framework.py 192.4 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
W
WangZhen 已提交
19
from collections import Iterable
Q
qiaolongfei 已提交
20
import contextlib
21
from .wrapped_decorator import signature_safe_contextmanager, wrap_decorator
P
peizhilin 已提交
22
import os
F
fengjiayi 已提交
23
import re
24
import traceback
25
import six
26

Y
Yu Yang 已提交
27
import numpy as np
28
import subprocess
S
sneaxiy 已提交
29
import multiprocessing
30
import sys
31
import logging
M
minqiyang 已提交
32
from .. import compat as cpt
33
from .proto import framework_pb2
34 35

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

41
__all__ = [
42 43 44 45
    'Program',
    'default_startup_program',
    'default_main_program',
    'program_guard',
46
    'name_scope',
S
sneaxiy 已提交
47 48 49
    'cuda_places',
    'cpu_places',
    'cuda_pinned_places',
L
lujun 已提交
50
    'in_dygraph_mode',
C
chengduo 已提交
51
    'is_compiled_with_cuda',
52
    'is_compiled_with_xpu',
53
    'Variable',
54
    'ComplexVariable',
55
    'load_op_library',
56
    'require_version',
57
    'device_guard',
G
guofei 已提交
58 59
    'set_flags',
    'get_flags',
60
]
Y
Yu Yang 已提交
61

Q
qiaolongfei 已提交
62 63 64 65
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
66 67
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()

L
lujun 已提交
68
_dygraph_tracer_ = None
69
_global_expected_place_ = None
70
_current_device = None
71 72
global_prog_seed = 0

73

74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180
def require_version(min_version, max_version=None):
    """
        Check if the installed version of PaddlePaddle is in [min_version, max_version],
        if the installed version is lower than ``min_version`` or higher than ``max_version``,
        an exception will be thrown, NO returns if the installed version is satisfied.

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

        Returns:
            None.

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

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

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

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

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

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

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

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

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

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

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

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

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


L
lujun 已提交
181
def in_dygraph_mode():
L
lujun 已提交
182
    """
183 184 185 186
    :alias_main: paddle.in_dygraph_mode
	:alias: paddle.in_dygraph_mode
	:old_api: paddle.fluid.framework.in_dygraph_mode

Y
Youwei Song 已提交
187
    This function checks whether the program runs in dynamic graph mode or not.
188 189 190
    You can enter dynamic graph mode with :ref:`api_fluid_dygraph_guard` api,
    or enable and disable dynamic graph mode with :ref:`api_fluid_dygraph_enable`
    and :ref:`api_fluid_dygraph_disable` api .
L
lujun 已提交
191 192

    Returns:
Y
Youwei Song 已提交
193
        bool: Whether the program is running in dynamic graph mode.
L
lujun 已提交
194 195 196 197

    Examples:
        .. code-block:: python

198
            import paddle.fluid as fluid
L
lujun 已提交
199

200 201 202 203
            fluid.enable_dygraph()  # Now we are in dygragh mode
            print(fluid.in_dygraph_mode())  # True
            fluid.disable_dygraph()
            print(fluid.in_dygraph_mode())  # False
L
lujun 已提交
204
    """
L
lujun 已提交
205
    return _dygraph_tracer_ is not None
206 207


208 209 210
def _dygraph_not_support_(func):
    def __impl__(*args, **kwargs):
        assert not in_dygraph_mode(
211
        ), "We don't support %s in imperative mode" % func.__name__
212 213 214 215 216 217 218 219
        return func(*args, **kwargs)

    return __impl__


def _dygraph_only_(func):
    def __impl__(*args, **kwargs):
        assert in_dygraph_mode(
220
        ), "We Only support %s in imperative mode, please use fluid.dygraph.guard() as context to run it in imperative Mode" % func.__name__
221 222 223 224 225
        return func(*args, **kwargs)

    return __impl__


226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241
# 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
# same base class. 
def _fake_interface_only_(func):
    def __impl__(*args, **kwargs):
        raise AssertionError(
            "'%s' should be called by imperative Varible in imperative mode, please use fluid.dygraph.guard() as context to run it in imperative mode"
            % func.__name__)

    return __impl__


242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
# 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 
# introducing compatibility issues, add this decorator
# NOTE(chenweihang): not using `wrap_decorator` here is because `wrap_decorator` will
# move kwargs to args, which doesn't work in this decorate case
def deprecate_stat_dict(func):
    @functools.wraps(func)
    def wrapper(*args, **kwargs):
        if 'stat_dict' in kwargs:
            warnings.warn(
                "The argument `stat_dict` has deprecated, please change it to `state_dict`.",
                DeprecationWarning)
            kwargs['state_dict'] = kwargs['stat_dict']
            kwargs.pop('stat_dict')
        return func(*args, **kwargs)

    return wrapper


261 262
dygraph_not_support = wrap_decorator(_dygraph_not_support_)
dygraph_only = wrap_decorator(_dygraph_only_)
263
fake_interface_only = wrap_decorator(_fake_interface_only_)
264 265


L
lujun 已提交
266 267
def _dygraph_tracer():
    return _dygraph_tracer_
268

W
Wu Yi 已提交
269

M
minqiyang 已提交
270
def _current_expected_place():
271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290
    global _global_expected_place_
    if _global_expected_place_ is None:
        if core.is_compiled_with_cuda():
            _global_expected_place_ = core.CUDAPlace(0)
        else:
            _global_expected_place_ = core.CPUPlace()

    return _global_expected_place_


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


def _set_expected_place(place):
    global _global_expected_place_
    _global_expected_place_ = place
    _set_dygraph_tracer_expected_place(place)
M
minqiyang 已提交
291 292


L
Leo Chen 已提交
293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
# TODO(zhiqiu): remove this function.
def _var_base_to_np(var_base):
    """	
    convert VarBase tp numpy	
    	
    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
310
def _cpu_num():
311
    if "CPU_NUM" not in os.environ.keys():
C
chengduo 已提交
312 313 314 315 316 317 318 319
        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
320
        os.environ['CPU_NUM'] = str(1)
321
    cpu_num = os.environ.get('CPU_NUM')
C
chengduo 已提交
322 323 324 325 326 327 328 329 330 331
    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
332 333


334 335 336 337 338 339 340 341 342 343 344 345 346 347 348
def is_compiled_with_xpu():
    """
    Whether this whl package can be used to run the model on XPU.

    Returns (bool): support xpu or not.

    Examples:
        .. code-block:: python

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


C
chengduo 已提交
349 350 351 352 353 354 355 356 357 358 359 360 361 362 363
def is_compiled_with_cuda():
    """
    Whether this whl package can be used to run the model on GPU.

    Returns (bool): support gpu or not.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid
            support_gpu = fluid.is_compiled_with_cuda()
    """
    return core.is_compiled_with_cuda()


S
sneaxiy 已提交
364
def cuda_places(device_ids=None):
L
lujun 已提交
365
    """
366 367 368 369 370
    **Note**:
        For multi-card tasks, please use `FLAGS_selected_gpus` environment variable to set the visible GPU device.
        The next version will fix the problem with `CUDA_VISIBLE_DEVICES` environment variable.

    This function creates a list of :code:`fluid.CUDAPlace` objects.
S
add doc  
sneaxiy 已提交
371 372

    If :code:`device_ids` is None, environment variable of
373
    :code:`FLAGS_selected_gpus` would be checked first. For example, if
S
add doc  
sneaxiy 已提交
374 375 376
    :code:`FLAGS_selected_gpus=0,1,2`, the returned list would
    be [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
    If :code:`FLAGS_selected_gpus` is not set, all visible
377
    gpu places would be returned according to the :code:`CUDA_VISIBLE_DEVICES` environment variable.
S
add doc  
sneaxiy 已提交
378 379

    If :code:`device_ids` is not None, it should be the device
380
    ids of GPUs. For example, if :code:`device_ids=[0,1,2]`,
S
add doc  
sneaxiy 已提交
381 382 383
    the returned list would be 
    [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
    
384 385
    Parameters:
        device_ids (list or tuple of int, optional): list of GPU device ids.
S
add doc  
sneaxiy 已提交
386 387

    Returns:
388
        list of fluid.CUDAPlace: Created GPU place list.
L
lujun 已提交
389 390 391 392

    Examples:
        .. code-block:: python

393
            import paddle.fluid as fluid
L
lujun 已提交
394 395 396
            cuda_places = fluid.cuda_places()

    """
S
sneaxiy 已提交
397 398 399
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_ids is None:
C
chengduo 已提交
400
        device_ids = _cuda_ids()
S
sneaxiy 已提交
401 402 403 404 405 406
    elif not isinstance(device_ids, (list, tuple)):
        device_ids = [device_ids]
    return [core.CUDAPlace(dev_id) for dev_id in device_ids]


def cpu_places(device_count=None):
L
lujun 已提交
407
    """
408
    This function creates a list of :code:`fluid.CPUPlace` objects, and returns the created list.
S
add doc  
sneaxiy 已提交
409 410 411
    
    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
C
chengduo 已提交
412 413
    If :code:`CPU_NUM` is not set, the default value is 1,
    i.e. CPU_NUM=1.
414 415
    :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
416

417 418
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
419 420

    Returns:
421
        list of fluid.CPUPlace: Created list of CPU places.
L
lujun 已提交
422 423 424 425

    Examples:
        .. code-block:: python

426
            import paddle.fluid as fluid
L
lujun 已提交
427 428 429
            cpu_places = fluid.cpu_places()
    """

S
sneaxiy 已提交
430 431 432 433 434 435
    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
436
    """
437
    This function creates a list of :code:`fluid.CUDAPinnedPlace` objects.
S
add doc  
sneaxiy 已提交
438 439 440

    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
441 442 443 444
    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
445

446 447
    Parameters:
        device_count (int, optional): device number. Default: None.
S
add doc  
sneaxiy 已提交
448 449

    Returns:
450
        list of fluid.CUDAPinnedPlace: Created list of CUDA pinned places.
L
lujun 已提交
451 452 453 454

    Examples:
        .. code-block:: python

455
            import paddle.fluid as fluid
L
lujun 已提交
456 457 458 459 460
            cuda_pinned_places_cpu_num = fluid.cuda_pinned_places()
            # or
            cuda_pinned_places = fluid.cuda_pinned_places(1)

    """
S
sneaxiy 已提交
461 462 463
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_count is None:
464 465
        device_count = len(_cuda_ids())
    return [core.CUDAPinnedPlace()] * device_count
S
sneaxiy 已提交
466 467


468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493
class NameScope(object):
    def __init__(self, name="", parent=None):
        self._children = dict()
        self._name = name
        self._parent = parent

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

    def parent(self):
        return self._parent

    def name(self):
        return self._name


_name_scope = NameScope()


S
rename  
sneaxiy 已提交
494
@signature_safe_contextmanager
495 496
def name_scope(prefix=None):
    """
497 498
    :api_attr: Static Graph

499 500
    Generate hierarchical name prefix for the operators.

