143

        """
F
fengjiayi 已提交
3144 3145 3146
        return self.desc.has_attr(name)

    def attr_type(self, name):
3147
        """
3148
        Get the type of attribute by attribute's name.
3149

3150 3151
        Args:
            name(str): the attribute name.
3152

3153 3154
        Returns:
            core.AttrType: the attribute type.
3155
        """
3156
        return self.desc.attr_type(name, True)
F
fengjiayi 已提交
3157

W
Wu Yi 已提交
3158
    def _set_attr(self, name, val):
3159 3160 3161 3162 3163 3164 3165 3166 3167 3168
        """
        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
3169 3170
        self._update_desc_attr(name, val)

3171 3172 3173
    def _remove_attr(self, name):
        self.desc.remove_attr(name)

G
gongweibao 已提交
3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184
    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).
        """
3185 3186 3187 3188 3189
        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
3190
            self.desc.set_block_attr(name, val.desc)
3191
        elif isinstance(val, list) and val and _all_is_type(val, Block):
3192
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
3193 3194 3195 3196
        elif isinstance(val, core.BlockDesc) or \
                isinstance(val, core.ProgramDesc):
            self.desc.set_serialized_attr(name, val.serialize_to_string())
        else:
3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232
            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
3233

F
fengjiayi 已提交
3234 3235
    @property
    def attr_names(self):
3236
        return self.desc.attr_names(True)
F
fengjiayi 已提交
3237 3238

    def attr(self, name):
3239
        """
3240 3241
        Get the attribute by name.

3242
        Args:
3243
            name(str): the attribute name.
3244

3245 3246
        Returns:
            bool|int|str|float|list: The attribute value. The return value
3247 3248
            can be any valid attribute type.
        """
F
fengjiayi 已提交
3249
        return self.desc.attr(name)
Y
Yu Yang 已提交
3250

W
Wu Yi 已提交
3251
    def _block_attr_id(self, name):
3252
        """
G
gongweibao 已提交
3253
        Get the block attribute's id by name.
3254

3255 3256
        Args:
            name(str): the attribute name.
3257

3258 3259
        Returns:
            int: the block index.
3260
        """
W
Wu Yi 已提交
3261
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
3262

W
Wu Yi 已提交
3263
    def _block_attr(self, name):
G
gongweibao 已提交
3264 3265 3266 3267 3268 3269 3270 3271 3272 3273
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
3274
        id = self._block_attr_id(name)
G
gongweibao 已提交
3275 3276 3277
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
3278
    def _blocks_attr(self, name):
G
gongweibao 已提交
3279 3280 3281 3282 3283 3284 3285 3286 3287 3288
        """
        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
3289
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
3290 3291 3292 3293 3294
            assert (i >= 0 and i < len(self.block.program.blocks))
            attrs.append(self.block.program.blocks[i])

        return attrs

W
Wu Yi 已提交
3295
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
3296 3297 3298 3299 3300 3301 3302 3303 3304 3305
        """
        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
3306
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
3307

3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342
    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)
        assert attr_type == core.AttrType.VAR, "Required type attr({}) is Variable, but received {}".format(
            name, attr_type)
        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)
        assert attr_type == core.AttrType.VARS, "Required type attr({}) is list[Variable], but received {}".format(
            name, attr_type)
        attr_vars = [
            self.block._var_recursive(var.name())
            for var in self.desc.attr(name, True)
        ]
        return attr_vars

J
JiayiFeng 已提交
3343
    def all_attrs(self):
F
fengjiayi 已提交
3344
        """
3345 3346 3347
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
3348
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
3349 3350 3351 3352
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
3353
            attr_type = self.desc.attr_type(n, True)
G
gongweibao 已提交
3354
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
3355
                attr_map[n] = self._block_attr(n)
3356
            elif attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
3357
                attr_map[n] = self._blocks_attr(n)
3358 3359 3360 3361 3362 3363
            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
3364

F
fengjiayi 已提交
3365 3366
        return attr_map

3367 3368 3369
    def _is_optimize_op(self):
        op_maker = core.op_proto_and_checker_maker
        OPTIMIZE = core.op_proto_and_checker_maker.OpRole.Optimize
3370 3371 3372 3373

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

3374 3375 3376
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(OPTIMIZE):
            return True
3377 3378 3379 3380 3381 3382 3383 3384

        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()):
3385 3386
            return False

3387 3388 3389 3390 3391 3392
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(BACKWARD):
            return True

        return False

3393
    @property
3394
    def dist_attr(self):
3395
        """
3396
        Get distributed attribute of this Variable.
3397
        """
3398
        return self.desc.dist_attr
3399

3400 3401
    @dist_attr.setter
    def dist_attr(self, dist_attr):
3402
        """
3403
        Set distributed attribute of this Variable.
3404
        """
3405
        self.desc.dist_attr = dist_attr
3406

Y
Yu Yang 已提交
3407

Y
Yu Yang 已提交
3408
class Block(object):
3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422
    """
    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
3423
        use `Program._create_block()` to create a block.
3424 3425 3426 3427

    Examples:
        .. code-block:: python

3428 3429 3430
            import paddle.fluid as fluid

            cur_program = fluid.Program()
3431 3432 3433 3434 3435 3436 3437 3438 3439
            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
3440
    def __init__(self, program, idx):
Y
Yu Yang 已提交
3441
        self.desc = program.desc.block(idx)
3442
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
3443
        self.ops = list()  # operator list
Y
Yu Yang 已提交
3444
        self.program = program
3445
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
3446

3447
    def __str__(self):
3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481
        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
3482
        ), "skip_op_callstack parameter's type is error, expect bool, received {}".format(
3483 3484 3485 3486 3487 3488 3489 3490 3491 3492 3493
            type(skip_op_callstack))
        block_str = "{ // block "
        block_str += "{}\n".format(self.idx)
        for var in list(self.vars.values()):
            block_str += "    {}\n".format(var._to_readable_code())
        block_str += "\n"
        for op in self.ops:
            block_str += "    {}\n".format(
                op._to_readable_code(skip_op_callstack))
        block_str += "}"
        return block_str
Y
Yang Yang(Tony) 已提交
3494

F
fengjiayi 已提交
3495 3496
    def to_string(self, throw_on_error, with_details=False):
        """
3497 3498
        Get debug string.

F
fengjiayi 已提交
3499 3500
        Args:
            throw_on_error(bool): raise exception when self is not initialized
3501
                when throw_on_error is True.
F
update  
fengjiayi 已提交
3502
            with_details(bool): more details about variables and parameters
3503 3504
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
3505

3506 3507
        Returns:
            str: The debug string.
F
fengjiayi 已提交
3508
        """
3509 3510
        assert isinstance(throw_on_error, bool) and isinstance(
            with_details, bool)
F
fengjiayi 已提交
3511
        if with_details:
F
fengjiayi 已提交
3512
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
3513 3514
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
3515
            for var in list(self.vars.values()):
F
fengjiayi 已提交
3516
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
3517
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
3518
            for op in self.ops:
F
fengjiayi 已提交
3519 3520
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
3521 3522 3523
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
3524 3525
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
3526 3527
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
3528 3529 3530

    __repr__ = __str__

Y
Yu Yang 已提交
3531 3532
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
3533
        return self.desc.parent
Y
Yu Yang 已提交
3534

Y
Yu Yang 已提交
3535 3536 3537 3538
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
3539
    def _set_forward_block_idx(self, idx):
3540 3541 3542 3543 3544 3545 3546 3547 3548
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
3549
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
3550

3551 3552 3553 3554 3555 3556 3557 3558
    @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
3559 3560
    @property
    def idx(self):
Y
Yu Yang 已提交
3561
        return self.desc.id
Y
Yu Yang 已提交
3562

Q
Qiao Longfei 已提交
3563
    def var(self, name):
3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576
        """
        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.
        """
3577
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
3578 3579 3580
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
3581 3582
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
3583
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
3584
        return v
Q
Qiao Longfei 已提交
3585

X
Xin Pan 已提交
3586
    def _find_var_recursive(self, name):
3587 3588 3589 3590 3591 3592 3593
        """
        Get a Variable by name from this block recursively.

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

        Returns:
X
Xin Pan 已提交
3594
            Variable: the Variable with the giving name. Or None if not found.
3595
        """
Y
Yu Yang 已提交
3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619
        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
3620
        return None
Y
Yu Yang 已提交
3621

X
Xin Pan 已提交
3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640
    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
3641

Q
Qiao Longfei 已提交
3642
    def all_parameters(self):
3643
        return list(self.iter_parameters())
3644

3645
    def iter_parameters(self):
M
minqiyang 已提交
3646
        return (item[1] for item in six.iteritems(self.vars)
3647
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
3648

Y
Yu Yang 已提交
3649
    def create_var(self, *args, **kwargs):
J
Jiabin Yang 已提交
3650
        if _non_static_mode():
L
Leo Chen 已提交
3651 3652
            var = _varbase_creator(*args, **kwargs)
        else:
3653 3654 3655
            var = Variable(block=self, *args, **kwargs)
            if 'initializer' in kwargs:
                kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
3656
        return var
Y
Yu Yang 已提交
3657

Q
Qiao Longfei 已提交
3658 3659 3660
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
3661
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
3662 3663
        """
        Rename variable in vars and ops' inputs and outputs
3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675

        Args:
            name(str): the name that need to be renamed.
            new_name(str): the name that need to rename to.

        Raises:
            ValueError: If this block doesn't have this the giving name,
                or the type of the var with the giving name is not Parameter
                or Variable.

        Returns:
            Variable: the Variable with the giving name.
T
typhoonzero 已提交
3676
        """
M
minqiyang 已提交
3677 3678
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
3679

T
typhoonzero 已提交
3680
        if not self.has_var(name):
3681
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
3682 3683
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
3684
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
3685 3686 3687 3688 3689 3690
            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
3691
            var_type = "Variable"
T
wip  
typhoonzero 已提交
3692 3693 3694 3695
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
3696
        orig_var_type = v.type
M
minqiyang 已提交
3697
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
3698
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
3699
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
3700
        if var_type == "Parameter":
L
Leo Chen 已提交
3701
            if in_dygraph_mode():
3702 3703 3704 3705 3706 3707 3708 3709 3710
                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)
3711
            else:
J
Jiabin Yang 已提交
3712
                if _in_legacy_dygraph():
3713 3714 3715 3716 3717 3718 3719 3720 3721
                    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
3722
                else:
3723 3724 3725 3726 3727 3728 3729 3730 3731 3732
                    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
3733
        elif var_type == "Variable":
3734 3735 3736 3737 3738
            var = Variable(self,
                           type=orig_var_type,
                           name=new_name,
                           error_clip=error_clip,
                           stop_gradient=stop_gradient)
T
wip  
typhoonzero 已提交
3739

W
Wu Yi 已提交
3740
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
3741 3742 3743
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
3744
        self._sync_with_cpp()
3745
        return var
T
typhoonzero 已提交
3746

3747 3748 3749
    def _remove_var(self, name, sync=True):
        if sync == True:
            self._sync_with_cpp()
M
minqiyang 已提交
3750
        self.desc._remove_var(cpt.to_bytes(name))
3751 3752
        del self.vars[name]

Y
Yu Yang 已提交
3753 3754
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
3755
        param = None
L
Leo Chen 已提交
3756
        if in_dygraph_mode():
J
Jiabin Yang 已提交
3757
            param = EagerParamBase(*args, **kwargs)
L
Leo Chen 已提交
3758
        else:
J
Jiabin Yang 已提交
3759 3760 3761 3762
            if _in_legacy_dygraph():
                param = ParamBase(*args, **kwargs)
            else:
                param = Parameter(global_block, *args, **kwargs)
3763

3764
        if 'initializer' in kwargs:
3765 3766 3767 3768 3769

            def _is_inited_by(block, var):
                init_ops = []
                for op in block.ops:
                    if var.name in op.output_arg_names:
3770
                        # In startup_program, "c_broadcast" and "c_sync_comm_stream"
T
tangwei12 已提交
3771
                        # are treated as initialization ops that cause error.
3772
                        # Think of "c_broadcast" and "c_sync_comm_stream" as a special case here.
3773 3774 3775 3776 3777
                        # NOTE: "coalesce_tensor" is a special case for rnn with cudnn support
                        if op.type in [
                                "c_broadcast", "c_sync_comm_stream",
                                "coalesce_tensor"
                        ]:
3778
                            continue
3779 3780 3781 3782 3783 3784 3785 3786
                        init_ops.append(op)
                return init_ops

            initializer = kwargs['initializer']
            init_ops = _is_inited_by(global_block, param)
            init_ops_len = len(init_ops)
            if init_ops_len > 1:
                raise RuntimeError("param " + param.name +
3787 3788
                                   " is inited by multiple init ops " +
                                   str(init_ops))
3789
            elif init_ops_len == 1:
3790
                # TODO already inited, do nothing, should log a warning
3791 3792 3793
                pass
            else:
                initializer(param, self)
Q
Qiao Longfei 已提交
3794
        return param
Y
Yu Yang 已提交
3795