T
Tao Luo 已提交
501 502 503
    Note: 
        This should only used for debugging and visualization purpose.
        Don't use it for serious analysis such as graph/program transformations.
504 505

    Args:
T
Tao Luo 已提交
506
        prefix(str, optional): prefix. Default is none.
507 508 509

    Examples:
        .. code-block:: python
T
Tink_Y 已提交
510

511
          import paddle.fluid as fluid
512
          with fluid.name_scope("s1"):
T
Tao Luo 已提交
513 514 515 516 517 518
             a = fluid.data(name='data', shape=[None, 1], dtype='int32')
             b = a + 1
             with fluid.name_scope("s2"):
                c = b * 1
             with fluid.name_scope("s3"):
                d = c / 1
519
          with fluid.name_scope("s1"):
T
Tao Luo 已提交
520
                f = fluid.layers.pow(d, 2.0)
521
          with fluid.name_scope("s4"):
T
Tao Luo 已提交
522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540
                g = f - 1

          # Op are created in the default main program.  
          for op in fluid.default_main_program().block(0).ops:
              # 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/'
541 542
    """
    # TODO(panyx0718): Only [0-9a-z].
543
    # in dygraph we don't need namescope since it will cause mem leak
L
Leo Chen 已提交
544 545 546
    if in_dygraph_mode():
        yield
    else:
T
tianshuo78520a 已提交
547
        assert prefix, "namescope prefix can not be empty."
548 549
        global _name_scope
        _name_scope = _name_scope.child(prefix)
550 551 552 553
        try:
            yield
        finally:
            _name_scope = _name_scope.parent()
554 555 556 557 558 559 560 561 562 563 564 565


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
566 567 568
def generate_control_dev_var_name():
    import random
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
569 570 571 572


def grad_var_name(var_name):
    """
573 574
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
575 576 577
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
578

579
def convert_np_dtype_to_dtype_(np_dtype):
580 581
    """
    Convert the data type in numpy to the data type in Paddle
582

583
    Args:
584
        np_dtype(np.dtype): the data type in numpy.
585

586 587
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
588 589

    """
590 591
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
592
        return core.VarDesc.VarType.FP32
593
    elif dtype == np.float64:
594
        return core.VarDesc.VarType.FP64
595
    elif dtype == np.float16:
596
        return core.VarDesc.VarType.FP16
597
    elif dtype == np.int32:
598
        return core.VarDesc.VarType.INT32
599
    elif dtype == np.int16:
600
        return core.VarDesc.VarType.INT16
601
    elif dtype == np.int64:
602
        return core.VarDesc.VarType.INT64
603
    elif dtype == np.bool:
604
        return core.VarDesc.VarType.BOOL
605 606
    elif dtype == np.uint16:
        return core.VarDesc.VarType.INT16
607 608
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
609 610
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
611
    else:
M
minqiyang 已提交
612
        raise ValueError("Not supported numpy dtype %s" % dtype)
613 614 615


def dtype_is_floating(dtype):
616 617 618
    """
    Check the data type is floating or not.
    Args:
619
        dtype(np.dtype|core.VarDesc.VarType): data type.
620 621 622 623 624
            Could be numpy format or Paddle format

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

    """
625
    if not isinstance(dtype, core.VarDesc.VarType):
626 627
        dtype = convert_np_dtype_to_dtype_(dtype)

628 629 630 631
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
632 633


Y
Yang Yang(Tony) 已提交
634
def _debug_string_(proto, throw_on_error=True):
635 636 637 638 639 640 641 642 643 644 645
    """
    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
646
    error_fields = list()
Y
Yang Yang(Tony) 已提交
647
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
648 649
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
650 651 652
    return proto.__str__()


653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709
def _varbase_creator(type=core.VarDesc.VarType.LOD_TENSOR,
                     name=None,
                     shape=None,
                     dtype=None,
                     persistable=None,
                     **kwargs):
    if dtype is not None:
        if not isinstance(dtype, core.VarDesc.VarType):
            dtype = convert_np_dtype_to_dtype_(dtype)

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


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


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


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

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

    Returns:
        Sliced variable
    """

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

    decrease_axis = []
    slice_axis = []
    slice_start = []
    slice_end = []
    slice_step = []
    use_strided_slice = False
    reverse_axis = []
710
    target_block = default_main_program().current_block()
711 712 713 714 715 716 717 718 719 720 721

    def fill_constant(shape, value, force_cpu=False, out=None):
        var.block.append_op(
            type='fill_constant',
            inputs={},
            outputs={'Out': [out]},
            attrs={
                'shape': shape,
                'dtype': out.dtype,
                'value': float(value),
                'force_cpu': force_cpu
722
            })
723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761
        out.stop_gradient = True
        return out

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

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

            if step is None:
                step = 1

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

            if start is None:
                start = 0

            if end is None:
                end = 10000000

            if step != 1:
                use_strided_slice = True

            slice_axis.append(dim)
            slice_start.append(start)
            slice_end.append(end)
            slice_step.append(step)
        else:
            decrease_axis.append(dim)
            slice_axis.append(dim)
            slice_start.append(slice_item)
            slice_step.append(1)
            if isinstance(slice_item, Variable):
762
                temp_1 = var.block.create_var(dtype=slice_item.dtype)
763
                fill_constant([1], 1, force_cpu=True, out=temp_1)
764
                temp_end = target_block.create_var(dtype=slice_item.dtype)
765
                target_block.append_op(
766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804
                    type='elementwise_add',
                    inputs={'X': slice_item,
                            'Y': temp_1},
                    outputs={'Out': temp_end},
                    attrs={'axis': -1})
                slice_end.append(temp_end)
            else:
                slice_end.append(slice_item + 1
                                 if slice_item != -1 else 10000000)

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

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

    inputs = {'Input': [var]}
    attrs = {
        'axes': slice_axis,
        'starts': [],
        'ends': [],
        'decrease_axis': decrease_axis
    }
    if (use_strided_slice == True):
        attrs['strides'] = []
    infer_flags = list(1 for i in range(len(slice_axis)))
L
Leo Chen 已提交
805

806
    # starts
L
Leo Chen 已提交
807
    if contain_var(slice_start):
808 809 810 811 812 813 814 815
        inputs['StartsTensorList'] = get_new_list_tensor(slice_start)
        for i, dim in enumerate(slice_start):
            if isinstance(dim, Variable):
                attrs['starts'].append(-1)
                infer_flags[i] = -1
            else:
                attrs['starts'].append(dim)
    else:
L
Leo Chen 已提交
816 817 818 819
        attrs['starts'] = slice_start

    # ends
    if contain_var(slice_end):
820 821 822 823 824 825 826
        inputs['EndsTensorList'] = get_new_list_tensor(slice_end)
        for i, dim in enumerate(slice_end):
            if isinstance(dim, Variable):
                attrs['ends'].append(-1)
                infer_flags[i] = -1
            else:
                attrs['ends'].append(dim)
L
Leo Chen 已提交
827 828 829
    else:
        attrs['ends'] = slice_end

830 831
    # strides
    if use_strided_slice == True:
L
Leo Chen 已提交
832
        if contain_var(slice_step):
833 834 835 836 837 838 839
            inputs['StridesTensorList'] = get_new_list_tensor(slice_step)
            for i, dim in enumerate(slice_step):
                if isinstance(dim, Variable):
                    attrs['strides'].append(-1)
                    infer_flags[i] = -1
                else:
                    attrs['strides'].append(dim)
L
Leo Chen 已提交
840 841
        else:
            attrs['strides'] = slice_step
842 843 844 845 846 847
    # infer_flags
    attrs['infer_flags'] = infer_flags

    out = var
    if use_strided_slice == False and len(slice_axis) > 0:
        # append slice_op here
848
        slice_out_var = target_block.create_var(
849 850 851
            name=unique_name.generate_with_ignorable_key(var.name + "_slice"),
            dtype=var.dtype)

852
        target_block.append_op(
853 854 855 856 857 858 859
            type="slice",
            inputs=inputs,
            outputs={'Out': [slice_out_var]},
            attrs=attrs)

        out = slice_out_var
    elif use_strided_slice == True and len(slice_axis) > 0:
860
        strided_slice_out_var = target_block.create_var(
861 862 863
            name=unique_name.generate_with_ignorable_key(var.name +
                                                         "_strided_slice"),
            dtype=var.dtype)
864
        target_block.append_op(
865 866 867 868 869 870 871 872
            type="strided_slice",
            inputs=inputs,
            outputs={'Out': [strided_slice_out_var]},
            attrs=attrs)

        out = strided_slice_out_var

    if len(reverse_axis) > 0:
873
        reverse_out_var = target_block.create_var(
874 875 876
            name=unique_name.generate_with_ignorable_key(var.name +
                                                         "_slice_reverse"),
            dtype=var.dtype)
877
        target_block.append_op(
878 879 880 881 882 883 884 885 886 887 888
            type="reverse",
            inputs={'X': out},
            outputs={'Out': [reverse_out_var]},
            attrs={'axis': reverse_axis})

        out = reverse_out_var

    return out


@six.add_metaclass(VariableMetaClass)
X
Xin Pan 已提交
889
class Variable(object):
890
    """
J
Jiabin Yang 已提交
891
    **Notes**:
892
        **The constructor of Variable should not be invoked directly.**
J
Jiabin Yang 已提交
893

894 895
        **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
896 897 898
        **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
899
    cases, variables are used for holding different kinds of data or training
J
Jiabin Yang 已提交
900 901
    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.
902

903
    There are many kinds of variables. Each kind of them has its own attributes
J
Jiabin Yang 已提交
904
    and usages. Please refer to the `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_ for details.
905

T
tianshuo78520a 已提交
906
    Most of a Variable's member variables can be set to be None. It mean
907
    it is not available or will be specified later.
908

909
    Examples:
910 911
        In Static Graph Mode:

912 913
        .. code-block:: python

914
            import paddle.fluid as fluid
915
            cur_program = fluid.Program()
916 917 918 919
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
J
Jiabin Yang 已提交
920
        In `Dygraph <../../user_guides/howto/dygraph/DyGraph.html>`_  Mode:
921 922 923 924 925 926 927 928 929

        .. 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))

930 931
    """

Y
Yu Yang 已提交
932 933
    def __init__(self,
                 block,
Y
Yu Yang 已提交
934
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
935 936 937 938
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
939
                 capacity=None,
Q
QI JUN 已提交
940
                 persistable=None,
F
fengjiayi 已提交
941
                 error_clip=None,
Y
Yu Yang 已提交
942
                 stop_gradient=False,
F
fengjiayi 已提交
943
                 is_data=False,
H
Huihuang Zheng 已提交
944
                 need_check_feed=False,
H
hong 已提交
945
                 belong_to_optimizer=False,
Y
Yu Yang 已提交
946
                 **kwargs):
Y
Yu Yang 已提交
947 948
        self.block = block
        if name is None:
Y
Yu Yang 已提交
949
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
950

Y
Yu Yang 已提交
951
        if dtype is not None:
952
            if not isinstance(dtype, core.VarDesc.VarType):
953
                dtype = convert_np_dtype_to_dtype_(dtype)
954

H
hong 已提交
955 956
        self.belong_to_optimizer = belong_to_optimizer

957 958 959 960 961
        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))
962

963 964 965
        if self.desc is None:
            self.desc = self.block.desc.var(cpt.to_bytes(name))
            is_new_var = True
966

967 968 969 970 971 972 973
        if is_new_var:
            self.desc.set_type(type)
        elif self.desc.type() != type:
            raise ValueError("Variable {0} has been created before. The "
                             "previous type is {1}; the new type is {2}. They"
                             " are not matched".format(self.name,
                                                       self.desc.type(), type))
974

975
        if shape is not None:
976
            if is_new_var:
977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017
                self.desc.set_shape(shape)
            else:
                old_shape = self.shape
                shape = tuple(shape)
                if shape != old_shape:
                    raise ValueError(
                        "Variable {0} has been created before. the previous "
                        "shape is {1}; the new shape is {2}. They are not "
                        "matched.".format(self.name, old_shape, shape))
        if dtype is not None:
            if is_new_var:
                self.desc.set_dtype(dtype)
            else:
                old_dtype = self.dtype
                if dtype != old_dtype:
                    raise ValueError("Variable {0} has been created before. "
                                     "The previous data type is {1}; the new "
                                     "data type is {2}. They are not "
                                     "matched.".format(self.name, old_dtype,
                                                       dtype))

        if lod_level is not None:
            if is_new_var:
                self.desc.set_lod_level(lod_level)
            else:
                if lod_level != self.lod_level:
                    raise ValueError("Variable {0} has been created before. "
                                     "The previous lod_level is {1}; the new "
                                     "lod_level is {2}. They are not "
                                     "matched".format(self.name, self.lod_level,
                                                      lod_level))
        if persistable is not None:
            if is_new_var:
                self.desc.set_persistable(persistable)
            else:
                if persistable != self.persistable:
                    raise ValueError(
                        "Variable {0} has been created before."
                        "The previous persistable is {1}; the new "
                        "persistable is {2}. They are not matched".format(
                            self.name, self.persistable, persistable))
1018

1019 1020
        if need_check_feed and is_new_var:
            self.desc.set_need_check_feed(need_check_feed)
H
Huihuang Zheng 已提交
1021

1022 1023 1024 1025 1026 1027 1028
        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
1029

1030 1031 1032 1033
        self.block.vars[name] = self
        self.op = None
        self._stop_gradient = stop_gradient
        self.is_data = is_data
Y
Yu Yang 已提交
1034

1035
    @fake_interface_only
1036 1037
    def detach(self):
        """
J
Jiabin Yang 已提交
1038
        **Notes**:
T
tianshuo78520a 已提交
1039
            **This API is ONLY available in Dygraph mode**
1040

1041
        Returns a new Variable, detached from the current graph.
1042

1043
        Returns:
J
Jiabin Yang 已提交
1044
             ( :ref:`api_guide_Variable_en` | dtype is same as current Variable): The detached Variable.
1045

1046

1047 1048 1049 1050 1051
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1052
                from paddle.fluid.dygraph import Linear
1053 1054 1055 1056
                import numpy as np

                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
                with fluid.dygraph.guard():
1057
                    linear = Linear(32, 64)
1058
                    data = to_variable(data)
1059
                    x = linear(data)
1060 1061 1062
                    y = x.detach()

        """
1063
        pass
1064

1065
    @fake_interface_only
1066
    def numpy(self):
1067
        """
J
Jiabin Yang 已提交
1068
        **Notes**:
T
tianshuo78520a 已提交
1069
            **This API is ONLY available in Dygraph mode**
1070

J
Jiabin Yang 已提交
1071
        Returns a numpy array shows the value of current :ref:`api_guide_Variable_en`
1072 1073 1074 1075 1076

        Returns:
            ndarray: The numpy value of current Variable.

        Returns type:
J
Jiabin Yang 已提交
1077
            ndarray: dtype is same as current Variable
1078 1079 1080 1081 1082 1083

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1084
                from paddle.fluid.dygraph import Linear
1085 1086 1087 1088
                import numpy as np

                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
                with fluid.dygraph.guard():
1089
                    linear = Linear(32, 64)
1090
                    data = to_variable(data)
1091
                    x = linear(data)
1092 1093 1094
                    print(x.numpy())

        """
1095
        pass
1096

1097
    @fake_interface_only
1098 1099
    def set_value(self, value):
        """
J
Jiabin Yang 已提交
1100
        **Notes**:
T
tianshuo78520a 已提交
1101
            **This API is ONLY available in Dygraph mode**
J
Jiabin Yang 已提交
1102

1103 1104 1105 1106 1107 1108 1109 1110 1111 1112
        Set a new value for this Variable.

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

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                from paddle.fluid.dygraph.base import to_variable
1113
                from paddle.fluid.dygraph import Linear
1114 1115
                import numpy as np

1116
                data = np.ones([3, 1024], dtype='float32')
1117
                with fluid.dygraph.guard():
1118
                    linear = fluid.dygraph.Linear(1024, 4)
1119
                    t = to_variable(data)
1120
                    linear(t)  # call with default weight
1121
                    custom_weight = np.random.randn(1024, 4).astype("float32")
1122 1123
                    linear.weight.set_value(custom_weight)  # change existing weight
                    out = linear(t)  # call with different weight
1124 1125

        """
1126
        pass
1127

1128
    @fake_interface_only
1129
    def backward(self, retain_graph=False):
1130
        """
J
Jiabin Yang 已提交
1131
        **Notes**:
T
tianshuo78520a 已提交
1132
            **This API is ONLY available in Dygraph mode**
1133

1134
        Run backward of current Graph which starts from current Tensor.
1135

J
Jiabin Yang 已提交
1136
        Args:
1137 1138 1139 1140
            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.
1141

J
Jiabin Yang 已提交
1142 1143
        Returns:
            NoneType: None
1144 1145 1146 1147 1148

        Examples:
            .. code-block:: python

                import numpy as np
1149 1150
                import paddle
                paddle.disable_static()
1151 1152

                x = np.ones([2, 2], np.float32)
1153 1154 1155 1156 1157 1158 1159 1160 1161 1162
                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)
                ret = paddle.sums(inputs)
                loss = paddle.reduce_sum(ret)
                loss.backward()
1163 1164

        """
1165
        pass
1166

1167
    @fake_interface_only
1168
    def gradient(self):
1169
        """
J
Jiabin Yang 已提交
1170
        **Notes**:
T
tianshuo78520a 已提交
1171
            **This API is ONLY available in Dygraph mode**
1172 1173 1174

        Get the Gradient of Current Variable

J
Jiabin Yang 已提交
1175
        Returns:
1176
            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.
1177 1178 1179 1180 1181 1182 1183

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

1184
                # example1: return ndarray
1185 1186 1187 1188 1189 1190 1191 1192 1193
                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)
1194
                    loss2.backward()
1195 1196
                    print(loss2.gradient())

1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209
                # 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())

1210
        """
1211
        pass
1212

1213
    @fake_interface_only
1214
    def clear_gradient(self):
1215
        """
J
Jiabin Yang 已提交
1216
        **Notes**:
T
tianshuo78520a 已提交
1217
            **1. This API is ONLY available in Dygraph mode**
J
Jiabin Yang 已提交
1218 1219

            **2. Use it only Variable has gradient, normally we use this for Parameters since other temporal Variable will be deleted by Python's GC**
1220

J
Jiabin Yang 已提交
1221
        Clear  (set to ``0`` ) the Gradient of Current Variable
1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239

        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)
1240
                    loss2.backward()
1241 1242 1243 1244 1245
                    print(loss2.gradient())
                    loss2.clear_gradient()
                    print("After clear {}".format(loss2.gradient()))

        """
1246
        pass
X
Xin Pan 已提交
1247

1248
    def __str__(self):
1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292
        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

                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(new_variable._to_readable_code())
        """
        if self.type == core.VarDesc.VarType.SELECTED_ROWS or self.type == core.VarDesc.VarType.LOD_TENSOR:
            var_str = "{name} : fluid.{type}.shape{shape}.astype({dtype})".\
                format(i="{", e="}", name=self.name, type=self.type, shape=self.shape, dtype=self.dtype)
        else:
            var_str = "{name} : fluid.{type})".\
                format(i="{", e="}", name=self.name, type=self.type)

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

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

        return var_str
Y
Yang Yang(Tony) 已提交
1293

F
update  
fengjiayi 已提交
1294
    def to_string(self, throw_on_error, with_details=False):
1295 1296 1297
        """
        Get debug string.

J
Jiabin Yang 已提交
1298 1299 1300 1301 1302
        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;
1303

1304 1305
        Returns:
            str: The debug string.
1306 1307 1308 1309 1310

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
1311

1312 1313 1314 1315 1316
                cur_program = fluid.Program()
                cur_block = cur_program.current_block()
                new_variable = cur_block.create_var(name="X",
                                                    shape=[-1, 23, 48],
                                                    dtype='float32')
1317
                print(new_variable.to_string(True))
J
Jiabin Yang 已提交
1318
                print("=============with detail===============")
1319
                print(new_variable.to_string(True, True))
1320
        """
F
update  
fengjiayi 已提交
1321 1322
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
1323
        protostr = self.desc.serialize_to_string()
1324
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
1325 1326 1327 1328
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
1329 1330 1331
                res_str += "%s: %s\n" % (attr_name,
                                         cpt.to_text(getattr(self, attr_name)))

F
update  
fengjiayi 已提交
1332
        return res_str
1333 1334 1335

    __repr__ = __str__

1336
    @property
1337
    def stop_gradient(self):
J
Jiabin Yang 已提交
1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352
        """
        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")
1353 1354
                linear = fluid.Linear(13, 5, dtype="float32")
                linear2 = fluid.Linear(3, 3, dtype="float32")
J
Jiabin Yang 已提交
1355 1356 1357
                a = fluid.dygraph.to_variable(value0)
                b = fluid.dygraph.to_variable(value1)
                c = fluid.dygraph.to_variable(value2)
1358 1359
                out1 = linear(a)
                out2 = linear2(b)
J
Jiabin Yang 已提交
1360 1361 1362 1363
                out1.stop_gradient = True
                out = fluid.layers.concat(input=[out1, out2, c], axis=1)
                out.backward()

1364
                assert linear.weight.gradient() is None
J
Jiabin Yang 已提交
1365 1366
                assert (out1.gradient() == 0).all()
        """
1367
        return self._stop_gradient
1368

1369 1370
    @stop_gradient.setter
    def stop_gradient(self, s):
1371
        self._stop_gradient = s
1372

1373 1374
    @property
    def persistable(self):
J
Jiabin Yang 已提交
1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395
        """
        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))
        """
1396
        return self.desc.persistable()
1397

Y
Yu Yang 已提交
1398 1399
    @persistable.setter
    def persistable(self, p):
1400
        self.desc.set_persistable(p)
Y
Yu Yang 已提交
1401

Y
Yu Yang 已提交
1402 1403
    @property
    def name(self):
J
Jiabin Yang 已提交
1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419
        """
        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))
        """
1420
        return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
1421

1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441
    @property
    def grad_name(self):
        """
        Indicating name of the gradient Variable of current Variable.

        **Notes: This is a read-only property. It simply returns name of
          gradient Variable from a naming convention but doesn't guarantee
          the gradient exists.**
       
        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
1442 1443
    @name.setter
    def name(self, new_name):
1444
        self.desc.set_name(new_name)
T
typhoonzero 已提交
1445

Y
Yu Yang 已提交
1446 1447
    @property
    def shape(self):
J
Jiabin Yang 已提交
1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464
        """
        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
1465
        # convert to tuple, make it as same as numpy API.
1466
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
1467 1468

    @property
F
fengjiayi 已提交
1469
    def dtype(self):
J
Jiabin Yang 已提交
1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485
        """
        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))
        """
1486
        return self.desc.dtype()
Y
Yu Yang 已提交
1487 1488 1489

    @property
    def lod_level(self):
J
Jiabin Yang 已提交
1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510
        """
        Indicating ``LoD`` info of current Variable, please refer to  :ref:`api_fluid_LoDTensor_en` to check the meaning
        of ``LoD``

        **Notes**:

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

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

        Examples:
          .. code-block:: python

            import paddle.fluid as fluid
            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
            print("LoD Level of current Var is: {}".format(new_variable.lod_level))
        """
1511 1512 1513
        if self.type == core.VarDesc.VarType.SELECTED_ROWS:
            raise Exception("SelectedRows DO NOT supprt lod")

1514
        return self.desc.lod_level()
Y
Yu Yang 已提交
1515

Y
Yu Yang 已提交
1516 1517
    @property
    def type(self):
J
Jiabin Yang 已提交
1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533
        """
        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))
        """
1534
        return self.desc.type()
Y
Yu Yang 已提交
1535

W
Wu Yi 已提交
1536
    def _set_error_clip(self, error_clip):
1537 1538 1539 1540 1541 1542 1543 1544 1545
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
1546 1547
        self.error_clip = error_clip

1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576
    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

1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587
    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
1588
            raise ValueError("slice step can not be zero")
1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663