Y
Yu Yang 已提交
3796
    def append_op(self, *args, **kwargs):
3797 3798 3799 3800 3801 3802
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
J
Jiabin Yang 已提交
3803
        if _non_static_mode():
3804
            attrs = kwargs.get("attrs", {})
Z
zyfncg 已提交
3805
            inplace_map = kwargs.get("inplace_map", None)
J
Jiabin Yang 已提交
3806
            type = kwargs.get("type", None)
3807 3808 3809 3810
            warnings.warn(
                "Op `%s` is executed through `append_op` under the dynamic mode, "
                "the corresponding API implementation needs to be upgraded to "
                "using `_C_ops` method." % type, DeprecationWarning)
3811 3812 3813 3814 3815 3816
            op = Operator(block=self,
                          desc=None,
                          type=type,
                          inputs=None,
                          outputs=None,
                          attrs=attrs)
3817

M
minqiyang 已提交
3818 3819 3820
            # record ops in tracer rather than blocks
            #
            # TODO(minqiyang): add op stop_gradient support in static mode too.
L
lujun 已提交
3821
            # currently, we only support stop_gradient in dygraph mode.
J
Jiabin Yang 已提交
3822

3823 3824 3825
            _dygraph_tracer().trace_op(type, kwargs.get("inputs", {}),
                                       kwargs.get("outputs",
                                                  {}), attrs if attrs else {},
Z
zyfncg 已提交
3826 3827
                                       kwargs.get("stop_gradient", False),
                                       inplace_map)
M
minqiyang 已提交
3828
        else:
3829 3830
            from paddle.fluid.dygraph.base import param_guard

3831
            op_desc = self.desc.append_op()
3832 3833 3834 3835 3836 3837
            # 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):
3838 3839 3840 3841 3842 3843
                op = Operator(block=self,
                              desc=op_desc,
                              type=kwargs.get("type", None),
                              inputs=inputs,
                              outputs=outputs,
                              attrs=kwargs.get("attrs", None))
3844

M
minqiyang 已提交
3845
            self.ops.append(op)
M
minqiyang 已提交
3846

3847 3848
        return op

W
Wu Yi 已提交
3849
    def _insert_op(self, index, *args, **kwargs):
3850 3851 3852 3853 3854 3855 3856 3857 3858
        """
        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
3859
        self._sync_with_cpp()
F
fangshuixun007 已提交
3860
        return self._insert_op_without_sync(index, *args, **kwargs)
Q
qiaolongfei 已提交
3861

3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878
    def _insert_op_without_sync(self, index, *args, **kwargs):
        """
        Insert an Operator according to the giving arguments, 
        without sync_with_cpp to meke the compilation faster.

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

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

    def _remove_op(self, index, sync=True):
3879 3880 3881 3882 3883 3884 3885 3886 3887
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
3888 3889
        if sync == True:
            self._sync_with_cpp()
W
Wu Yi 已提交
3890
        self.desc._remove_op(index, index + 1)
3891 3892
        del self.ops[index]

W
Wu Yi 已提交
3893
    def _slice_ops(self, start, end):
3894 3895 3896 3897 3898 3899 3900 3901 3902 3903
        """
        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
3904
        return self.ops[start:end]
Y
Yancey1989 已提交
3905

W
Wu Yi 已提交
3906
    def _prepend_op(self, *args, **kwargs):
J
Jiabin Yang 已提交
3907
        if _non_static_mode():
J
Jiabin Yang 已提交
3908 3909
            type = kwargs.get("type", None)
            attrs = kwargs.get("attrs", {})
3910 3911 3912 3913 3914 3915 3916 3917 3918 3919
            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 {},
M
minqiyang 已提交
3920
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
3921
        else:
3922
            op_desc = self.desc._prepend_op()
3923 3924 3925 3926 3927 3928
            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
3929
            self.ops.insert(0, op)
3930

Y
Yu Yang 已提交
3931 3932
        return op

W
Wu Yi 已提交
3933
    def _sync_with_cpp(self):
3934
        """
3935 3936
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
3937
        """
Q
Qiao Longfei 已提交
3938 3939 3940
        # sync variables from cpp
        for var in self.desc.all_vars():
            if not self.has_var(var.name()):
3941 3942 3943 3944
                is_stop_gradient = False
                if var.has_stop_gradient():
                    is_stop_gradient = var.stop_gradient()
                if var.has_is_parameter() and var.is_parameter():
3945 3946 3947 3948 3949 3950
                    self.create_parameter(name=var.name(),
                                          desc=var,
                                          type=var.type(),
                                          shape=var.shape(),
                                          dtype=var.dtype(),
                                          stop_gradient=is_stop_gradient)
3951
                else:
3952 3953 3954 3955
                    self.create_var(name=var.name(),
                                    desc=var,
                                    type=var.type(),
                                    stop_gradient=is_stop_gradient)
Q
Qiao Longfei 已提交
3956

3957
        # sync variables removed from c++ end
3958
        for var in list(self.vars.keys()):
M
minqiyang 已提交
3959
            if not self.desc.find_var(cpt.to_bytes(var)):
3960 3961
                self.vars.pop(var)

Q
Qiao Longfei 已提交
3962
        # sync operators from cpp
3963 3964 3965 3966
        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
3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982
        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
3983 3984 3985 3986 3987

        # 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
3988
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
3989 3990 3991 3992 3993 3994 3995

        # 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)

3996 3997 3998 3999 4000 4001 4002 4003 4004 4005 4006 4007 4008
        # sync ops removed from c++ end
        if end_index != -1 and end_index < len(self.ops):
            ops_in_cpp_index = 0
            ops_in_python_index = 0
            while ops_in_python_index < len(
                    self.ops) and ops_in_cpp_index < len(ops_in_cpp):
                if self.ops[ops_in_python_index].desc != ops_in_cpp[
                        ops_in_cpp_index]:
                    del self.ops[ops_in_python_index]
                else:
                    ops_in_cpp_index += 1
                    ops_in_python_index += 1

Q
Qiao Longfei 已提交
4009 4010 4011 4012
        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
4013
    def _copy_param_info_from(self, other):
4014
        """
4015 4016
        Copy the information of parameters from the other block.

4017
        Args:
4018 4019 4020 4021 4022
            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.
4023 4024 4025 4026 4027

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
4028 4029
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
4030
        for p in other.iter_parameters():
4031 4032 4033
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
4034 4035
                # if the Parameter is pruned, v may be None
                continue
4036
            assert isinstance(v, Variable)
4037
            new_p = None
L
Leo Chen 已提交
4038
            if in_dygraph_mode():
4039 4040 4041 4042 4043 4044 4045 4046 4047 4048
                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)
4049
            else:
J
Jiabin Yang 已提交
4050
                if _in_legacy_dygraph():
4051 4052 4053 4054 4055 4056 4057 4058 4059 4060
                    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
4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074
                else:
                    new_p = Parameter(
                        block=self,
                        shape=v.shape,
                        dtype=v.dtype,
                        type=v.type,
                        lod_level=v.lod_level
                        if v.type == core.VarDesc.VarType.LOD_TENSOR else None,
                        stop_gradient=p.stop_gradient,
                        trainable=p.trainable,
                        optimize_attr=p.optimize_attr,
                        regularizer=p.regularizer,
                        error_clip=p.error_clip,
                        name=v.name)
4075 4076
            self.vars[new_p.name] = new_p

4077
    def _clone_variable(self, var, force_persistable=True):
4078 4079
        """
        Clone a variable into current block.
4080

4081 4082
        Args:
            var: the variable to be cloned.
4083 4084 4085
            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.
4086 4087

        Returns:
4088
            Variable: the new  variable cloned from 'var' in current block.
4089 4090
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
4091 4092 4093
        ret_var = None
        # make STEP_SCOPES var can be safely cloned.
        if var.type == core.VarDesc.VarType.STEP_SCOPES:
4094 4095 4096
            ret_var = self.create_var(name=var.name,
                                      persistable=var.persistable,
                                      type=var.type)
T
tangwei12 已提交
4097
        elif var.type == core.VarDesc.VarType.RAW:
4098 4099 4100
            ret_var = self.create_var(name=var.name,
                                      persistable=var.persistable,
                                      type=var.type)
T
typhoonzero 已提交
4101 4102 4103 4104 4105 4106
        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,
4107
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
4108 4109
                is_data=var.is_data,
                need_check_feed=var.desc.need_check_feed())
T
update  
typhoonzero 已提交
4110 4111 4112 4113 4114 4115 4116
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
4117
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
4118 4119
                is_data=var.is_data,
                need_check_feed=var.desc.need_check_feed())
T
update  
typhoonzero 已提交
4120
        return ret_var
4121

Y
Yu Yang 已提交
4122

4123 4124 4125 4126
# 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)
4127
# of some old Python Variables(all old Python Operators) may have
4128
# been destructed.
4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139 4140 4141 4142 4143 4144
def _apply_pass(main_program,
                startup_program,
                pass_name,
                pass_attrs={},
                pass_attr_types={}):
    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)
    attrs = core.apply_pass(tmp_main_program, tmp_startup_program, pass_name,
                            pass_attrs, pass_attr_types)
    main_program._rebuild_from_desc(tmp_main_program)
    startup_program._rebuild_from_desc(tmp_startup_program)
    return attrs


4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171 4172 4173 4174 4175 4176 4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222 4223 4224 4225 4226 4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238 4239
class IrNode(object):
    """
    Python IrNode. Beneath it is a core.Node, which is used for Ir Pass.
    """

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

4240
    def remove_input_by_id(self, node_id):
4241 4242 4243 4244 4245 4246
        """
        Remove a node from inputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
4247
        self.node.remove_input(node_id)
4248

4249
    def remove_input(self, node):
4250 4251 4252 4253
        """
        Remove a node from inputs.

        Args:
4254
            node(IrNode): the node being removed.
4255
        """
4256
        self.node.remove_input(node.node)
4257

4258
    def append_input(self, node):
4259 4260 4261 4262
        """
        Append a node in inputs.

        Args:
4263
            node(IrNode): the node being appended.
4264
        """
4265
        self.node.append_input(node.node)
4266 4267 4268 4269 4270 4271 4272 4273

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

4274
    def remove_output_by_id(self, node_id):
4275 4276 4277 4278 4279 4280
        """
        Remove a node from outputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
4281
        self.node.remove_output(node_id)
4282

4283
    def remove_output(self, node):
4284 4285 4286 4287
        """
        Remove a node from outputs.

        Args:
4288
            node(IrNode): the node being removed.
4289
        """
4290
        self.node.remove_output(node.node)
4291

4292
    def append_output(self, node):
4293 4294 4295 4296
        """
        Append a node in outputs.

        Args:
4297
            node(IrNode): the node being appended.
4298
        """
4299
        self.node.append_output(node.node)
4300 4301 4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321 4322 4323 4324 4325 4326 4327 4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346

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

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

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

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


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

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

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

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

        Args:
            shape(list): shape to be set.
        """
        assert self.node.var() is not None, \
T
tianshuo78520a 已提交
4347
            "The node variable description can not be None."
4348 4349 4350 4351 4352 4353 4354 4355 4356 4357
        self.node.var().set_shape(shape)

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

        Returns:
            bool: indicate whether the variable is persistable.
        """
        assert self.node.var() is not None, \
T
tianshuo78520a 已提交
4358
            "The node variable description can not be None."
4359 4360
        return self.node.var().persistable()

4361 4362 4363 4364 4365 4366 4367 4368
    def type(self):
        """
        Return the variable type.

        Returns:
            core.VarDesc.VarType: the variable type.
        """
        assert self.node.var() is not None, \
T
tianshuo78520a 已提交
4369
            "The node variable description can not be None."
4370 4371 4372 4373 4374 4375 4376 4377 4378 4379
        return self.node.var().type()

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

        Returns:
            core.VarDesc.VarType: the variable data type.
        """
        assert self.node.var() is not None, \
T
tianshuo78520a 已提交
4380
            "The node variable description can not be None."
4381 4382 4383 4384 4385 4386 4387 4388 4389 4390
        return self.node.var().dtype()

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

        Returns:
            list: the variable shape.
        """
        assert self.node.var() is not None, \
T
tianshuo78520a 已提交
4391
            "The node variable description can not be None."
4392 4393
        return self.node.var().shape()

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

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

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

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


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

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

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

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

        Args:
            old_input_name(str): the old input name.
            new_input_name(str): the new input name.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
4441
            "The node operator description can not be None."
4442 4443
        self.node.op()._rename_input(old_input_name, new_input_name)

4444 4445 4446 4447 4448 4449 4450 4451 4452
    def rename_output(self, old_output_name, new_output_name):
        """
        Rename the output of this node.

        Args:
            old_output_name(str): the old output name.
            new_output_name(str): the new output name.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
4453
            "The node operator description can not be None."
4454 4455
        self.node.op()._rename_output(old_output_name, new_output_name)

4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466
    def input(self, name):
        """
        Get the argument name list by the parameter name for input.

        Args:
            name(str): the parameter name.

        Returns:
            list(str): the argument name list.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
4467
            "The node operator description can not be None."
4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480
        return self.node.op().input(name)

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

        Args:
            name(str): the parameter name.