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

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

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

        return start, stop, step

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

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

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

L
lujun 已提交
1664
    def _cloneVar(self, copy=False):
1665 1666
        if not copy:
            return self.block.create_var(
H
Hongyu Liu 已提交
1667 1668
                name=unique_name.generate_with_ignorable_key(self.name),
                dtype=self.dtype)
1669 1670 1671 1672
        else:
            return self

    def _sliceVar(self, axes, starts, ends):
L
lujun 已提交
1673
        new_var = self._cloneVar()
1674 1675 1676 1677 1678 1679 1680 1681 1682 1683
        self.block.append_op(
            type="slice",
            inputs={'Input': [self]},
            outputs={'Out': [new_var]},
            attrs={'axes': axes,
                   'starts': starts,
                   'ends': ends})
        return new_var

    def _concatVar(self, inputs, axis):
L
lujun 已提交
1684
        new_var = self._cloneVar()
1685 1686 1687 1688 1689 1690 1691 1692 1693 1694
        self.block.append_op(
            type="concat",
            inputs={'X': inputs},
            outputs={'Out': [new_var]},
            attrs={'axis': axis, })
        return new_var

    def _sliceAndConcatVar(self, item, axis):
        if isinstance(item, slice):
            if self.shape[axis] < 0:
L
lujun 已提交
1695
                return self._cloneVar(True)
1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713
            start, stop, step = self._slice_indices(item, self.shape[axis])
            if step == 1:
                return self._sliceVar([axis], [start], [stop])
            else:
                vars = []
                if step > 0:
                    while start < stop:
                        vars.append(
                            self._sliceVar([axis], [start], [start + 1]))
                        start += step
                else:
                    while start > stop:
                        vars.append(
                            self._sliceVar([axis], [start], [start + 1]))
                        start += step
                return self._concatVar(vars, axis)
        elif isinstance(item, int):
            if self.shape[axis] < 0:
L
lujun 已提交
1714
                return self._cloneVar(True)
1715
            index = int(item)
1716
            if (index > 0 and index >= self.shape[axis]) \
1717 1718 1719 1720 1721 1722 1723
                    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):
1724
        return _getitem_impl_(self, item)
1725

Y
Yu Yang 已提交
1726

F
fengjiayi 已提交
1727 1728 1729
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
1730

1731 1732
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
1733 1734 1735 1736
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
1737
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
1738 1739 1740 1741
        ret_values.append(op_proto)
    return ret_values


1742 1743
class ComplexVariable(object):
    """
1744 1745
    The ComplexTensor defined on the complex number domain. It contains two common 
    real number Tensor as its members, :attr:`real` and :attr:`imag` 
1746 1747 1748
    holding the real part and imaginary part of complex numbers respectively.
    
    **Notes**:
1749
        **The constructor of ComplexTensor should not be invoked directly.**
1750

1751
        **Only support dygraph mode at present. Please use** :ref:`api_fluid_dygraph_to_variable` **to create a dygraph ComplexTensor with complex number data.**
1752 1753

    Args:
1754 1755
        real (Tensor): The Tensor holding real-part data.
        imag (Tensor): The Tensor holding imaginery-part data.
1756 1757 1758 1759
    
    Examples:
        .. code-block:: python

1760
            import paddle
1761 1762
            import numpy as np

1763 1764 1765 1766 1767 1768 1769 1770
            paddle.enable_imperative()
            x = paddle.to_tensor([1.0+2.0j, 0.2])
            print(x.name, x.dtype, x.shape)
            # ({'real': 'generated_tensor_0.real', 'imag': 'generated_tensor_0.imag'}, 'complex128', [2L])
            print(x.numpy())
            # [1. +2.j 0.2+0.j]
            print(type(x))
            # <class 'paddle.ComplexTensor'>
1771 1772
    """

1773 1774 1775 1776 1777
    def __new__(cls, *arg, **kwargs):
        cls.__module__ = "paddle"
        cls.__name__ = "ComplexTensor"
        return super(ComplexVariable, cls).__new__(cls)

1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823
    def __init__(self, real, imag):
        assert real.shape == imag.shape, "The real part and imaginary part " \
            "of a ComplexVariable should have the same shape!"
        assert real.dtype == imag.dtype, "The real part and imaginary part " \
            "of a ComplexVariable should have the same data type!"

        self.real = real
        self.imag = imag
        if self.real.dtype in [
                core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32
        ]:
            self._dtype = "complex64"
        else:
            self._dtype = "complex128"
        self._shape = self.real.shape

    @property
    def dtype(self):
        return self._dtype

    @property
    def shape(self):
        return self._shape

    @property
    def name(self):
        return {"real": self.real.name, "imag": self.imag.name}

    @name.setter
    def name(self, name):
        # rename
        if isinstance(name, str):
            self.real.name = name + ".real"
            self.imag.name = name + ".imag"
        elif (isinstance(name, tuple) or isinstance(name,
                                                    list)) and len(name) == 2:
            self.real.name, self.imag.name = name[0], name[1]
        else:
            raise ValueError(
                "An invalid name assigned to the ComplexVariable, "
                "which must be a string, or a tuple or a list with length 2!")

    def numpy(self):
        return self.real.numpy() + 1j * self.imag.numpy()

    def __str__(self):
1824 1825 1826
        return "ComplexTensor[real]: %s\n%s\nComplexTensor[imag]: %s\n%s" % (
            self.real.name, str(self.real.value().get_tensor()), self.imag.name,
            str(self.imag.value().get_tensor()))
1827 1828 1829 1830

    __repr__ = __str__


F
fengjiayi 已提交
1831
class OpProtoHolder(object):
1832 1833 1834 1835
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
1836 1837 1838 1839 1840 1841 1842 1843 1844
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
1845
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
1846 1847 1848 1849 1850 1851
        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):
1852 1853 1854 1855 1856 1857 1858 1859
        """
        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
1860 1861
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
1862 1863
        return self.op_proto_map[type]

1864 1865 1866 1867 1868 1869
    def update_op_proto(self):
        op_protos = get_all_op_protos()
        for proto in op_protos:
            if proto.type not in self.op_proto_map:
                self.op_proto_map[proto.type] = proto

1870 1871 1872 1873
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
1874
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
1875
            core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
1876 1877
            core.op_proto_and_checker_maker.kOpCreationCallstackAttrName(),
            core.op_proto_and_checker_maker.kOpDeviceAttrName()
1878 1879
        }

F
fengjiayi 已提交
1880

X
Xin Pan 已提交
1881
class Operator(object):
1882
    """
1883 1884 1885 1886 1887 1888 1889
    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
1890
        type(str): The type of operator. Default None.
1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910
        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
1911
        Block.append_op or Block._prepend_op instead.
1912 1913 1914 1915

    Examples:
        .. code-block:: python

1916
            import paddle.fluid as fluid
1917
            cur_program = fluid.Program()
1918 1919 1920 1921 1922
            cur_block = cur_program.current_block()
            # var1 += var2 + var3
            cur_block.append_op(type="sum",
                                inputs={"X": [var1, var2, var3]},
                                outputs={"Out": [var1]})
1923
    """
1924
    OP_WITHOUT_KERNEL_SET = {
1925 1926
        'feed', 'fetch', 'recurrent', 'go', 'rnn_memory_helper_grad',
        'conditional_block', 'while', 'send', 'recv', 'listen_and_serv',
1927 1928
        'fl_listen_and_serv', 'ncclInit', 'select', 'checkpoint_notify',
        'gen_nccl_id', 'c_gen_nccl_id', 'c_comm_init', 'c_sync_calc_stream',
1929
        'c_sync_comm_stream', 'queue_generator', 'dequeue', 'enqueue'
1930
    }
1931

Y
Yu Yang 已提交
1932 1933
    def __init__(self,
                 block,
Y
Yu Yang 已提交
1934
                 desc,
Y
Yu Yang 已提交
1935 1936 1937
                 type=None,
                 inputs=None,
                 outputs=None,
M
minqiyang 已提交
1938
                 attrs=None):
L
lujun 已提交
1939
        if in_dygraph_mode():
1940 1941
            if type is None:
                raise ValueError(
1942
                    "`type` to initialized an Operator can not be None.")
J
Jiabin Yang 已提交
1943
            self._type = type
M
minqiyang 已提交
1944
            self.attrs = attrs if attrs else {}
1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958
        else:
            self.block = block
            self.desc = desc
            # note: not add self.attrs here:
            # https://github.com/PaddlePaddle/Paddle/pull/12583#pullrequestreview-145093173
            op_attrs = attrs
            if op_attrs is None:
                op_attrs = dict()
            del attrs

            op_maker = core.op_proto_and_checker_maker

            if op_maker.kOpRoleAttrName() not in op_attrs:
                op_attrs[op_maker.kOpRoleAttrName(
1959
                )] = self.block.program._op_role
1960 1961 1962

            role_var_name = op_maker.kOpRoleVarAttrName()
            if len(self.block.program.
1963 1964
                   _op_role_var) != 0 and role_var_name not in op_attrs:
                op_attrs[role_var_name] = self.block.program._op_role_var
1965 1966 1967 1968 1969 1970 1971 1972

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

            if len(self.desc.type()) != 0:
                return
            if type is None:
                raise ValueError(
1973
                    "`type` to initialized an Operator can not be None.")
1974 1975
            else:
                callstack_var_name = op_maker.kOpCreationCallstackAttrName()
1976 1977 1978 1979 1980 1981 1982
                op_attrs[callstack_var_name] = []
                for frame in traceback.extract_stack():
                    op_attrs[callstack_var_name].append(
                        '  File "{}", line {}, in {}'.format(frame[0], frame[1],
                                                             frame[2]))
                    op_attrs[callstack_var_name].append('    {}'.format(frame[
                        3]))
1983 1984 1985 1986 1987 1988 1989

            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()

1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
            # set device for op with kernels, give warning for op without kernels
            # when force_cpu and device_guard are used at the same time, a warning will be given.
            # TODO(zhangting2020): when force_cpu is removed, clear warning below.
            if _current_device is not None:
                if self._has_kernel(type):
                    op_device = op_maker.kOpDeviceAttrName()
                    op_attrs[op_device] = _current_device
                else:
                    warnings.warn("The Op(%s) is not support to set device." %
                                  type)
                if 'force_cpu' in op_attrs:
                    if (type is 'less_than' and op_attrs['force_cpu'] != None
                        ) or op_attrs['force_cpu'] != False:
                        warnings.warn(
                            "The Attr(force_cpu) of Op(%s) will be deprecated in the future, "
                            "please use 'device_guard' instead. 'device_guard' has higher priority when they are "
                            "used at the same time." % type)

2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
            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]
2021
                        if not isinstance(in_args, (list, tuple)):
2022 2023 2024 2025 2026 2027
                            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 = []
2028
                        for index, arg in enumerate(in_args):
2029 2030 2031 2032
                            if isinstance(arg, six.string_types):
                                in_arg_names.append(arg)
                            elif isinstance(arg, six.binary_type):
                                in_arg_names.append(arg.decode())
2033
                            elif isinstance(arg, (Variable, core.VarBase)):
2034
                                in_arg_names.append(cpt.to_text(arg.name))
2035
                            else:
2036 2037 2038 2039
                                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."
2040 2041
                                    "but received : %s" %
                                    (in_proto.name, type, arg))
2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067
                        self.desc.set_input(in_proto.name, in_arg_names)
                    else:
                        self.desc.set_input(in_proto.name, [])

            if outputs is not None:
                for m in proto.outputs:
                    if (m.name not in outputs) and m.dispensable:
                        continue
                    if not ((m.name in outputs) or m.dispensable):
                        raise ValueError(("Incorrect setting for output(s) of "
                                          "operator \"%s\", should set: [%s].")
                                         % (type, m.name))
                for out_proto in proto.outputs:
                    if out_proto.name not in outputs:
                        continue
                    out_args = outputs[out_proto.name]
                    if not isinstance(out_args, list):
                        out_args = [out_args]
                    if not out_proto.duplicable and len(out_args) > 1:
                        raise ValueError(
                            "Output %s expects only one output, but %d are given."
                            % (out_proto.name, len(out_args)))
                    out_arg_names = []
                    for arg in out_args:
                        out_arg_names.append(cpt.to_text(arg.name))
                        # TODO(minqiyang): could we remove variable's op in static mode?
L
lujun 已提交
2068
                        if not in_dygraph_mode():
2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087
                            arg.op = self
                    self.desc.set_output(out_proto.name, out_arg_names)

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

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

W
Wu Yi 已提交
2088
    def _has_kernel(self, op_type):
2089 2090
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
2091
    def to_string(self, throw_on_error):
2092
        """
2093 2094
        Get debug string.