        Returns:
            list(str): the argument name list.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
4481
            "The node operator description can not be None."
4482 4483 4484 4485 4486 4487 4488 4489 4490 4491
        return self.node.op().output(name)

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

        Args:
            new_type(str): new operator type to be set.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
4492
            "The node operator description can not be None."
4493 4494
        return self.node.op().set_type(new_type)

4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509
    def set_attr(self, name, val):
        """
        Set the value of attribute by attribute's name.

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

    def _update_desc_attr(self, name, val):
        """
        Update the value of the op desc's attribute by attribute's name.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
4510
            "The node operator description can not be None."
4511
        desc = self.node.op()
4512 4513 4514 4515 4516
        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):
4517
            desc.set_block_attr(name, val.desc)
4518
        elif isinstance(val, list) and val and _all_is_type(val, Block):
4519 4520
            desc.set_blocks_attr(name, [v.desc for v in val])
        elif isinstance(val, core.BlockDesc) or \
4521
                isinstance(val, core.ProgramDesc):
4522 4523 4524 4525
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)

4526 4527 4528 4529 4530 4531 4532 4533
    def input_arg_names(self):
        """
        Return input arguments' names of this op node.

        Returns:
            list(str): input arguments' names of this op node.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
4534
            "The node operator description can not be None."
4535 4536 4537 4538 4539 4540 4541 4542 4543 4544
        return self.node.op().input_arg_names()

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

        Returns:
            list(str): output arguments' names of this op node.
        """
        assert self.node.op() is not None, \
T
tianshuo78520a 已提交
4545
            "The node operator description can not be None."
4546 4547
        return self.node.op().output_arg_names()

4548 4549 4550 4551 4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567 4568
    @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]


4569 4570
class IrGraph(object):
    """
4571
    Python IrGraph. Beneath it is a core.Graph, which is used for
4572
    creating a c++ Ir Pass Graph. An IrGraph is just a graph view of
4573 4574
    a Program. In an IrGraph, both Variables and Operators are graph
    nodes.
4575 4576 4577 4578
    """

    def __init__(self, graph, for_test=False):
        """
4579 4580
        Construct an IrGraph using core.Graph.

4581 4582 4583 4584 4585 4586 4587 4588 4589
        Args:
            graph(core.Graph): C++ Graph.
            for_test(bool): True for the test graph and false for the train graph.
        """
        assert isinstance(
            graph, core.Graph), 'graph must be the instance of core.Graph.'
        self.graph = graph
        self._for_test = for_test

4590 4591 4592 4593
    def clone(self):
        """
        Create a new and duplicated IrGraph.

4594 4595 4596
        Warns:
            The method only clones the graph structure, not its attributes.

4597 4598 4599
        Returns:
            IrGraph: A new and duplicated graph.
        """
4600
        g = self.graph.clone()
4601 4602
        return IrGraph(g, self._for_test)

4603
    def is_test(self):
4604 4605 4606
        """
        If the graph is used for testing, the function returns true. Otherwise, returns false.
        """
4607 4608
        return self._for_test

W
WangZhen 已提交
4609
    def all_nodes(self):
4610 4611 4612
        """
        Return all nodes included in the graph as a set.
        """
4613
        return {IrNode(node) for node in self.graph.nodes()}
4614

4615
    def all_var_nodes(self):
4616 4617 4618
        """
        Return all variable nodes included in the graph as a set.
        """
4619
        return {IrVarNode(node) for node in self.graph.nodes() if node.is_var()}
4620

4621
    def all_persistable_nodes(self):
4622 4623 4624
        """
        Return all persistable variable nodes included in the graph as a set.
        """
W
WangZhen 已提交
4625 4626 4627 4628 4629
        persistable_nodes = set()
        for node in self.graph.nodes():
            if node.is_var() and node.var() is not None and node.var(
            ).persistable():
                persistable_nodes.add(node)
4630
        return {IrVarNode(p) for p in persistable_nodes}
W
WangZhen 已提交
4631

4632
    def all_op_nodes(self):
4633 4634 4635
        """
        Return all operator nodes included in the graph as a set.
        """
4636
        return {IrOpNode(node) for node in self.graph.nodes() if node.is_op()}
4637

4638 4639 4640 4641 4642 4643
    def all_sub_graphs(self, for_test=False):
        """
        Return all sub_graphs included in the main graph as a set.
        """

        return [
4644
            IrGraph(self.graph.get_sub_graph(i), for_test=for_test)
4645 4646 4647 4648 4649 4650 4651 4652 4653
            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)

4654
    def create_persistable_node(self, name, var_type, shape, var_dtype):
4655 4656 4657 4658 4659 4660 4661 4662 4663 4664 4665
        """
        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:
4666
            IrVarNode: the created persistable variable node.
4667
        """
4668 4669 4670 4671 4672
        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)
4673
        return IrVarNode(self.graph.create_var_node(var_desc))
4674 4675

    def create_var_node(self, name, var_type, shape, var_dtype):
4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686
        """
        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:
4687
            IrVarNode: the created variable node.
4688 4689
        """

4690 4691 4692 4693
        var_desc = core.VarDesc(name)
        var_desc.set_type(var_type)
        var_desc.set_shape(shape)
        var_desc.set_dtype(var_dtype)
4694
        return IrVarNode(self.graph.create_var_node(var_desc))
4695

4696 4697 4698 4699 4700 4701
    def create_control_dep_var(self):
        """
        create a control var
        """
        return IrVarNode(self.graph.create_control_dep_var())

4702
    def create_var_node_from_desc(self, var_desc):
4703 4704 4705 4706 4707 4708 4709 4710
        """
        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:
4711
            IrVarNode: the created variable node.
4712
        """
4713
        return IrVarNode(self.graph.create_var_node(var_desc))
4714 4715

    def create_op_node(self, op_type, attrs, inputs, outputs):
4716 4717 4718 4719 4720 4721 4722
        """
        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
4723
            outputs(dict): the outputs of the operator node.
4724 4725

        Returns:
4726
            IrOpNode: the created operator node.
4727
        """
4728 4729
        op_desc = core.OpDesc()
        op_desc.set_type(op_type)
4730
        for attr, value in six.iteritems(attrs):
4731
            self._update_desc_attr(op_desc, attr, value)
4732
        for input_name, var_nodes in six.iteritems(inputs):
4733 4734 4735 4736
            if not isinstance(var_nodes, list):
                var_nodes = [var_nodes]
            op_desc.set_input(input_name,
                              [var_node.name() for var_node in var_nodes])
4737
        for output_name, var_nodes in six.iteritems(outputs):
4738 4739 4740 4741
            if not isinstance(var_nodes, list):
                var_nodes = [var_nodes]
            op_desc.set_output(output_name,
                               [var_node.name() for var_node in var_nodes])
4742
        return IrOpNode(self.graph.create_op_node(op_desc))
4743 4744

    def create_op_node_from_desc(self, op_desc):
4745 4746 4747 4748 4749 4750 4751
        """
        Create a operator node by using an existing OpDesc in the graph.

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

        Returns:
4752
            IrOpNode: the created operator node.
4753
        """
4754
        return IrOpNode(self.graph.create_op_node(op_desc))
4755 4756

    def update_input_link(self, old_input_node, new_input_node, op_node):
4757 4758 4759 4760
        """
        Update the input's link of a operator node.

        Args:
4761 4762 4763
            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.
4764
        """
4765
        assert old_input_node.node in self.graph.nodes() and new_input_node.node in \
T
tangwei12 已提交
4766
            self.graph.nodes() and op_node.node in self.graph.nodes(), \
4767
            'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.'
4768 4769 4770 4771
        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)
4772
        op_node.rename_input(old_input_node.name(), new_input_node.name())
4773

4774 4775 4776 4777 4778 4779 4780 4781 4782 4783
    def update_output_link(self, old_output_node, new_output_node, op_node):
        """
        Update the output's link of an operator node.

        Args:
            old_output_node(IrNode): the old output node of the giving op_node.
            new_output_node(IrNode): the new output node of the giving op_node.
            op_node(IrOpNode): the operator node that is needed to update input's link.
        """
        assert old_output_node.node in self.graph.nodes() and new_output_node.node in \
T
tangwei12 已提交
4784
            self.graph.nodes() and op_node.node in self.graph.nodes(), \
4785
            'The three arguments(old_output_node &new_output_node &op_node) must be in the graph nodes.'
4786 4787 4788 4789 4790 4791
        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())

4792
    def link_to(self, node_in, node_out):
4793 4794 4795 4796
        """
        Connect two nodes.

        Args:
4797 4798
            node_in(IrNode): the input node.
            node_out(IrNode): the output node.
4799
        """
4800 4801 4802 4803
        assert node_in.node in self.graph.nodes(), (
            'node_in(%s) must be in the graph nodes.' % node_in.node.name())
        assert node_out.node in self.graph.nodes(), (
            'node_out(%s) must be in the graph nodes.' % node_out.node.name())
4804 4805
        node_in.append_output(node_out)
        node_out.append_input(node_in)
4806 4807

    def safe_remove_nodes(self, remove_nodes):
4808 4809 4810 4811 4812 4813 4814
        """
        Remove nodes safely since links connected to these removed nodes are
        also removed.

        Args:
            remove_nodes(set): the nodes prepared to be removed.
        """
4815
        if not isinstance(remove_nodes, set):
W
WangZhen 已提交
4816 4817 4818 4819
            if isinstance(remove_nodes, Iterable):
                remove_nodes = set(remove_nodes)
            else:
                remove_nodes = {remove_nodes}
4820 4821
        original_nodes = {n.node for n in remove_nodes}
        core.graph_safe_remove_nodes(self.graph, original_nodes)
4822

Z
Zhen Wang 已提交
4823 4824 4825 4826 4827 4828 4829 4830
    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] = [
4831
                            self._find_node_by_name(node.inputs, each_var_name)
Z
Zhen Wang 已提交
4832 4833 4834 4835
                        ]
                for each_var_name in node.op().output_arg_names():
                    if each_var_name not in var_nodes:
                        var_nodes[each_var_name] = [
4836
                            self._find_node_by_name(node.outputs, each_var_name)
Z
Zhen Wang 已提交
4837 4838 4839
                        ]
                    else:
                        var_nodes[each_var_name].append(
4840 4841
                            self._find_node_by_name(node.outputs,
                                                    each_var_name))
Z
Zhen Wang 已提交
4842 4843
        self.graph.resolve_hazard(var_nodes)

W
WangZhen 已提交
4844
    def has_circle(self):
4845 4846 4847 4848 4849 4850
        """
        Check if the graph has a circle.

        Returns:
            bool: True if the graph has a circle else False.
        """
W
WangZhen 已提交
4851 4852 4853
        return core.has_circle(self.graph)

    def graph_num(self):
4854 4855 4856 4857 4858 4859
        """
        Count the number of unconnected graphs in this graph.

        Returns:
            int: the number of unconnected graphs.
        """
W
WangZhen 已提交
4860 4861 4862
        return core.graph_num(self.graph)

    def topology_sort(self):
4863 4864 4865
        """
        Perform the topology sort operation on the graph.

T
tianshuo78520a 已提交
4866
        Notes: the `graph` can not contain a circle.
4867 4868

        Returns:
Z
Zhen Wang 已提交
4869
            list(IrNode): nodes in topology order.
4870
        """
4871
        ordered_nodes = core.topology_sort(self.graph)
Z
Zhen Wang 已提交
4872
        return [IrNode(n) for n in ordered_nodes]
W
WangZhen 已提交
4873 4874

    def build_adjacency_list(self):
4875 4876 4877 4878
        """
        Build an adjacency list of operations for the `graph`.

        Returns:
4879
            dict{IrNode: set(IrNode)}: the adjacency list.
4880
        """
4881 4882 4883 4884 4885
        adj_list = core.build_adjacency_list(self.graph)
        wrapped_adj_list = dict()
        for k, v in six.iteritems(adj_list):
            wrapped_adj_list[IrNode(k)] = {IrNode(n) for n in v}
        return wrapped_adj_list
W
WangZhen 已提交
4886

4887 4888 4889 4890 4891 4892 4893 4894
    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.
4895
            marked_nodes(set(IrNode)): nodes that are needed to be marked.
4896 4897 4898 4899 4900
            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.
        """

4901 4902
        def _convert_to_pdf(dot_file_path):
            pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf'
4903 4904 4905
            exited_code = subprocess.call('dot -Tpdf ' + dot_file_path +
                                          ' -o ' + pdf_save_path,
                                          shell=True)
4906 4907 4908 4909 4910
            if exited_code != 0:
                print('The dot command is needed for creating pdf files.')
                print('The {} is saved as the dot filetype.'.format(
                    dot_file_path))

4911
        remove_ctr_vars = set()
4912
        if remove_ctr_var:
4913
            for node in self.all_var_nodes():
4914 4915 4916
                if node.is_ctrl_var():
                    remove_ctr_vars.add(node)
            self.safe_remove_nodes(remove_ctr_vars)
4917 4918
        print('Total ops num = {}.'.format(len(self.all_op_nodes())))

4919 4920
        if marked_nodes is not None:
            if not isinstance(marked_nodes, set):
4921 4922 4923 4924 4925 4926
                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}
4927 4928 4929 4930
            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)
4931 4932
        if not os.path.exists(save_path):
            os.makedirs(save_path)
4933 4934 4935 4936 4937 4938 4939
        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):
4940 4941 4942
        """
        Convert the graph into a Program.

Z
Zhen Wang 已提交
4943
        WARN: When the graph includes backward operator nodes, the
4944 4945 4946 4947 4948 4949
        conversion process may be failed. Usually, this function is
        only used to convert a test graph.

        Returns:
            Program: a program converted from the graph.
        """
4950
        convert_pass = core.get_pass('graph_to_program_pass')
4951 4952
        desc = core.ProgramDesc()
        convert_pass.set_not_owned('program', desc)
4953 4954 4955 4956
        convert_pass.apply(self.graph)
        program = Program._construct_from_desc(desc)
        return program

4957 4958 4959 4960 4961 4962 4963 4964
    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
4965 4966
        assert target_node is not None, (
            "Cannot find the target node (%s)in the giving set." % node_name)
4967 4968
        return target_node

4969 4970 4971 4972
    def _update_desc_attr(self, desc, name, val):
        """
        Update the value of desc's attribute by attribute's name.
        """
4973 4974 4975 4976 4977
        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):
4978
            desc.set_block_attr(name, val.desc)
4979
        elif isinstance(val, list) and val and _all_is_type(val, Block):
4980 4981 4982 4983 4984 4985 4986 4987
            desc.set_blocks_attr(name, [v.desc for v in val])
        elif isinstance(val, core.BlockDesc) or \
                isinstance(val, core.ProgramDesc):
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)


Y
Yu Yang 已提交
4988
class Program(object):
D
dzhwinter 已提交
4989
    """
4990
    Create Python Program.  It has at least one :ref:`api_guide_Block_en`, when the
4991
    control flow op like conditional_block, while :ref:`api_paddle_fluid_layers_While` is included,
J
Jiabin Yang 已提交
4992
    it will contain nested block.
4993

J
Jiabin Yang 已提交
4994 4995 4996
    Please reference the
    `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_
    for details.
D
dzhwinter 已提交
4997

J
Jiabin Yang 已提交
4998
    A set of Program usually contains startup program and main program.
J
Jiabin Yang 已提交
4999
    A startup program is set to contain some initial work, eg. initialize the ``Parameter``, and the main
J
Jiabin Yang 已提交
5000 5001 5002 5003 5004 5005 5006
    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
5007
    **Notes**:
5008 5009 5010
        **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
5011 5012

    Returns:
J
Jiabin Yang 已提交
5013
        Program: An empty Program.
D
dzhwinter 已提交
5014 5015

    Examples:
5016 5017
        .. code-block:: python

5018 5019 5020 5021
            import paddle
            import paddle.static as static

            paddle.enable_static()
5022

5023 5024 5025 5026 5027
            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')
5028
                z = static.nn.fc(name="fc", x=x, size=10, activation="relu")
5029 5030 5031

            print("main program is: {}".format(main_program))
            print("start up program is: {}".format(startup_program))
D
dzhwinter 已提交
5032 5033 5034

    """

5035 5036
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
5037 5038
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
5039 5040
        global global_prog_seed
        self._seed = global_prog_seed
Y
yuyang18 已提交
5041
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
5042
        self.__op_role_var = []
T
tangwei12 已提交
5043

5044 5045
        # for distribute training
        # _is_distributed = True if under distributed training
T
tangwei12 已提交
5046
        self._is_distributed = False
5047
        # _is_chief = True if the trainer is the first one, usually No.0
T
tangwei12 已提交
5048
        self._is_chief = False
5049 5050 5051
        # _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
5052
        self._endpoints = []
5053 5054 5055
        # if current role is parameter server, the _ps_endpoint is its "ip:port"
        self._ps_endpoint = None
        # trainers_endpoints, it is used for distribution.
5056
        self._trainers_endpoints = []
5057
        # the distributed lookup table names
T
tangwei12 已提交
5058
        self._distributed_lookup_table = None
5059 5060 5061

        # use Deep gradient comrepssion or not
        self._enable_dgc = False
5062 5063
        self._use_lamb = False

5064 5065 5066
        self._nccl_comm_num = 1
        self._use_hierarchical_allreduce = False
        self._hierarchical_allreduce_inter_nranks = 0
5067

5068 5069 5070
        # 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
5071
        self._program_config = None
5072

H
hutuxian 已提交
5073 5074 5075
        # assigned if this program has been parsed by a pipeline optimizer
        self._pipeline_opt = None

5076 5077 5078
        # assigned if this program has been parsed by a heter pipeline parameter server optimizer
        self._heter_pipeline_opt = None

5079 5080 5081
        # appending gradients times
        self._appending_grad_times = 0

5082 5083 5084 5085
        # identifier for auto checkpoint
        self._auto_checkpoint_name = unique_name.generate(
            "__auto_checkpoint_program__")

5086 5087
        # compiled program, i.e. Graph
        self._graph = None
5088 5089
        # to tag whether is startup_program
        self._is_start_up_program_ = False
5090

5091
    def _find_var_class_kwargs(self, new_desc):
5092 5093 5094 5095 5096 5097 5098 5099
        # 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

5100 5101 5102 5103
        old_desc = self.desc
        all_new_vars = []
        block_num = new_desc.num_blocks()
        for idx in range(block_num):
5104 5105
            if (idx > (len(self.blocks) - 1)):
                self._create_block()
5106 5107 5108 5109 5110 5111 5112 5113 5114 5115
            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 = {
5116 5117 5118 5119 5120 5121
                    'type':
                    new_var_desc.type(),
                    'name':
                    new_var_desc.name(),
                    'shape':
                    get_var_desc_attr_or_none(new_var_desc, "shape", [
5122 5123 5124 5125
                        core.VarDesc.VarType.LOD_TENSOR,
                        core.VarDesc.VarType.SELECTED_ROWS,
                        core.VarDesc.VarType.LOD_TENSOR_ARRAY,
                    ]),
5126 5127
                    'dtype':
                    get_var_desc_attr_or_none(new_var_desc, "dtype", [
5128 5129 5130 5131 5132 5133 5134 5135 5136
                        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,
                    ]),
5137 5138 5139 5140 5141 5142 5143 5144 5145 5146
                    '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
5147 5148 5149 5150 5151 5152 5153 5154 5155 5156 5157 5158 5159 5160 5161 5162 5163 5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176
                    if old_var is not None else False,
                }

                if isinstance(old_var, Parameter):
                    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),
                    })
                else:
                    kwargs['persistable'] = new_var_desc.persistable()
                    block_new_vars.append({
                        'class': Variable,
                        'kwargs': copy.deepcopy(kwargs),
                    })

        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)
5177
        assert block_num == self.desc.num_blocks()
5178 5179

        # clear old blocks and desc
5180 5181 5182 5183 5184 5185 5186 5187 5188
        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)
5189

5190
        del desc
5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205 5206 5207 5208 5209

        # 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)

5210 5211 5212 5213 5214 5215 5216 5217 5218 5219
    def global_seed(self, seed=0):
        """
        Set global seed for Program

        Returns:
            None.

        Examples:
            .. code-block:: python

5220 5221
                import paddle
                import paddle.static as static
5222

5223 5224 5225
                paddle.enable_static()

                prog = static.default_main_program()
5226 5227 5228 5229 5230
                print(prog.random_seed)
                ## 0
                ## the default random seed is 0

                prog.global_seed(102)
5231
                prog1 = static.default_main_program()
5232 5233 5234 5235 5236 5237 5238 5239
                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
5240
    @property
5241
    def _op_role(self):
Y
yuyang18 已提交
5242 5243 5244 5245 5246 5247 5248 5249
        """
        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
5250
        parameter gradient of backward (use :code:`_op_role_var` to get this
Y
yuyang18 已提交
5251 5252 5253 5254
        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
5255 5256
        return self._current_role

5257 5258
    @_op_role.setter
    def _op_role(self, role):
Y
yuyang18 已提交
5259 5260 5261
        self._current_role = role

    @property
5262
    def _op_role_var(self):
Y
yuyang18 已提交
5263
        """
5264
        The auxiliary variables for :code:`_op_role` property.
Y
yuyang18 已提交
5265

5266
        See Also: :code:`Program._op_role`'s documentation for details.
Y
yuyang18 已提交
5267 5268 5269

        Notes: This is a very low-level API. Users should not use it directly.
        """
5270
        return self.__op_role_var
Y
yuyang18 已提交
5271

5272
    @signature_safe_contextmanager
5273 5274 5275 5276 5277
    def _backward_role_guard(self):
        tmp_role = self._current_role

        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Backward
5278 5279 5280 5281
        try:
            yield
        finally:
            self._current_role = tmp_role
5282

S
rename  
sneaxiy 已提交
5283
    @signature_safe_contextmanager
W
Wu Yi 已提交
5284
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
5285 5286 5287 5288 5289 5290 5291
        """
        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:
5292
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
5293 5294 5295

        Examples:

5296
            >>> import paddle.fluid as fluid
Y
yuyang18 已提交
5297
            >>> p, g = backward(...)
W
Wu Yi 已提交
5298
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
5299 5300
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
5301
        tmp_role = self._current_role
5302
        tmp_var = self.__op_role_var
X
Xin Pan 已提交
5303

Y
yuyang18 已提交
5304 5305
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
5306
        self.__op_role_var = [
5307 5308 5309
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
5310 5311 5312 5313 5314
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
Y
Yu Yang 已提交
5315

S
rename  
sneaxiy 已提交
5316
    @signature_safe_contextmanager
X
Xin Pan 已提交
5317
    def _lr_schedule_guard(self, is_with_opt=False):
5318 5319 5320 5321 5322 5323 5324
        """
        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
5325 5326 5327 5328
        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.
5329 5330 5331

        Examples:

5332
            >>> import paddle.fluid as fluid
5333 5334 5335 5336
            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
5337 5338

        tmp_role = self._current_role
5339
        tmp_var = self.__op_role_var
5340

5341 5342
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
5343 5344
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
5345
        # TODO(typhoonzero): how to set target learning rate var
5346
        self.__op_role_var = []
5347 5348 5349 5350 5351
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
5352

5353
    def __str__(self):
Y
yuyang18 已提交
5354 5355 5356 5357 5358 5359 5360 5361 5362
        """
        Get the protobuf debug string of this Program.

        Returns:
            (str): The protobuf debug string.

        Raises:
            ValueError: If any of required fields is not set.
        """
5363 5364 5365 5366 5367 5368 5369 5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381 5382
        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

5383 5384
            import paddle
            import paddle.static as static
5385

5386 5387 5388
            paddle.enable_static()

            cur_program = static.Program()
5389 5390 5391 5392 5393 5394 5395 5396 5397 5398 5399
            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
5400
        ), "skip_op_callstack parameter's type is error, expect bool, received {}".format(
5401 5402 5403 5404
            type(skip_op_callstack))
        program_str = ""
        for block in self.blocks:
            program_str += block._to_readable_code(skip_op_callstack)
5405
            program_str += '\n'
5406
        return program_str
Y
Yang Yang(Tony) 已提交
5407

F
fengjiayi 已提交
5408 5409 5410
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
5411

J
Jiabin Yang 已提交
5412 5413 5414
        Args:

            throw_on_error (bool): raise Value error when any of required fields is not set.
F
fengjiayi 已提交
5415

J
Jiabin Yang 已提交
5416
            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
5417

H
haowang101779990 已提交
5418
        Returns:
J
Jiabin Yang 已提交
5419
            str: The debug string describe current Program.
Y
yuyang18 已提交
5420 5421

        Raises:
J
Jiabin Yang 已提交
5422
            ValueError: If any of required fields is not set and throw_on_error is True.
F
fengjiayi 已提交
5423

5424 5425 5426
        Examples:
            .. code-block:: python

5427 5428 5429 5430
                import paddle
                import paddle.static as static

                paddle.enable_static()
5431

5432 5433 5434
                prog = static.default_main_program()
                x = static.data(name="X", shape=[2,3], dtype="float32")
                pred = static.nn.fc(x, size=3)
5435
                prog_string = prog.to_string(throw_on_error=True, with_details=False)
5436
                prog_string_with_details = prog.to_string(throw_on_error=False, with_details=True)
T
tianshuo78520a 已提交
5437
                print("program string without detail: {}".format(prog_string))
5438
                print("program string with detail: {}".format(prog_string_with_details))
F
fengjiayi 已提交
5439
        """
5440 5441 5442 5443 5444 5445 5446 5447 5448
        assert isinstance(
            throw_on_error, bool
        ), "The type of throw_on_error parameter is wrong, expected bool, but received {}.".format(
            type(throw_on_error))
        assert isinstance(
            with_details, bool
        ), "The type of with_details parameter is wrong, expected bool, but received {}.".format(
            type(with_details))

F
fengjiayi 已提交
5449 5450 5451 5452 5453 5454
        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()
5455 5456
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
5457 5458
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
5459