2095
        Args:
2096 2097
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
2098

2099 2100
        Returns:
            str: The debug string.
2101 2102

        """
2103
        protostr = self.desc.serialize_to_string()
2104
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
2105 2106
        return _debug_string_(proto, throw_on_error)

2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 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 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199
    def _to_readable_code(self, skip_op_callstack=True):
        """
        Get readable debug string of Operator.

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

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

        Returns:
            string: The formatted Operator string.

        Examples:
            .. code-block:: python

            import paddle.fluid as fluid

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

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

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

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

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

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

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

        if outputs_str != "{}":
            op_str = "{outputs} = {op_type}(inputs={inputs}, {attrs})".\
                format(outputs = outputs_str, op_type=self.type, inputs=inputs_str, attrs=attrs_str)
        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
2200
    def __str__(self):
2201
        return self._to_readable_code()
2202 2203 2204

    __repr__ = __str__

F
fengjiayi 已提交
2205 2206
    @property
    def type(self):
2207
        return self.desc.type()
F
fengjiayi 已提交
2208 2209

    def input(self, name):
2210
        """
2211
        Get the input arguments according to the input parameter name.
2212

2213 2214
        Args:
            name(str): The input parameter name.
2215

2216 2217 2218
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
2219
        """
F
fengjiayi 已提交
2220 2221
        return self.desc.input(name)

W
Wu Yi 已提交
2222
    def _rename_input(self, old_name, new_name):
2223 2224 2225 2226 2227 2228 2229 2230 2231 2232
        """
        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
2233
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
2234

W
Wu Yi 已提交
2235
    def _rename_output(self, old_name, new_name):
2236 2237 2238 2239 2240 2241 2242 2243 2244 2245
        """
        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
2246
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
2247

F
fengjiayi 已提交
2248 2249 2250 2251
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
2252 2253 2254 2255 2256 2257 2258 2259
    @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
2260
    def output(self, name):
2261
        """
2262
        Get output arguments by the output parameter name.
2263

2264 2265
        Args:
            name(str): The output parameter name.
2266

2267 2268 2269
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
2270
        """
F
fengjiayi 已提交
2271 2272 2273 2274 2275 2276
        return self.desc.output(name)

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

2277 2278 2279 2280 2281 2282 2283 2284
    @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
2285
    def has_attr(self, name):
2286
        """
2287 2288
        Whether this Operator has the attribute with name or not.

2289
        Args:
2290
            name(str): the attribute name.
2291

2292 2293
        Returns:
            bool: True if has this attribute.
2294 2295

        """
F
fengjiayi 已提交
2296 2297 2298
        return self.desc.has_attr(name)

    def attr_type(self, name):
2299
        """
2300
        Get the type of attribute by attribute's name.
2301

2302 2303
        Args:
            name(str): the attribute name.
2304

2305 2306
        Returns:
            core.AttrType: the attribute type.
2307
        """
F
fengjiayi 已提交
2308 2309
        return self.desc.attr_type(name)

W
Wu Yi 已提交
2310
    def _set_attr(self, name, val):
2311 2312 2313 2314 2315 2316 2317 2318 2319 2320
        """
        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
2321 2322
        self._update_desc_attr(name, val)

2323 2324 2325
    def _remove_attr(self, name):
        self.desc.remove_attr(name)

G
gongweibao 已提交
2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336
    def _update_desc_attr(self, name, val):
        """
        Update the value of desc's attribute by attribute's name.

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

        Raises:
            ValueError: If the type of value doesn't match with desc.attr_type(name).
        """
Q
Qiyang Min 已提交
2337 2338
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
2339 2340
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
2341
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
2342 2343 2344 2345
        elif isinstance(val, core.BlockDesc) or \
                isinstance(val, core.ProgramDesc):
            self.desc.set_serialized_attr(name, val.serialize_to_string())
        else:
W
Wu Yi 已提交
2346
            self.desc._set_attr(name, val)
Y
yuyang18 已提交
2347

F
fengjiayi 已提交
2348 2349 2350 2351 2352
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
2353
        """
2354 2355
        Get the attribute by name.

2356
        Args:
2357
            name(str): the attribute name.
2358

2359 2360
        Returns:
            bool|int|str|float|list: The attribute value. The return value
2361 2362
            can be any valid attribute type.
        """
F
fengjiayi 已提交
2363
        return self.desc.attr(name)
Y
Yu Yang 已提交
2364

W
Wu Yi 已提交
2365
    def _block_attr_id(self, name):
2366
        """
G
gongweibao 已提交
2367
        Get the block attribute's id by name.
2368

2369 2370
        Args:
            name(str): the attribute name.
2371

2372 2373
        Returns:
            int: the block index.
2374
        """
W
Wu Yi 已提交
2375
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
2376

W
Wu Yi 已提交
2377
    def _block_attr(self, name):
G
gongweibao 已提交
2378 2379 2380 2381 2382 2383 2384 2385 2386 2387
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
2388
        id = self._block_attr_id(name)
G
gongweibao 已提交
2389 2390 2391
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
2392
    def _blocks_attr(self, name):
G
gongweibao 已提交
2393 2394 2395 2396 2397 2398 2399 2400 2401 2402
        """
        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
2403
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
2404 2405 2406 2407 2408
            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
2409
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
2410 2411 2412 2413 2414 2415 2416 2417 2418 2419
        """
        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
2420
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
2421

J
JiayiFeng 已提交
2422
    def all_attrs(self):
F
fengjiayi 已提交
2423
        """
2424 2425 2426
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
2427
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
2428 2429 2430 2431
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
G
gongweibao 已提交
2432 2433
            attr_type = self.desc.attr_type(n)
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
2434
                attr_map[n] = self._block_attr(n)
G
gongweibao 已提交
2435 2436 2437
                continue

            if attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
2438
                attr_map[n] = self._blocks_attr(n)
G
gongweibao 已提交
2439 2440 2441 2442
                continue

            attr_map[n] = self.attr(n)

F
fengjiayi 已提交
2443 2444
        return attr_map

2445 2446 2447
    def _is_optimize_op(self):
        op_maker = core.op_proto_and_checker_maker
        OPTIMIZE = core.op_proto_and_checker_maker.OpRole.Optimize
2448 2449 2450 2451

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

2452 2453 2454
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(OPTIMIZE):
            return True
2455 2456 2457 2458 2459 2460 2461 2462

        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()):
2463 2464
            return False

2465 2466 2467 2468 2469 2470
        op_role = self.desc.attr(op_maker.kOpRoleAttrName())
        if op_role & int(BACKWARD):
            return True

        return False

Y
Yu Yang 已提交
2471

Y
Yu Yang 已提交
2472
class Block(object):
2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486
    """
    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
2487
        use `Program._create_block()` to create a block.
2488 2489 2490 2491

    Examples:
        .. code-block:: python

2492 2493 2494
            import paddle.fluid as fluid

            cur_program = fluid.Program()
2495 2496 2497 2498 2499 2500 2501 2502 2503
            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
2504
    def __init__(self, program, idx):
Y
Yu Yang 已提交
2505
        self.desc = program.desc.block(idx)
2506
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
2507
        self.ops = list()  # operator list
Y
Yu Yang 已提交
2508
        self.program = program
2509
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
2510

2511
    def __str__(self):
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 2551 2552 2553 2554 2555 2556 2557
        return self._to_readable_code()

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

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

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

        Returns:
            string: The formatted Block string.

        Examples:
            .. code-block:: python

            import paddle.fluid as fluid

            cur_program = fluid.Program()
            cur_block = cur_program.current_block()
            new_var = cur_block.create_var(name="X",
                                           shape=[-1, 23, 48],
                                           dtype='float32')
            new_op = cur_block.append_op(type="abs",
                                inputs={"X": [new_var]},
                                outputs={"Out": [new_var]})
            print(cur_block._to_readable_code())
        """
        assert isinstance(
            skip_op_callstack, bool
        ), "skip_op_callstack parameter's type is error, expect bool, received %s".format(
            type(skip_op_callstack))
        block_str = "{ // block "
        block_str += "{}\n".format(self.idx)
        for var in list(self.vars.values()):
            block_str += "    {}\n".format(var._to_readable_code())
        block_str += "\n"
        for op in self.ops:
            block_str += "    {}\n".format(
                op._to_readable_code(skip_op_callstack))
        block_str += "}"
        return block_str
Y
Yang Yang(Tony) 已提交
2558

F
fengjiayi 已提交
2559 2560
    def to_string(self, throw_on_error, with_details=False):
        """
2561 2562
        Get debug string.

F
fengjiayi 已提交
2563 2564
        Args:
            throw_on_error(bool): raise exception when self is not initialized
2565
                when throw_on_error is True.
F
update  
fengjiayi 已提交
2566
            with_details(bool): more details about variables and parameters
2567 2568
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
2569

2570 2571
        Returns:
            str: The debug string.
F
fengjiayi 已提交
2572 2573 2574 2575
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
2576
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
2577 2578
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
2579
            for var in list(self.vars.values()):
F
fengjiayi 已提交
2580
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
2581
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
2582
            for op in self.ops:
F
fengjiayi 已提交
2583 2584
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
2585 2586 2587
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
2588 2589
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
2590 2591
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
2592 2593 2594

    __repr__ = __str__

Y
Yu Yang 已提交
2595 2596
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
2597
        return self.desc.parent
Y
Yu Yang 已提交
2598

Y
Yu Yang 已提交
2599 2600 2601 2602
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
2603
    def _set_forward_block_idx(self, idx):
2604 2605 2606 2607 2608 2609 2610 2611 2612
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
2613
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
2614

2615 2616 2617 2618 2619 2620 2621 2622
    @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
2623 2624
    @property
    def idx(self):
Y
Yu Yang 已提交
2625
        return self.desc.id
Y
Yu Yang 已提交
2626

Q
Qiao Longfei 已提交
2627
    def var(self, name):
2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640
        """
        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.
        """
2641
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
2642 2643 2644
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
2645 2646
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
2647
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
2648
        return v
Q
Qiao Longfei 已提交
2649

X
Xin Pan 已提交
2650
    def _find_var_recursive(self, name):
2651 2652 2653 2654 2655 2656 2657
        """
        Get a Variable by name from this block recursively.

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

        Returns:
X
Xin Pan 已提交
2658
            Variable: the Variable with the giving name. Or None if not found.
2659
        """
Y
Yu Yang 已提交
2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683
        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
2684
        return None
Y
Yu Yang 已提交
2685

X
Xin Pan 已提交
2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704
    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
2705

Q
Qiao Longfei 已提交
2706
    def all_parameters(self):
2707
        return list(self.iter_parameters())
2708

2709
    def iter_parameters(self):
M
minqiyang 已提交
2710
        return (item[1] for item in six.iteritems(self.vars)
2711
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
2712

Y
Yu Yang 已提交
2713
    def create_var(self, *args, **kwargs):
L
Leo Chen 已提交
2714 2715 2716
        if in_dygraph_mode():
            var = _varbase_creator(*args, **kwargs)
        else:
2717 2718 2719
            var = Variable(block=self, *args, **kwargs)
            if 'initializer' in kwargs:
                kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
2720
        return var
Y
Yu Yang 已提交
2721

Q
Qiao Longfei 已提交
2722 2723 2724
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
2725
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
2726 2727
        """
        Rename variable in vars and ops' inputs and outputs
2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739

        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
2740
        """
M
minqiyang 已提交
2741 2742
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
2743

T
typhoonzero 已提交
2744
        if not self.has_var(name):
2745
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
2746 2747
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
2748
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
2749 2750 2751 2752 2753 2754
            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
2755
            var_type = "Variable"
T
wip  
typhoonzero 已提交
2756 2757 2758 2759
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
2760
        orig_var_type = v.type
M
minqiyang 已提交
2761
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
2762
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
2763
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
2764
        if var_type == "Parameter":
L
Leo Chen 已提交
2765 2766
            if in_dygraph_mode():
                var = ParamBase(
2767 2768 2769 2770 2771 2772 2773 2774 2775 2776
                    d.shape(),
                    d.dtype(),
                    type=orig_var_type,
                    name=new_name,
                    stop_gradient=stop_gradient,
                    trainable=trainable,
                    optimize_attr=optimize_attr,
                    regularizer=regularizer,
                    error_clip=error_clip)
            else:
L
Leo Chen 已提交
2777 2778
                var = Parameter(
                    self,
2779 2780 2781 2782 2783 2784 2785 2786 2787
                    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
2788
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
2789 2790
            var = Variable(
                self,
T
typhoonzero 已提交
2791
                type=orig_var_type,
T
wip  
typhoonzero 已提交
2792 2793 2794 2795
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
2796
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
2797 2798 2799
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
2800
        self._sync_with_cpp()
2801
        return var
T
typhoonzero 已提交
2802

W
Wu Yi 已提交
2803 2804
    def _remove_var(self, name):
        self._sync_with_cpp()
M
minqiyang 已提交
2805
        self.desc._remove_var(cpt.to_bytes(name))
2806 2807
        del self.vars[name]

Y
Yu Yang 已提交
2808 2809
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
2810
        param = None
L
Leo Chen 已提交
2811
        if in_dygraph_mode():
2812
            param = ParamBase(*args, **kwargs)
L
Leo Chen 已提交
2813 2814
        else:
            param = Parameter(global_block, *args, **kwargs)
2815
        if 'initializer' in kwargs:
2816 2817 2818 2819 2820

            def _is_inited_by(block, var):
                init_ops = []
                for op in block.ops:
                    if var.name in op.output_arg_names:
2821 2822 2823 2824 2825
                        # In startup_program, "c_broadcast" and "c_sync_comm_stream"
                        # are treated as initialization ops that cause error. 
                        # Think of "c_broadcast" and "c_sync_comm_stream" as a special case here.
                        if op.type in ["c_broadcast", "c_sync_comm_stream"]:
                            continue
2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836
                        init_ops.append(op)
                return init_ops

            initializer = kwargs['initializer']
            init_ops = _is_inited_by(global_block, param)
            init_ops_len = len(init_ops)
            if init_ops_len > 1:
                raise RuntimeError("param " + param.name +
                                   " is inited by multiple init ops " + str(
                                       init_ops))
            elif init_ops_len == 1:
2837
                # TODO already inited, do nothing, should log a warning
2838 2839 2840
                pass
            else:
                initializer(param, self)
2841
        param.stop_gradient = False
Q
Qiao Longfei 已提交
2842
        return param
Y
Yu Yang 已提交
2843

Y
Yu Yang 已提交
2844
    def append_op(self, *args, **kwargs):
2845 2846 2847 2848 2849 2850
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
L
lujun 已提交
2851
        if in_dygraph_mode():
2852
            attrs = kwargs.get("attrs", {})
J
Jiabin Yang 已提交
2853
            type = kwargs.get("type", None)
2854 2855 2856
            op = Operator(
                block=self,
                desc=None,
J
Jiabin Yang 已提交
2857
                type=type,
M
minqiyang 已提交
2858 2859
                inputs=None,
                outputs=None,
2860
                attrs=attrs)
2861

M
minqiyang 已提交
2862 2863 2864
            # record ops in tracer rather than blocks
            #
            # TODO(minqiyang): add op stop_gradient support in static mode too.
L
lujun 已提交
2865
            # currently, we only support stop_gradient in dygraph mode.
J
Jiabin Yang 已提交
2866 2867

            _dygraph_tracer().trace_op(type,
M
minqiyang 已提交
2868
                                       kwargs.get("inputs", {}),
J
Jiabin Yang 已提交
2869 2870
                                       kwargs.get("outputs", {}), attrs
                                       if attrs else {},
M
minqiyang 已提交
2871
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
2872
        else:
2873 2874 2875 2876 2877 2878 2879 2880 2881
            op_desc = self.desc.append_op()
            op = Operator(
                block=self,
                desc=op_desc,
                type=kwargs.get("type", None),
                inputs=kwargs.get("inputs", None),
                outputs=kwargs.get("outputs", None),
                attrs=kwargs.get("attrs", None))

M
minqiyang 已提交
2882
            self.ops.append(op)
M
minqiyang 已提交
2883

2884 2885
        return op

W
Wu Yi 已提交
2886
    def _insert_op(self, index, *args, **kwargs):
2887 2888 2889 2890 2891 2892 2893 2894 2895
        """
        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
2896 2897
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
2898 2899 2900 2901
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

W
Wu Yi 已提交
2902
    def _remove_op(self, index):
2903 2904 2905 2906 2907 2908 2909 2910 2911
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
W
Wu Yi 已提交
2912 2913
        self._sync_with_cpp()
        self.desc._remove_op(index, index + 1)
2914 2915
        del self.ops[index]

W
Wu Yi 已提交
2916
    def _slice_ops(self, start, end):
2917 2918 2919 2920 2921 2922 2923 2924 2925 2926
        """
        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
2927
        return self.ops[start:end]
Y
Yancey1989 已提交
2928

W
Wu Yi 已提交
2929
    def _prepend_op(self, *args, **kwargs):
L
lujun 已提交
2930
        if in_dygraph_mode():
J
Jiabin Yang 已提交
2931 2932
            type = kwargs.get("type", None)
            attrs = kwargs.get("attrs", {})
2933
            op = Operator(
J
Jiabin Yang 已提交
2934
                self, None, type=type, inputs=None, outputs=None, attrs=attrs)
M
minqiyang 已提交
2935

J
Jiabin Yang 已提交
2936
            _dygraph_tracer().trace_op(type,
M
minqiyang 已提交
2937
                                       kwargs.get("inputs", {}),
J
Jiabin Yang 已提交
2938 2939
                                       kwargs.get("outputs", {}), attrs
                                       if attrs else {},
M
minqiyang 已提交
2940
                                       kwargs.get("stop_gradient", False))
M
minqiyang 已提交
2941
        else:
2942 2943 2944 2945 2946 2947 2948 2949
            op_desc = self.desc._prepend_op()
            op = Operator(
                self,
                op_desc,
                type=kwargs.get("type", None),
                inputs=kwargs.get("inputs", None),
                outputs=kwargs.get("outputs", None),
                attrs=kwargs.get("attrs", None))
M
minqiyang 已提交
2950
            self.ops.insert(0, op)
2951

Y
Yu Yang 已提交
2952 2953
        return op

W
Wu Yi 已提交
2954
    def _sync_with_cpp(self):
2955
        """
2956 2957
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
2958
        """
Q
Qiao Longfei 已提交
2959 2960 2961 2962 2963
        # sync variables from cpp
        for var in self.desc.all_vars():
            if not self.has_var(var.name()):
                self.create_var(name=var.name(), desc=var, type=var.type())

2964
        # sync variables removed from c++ end
2965
        for var in list(self.vars.keys()):
M
minqiyang 已提交
2966
            if not self.desc.find_var(cpt.to_bytes(var)):
2967 2968
                self.vars.pop(var)

Q
Qiao Longfei 已提交
2969
        # sync operators from cpp
2970 2971 2972 2973
        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
2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989
        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
2990 2991 2992 2993 2994

        # 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
2995
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
2996 2997 2998 2999 3000 3001 3002

        # 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)

3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015
        # 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
3016 3017 3018 3019
        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
3020
    def _copy_param_info_from(self, other):
3021
        """
3022 3023
        Copy the information of parameters from the other block.

3024
        Args:
3025 3026 3027 3028 3029
            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.
3030 3031 3032 3033 3034

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
3035 3036
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
3037
        for p in other.iter_parameters():
3038 3039 3040
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
3041 3042
                # if the Parameter is pruned, v may be None
                continue
3043
            assert isinstance(v, Variable)
3044
            new_p = None
L
Leo Chen 已提交
3045 3046
            if in_dygraph_mode():
                new_p = ParamBase(
3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057
                    shape=v.shape,
                    dtype=v.dtype,
                    type=v.type,
                    lod_level=v.lod_level,
                    stop_gradient=p.stop_gradient,
                    trainable=p.trainable,
                    optimize_attr=p.optimize_attr,
                    regularizer=p.regularizer,
                    error_clip=p.error_clip,
                    name=v.name)
            else:
L
Leo Chen 已提交
3058 3059
                new_p = Parameter(
                    block=self,
3060 3061 3062
                    shape=v.shape,
                    dtype=v.dtype,
                    type=v.type,
3063 3064
                    lod_level=v.lod_level
                    if v.type == core.VarDesc.VarType.LOD_TENSOR else None,
3065 3066 3067 3068 3069 3070
                    stop_gradient=p.stop_gradient,
                    trainable=p.trainable,
                    optimize_attr=p.optimize_attr,
                    regularizer=p.regularizer,
                    error_clip=p.error_clip,
                    name=v.name)
3071 3072
            self.vars[new_p.name] = new_p

3073
    def _clone_variable(self, var, force_persistable=True):
3074 3075
        """
        Clone a variable into current block.
3076

3077 3078
        Args:
            var: the variable to be cloned.
3079 3080 3081
            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.
3082 3083

        Returns:
3084
            Variable: the new  variable cloned from 'var' in current block.
3085 3086
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
3087 3088 3089 3090 3091
        ret_var = None
        # make STEP_SCOPES var can be safely cloned.
        if var.type == core.VarDesc.VarType.STEP_SCOPES:
            ret_var = self.create_var(
                name=var.name, persistable=var.persistable, type=var.type)
T
tangwei12 已提交
3092 3093
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
3094
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
3095 3096 3097 3098 3099 3100
        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,
3101
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
3102 3103
                is_data=var.is_data,
                need_check_feed=var.desc.need_check_feed())
T
update  
typhoonzero 已提交
3104 3105 3106 3107 3108 3109 3110
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
3111
                persistable=True if force_persistable else var.persistable,
H
Huihuang Zheng 已提交
3112 3113
                is_data=var.is_data,
                need_check_feed=var.desc.need_check_feed())
T
update  
typhoonzero 已提交
3114
        return ret_var
3115

Y
Yu Yang 已提交
3116

3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211
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()

3212
    def remove_input_by_id(self, node_id):
3213 3214 3215 3216 3217 3218
        """
        Remove a node from inputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
3219
        self.node.remove_input(node_id)
3220

3221
    def remove_input(self, node):
3222 3223 3224 3225
        """
        Remove a node from inputs.

        Args:
3226
            node(IrNode): the node being removed.
3227
        """
3228
        self.node.remove_input(node.node)
3229

3230
    def append_input(self, node):
3231 3232 3233 3234
        """
        Append a node in inputs.

        Args:
3235
            node(IrNode): the node being appended.
3236
        """
3237
        self.node.append_input(node.node)
3238 3239 3240 3241 3242 3243 3244 3245

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

3246
    def remove_output_by_id(self, node_id):
3247 3248 3249 3250 3251 3252
        """
        Remove a node from outputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
3253
        self.node.remove_output(node_id)
3254

3255
    def remove_output(self, node):
3256 3257 3258 3259
        """
        Remove a node from outputs.

        Args:
3260
            node(IrNode): the node being removed.
3261
        """
3262
        self.node.remove_output(node.node)
3263

3264
    def append_output(self, node):
3265 3266 3267 3268
        """
        Append a node in outputs.