W
Wu Yi 已提交
5460
    def _get_desc(self):
Y
yuyang18 已提交
5461 5462 5463 5464 5465 5466 5467
        """
        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.
        """
5468 5469
        return self.desc

X
version  
Xin Pan 已提交
5470 5471 5472
    def _version(self):
        return self.desc._version()

5473
    def clone(self, for_test=False):
Y
yuyang18 已提交
5474
        """
5475 5476 5477 5478
        .. note:::
            1. :code:`Program.clone()` method DOES NOT clone :ref:`api_paddle_io_DataLoader` . 
            2. Recommend you to use :code:`clone` before using :code:`Opimizer.minimize` . 
            3. This API has no effect in Dygraph Mode.
Y
yuyang18 已提交
5479

5480
        Create a new Program with forward content of original one when ``for_test=True``.
5481
        Create a new Program as same as the original one when ``for_test=False``.
5482

5483
        Some operators, e.g., :ref:`api_paddle_fluid_layers_batch_norm` , behave differently between
Y
yuyang18 已提交
5484 5485 5486
        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`.
5487

5488 5489
        * 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.
5490 5491
          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
5492
          recommend you to use :code:`clone` before using :code:`Opimizer.minimize`.
Y
yuyang18 已提交
5493

J
Jiabin Yang 已提交
5494
        For Example:
5495
          ::
L
Luo Tao 已提交
5496

5497 5498 5499 5500 5501 5502
            import paddle
            import paddle.static as static

            paddle.enable_static()

            img = static.data(name='image', shape=[None, 784])
5503
            pred = static.nn.fc(x=img, size=10, actvation='relu')
5504
            loss = paddle.mean(pred)
5505
            # Here we use clone before Momentum
5506 5507
            test_program = static.default_main_program().clone(for_test=True)
            optimizer = paddle.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
5508
            optimizer.minimize(loss)
5509

J
Jiabin Yang 已提交
5510
        Args:
5511

5512 5513
            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` .
5514

J
Jiabin Yang 已提交
5515
        Returns:
5516
            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``
5517

Y
yuyang18 已提交
5518 5519 5520

        Examples:

5521 5522 5523 5524 5525 5526 5527
            .. 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`:

5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543
            .. code-block:: python

                import six

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


5544
            1. To clone a test program, the sample code is:
5545 5546 5547
                .. code-block:: python

                    import six
5548 5549 5550 5551 5552 5553
                    import paddle
                    import paddle.static as static
                    import paddle.utils as utils
                    import paddle.nn.functional as F

                    paddle.enable_static()
5554 5555 5556 5557 5558 5559 5560 5561 5562 5563 5564 5565

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

5566 5567
                    train_program = static.Program()
                    startup_program = static.Program()
J
Jiabin Yang 已提交
5568 5569 5570

                    # startup_program is used to do some parameter init work,
                    # and main program is used to hold the network
5571 5572 5573
                    with static.program_guard(train_program, startup_program):
                        with utils.unique_name.guard():
                            img = static.data(name='image', shape=[None, 784])
5574
                            hidden = static.nn.fc(x=img, size=200, activation='relu')
5575 5576
                            hidden = F.dropout(hidden, p=0.5)
                            loss = F.cross_entropy(
5577
                                input=static.nn.fc(x=hidden, size=10, activation='softmax'),
5578 5579
                                label=static.data(name='label', shape=[1], dtype='int64'))
                            avg_loss = paddle.mean(loss)
5580
                            test_program = train_program.clone(for_test=True)
5581
                    print_prog(test_program)
J
Jiabin Yang 已提交
5582 5583 5584 5585

                    # 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

5586
                    # In Paddle we will share weights by using the same Tensor name. In train and test program
J
Jiabin Yang 已提交
5587 5588 5589 5590
                    # 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.

5591 5592 5593
                    with static.program_guard(train_program, startup_program):
                        with utils.unique_name.guard():
                            sgd = paddle.optimizer.SGD(learning_rate=1e-3)
5594 5595 5596
                            sgd.minimize(avg_loss)


5597
            2. The clone method can be avoid if you create program for training and program for testing individually.
5598 5599 5600
                .. code-block:: python

                    import six
5601 5602 5603 5604 5605 5606
                    import paddle
                    import paddle.static as static
                    import paddle.utils as utils
                    import paddle.nn.functional as F

                    paddle.enable_static()
5607 5608 5609 5610 5611 5612 5613 5614 5615 5616 5617

                    def print_prog(prog):
                        for name, value in sorted(six.iteritems(prog.block(0).vars)):
                            print(value)
                        for op in prog.block(0).ops:
                            print("op type is {}".format(op.type))
                            print("op inputs are {}".format(op.input_arg_names))
                            print("op outputs are {}".format(op.output_arg_names))
                            for key, value in sorted(six.iteritems(op.all_attrs())):
                                if key not in ['op_callstack', 'op_role_var']:
                                    print(" [ attrs: {}:   {} ]".format(key, value))
5618

5619
                    def network():
5620
                        img = static.data(name='image', shape=[None, 784])
5621
                        hidden = static.nn.fc(x=img, size=200, activation='relu')
5622 5623
                        hidden = F.dropout(hidden, p=0.5)
                        loss = F.cross_entropy(
5624
                            input=static.nn.fc(x=hidden, size=10, activation='softmax'),
5625 5626
                            label=static.data(name='label', shape=[1], dtype='int64'))
                        avg_loss = paddle.mean(loss)
5627 5628
                        return avg_loss

5629 5630 5631 5632 5633
                    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():
5634
                            avg_loss = network()
5635
                            sgd = paddle.optimizer.SGD(learning_rate=1e-3)
5636
                            sgd.minimize(avg_loss)
5637
                    # the test startup program is not used.
5638 5639
                    with static.program_guard(test_program_2, startup_program_2):
                        with utils.unique_name.guard():
5640 5641
                            avg_loss = network()
                    print_prog(test_program_2)
5642

5643
            The two code snippets above will generate and print same programs.
5644
        """
5645

T
tangwei12 已提交
5646
        # NOTE(zhiqiu): we sync the original program first, since its program may diff with
5647 5648 5649
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

5650
        pruned_origin_block_id_map = None
5651
        if for_test:
5652 5653 5654 5655 5656 5657 5658 5659 5660
            forward_prog = Program()
            forward_prog.desc, pruned_origin_block_id_map = core.prune_backward(
                self.desc)
            forward_prog.blocks = [
                Block(forward_prog, i)
                for i in six.moves.range(forward_prog.desc.num_blocks())
            ]
            forward_prog._sync_with_cpp()
            p = forward_prog._inference_optimize(prune_read_op=False)
5661
        else:
5662
            p = Program()
G
gongweibao 已提交
5663 5664
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
5665
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
5666 5667 5668
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
5669 5670

            p._current_role = self._current_role
5671
            p.__op_role_var = self.__op_role_var
5672
            p._appending_grad_times = self._appending_grad_times
5673 5674
            if hasattr(self, 'lr_sheduler'):
                p.lr_sheduler = self.lr_sheduler
G
gongweibao 已提交
5675

T
tangwei12 已提交
5676
            # NOTE(zhiqiu): we sync the cloned program, to update its program by
5677
            # its desc.
W
Wu Yi 已提交
5678
            p._sync_with_cpp()
5679

W
Wu Yi 已提交
5680
        p._copy_param_info_from(self)
5681
        p._copy_data_info_from(self, pruned_origin_block_id_map)
5682
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
5683
        return p
5684

5685
    def _prune(self, targets):
Y
yuyang18 已提交
5686 5687 5688 5689 5690 5691 5692 5693
        """
        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:
5694
            targets(list|Variable|Operator): A list of variables, operators, or variable names
Y
yuyang18 已提交
5695 5696 5697 5698
                need to be pruned

        Returns:
            Program:  A new, pruned program.
5699
        """
5700
        return self._prune_with_input([], targets)
5701 5702

    def _prune_with_input(self, feeded_var_names, targets):
Y
yuyang18 已提交
5703
        """
5704 5705 5706 5707 5708 5709 5710 5711 5712 5713
        Prune operators and variables which are not needed to generate
        :code:`targets`. Prune operators and variables which are needed 
        to generate feeded_var 

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

        Args:
            feeded_var_names(list|str): A list of variable names from where
                pruning start. If it is set as [], this API works just like _prune()
5714
            targets(list|Variable|Operator): A list of variables, operators, or variable names
5715 5716 5717 5718 5719 5720
                need to be pruned

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

T
tangwei12 已提交
5721
        # NOTE(zhiqiu): we sync the original program first, since its program may diff with
5722 5723 5724
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

5725 5726
        if not isinstance(feeded_var_names, list):
            feeded_var_names = [feeded_var_names]
5727 5728
        if not isinstance(targets, list):
            targets = [targets]
5729 5730 5731

        for var in feeded_var_names:
            if not isinstance(var, six.string_types):
5732 5733 5734
                raise ValueError(
                    "All feeded_var_names of Program._prune_with_input() can only be "
                    "str, but received %s." % type(var))
5735

5736 5737 5738 5739 5740 5741 5742 5743 5744 5745 5746 5747 5748 5749 5750 5751
        # 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)

5752 5753 5754 5755
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
5756 5757 5758
                    name = t.name
                elif isinstance(t, six.string_types):
                    name = str(t)
5759
                else:
5760 5761 5762
                    raise ValueError(
                        "All targets of Program._prune_with_input() can only be "
                        "Variable or Operator, but received %s." % type(t))
5763 5764 5765 5766 5767 5768

                # 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:
5769 5770 5771
                    # however if the var is also updated by a runnable op, will shall keep it
                    if name not in generatable_vars:
                        continue
5772

5773 5774 5775 5776 5777 5778 5779 5780 5781
                # 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
5782
                        # Skip optimize op except for optimize op in targets,
5783 5784 5785 5786 5787
                        # since optimize op generates parameters.
                        if op._is_optimize_op() and op not in targets:
                            continue
                        else:
                            target_op = op
5788

5789
                if target_op is not None:
5790 5791 5792
                    targets_idx.append([target_op.block.idx, target_op.idx])
            else:
                targets_idx.append([t.block.idx, t.idx])
5793

5794
        res = Program()
5795 5796
        res.desc, pruned_origin_block_id_map = core.prune(
            self.desc, set(feeded_var_names), targets_idx)
M
minqiyang 已提交
5797 5798 5799
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
5800
        res._sync_with_cpp()
5801 5802 5803 5804 5805

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

5806 5807
        return res

X
Xin Pan 已提交
5808
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
5809
        """
F
fengjiayi 已提交
5810 5811 5812 5813 5814
        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.

5815
        3. change the :code:`is_test`
Y
yuyang18 已提交
5816 5817 5818
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

5819
        Args:
X
Xin Pan 已提交
5820 5821
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
5822

Y
yuyang18 已提交
5823 5824 5825 5826 5827 5828
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
5829
        res = Program()
5830
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
5831 5832 5833 5834

        # 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
5835
        if prune_read_op:
5836 5837 5838 5839 5840 5841 5842 5843 5844
            while True:
                if read_op_idx >= root_block.op_size() or root_block.op(
                        read_op_idx).type() == 'read':
                    break
                read_op_idx += 1
            if read_op_idx < root_block.op_size():
                root_block._remove_op(0, read_op_idx + 1)
            for var in root_block.all_vars():
                if var.type() == core.VarDesc.VarType.READER:
M
minqiyang 已提交
5845
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
5846 5847

        # change all `is_test` attributes to True
M
minqiyang 已提交
5848
        for i in six.moves.range(res.desc.num_blocks()):
5849
            block = res.desc.block(i)
M
minqiyang 已提交
5850
            for j in six.moves.range(block.op_size()):
5851 5852
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
5853
                    op._set_attr('is_test', True)
5854 5855 5856
                if op.type() == "batch_norm":
                    # Remove the output ReserveSpace of batch_norm if exists.
                    op.remove_output("ReserveSpace")
M
minqiyang 已提交
5857 5858 5859
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
5860
        res._sync_with_cpp()
5861 5862
        return res

5863
    def _remove_training_info(self, clip_extra=True):
5864 5865 5866 5867 5868 5869 5870 5871 5872 5873 5874 5875 5876 5877 5878 5879 5880 5881 5882
        """
        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)

        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
        res._sync_with_cpp()

5883 5884
        # Note: The op_role and op_role_var cann't be deleted currently,
        # and we will try to remove them in the future.
5885
        common_clipped_attrs_list = ['op_callstack', 'with_quant_attr']
5886

5887 5888 5889 5890 5891
        for i in six.moves.range(res.desc.num_blocks()):
            block = res.desc.block(i)
            for var in block.all_vars():
                var.clear_is_parameter()
                var.clear_stop_gradient()
5892 5893
            if not clip_extra:
                continue
5894 5895 5896 5897
            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
5898 5899 5900

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

5901 5902 5903 5904 5905 5906 5907 5908 5909 5910 5911 5912 5913
                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)
5914 5915 5916
                # The extra input of op will be removed in the future
                # for name in remove_input_list:
                #     op.remove_input(name)
5917 5918 5919 5920 5921 5922 5923 5924 5925 5926 5927 5928 5929

                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)
5930 5931 5932
                # The extra output of op will be removed in the future
                # for name in remove_output_list:
                #     op.remove_output(name)
5933 5934 5935 5936 5937 5938 5939 5940 5941 5942

                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
                quant_attrs = [
                    op_quant_name, "quantization_type", "skip_quant",
                    "activation_bits", "bit_length", "quantize_weight_bits",
                    "weight_quant_scale"
                ]
5943 5944 5945
                for extra_attr_name in extra_attrs_map.keys():
                    op.remove_attr(extra_attr_name)
                remove_attr_list = []
5946 5947 5948 5949 5950 5951
                for name in op.attr_names():
                    if quant:
                        if name in quant_attrs:
                            continue
                        if name.endswith("_threshold"):
                            continue
5952
                    if len(extra_attrs_map) > 0:
5953
                        if name in common_clipped_attrs_list:
5954
                            op.remove_attr(name)
5955
                        continue
5956 5957 5958 5959 5960 5961 5962 5963 5964 5965
                    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)
5966 5967
        return res

5968 5969
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
5970
        """
5971 5972 5973
        .. note::
            1. All information about parameters will be lost after serialization; 
            2. This API has no effect in Dygraph mode.
Y
yuyang18 已提交
5974

5975 5976
        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
5977

J
Jiabin Yang 已提交
5978
        Args:
Y
yuyang18 已提交
5979

J
Jiabin Yang 已提交
5980
            binary_str_type (str): the binary prootbuf string.
5981

J
Jiabin Yang 已提交
5982 5983
        Returns:
            Program: A deserialized Program.
5984 5985 5986 5987

        Examples:
            .. code-block:: python

5988 5989 5990 5991
                import paddle
                import paddle.static as static

                paddle.enable_static()
5992

5993 5994 5995 5996
                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')
5997

5998
                    y = static.data(name='Y', shape=[784, 100], dtype='float32')
5999

6000
                    z = paddle.matmul(x=x, y=y)
6001

6002 6003
                    binary_str = static.default_main_program().desc.serialize_to_string()
                    prog_restored = static.default_main_program().parse_from_string(binary_str)
6004

6005
                    print(static.default_main_program())
6006
                    print(prog_restored)
Y
yuyang18 已提交
6007
        """
6008 6009
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
6010
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
6011
        p._sync_with_cpp()
6012
        return p
Y
Yu Yang 已提交
6013

6014
    @staticmethod
6015
    def _construct_from_desc(desc):
6016 6017 6018 6019 6020 6021 6022 6023 6024 6025 6026 6027 6028 6029 6030
        """
        Construct a program from program desc.

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

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

D
dzhwinter 已提交
6031 6032
    @property
    def random_seed(self):
Y
yuyang18 已提交
6033
        """
J
Jiabin Yang 已提交
6034
        The default random seed for random operators in Program. ``0`` means get
Y
yuyang18 已提交
6035 6036
        the random seed from random device.

6037 6038
        .. note:: 
            It must be set before the operators have been added.
J
Jiabin Yang 已提交
6039 6040 6041

        Returns:
            int64: Random seed in current Program
6042

6043 6044 6045 6046

        Examples:
            .. code-block:: python

6047 6048 6049
                import paddle
                import paddle.static as static
                import paddle.nn.functional as F
6050

6051 6052 6053
                paddle.enable_static()

                prog = static.default_main_program()
6054
                random_seed = prog.random_seed
6055
                x_var = static.data(name="X", shape=[3,3], dtype="float32")
6056 6057 6058
                print(random_seed)
                ## 0
                ## the default random seed is 0
6059

6060
                # Here we need to set random seed before we use paddle.nn.functional.dropout
6061
                prog.random_seed = 1
6062
                z_var = F.dropout(x_var, 0.7)
6063

6064
                print(prog.random_seed)
6065 6066
                ## 1
                ## the random seed is change to 1
Y
yuyang18 已提交
6067
        """
D
dzhwinter 已提交
6068 6069
        return self._seed

Q
qiaolongfei 已提交
6070 6071
    @property
    def num_blocks(self):
Y
yuyang18 已提交
6072
        """
6073 6074
        The number of :ref:`api_guide_Block_en`  in this Program.

6075 6076
        .. note:: 
            This API has no effect in Dygraph mode.
J
Jiabin Yang 已提交
6077 6078 6079

        Returns:
            int(Platform-dependent size): num of :ref:`api_guide_Block_en`  in current Program
6080

6081 6082 6083 6084

        Examples:
            .. code-block:: python

6085 6086 6087 6088
                import paddle
                import paddle.static as static

                paddle.enable_static()
6089

6090
                prog = static.default_main_program()
6091 6092
                num_blocks = prog.num_blocks
                print(num_blocks)
6093

6094 6095
                # print result:
                # 1
Y
yuyang18 已提交
6096
        """
Q
qiaolongfei 已提交
6097 6098
        return self.desc.num_blocks()

D
dzhwinter 已提交
6099 6100 6101
    @random_seed.setter
    def random_seed(self, seed):
        if not isinstance(seed, int):
6102 6103 6104
            raise ValueError(
                "Program.random_seed's input seed must be an integer, but received %s."
                % type(seed))
D
dzhwinter 已提交
6105 6106
        self._seed = seed

Y
Yu Yang 已提交
6107
    def __repr__(self):
6108
        return self.__str__()
6109

Y
Yu Yang 已提交
6110
    def global_block(self):
Y
yuyang18 已提交
6111
        """
6112 6113
        .. note::
            This API has no effect in Dygraph mode.
6114 6115 6116

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

J
Jiabin Yang 已提交
6117 6118
        Returns:
            :ref:`api_guide_Block_en`: The first :ref:`api_guide_Block_en`  of this Program.
6119

6120 6121 6122 6123

        Examples:
            .. code-block:: python

6124 6125 6126 6127
                import paddle
                import paddle.static as static

                paddle.enable_static()
6128

6129
                prog = static.default_main_program()
6130 6131
                gb_block = prog.global_block()
                print(gb_block)
6132

Y
yuyang18 已提交
6133
        """
Y
Yu Yang 已提交
6134 6135
        return self.blocks[0]

Q
Qiao Longfei 已提交
6136
    def block(self, index):
Y
yuyang18 已提交
6137
        """
6138 6139
        .. note::
            This API has no effect in Dygraph mode.
Y
yuyang18 已提交
6140

6141 6142
        Get the :code:`index`  :ref:`api_guide_Block_en`  of this Program

J
Jiabin Yang 已提交
6143 6144
        Args:
            index (int) - The index of  :ref:`api_guide_Block_en`  to get
6145

J
Jiabin Yang 已提交
6146 6147
        Returns:
            :ref:`api_guide_Block_en`: The :code:`index` block
6148 6149 6150 6151

        Examples:
            .. code-block:: python

6152 6153 6154 6155
                import paddle
                import paddle.static as static

                paddle.enable_static()
6156

6157
                prog = static.default_main_program()
6158 6159
                block_0 = prog.block(0)
                print(block_0)
Y
yuyang18 已提交
6160
        """
Q
Qiao Longfei 已提交
6161 6162
        return self.blocks[index]

Y
Yu Yang 已提交
6163
    def current_block(self):
Y
yuyang18 已提交
6164
        """
6165 6166
        .. note::
            This API has no effect in Dygraph mode.
6167

J
Jiabin Yang 已提交
6168 6169
        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.
6170

J
Jiabin Yang 已提交
6171 6172
        Returns:
             :ref:`api_guide_Block_en`: The :code:`index`  :ref:`api_guide_Block_en`
6173

6174 6175 6176
        Examples:
            .. code-block:: python

6177 6178 6179 6180
                import paddle
                import paddle.static as static

                paddle.enable_static()
6181

6182
                prog = static.default_main_program()
6183 6184
                current_blk = prog.current_block()
                print(current_blk)
Y
yuyang18 已提交
6185
        """
Y
Yu Yang 已提交
6186 6187
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
6188
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
6189 6190 6191 6192 6193
        """
        Create a new block with the :code:`parent_idx` and change the current block
        to new block.

        Args:
J
Jiabin Yang 已提交
6194

Y
yuyang18 已提交
6195 6196 6197 6198 6199
            parent_idx(int): The parent block index.

        Returns:
            Block: The new block.
        """
Y
Yu Yang 已提交
6200
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
6201 6202 6203
        parent = self.current_block() if parent_idx is None else self.block(
            parent_idx)
        self.desc.append_block(parent.desc)
Y
Yu Yang 已提交
6204 6205 6206 6207
        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
6208
    def _rollback(self):
Y
yuyang18 已提交
6209 6210 6211 6212 6213
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
6214 6215
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
6216
    def _sync_with_cpp(self):
Y
yuyang18 已提交
6217 6218 6219 6220 6221 6222 6223 6224 6225 6226
        """
        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
6227 6228 6229
        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
6230
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
6231

W
Wu Yi 已提交
6232
    def _copy_param_info_from(self, other):
6233
        """
6234
        Copy the information of parameters from other program.
D
dzhwinter 已提交
6235

Y
yuyang18 已提交
6236 6237 6238
        Notes: This is a very low level API. Users should not invoke it
        directly.

6239 6240 6241 6242 6243 6244 6245
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
6246 6247 6248
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
6249

W
Wu Yi 已提交
6250
        self.global_block()._copy_param_info_from(other.global_block())
6251

6252 6253 6254 6255 6256 6257 6258 6259 6260 6261 6262
    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):
6263 6264 6265
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
6266 6267
        self._is_distributed = other._is_distributed
        self._is_chief = other._is_chief
6268
        self._parameters_on_pservers = other._parameters_on_pservers
6269
        self._endpoints = other._endpoints
6270
        self._ps_endpoint = other._ps_endpoint
6271 6272
        self._distributed_lookup_table = other._distributed_lookup_table

6273
    def _copy_data_info_from(self, other, pruned_origin_block_id_map=None):
F
fengjiayi 已提交
6274 6275
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
6276

Y
yuyang18 已提交
6277 6278 6279
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
6280 6281
        Args:
            other(Program): Other program
6282 6283 6284 6285
            pruned_origin_block_id_map(dict{int:int}): A dict which maps the block id in program
            self to the block id in program other. For example, {0:0, 1:1, 2:3} means block 0 in self is 
            cloned from block 0 in other, etc. Default is None, which means default mapped, 
            {0:0, 1:1,..., n:n}.
F
fengjiayi 已提交
6286 6287 6288 6289 6290

        Returns:
            None
        """
        if not isinstance(other, Program):
6291 6292 6293
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
F
fengjiayi 已提交
6294

6295 6296 6297 6298 6299
        if not pruned_origin_block_id_map:
            pruned_origin_block_id_map = {
                i: i
                for i in six.moves.range(self.desc.num_blocks())
            }
6300 6301 6302

        # NOTE(zhiqiu): All vars in cloned program exist in original program.
        # The reverse is not true, due to backward pruning.
6303 6304
        for i, block in enumerate(self.blocks):
            other_block = other.blocks[pruned_origin_block_id_map[i]]
6305
            for var in list(block.vars.values()):
6306 6307 6308 6309 6310 6311 6312
                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
6313

6314
    def list_vars(self):
Y
yuyang18 已提交
6315
        """
6316
        Get all Tensors from this Program. A iterable object is returned.
Y
yuyang18 已提交
6317

J
Jiabin Yang 已提交
6318
        Returns:
6319
            iterable Tensors: The Generator will yield every Tensor in this program.
6320 6321 6322 6323

        Examples:
            .. code-block:: python

6324 6325
                import paddle
                import paddle.static as static
6326

6327 6328 6329 6330 6331
                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')
6332 6333
                for var in prog.list_vars():
                    print(var)
T
tangwei12 已提交
6334

6335 6336
                # 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
6337
        """
6338
        for each_block in self.blocks:
6339
            for each_var in list(each_block.vars.values()):
6340 6341
                yield each_var

6342 6343 6344 6345 6346 6347 6348 6349 6350 6351
    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

6352 6353 6354 6355
                import paddle
                import paddle.static as static

                paddle.enable_static()
6356

6357 6358
                program = static.default_main_program()
                data = static.data(name='x', shape=[None, 13], dtype='float32')
6359
                hidden = static.nn.fc(x=data, size=10)
6360 6361
                loss = paddle.mean(hidden)
                paddle.optimizer.SGD(learning_rate=0.01).minimize(loss)
6362 6363 6364 6365 6366 6367 6368

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

                # Here will print all parameters in current program, in this example,
                # the result is like:
                #
6369 6370
                # 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)
6371 6372 6373 6374 6375 6376 6377 6378 6379 6380
                #
                # 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