        Args:
3269
            node(IrNode): the node being appended.
3270
        """
3271
        self.node.append_output(node.node)
3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318

    @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
3319
            "The node variable description can not be None."
3320 3321 3322 3323 3324 3325 3326 3327 3328 3329
        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
3330
            "The node variable description can not be None."
3331 3332
        return self.node.var().persistable()

3333 3334 3335 3336 3337 3338 3339 3340
    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
3341
            "The node variable description can not be None."
3342 3343 3344 3345 3346 3347 3348 3349 3350 3351
        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
3352
            "The node variable description can not be None."
3353 3354 3355 3356 3357 3358 3359 3360 3361 3362
        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
3363
            "The node variable description can not be None."
3364 3365
        return self.node.var().shape()

3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412
    @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
3413
            "The node operator description can not be None."
3414 3415
        self.node.op()._rename_input(old_input_name, new_input_name)

3416 3417 3418 3419 3420 3421 3422 3423 3424
    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
3425
            "The node operator description can not be None."
3426 3427
        self.node.op()._rename_output(old_output_name, new_output_name)

3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438
    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
3439
            "The node operator description can not be None."
3440 3441 3442 3443 3444 3445 3446 3447 3448 3449 3450 3451 3452
        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
3453
            "The node operator description can not be None."
3454 3455 3456 3457 3458 3459 3460 3461 3462 3463
        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
3464
            "The node operator description can not be None."
3465 3466
        return self.node.op().set_type(new_type)

3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481
    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
3482
            "The node operator description can not be None."
3483 3484 3485 3486
        desc = self.node.op()
        if isinstance(val, Block):
            desc.set_block_attr(name, val.desc)
        elif isinstance(val, list) and val and \
3487
                all(isinstance(v, Block) for v in val):
3488 3489
            desc.set_blocks_attr(name, [v.desc for v in val])
        elif isinstance(val, core.BlockDesc) or \
3490
                isinstance(val, core.ProgramDesc):
3491 3492 3493 3494
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)

3495 3496 3497 3498 3499 3500 3501 3502
    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
3503
            "The node operator description can not be None."
3504 3505 3506 3507 3508 3509 3510 3511 3512 3513
        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
3514
            "The node operator description can not be None."
3515 3516
        return self.node.op().output_arg_names()

3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537
    @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]


3538 3539
class IrGraph(object):
    """
3540
    Python IrGraph. Beneath it is a core.Graph, which is used for
3541
    creating a c++ Ir Pass Graph. An IrGraph is just a graph view of
3542 3543
    a Program. In an IrGraph, both Variables and Operators are graph
    nodes.
3544 3545 3546 3547
    """

    def __init__(self, graph, for_test=False):
        """
3548 3549
        Construct an IrGraph using core.Graph.

3550 3551 3552 3553 3554 3555 3556 3557 3558
        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

3559 3560 3561 3562
    def clone(self):
        """
        Create a new and duplicated IrGraph.

3563 3564 3565
        Warns:
            The method only clones the graph structure, not its attributes.

3566 3567 3568
        Returns:
            IrGraph: A new and duplicated graph.
        """
3569
        g = self.graph.clone()
3570 3571
        return IrGraph(g, self._for_test)

3572
    def is_test(self):
3573 3574 3575
        """
        If the graph is used for testing, the function returns true. Otherwise, returns false.
        """
3576 3577
        return self._for_test

W
WangZhen 已提交
3578
    def all_nodes(self):
3579 3580 3581
        """
        Return all nodes included in the graph as a set.
        """
3582
        return {IrNode(node) for node in self.graph.nodes()}
3583

3584
    def all_var_nodes(self):
3585 3586 3587
        """
        Return all variable nodes included in the graph as a set.
        """
3588
        return {IrVarNode(node) for node in self.graph.nodes() if node.is_var()}
3589

3590
    def all_persistable_nodes(self):
3591 3592 3593
        """
        Return all persistable variable nodes included in the graph as a set.
        """
W
WangZhen 已提交
3594 3595 3596 3597 3598
        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)
3599
        return {IrVarNode(p) for p in persistable_nodes}
W
WangZhen 已提交
3600

3601
    def all_op_nodes(self):
3602 3603 3604
        """
        Return all operator nodes included in the graph as a set.
        """
3605
        return {IrOpNode(node) for node in self.graph.nodes() if node.is_op()}
3606

3607
    def create_persistable_node(self, name, var_type, shape, var_dtype):
3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618
        """
        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:
3619
            IrVarNode: the created persistable variable node.
3620
        """
3621 3622 3623 3624 3625
        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)
3626
        return IrVarNode(self.graph.create_var_node(var_desc))
3627 3628

    def create_var_node(self, name, var_type, shape, var_dtype):
3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639
        """
        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:
3640
            IrVarNode: the created variable node.
3641 3642
        """

3643 3644 3645 3646
        var_desc = core.VarDesc(name)
        var_desc.set_type(var_type)
        var_desc.set_shape(shape)
        var_desc.set_dtype(var_dtype)
3647
        return IrVarNode(self.graph.create_var_node(var_desc))
3648

3649 3650 3651 3652 3653 3654
    def create_control_dep_var(self):
        """
        create a control var
        """
        return IrVarNode(self.graph.create_control_dep_var())

3655
    def create_var_node_from_desc(self, var_desc):
3656 3657 3658 3659 3660 3661 3662 3663
        """
        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:
3664
            IrVarNode: the created variable node.
3665
        """
3666
        return IrVarNode(self.graph.create_var_node(var_desc))
3667 3668

    def create_op_node(self, op_type, attrs, inputs, outputs):
3669 3670 3671 3672 3673 3674 3675
        """
        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
3676
            outputs(dict): the outputs of the operator node.
3677 3678

        Returns:
3679
            IrOpNode: the created operator node.
3680
        """
3681 3682
        op_desc = core.OpDesc()
        op_desc.set_type(op_type)
3683
        for attr, value in six.iteritems(attrs):
3684
            self._update_desc_attr(op_desc, attr, value)
3685
        for input_name, var_nodes in six.iteritems(inputs):
3686 3687 3688 3689
            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])
3690
        for output_name, var_nodes in six.iteritems(outputs):
3691 3692 3693 3694
            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])
3695
        return IrOpNode(self.graph.create_op_node(op_desc))
3696 3697

    def create_op_node_from_desc(self, op_desc):
3698 3699 3700 3701 3702 3703 3704
        """
        Create a operator node by using an existing OpDesc in the graph.

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

        Returns:
3705
            IrOpNode: the created operator node.
3706
        """
3707
        return IrOpNode(self.graph.create_op_node(op_desc))
3708 3709

    def update_input_link(self, old_input_node, new_input_node, op_node):
3710 3711 3712 3713
        """
        Update the input's link of a operator node.

        Args:
3714 3715 3716
            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.
3717
        """
3718
        assert old_input_node.node in self.graph.nodes() and new_input_node.node in \
3719 3720
               self.graph.nodes() and op_node.node in self.graph.nodes(), \
            'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.'
3721 3722 3723 3724
        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)
3725
        op_node.rename_input(old_input_node.name(), new_input_node.name())
3726

3727 3728 3729 3730 3731 3732 3733 3734 3735 3736
    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 \
3737 3738
               self.graph.nodes() and op_node.node in self.graph.nodes(), \
            'The three arguments(old_output_node &new_output_node &op_node) must be in the graph nodes.'
3739 3740 3741 3742 3743 3744
        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())

3745
    def link_to(self, node_in, node_out):
3746 3747 3748 3749
        """
        Connect two nodes.

        Args:
3750 3751
            node_in(IrNode): the input node.
            node_out(IrNode): the output node.
3752
        """
3753
        assert node_in.node in self.graph.nodes() and node_out.node in self.graph.nodes(), \
W
WangZhen 已提交
3754
            'The two arguments(node_in&node_out) must be in the graph nodes.'
3755 3756
        node_in.append_output(node_out)
        node_out.append_input(node_in)
3757 3758

    def safe_remove_nodes(self, remove_nodes):
3759 3760 3761 3762 3763 3764 3765
        """
        Remove nodes safely since links connected to these removed nodes are
        also removed.

        Args:
            remove_nodes(set): the nodes prepared to be removed.
        """
3766
        if not isinstance(remove_nodes, set):
W
WangZhen 已提交
3767 3768 3769 3770
            if isinstance(remove_nodes, Iterable):
                remove_nodes = set(remove_nodes)
            else:
                remove_nodes = {remove_nodes}
3771 3772
        original_nodes = {n.node for n in remove_nodes}
        core.graph_safe_remove_nodes(self.graph, original_nodes)
3773

Z
Zhen Wang 已提交
3774 3775 3776 3777 3778 3779 3780 3781
    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] = [
3782
                            self._find_node_by_name(node.inputs, each_var_name)
Z
Zhen Wang 已提交
3783 3784 3785 3786
                        ]
                for each_var_name in node.op().output_arg_names():
                    if each_var_name not in var_nodes:
                        var_nodes[each_var_name] = [
3787
                            self._find_node_by_name(node.outputs, each_var_name)
Z
Zhen Wang 已提交
3788 3789 3790
                        ]
                    else:
                        var_nodes[each_var_name].append(
3791 3792
                            self._find_node_by_name(node.outputs,
                                                    each_var_name))
Z
Zhen Wang 已提交
3793 3794
        self.graph.resolve_hazard(var_nodes)

W
WangZhen 已提交
3795
    def has_circle(self):
3796 3797 3798 3799 3800 3801
        """
        Check if the graph has a circle.

        Returns:
            bool: True if the graph has a circle else False.
        """
W
WangZhen 已提交
3802 3803 3804
        return core.has_circle(self.graph)

    def graph_num(self):
3805 3806 3807 3808 3809 3810
        """
        Count the number of unconnected graphs in this graph.

        Returns:
            int: the number of unconnected graphs.
        """
W
WangZhen 已提交
3811 3812 3813
        return core.graph_num(self.graph)

    def topology_sort(self):
3814 3815 3816
        """
        Perform the topology sort operation on the graph.

T
tianshuo78520a 已提交
3817
        Notes: the `graph` can not contain a circle.
3818 3819

        Returns:
Z
Zhen Wang 已提交
3820
            list(IrNode): nodes in topology order.
3821
        """
3822
        ordered_nodes = core.topology_sort(self.graph)
Z
Zhen Wang 已提交
3823
        return [IrNode(n) for n in ordered_nodes]
W
WangZhen 已提交
3824 3825

    def build_adjacency_list(self):
3826 3827 3828 3829
        """
        Build an adjacency list of operations for the `graph`.

        Returns:
3830
            dict{IrNode: set(IrNode)}: the adjacency list.
3831
        """
3832 3833 3834 3835 3836
        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
3837

3838 3839 3840 3841 3842 3843 3844 3845
    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.
3846
            marked_nodes(set(IrNode)): nodes that are needed to be marked.
3847 3848 3849 3850 3851
            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.
        """

3852 3853 3854
        def _convert_to_pdf(dot_file_path):
            pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf'
            exited_code = subprocess.call('dot -Tpdf ' + dot_file_path \
3855
                                          + ' -o ' + pdf_save_path, shell=True)
3856 3857 3858 3859 3860
            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))

3861
        remove_ctr_vars = set()
3862
        if remove_ctr_var:
3863
            for node in self.all_var_nodes():
3864 3865 3866
                if node.is_ctrl_var():
                    remove_ctr_vars.add(node)
            self.safe_remove_nodes(remove_ctr_vars)
3867 3868
        print('Total ops num = {}.'.format(len(self.all_op_nodes())))

3869 3870
        if marked_nodes is not None:
            if not isinstance(marked_nodes, set):
3871 3872 3873 3874 3875 3876
                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}
3877 3878 3879 3880
            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)
3881 3882
        if not os.path.exists(save_path):
            os.makedirs(save_path)
3883 3884 3885 3886 3887 3888 3889
        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):
3890 3891 3892
        """
        Convert the graph into a Program.