6381 6382 6383 6384 6385 6386 6387 6388 6389 6390 6391 6392 6393 6394 6395 6396 6397 6398 6399 6400 6401 6402 6403 6404 6405 6406 6407 6408 6409 6410 6411 6412 6413 6414 6415 6416 6417 6418 6419 6420 6421 6422
    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:
            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.  
                    'all' : The return value contains the variable in the network and optimizer.
                    Default: 'all'
            scope(Scope, optional) : If scope is None, state_dict will be set to global scope 
                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'
6423
        # can not be imported at the begainning of this file.
6424 6425 6426 6427
        # Therefore, the above two modules are dynamically imported.
        from .executor import global_scope
        if scope is not None and not isinstance(scope, core._Scope):
            raise TypeError(
6428 6429
                "`scope` should be None or `paddle.static.Scope'` type, but received {}."
                .format(type(scope)))
6430 6431 6432 6433 6434

        if scope is None:
            scope = global_scope()

        if not isinstance(mode, str):
6435 6436 6437
            raise TypeError(
                "Type of `mode` should be string, but received {}.".format(
                    type(mode)))
6438 6439 6440 6441 6442 6443 6444 6445 6446 6447 6448 6449 6450 6451 6452 6453 6454 6455 6456 6457 6458 6459 6460 6461 6462 6463

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

        def is_persistable(var):
            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:
                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(
6464 6465
                    "`mode` string should be 'param', 'opt' or 'all', but received {}."
                    .format(mode))
6466 6467 6468 6469 6470 6471 6472 6473

        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(
6474 6475
                    "Can not find Variable '{}' in the scope. Make sure it is initialized"
                    .format(var.name))
6476 6477 6478 6479 6480 6481 6482 6483 6484 6485 6486 6487 6488 6489 6490 6491 6492 6493 6494 6495 6496 6497 6498 6499 6500 6501 6502 6503 6504 6505 6506 6507 6508 6509 6510 6511 6512 6513 6514 6515 6516 6517 6518 6519 6520 6521 6522 6523 6524 6525 6526 6527 6528 6529 6530 6531 6532 6533 6534 6535 6536 6537 6538 6539 6540 6541 6542 6543 6544
            state_dict[var.name] = var_temp.get_tensor()

        return state_dict

    def set_state_dict(self, state_dict, scope=None):
        """
        Set parameters and persistable buffers in state_dict to program. 
        An exception will throw if shape or dtype of the parameters is not match.
        
        .. note::
            This function MUST called after run start_up_program

        Args:
            state_dict(dict): the dict store parameters and persistable buffers. 
                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.
            scope(Scope, optional) : If scope is None, state_dict will be set to global scope 
                obtained through 'paddle.static.global_scope()'. Otherwise, value will be set to scope.
                Default: None
        
        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(
                    type(state_dict)))

        vars_dict = {var.name: var for var in self.list_vars()}
        condition = True if 'StructuredToParameterName@@' in state_dict else False
        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(
                        ("Skip loading for '{}'. ".format(name) + str(err)))
                except TypeError as err:
                    warnings.warn(
                        ("Skip loading for '{}'. ".format(name) + str(err)))
            else:
6545 6546 6547
                warnings.warn(
                    ("Skip loading for '{0}'. Because '{0}' not in the program."
                     .format(name)))
6548

Y
Yu Yang 已提交
6549

6550
@six.add_metaclass(ParameterMetaClass)
Y
Yu Yang 已提交
6551
class Parameter(Variable):
6552
    """
6553
    Parameter is derived from Variable. A parameter is a persistable
6554
    Variable, and will be updated by optimizers after each iteration.
6555
    The training of a neural network is essentially the updating of
6556 6557
    its parameters.

6558
    Relative to a general Variable, a Parameter has several its own
6559 6560
    member variables:

6561 6562 6563 6564 6565 6566 6567 6568 6569 6570
    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.
6571 6572
        need_clip (bool): Whether the parameter gradient need to be cliped 
            in optimizer. Default is True.
6573 6574
    """

6575 6576 6577 6578 6579 6580
    def __init__(self,
                 block,
                 shape,
                 dtype,
                 type=core.VarDesc.VarType.LOD_TENSOR,
                 **kwargs):
6581 6582 6583 6584 6585
        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
6586
        if len(shape) == 0:
6587 6588
            raise ValueError(
                "The dimensions of shape for Parameter must be greater than 0")
Y
Yu Yang 已提交
6589 6590 6591

        for each in shape:
            if each < 0:
6592 6593 6594
                raise ValueError(
                    "Each dimension of shape for Parameter must be greater than 0, but received %s"
                    % list(shape))
6595

6596 6597 6598 6599 6600 6601 6602
        Variable.__init__(self,
                          block,
                          persistable=True,
                          shape=shape,
                          dtype=dtype,
                          type=type,
                          **kwargs)
Y
Yu Yang 已提交
6603 6604 6605 6606
        self.trainable = kwargs.get('trainable', True)

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

6607 6608
        self.regularizer = kwargs.get('regularizer', None)

W
wanghaoshuang 已提交
6609
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
6610

6611 6612
        self.need_clip = kwargs.get('need_clip', True)

6613 6614
        self.is_distributed = False

6615 6616
        self.is_parameter = True

F
fengjiayi 已提交
6617
    def __str__(self):
6618
        return self._to_readable_code()
F
fengjiayi 已提交
6619

F
update  
fengjiayi 已提交
6620 6621 6622
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
6623

F
update  
fengjiayi 已提交
6624 6625 6626 6627 6628 6629 6630 6631
        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.

6632 6633 6634 6635 6636 6637 6638 6639 6640
        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
6641
        """
6642 6643
        assert isinstance(throw_on_error, bool) and isinstance(
            with_details, bool)
F
update  
fengjiayi 已提交
6644 6645 6646
        if with_details:
            res_str = Variable.to_string(self, throw_on_error, True)
            additional_attr = ("trainable", "optimize_attr", "regularizer",
6647
                               "do_model_average", "need_clip")
F
update  
fengjiayi 已提交
6648
            for attr_name in additional_attr:
6649 6650
                res_str += "%s: %s\n" % (attr_name,
                                         cpt.to_text(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
6651 6652
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
6653 6654 6655 6656
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
6657

6658 6659
class ParamBase(core.VarBase):
    """
6660 6661 6662
    ParamBase is derived from Tensor( Which is the concept in Dygraph Mode). 
    A ParamBase is a persistable Tensor, and will be updated by optimizers 
    after each iteration.
6663 6664 6665
    The training of a neural network is essentially the updating of
    its ParamBase.

6666
    Relative to a general Tensor, a ParamBase has several its own
6667 6668 6669 6670 6671 6672 6673 6674 6675 6676 6677 6678
    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.
6679 6680
        need_clip (bool): Whether the parameter gradient need to be cliped 
            in optimizer. Default is True.
6681 6682 6683 6684 6685 6686 6687 6688 6689 6690 6691 6692 6693 6694 6695 6696 6697 6698 6699 6700 6701 6702 6703 6704 6705
    """

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

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

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

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

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

6706 6707 6708 6709
        super(ParamBase,
              self).__init__(dtype if dtype else core.VarDesc.VarType.FP32,
                             list(shape) if shape else [], name,
                             core.VarDesc.VarType.LOD_TENSOR, True)
6710

6711 6712
        trainable = kwargs.get('trainable', True)
        self.stop_gradient = not trainable
6713 6714 6715 6716 6717 6718 6719

        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)

6720 6721
        self.need_clip = kwargs.get('need_clip', True)

6722
        self.is_distributed = kwargs.get('is_distributed', False)
6723
        # self.block = default_main_program().global_block()
6724

6725 6726 6727 6728 6729 6730 6731 6732 6733 6734 6735 6736 6737
    @property
    def trainable(self):
        return not self.stop_gradient

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

6738
    def __str__(self):
6739
        """
6740
        Convert a ParamBase object to a readable string.
6741

6742
        Returns(str): A readable string.
6743 6744 6745 6746

        Examples:
            .. code-block:: python

6747
                import paddle
6748 6749 6750 6751 6752 6753 6754
                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]])
6755
        """
6756 6757
        return "Parameter containing:\n{tensor}".format(
            tensor=super(ParamBase, self).__str__())
6758

6759 6760 6761 6762 6763 6764 6765 6766 6767 6768 6769
    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
6770

6771 6772 6773 6774 6775 6776 6777 6778 6779 6780 6781 6782 6783 6784 6785 6786 6787 6788
                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

6789 6790 6791 6792
    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)
6793 6794 6795 6796 6797 6798
        return new_param

    __repr__ = __str__


if hasattr(core, "eager"):
6799
    _core_eager_eagertensor = core.eager.Tensor
6800 6801 6802 6803 6804 6805 6806 6807 6808 6809 6810 6811 6812 6813 6814 6815 6816 6817 6818 6819 6820 6821 6822 6823 6824 6825 6826 6827 6828 6829 6830 6831 6832 6833 6834 6835 6836 6837 6838 6839 6840 6841 6842 6843 6844 6845 6846 6847 6848 6849 6850 6851
else:
    _core_eager_eagertensor = object


class EagerParamBase(_core_eager_eagertensor):
    """
    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 
    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.
        need_clip (bool): Whether the parameter gradient need to be cliped 
            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")

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

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

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

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

6852 6853 6854
        if isinstance(shape, core.eager.Tensor):
            shape = shape.numpy()

6855 6856 6857 6858
        super(EagerParamBase,
              self).__init__(dtype if dtype else core.VarDesc.VarType.FP32,
                             list(shape) if shape else [], name,
                             core.VarDesc.VarType.LOD_TENSOR, True)
6859 6860 6861 6862 6863 6864 6865 6866 6867 6868 6869 6870 6871 6872
        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)
6873 6874 6875
        # hook functions for lazy initialization
        self._init_func = None
        self._init_op_creator = None
6876 6877

    def set_init_func(self, obj):
6878
        self._init_func = obj
6879 6880 6881

    @dygraph_only
    def initialize(self):
6882 6883
        assert self._init_func is not None, "Required self._init_func is not None, but received None."
        self._init_func()
6884
        # clear function handle to release resource
6885
        self._init_func = None
6886 6887 6888 6889 6890 6891 6892 6893 6894 6895 6896 6897 6898 6899

    @property
    def trainable(self):
        return not self.stop_gradient

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

6900 6901 6902 6903 6904 6905 6906
    def _create_init_op(self, block):
        """
        Call init_op_creator function to create initializer operation in block.
        """
        assert self._init_op_creator is not None, "Required self._init_op_creator is not None, but received None."
        self._init_op_creator(block)

6907 6908 6909 6910 6911 6912 6913 6914 6915 6916 6917 6918 6919 6920 6921 6922 6923 6924 6925 6926 6927 6928 6929 6930 6931 6932 6933 6934 6935 6936 6937 6938 6939 6940 6941 6942 6943 6944 6945 6946 6947 6948 6949 6950 6951 6952 6953 6954 6955 6956 6957 6958 6959 6960 6961
    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(
            tensor=super(EagerParamBase, self).__str__())

    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)
6962 6963
        return new_param

6964 6965 6966
    __repr__ = __str__


Y
Yu Yang 已提交
6967
# program is a global instance.
Y
Yu Yang 已提交
6968 6969
_main_program_ = Program()
_startup_program_ = Program()
6970
_startup_program_._is_start_up_program_ = True
6971

6972

6973
def default_startup_program():
Y
Yu Yang 已提交
6974
    """
Y
yuyang18 已提交
6975 6976
    Get default/global startup program.

6977 6978
    The :code:`paddle.nn` function will append the initialization operators into startup program.
    The :code:`startup_program` will initialize the parameters by the OPs. 
T
tangwei12 已提交
6979

6980 6981
    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
6982

6983 6984
    Returns:
        Program: current default startup program.
6985

6986
    Returns type: 
6987 6988 6989 6990

    Examples:
        .. code-block:: python

6991
            import paddle
6992

6993
            paddle.enable_static()
6994 6995 6996 6997
            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
6998
    """
Y
Yu Yang 已提交
6999
    return _startup_program_
7000

7001

7002
def default_main_program():
Y
Yu Yang 已提交
7003
    """
7004
    This API can be used to get ``default main program`` which store the 
7005
    descriptions of Ops and tensors.
T
tangwei12 已提交
7006

7007
    For example ``z = paddle.add(x, y)`` will create a new ``add`` 
7008
    Op and a new ``z`` tensor, and they will be recorded in ``default main program`` . 
Y
yuyang18 已提交
7009

7010 7011
    The ``default main program`` is the default value for ``Program`` parameter in 
    a lot of APIs. For example, the :code:`Executor.run()` will execute the
Y
yuyang18 已提交
7012
    :code:`default_main_program` when the program is not specified.
7013

7014
    If you want to switch the ``default main program``, you can use :ref:`api_paddle_fluid_framework_program_guard` .
T
tangwei12 已提交
7015

Y
Yu Yang 已提交
7016
    Returns:
7017
        Program: A ``Program`` which holding the descriptions of OPs and tensors in the network.
7018 7019 7020 7021

    Examples:
        ..  code-block:: python

7022
            import paddle
7023

7024
            paddle.enable_static()
7025
            # Sample Network:
7026 7027 7028
            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)
7029

7030 7031 7032
            #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
7033
            print(paddle.static.default_main_program())
Y
Yu Yang 已提交
7034
    """
Y
Yu Yang 已提交
7035
    return _main_program_
Y
Yu Yang 已提交
7036 7037 7038 7039 7040


def switch_main_program(program):
    """
    Switch the main program to a new program.
7041

Y
Yu Yang 已提交
7042 7043 7044 7045 7046 7047 7048 7049 7050 7051 7052 7053 7054 7055
    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):
    """
7056
    Switch the startup program to a new program
Y
Yu Yang 已提交
7057 7058 7059 7060 7061 7062 7063 7064 7065 7066 7067 7068
    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
7069
@signature_safe_contextmanager
Y
Yu Yang 已提交
7070 7071
def program_guard(main_program, startup_program=None):
    """
7072 7073
    :api_attr: Static Graph

7074 7075 7076
    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.
7077

G
guofei 已提交
7078
    Args:
7079 7080
        main_program(Program): New main program inside ``with`` statement.
        startup_program(Program, optional): New startup program inside ``with`` 
G
guofei 已提交
7081 7082 7083 7084
            statement. :code:`None` means not changing startup program, 
            default_startup_program is still used.
            Default: None.

Y
Yu Yang 已提交
7085
    Examples:
7086
       .. code-block:: python
T
tangwei12 已提交
7087

7088
          import paddle
Y
yuyang18 已提交
7089

7090 7091 7092 7093 7094
          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')
7095
              hidden = paddle.static.nn.fc(x=data, size=10, activation='relu')
Y
yuyang18 已提交
7096 7097 7098

    Notes: The temporary :code:`Program` can be used if the user does not need
    to construct either of startup program or main program.
7099

Y
Yu Yang 已提交
7100
    Examples:
7101
       .. code-block:: python
Y
yuyang18 已提交
7102

7103
          import paddle
7104

7105 7106 7107 7108 7109
          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
7110

Y
Yu Yang 已提交
7111
    """
7112
    from .data_feeder import check_type
7113 7114
    check_type(main_program, 'main_program', Program,
               'paddle.static.program_guard')
Y
Yu Yang 已提交
7115 7116
    main_program = switch_main_program(main_program)
    if startup_program is not None:
7117
        check_type(startup_program, 'startup_program', Program,
7118
                   'paddle.static.program_guard')
7119 7120
        # Tag the program __is_start_up as True
        startup_program._is_start_up_program_ = True
Y
Yu Yang 已提交
7121
        startup_program = switch_startup_program(startup_program)
7122 7123 7124 7125 7126 7127
    try:
        yield
    finally:
        switch_main_program(main_program)
        if startup_program is not None:
            switch_startup_program(startup_program)
X
xuwei06 已提交
7128 7129


W
Wu Yi 已提交
7130
def _get_var(name, program=None):
X
xuwei06 已提交
7131
    """
Y
yuyang18 已提交
7132
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
7133

X
xuwei06 已提交
7134 7135 7136
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
7137
        If None, default_global_program() will be used.
X
xuwei06 已提交
7138 7139 7140 7141 7142 7143 7144

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
7145
    assert isinstance(program, Program)
X
xuwei06 已提交
7146 7147

    return program.global_block().var(name)
7148 7149


S
rename  
sneaxiy 已提交
7150
@signature_safe_contextmanager
L
lujun 已提交
7151 7152
def _dygraph_guard(tracer):
    global _dygraph_tracer_
7153
    tmp_tracer = _dygraph_tracer_
L
lujun 已提交
7154
    _dygraph_tracer_ = tracer
7155
    core._switch_tracer(tracer)
M
minqiyang 已提交
7156

7157 7158 7159
    try:
        yield
    finally:
7160 7161
        core._switch_tracer(tmp_tracer)
        _dygraph_tracer_ = tmp_tracer
P
Paddle CI 已提交
7162 7163


S
rename  
sneaxiy 已提交
7164
@signature_safe_contextmanager
L
lujun 已提交
7165
def _dygraph_place_guard(place):
7166 7167 7168
    global _global_expected_place_
    tmp_place = _global_expected_place_
    _global_expected_place_ = place
7169 7170
    _set_dygraph_tracer_expected_place(place)

7171 7172 7173
    try:
        yield
    finally:
7174
        _global_expected_place_ = tmp_place
J
Jiabin Yang 已提交
7175
        _set_dygraph_tracer_expected_place(_global_expected_place_)
7176 7177


7178 7179 7180 7181 7182 7183 7184 7185 7186 7187
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):
    """
7188 7189 7190
    
    Note:
        The API only supports static mode.
7191 7192 7193 7194

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

    Args:
7195 7196
        device(str|None): Specify the device to use in the context. It should be ``cpu``,
            ``gpu`` or ``gpu:x``, where ``x`` is the index of the GPUs. 
7197 7198 7199 7200 7201 7202 7203
            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:
7204
    
7205
        .. code-block:: python
7206 7207
            
            # required: gpu
Z
Zhang Ting 已提交
7208
            import paddle
7209

Z
Zhang Ting 已提交
7210 7211 7212
            paddle.enable_static()
            support_gpu = paddle.is_compiled_with_cuda()
            place = paddle.CPUPlace()
7213
            if support_gpu:
Z
Zhang Ting 已提交
7214
                place = paddle.CUDAPlace(0)
7215 7216

            # if GPU is supported, the three OPs below will be automatically assigned to CUDAPlace(0)
Z
Zhang Ting 已提交
7217 7218 7219
            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)
7220

Z
Zhang Ting 已提交
7221
            with paddle.static.device_guard("cpu"):
7222
                # Ops created here will be placed on CPUPlace
Z
Zhang Ting 已提交
7223 7224
                shape = paddle.slice(shape, axes=[0], starts=[0], ends=[4])
            with paddle.static.device_guard('gpu'):
7225
                # if GPU is supported, OPs created here will be placed on CUDAPlace(0), otherwise on CPUPlace
Z
Zhang Ting 已提交
7226
                out = paddle.reshape(data1, shape=shape)
7227

Z
Zhang Ting 已提交
7228 7229
            exe = paddle.static.Executor(place)
            exe.run(paddle.static.default_startup_program())
7230 7231 7232
            result = exe.run(fetch_list=[out])
    """

7233 7234 7235 7236 7237
    index = None
    if device and ':' in device:
        device, index = device.split(':')
        if device == 'cpu':
            raise ValueError("Should not set device id for cpu.")
7238
    if device not in ['cpu', 'gpu', 'npu', 'xpu', 'mlu', '', None]:
7239
        raise ValueError(
7240
            "The Attr(device) should be 'cpu' 'npu' 'xpu' 'mlu' or 'gpu', and it can also be empty string or None "
7241
            "when there is no need to specify device. But received %s" % device)
7242 7243
    if index:
        device = ":".join([device, index])
7244
    pre_device = switch_device(device)
7245 7246 7247 7248
    try:
        yield
    finally:
        switch_device(pre_device)
G
guofei 已提交
7249 7250


7251 7252 7253 7254 7255 7256 7257 7258 7259 7260 7261 7262 7263 7264 7265 7266 7267 7268 7269 7270 7271 7272 7273 7274 7275 7276 7277 7278 7279 7280 7281
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
    """
    assert not _non_static_mode(
    ), "cuda_graph_guard only works under static mode"
    assert core.is_compiled_with_cuda(
    ), "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
7282 7283 7284
def set_flags(flags):
    """
    This function sets the GFlags value in Paddle.
7285
    For FLAGS please refer to :ref:`en_guides_flags_flags`
G
guofei 已提交
7286 7287 7288 7289 7290 7291 7292

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

    Examples:
            .. code-block:: python

7293 7294
                import paddle
                paddle.set_flags({'FLAGS_eager_delete_tensor_gb': 1.0})
G
guofei 已提交
7295 7296 7297 7298
    """
    if not isinstance(flags, dict):
        raise TypeError('flags in set_flags should be a dict')
    for key, value in flags.items():
7299 7300
        if _global_flags().is_public(key):
            _global_flags()[key] = value
G
guofei 已提交
7301 7302 7303 7304 7305 7306 7307 7308
        else:
            raise ValueError(
                "Flag %s cannot set its value through this function." % (key))


def get_flags(flags):
    """
    This function gets the GFlags value in Paddle.
7309
    For FLAGS please refer to :ref:`en_guides_flags_flags`
G
guofei 已提交
7310 7311 7312 7313 7314 7315 7316 7317 7318 7319

    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

7320
            import paddle
G
guofei 已提交
7321 7322

            flags = ['FLAGS_eager_delete_tensor_gb', 'FLAGS_check_nan_inf']
7323
            res = paddle.get_flags(flags)
G
guofei 已提交
7324 7325 7326 7327 7328 7329
            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:
7330 7331
            if (_global_flags().is_public(key)):
                value = _global_flags()[key]
G
guofei 已提交
7332 7333 7334 7335 7336 7337 7338
                temp = {key: value}
                flags_value.update(temp)
            else:
                raise ValueError(
                    'Flag %s cannot get its value through this function.' %
                    (key))
    elif isinstance(flags, str):
7339 7340
        if (_global_flags().is_public(flags)):
            value = _global_flags()[flags]
G
guofei 已提交
7341 7342 7343 7344 7345 7346 7347 7348
            temp = {flags: value}
            flags_value.update(temp)
        else:
            raise ValueError(
                'Flag %s cannot get its value through this function.' % (flags))
    else:
        raise TypeError('Flags in get_flags should be a list, tuple or string.')
    return flags_value
7349 7350 7351 7352 7353 7354 7355


def _get_paddle_place(place):
    "convert the string to paddle Place"
    if place is None:
        return place
    if isinstance(place, (core.Place, core.XPUPlace, core.CPUPlace,
7356
                          core.CUDAPinnedPlace, core.CUDAPlace, core.NPUPlace,
7357
                          core.IPUPlace, core.MLUPlace, core.CustomPlace)):
7358 7359 7360 7361 7362 7363 7364 7365 7366
        return place

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

    place = place.lower()
    if (place == "cpu"):
        return core.CPUPlace()
7367

7368 7369 7370
    if (place == "device"):
        return core.Place()

7371
    # GPU
7372 7373 7374 7375 7376 7377 7378 7379 7380 7381 7382 7383 7384 7385 7386
    avaliable_gpu_place = re.match(r'gpu:\d+', place)
    if place == "gpu_pinned" or place == "gpu" or avaliable_gpu_place:
        if not core.is_compiled_with_cuda():
            raise ValueError(
                "The device should not be {}, since PaddlePaddle is " \
                "not compiled with CUDA".format(avaliable_gpu_place))
        if place == "gpu_pinned":
            return core.CUDAPinnedPlace()
        elif place == "gpu":
            return core.CUDAPlace(0)
        else:
            place_info_list = place.split(':', 1)
            device_id = place_info_list[1]
            device_id = int(device_id)
            return core.CUDAPlace(device_id)
7387 7388

    # XPU
7389 7390 7391 7392 7393 7394 7395 7396 7397 7398
    avaliable_xpu_place = re.match(r'xpu:\d+', place)
    if avaliable_xpu_place:
        if not core.is_compiled_with_xpu():
            raise ValueError(
                "The device should not be {}, since PaddlePaddle is " \
                "not compiled with XPU".format(avaliable_xpu_place))
        place_info_list = place.split(':', 1)
        device_id = place_info_list[1]
        device_id = int(device_id)
        return core.XPUPlace(device_id)
7399 7400 7401 7402 7403 7404 7405 7406 7407 7408 7409 7410 7411

    # NPU
    avaliable_npu_place = re.match(r'npu:\d+', place)
    if avaliable_npu_place:
        if not core.is_compiled_with_npu():
            raise ValueError(
                "The device should not be {}, since PaddlePaddle is " \
                "not compiled with NPU".format(avaliable_npu_place))
        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
7412 7413 7414 7415 7416 7417 7418 7419 7420 7421 7422 7423
    # IPU
    avaliable_ipu_place = re.match(r'ipu:\d+', place)
    if avaliable_ipu_place:
        if not core.is_compiled_with_ipu():
            raise ValueError(
                "The device should not be {}, since PaddlePaddle is " \
                "not compiled with IPU".format(avaliable_ipu_place))
        place_info_list = place.split(':', 1)
        device_id = place_info_list[1]
        device_id = int(device_id)
        return core.IPUPlace(device_id)

7424 7425 7426 7427 7428 7429 7430 7431 7432 7433 7434 7435
    # MLU
    avaliable_mlu_place = re.match(r'mlu:\d+', place)
    if avaliable_mlu_place:
        if not core.is_compiled_with_mlu():
            raise ValueError(
                "The device should not be {}, since PaddlePaddle is " \
                "not compiled with MLU".format(avaliable_mlu_place))
        place_info_list = place.split(':', 1)
        device_id = place_info_list[1]
        device_id = int(device_id)
        return core.MLUPlace(device_id)

7436
    raise ValueError(
7437 7438
        "Paddle supports CPUPlace, CUDAPlace,CUDAPinnedPlace, XPUPlace, IPUPlace, MLUPlace and NPUPlace, but received {}."
        .format(place))
7439 7440 7441 7442 7443 7444 7445 7446 7447 7448 7449 7450 7451


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