Z
Zhen Wang 已提交
3893
        WARN: When the graph includes backward operator nodes, the
3894 3895 3896 3897 3898 3899
        conversion process may be failed. Usually, this function is
        only used to convert a test graph.

        Returns:
            Program: a program converted from the graph.
        """
3900
        convert_pass = core.get_pass('graph_to_program_pass')
3901 3902
        desc = core.ProgramDesc()
        convert_pass.set_not_owned('program', desc)
3903 3904 3905 3906
        convert_pass.apply(self.graph)
        program = Program._construct_from_desc(desc)
        return program

3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917
    def _find_node_by_name(self, nodes, node_name):
        """
        Find a node in the giving nodes set by the name.
        """
        target_node = None
        for n in nodes:
            if n.name() == node_name:
                target_node = n
        assert target_node is not None, "Cannot find the target node in the giving set."
        return target_node

3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933
    def _update_desc_attr(self, desc, name, val):
        """
        Update the value of desc's attribute by attribute's name.
        """
        if isinstance(val, Block):
            desc.set_block_attr(name, val.desc)
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
            desc.set_blocks_attr(name, [v.desc for v in val])
        elif isinstance(val, core.BlockDesc) or \
                isinstance(val, core.ProgramDesc):
            desc.set_serialized_attr(name, val.serialize_to_string())
        else:
            desc._set_attr(name, val)


Y
Yu Yang 已提交
3934
class Program(object):
D
dzhwinter 已提交
3935
    """
3936 3937
    Create Python Program.  It has at least one :ref:`api_guide_Block_en`, when the
    control flow op like conditional_block, while :ref:`api_fluid_layers_While` is included,
J
Jiabin Yang 已提交
3938
    it will contain nested block.
3939

J
Jiabin Yang 已提交
3940 3941 3942
    Please reference the
    `framework.proto <https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/framework.proto>`_
    for details.
D
dzhwinter 已提交
3943

J
Jiabin Yang 已提交
3944
    A set of Program usually contains startup program and main program.
J
Jiabin Yang 已提交
3945
    A startup program is set to contain some initial work, eg. initialize the ``Parameter``, and the main
J
Jiabin Yang 已提交
3946 3947 3948 3949 3950 3951 3952
    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
3953 3954 3955 3956
    **Notes**:
        **we have** :ref:`api_fluid_default_startup_program` **and** :ref:`api_fluid_default_main_program`
        **by default, a pair of them will shared the parameters. The** :ref:`api_fluid_default_startup_program` **only run once to initialize parameters,**
        :ref:`api_fluid_default_main_program` **run in every mini batch and adjust the weights.**
D
dzhwinter 已提交
3957 3958

    Returns:
J
Jiabin Yang 已提交
3959
        Program: An empty Program.
D
dzhwinter 已提交
3960 3961

    Examples:
3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974
        .. code-block:: python

            import paddle.fluid as fluid

            main_program = fluid.Program()
            startup_program = fluid.Program()
            with fluid.program_guard(main_program=main_program, startup_program=startup_program):
                x = fluid.layers.data(name="x", shape=[-1, 784], dtype='float32')
                y = fluid.layers.data(name="y", shape=[-1, 1], dtype='int32')
                z = fluid.layers.fc(name="fc", input=x, size=10, act="relu")

            print("main program is: {}".format(main_program))
            print("start up program is: {}".format(startup_program))
D
dzhwinter 已提交
3975 3976 3977

    """

3978 3979
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
3980 3981
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
3982 3983
        global global_prog_seed
        self._seed = global_prog_seed
Y
yuyang18 已提交
3984
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
3985
        self.__op_role_var = []
T
tangwei12 已提交
3986

3987 3988
        # for distribute training
        # _is_distributed = True if under distributed training
T
tangwei12 已提交
3989
        self._is_distributed = False
3990
        # _is_chief = True if the trainer is the first one, usually No.0
T
tangwei12 已提交
3991
        self._is_chief = False
3992 3993 3994
        # _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
3995
        self._endpoints = []
3996 3997 3998
        # if current role is parameter server, the _ps_endpoint is its "ip:port"
        self._ps_endpoint = None
        # trainers_endpoints, it is used for distribution.
3999
        self._trainers_endpoints = []
4000
        # the distributed lookup table names
T
tangwei12 已提交
4001
        self._distributed_lookup_table = None
4002 4003 4004

        # use Deep gradient comrepssion or not
        self._enable_dgc = False
4005 4006
        self._use_lamb = False

4007 4008 4009
        self._nccl_comm_num = 1
        self._use_hierarchical_allreduce = False
        self._hierarchical_allreduce_inter_nranks = 0
4010

4011 4012 4013
        # 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
4014
        self._program_config = None
4015

H
hutuxian 已提交
4016 4017 4018
        # assigned if this program has been parsed by a pipeline optimizer
        self._pipeline_opt = None

4019 4020 4021
        # appending gradients times
        self._appending_grad_times = 0

4022 4023 4024 4025
        # identifier for auto checkpoint
        self._auto_checkpoint_name = unique_name.generate(
            "__auto_checkpoint_program__")

4026 4027 4028
        # compiled program, i.e. Graph
        self._graph = None

4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051 4052 4053 4054 4055
    def global_seed(self, seed=0):
        """
        Set global seed for Program

        Returns:
            None.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
                print(prog.random_seed)
                ## 0
                ## the default random seed is 0

                prog.global_seed(102)
                prog1 = fluid.default_main_program()
                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
4056
    @property
4057
    def _op_role(self):
Y
yuyang18 已提交
4058 4059 4060 4061 4062 4063 4064 4065
        """
        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
4066
        parameter gradient of backward (use :code:`_op_role_var` to get this
Y
yuyang18 已提交
4067 4068 4069 4070
        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
4071 4072
        return self._current_role

4073 4074
    @_op_role.setter
    def _op_role(self, role):
Y
yuyang18 已提交
4075 4076 4077
        self._current_role = role

    @property
4078
    def _op_role_var(self):
Y
yuyang18 已提交
4079
        """
4080
        The auxiliary variables for :code:`_op_role` property.
Y
yuyang18 已提交
4081

4082
        See Also: :code:`Program._op_role`'s documentation for details.
Y
yuyang18 已提交
4083 4084 4085

        Notes: This is a very low-level API. Users should not use it directly.
        """
4086
        return self.__op_role_var
Y
yuyang18 已提交
4087

4088
    @signature_safe_contextmanager
4089 4090 4091 4092 4093
    def _backward_role_guard(self):
        tmp_role = self._current_role

        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Backward
4094 4095 4096 4097
        try:
            yield
        finally:
            self._current_role = tmp_role
4098

S
rename  
sneaxiy 已提交
4099
    @signature_safe_contextmanager
W
Wu Yi 已提交
4100
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
4101 4102 4103 4104 4105 4106 4107
        """
        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:
4108
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
4109 4110 4111

        Examples:

4112
            >>> import paddle.fluid as fluid
Y
yuyang18 已提交
4113
            >>> p, g = backward(...)
W
Wu Yi 已提交
4114
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
4115 4116
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
4117
        tmp_role = self._current_role
4118
        tmp_var = self.__op_role_var
X
Xin Pan 已提交
4119

Y
yuyang18 已提交
4120 4121
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
4122
        self.__op_role_var = [
4123 4124 4125
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
4126 4127 4128 4129 4130
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
Y
Yu Yang 已提交
4131

S
rename  
sneaxiy 已提交
4132
    @signature_safe_contextmanager
X
Xin Pan 已提交
4133
    def _lr_schedule_guard(self, is_with_opt=False):
4134 4135 4136 4137 4138 4139 4140
        """
        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
4141 4142 4143 4144
        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.
4145 4146 4147

        Examples:

4148
            >>> import paddle.fluid as fluid
4149 4150 4151 4152
            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
4153 4154

        tmp_role = self._current_role
4155
        tmp_var = self.__op_role_var
4156

4157 4158
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
4159 4160
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
4161
        # TODO(typhoonzero): how to set target learning rate var
4162
        self.__op_role_var = []
4163 4164 4165 4166 4167
        try:
            yield
        finally:
            self.__op_role_var = tmp_var
            self._current_role = tmp_role
4168

4169
    def __str__(self):
Y
yuyang18 已提交
4170 4171 4172 4173 4174 4175 4176 4177 4178
        """
        Get the protobuf debug string of this Program.

        Returns:
            (str): The protobuf debug string.

        Raises:
            ValueError: If any of required fields is not set.
        """
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
        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

            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_program._to_readable_code())
        """
        assert isinstance(
            skip_op_callstack, bool
        ), "skip_op_callstack parameter's type is error, expect bool, received %s".format(
            type(skip_op_callstack))
        program_str = ""
        for block in self.blocks:
            program_str += block._to_readable_code(skip_op_callstack)
4218
            program_str += '\n'
4219
        return program_str
Y
Yang Yang(Tony) 已提交
4220

F
fengjiayi 已提交
4221 4222 4223
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
4224

J
Jiabin Yang 已提交
4225 4226 4227
        Args:

            throw_on_error (bool): raise Value error when any of required fields is not set.
F
fengjiayi 已提交
4228

J
Jiabin Yang 已提交
4229
            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
4230

H
haowang101779990 已提交
4231
        Returns:
J
Jiabin Yang 已提交
4232
            str: The debug string describe current Program.
Y
yuyang18 已提交
4233 4234

        Raises:
J
Jiabin Yang 已提交
4235
            ValueError: If any of required fields is not set and throw_on_error is True.
F
fengjiayi 已提交
4236

4237 4238 4239 4240 4241 4242
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
4243 4244
                x = fluid.layers.data(name="X", shape=[2,3], dtype="float32", append_batch_size=False)
                pred = fluid.layers.fc(x, size=3)
4245
                prog_string = prog.to_string(throw_on_error=True, with_details=False)
4246
                prog_string_with_details = prog.to_string(throw_on_error=False, with_details=True)
T
tianshuo78520a 已提交
4247
                print("program string without detail: {}".format(prog_string))
4248
                print("program string with detail: {}".format(prog_string_with_details))
F
fengjiayi 已提交
4249
        """
4250 4251 4252 4253 4254 4255 4256 4257 4258
        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
4259 4260 4261 4262 4263 4264
        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()
4265 4266
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
4267 4268
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
4269

W
Wu Yi 已提交
4270
    def _get_desc(self):
Y
yuyang18 已提交
4271 4272 4273 4274 4275 4276 4277
        """
        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.
        """
4278 4279
        return self.desc

X
version  
Xin Pan 已提交
4280 4281 4282
    def _version(self):
        return self.desc._version()

4283
    def clone(self, for_test=False):
Y
yuyang18 已提交
4284
        """
4285
        **Notes**:
J
Jiabin Yang 已提交
4286 4287 4288 4289
            **1.** :code:`Program.clone()` **method DOES NOT clone** :ref:`api_fluid_io_DataLoader` .

            **2. Recommend you to use** :code:`clone` **before using** :code:`Opimizer.minimize`.

4290
            **3. This API has no effect in Dygraph Mode**
Y
yuyang18 已提交
4291

4292
        Create a new Program with forward content of original one when ``for_test=True``.
4293
        Create a new Program as same as the original one when ``for_test=False``.
4294

J
Jiabin Yang 已提交
4295
        Some operators, e.g., :ref:`api_fluid_layers_batch_norm` , behave differently between
Y
yuyang18 已提交
4296 4297 4298
        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`.
4299

4300 4301
        * 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.
4302 4303
          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
4304
          recommend you to use :code:`clone` before using :code:`Opimizer.minimize`.
Y
yuyang18 已提交
4305

J
Jiabin Yang 已提交
4306
        For Example:
4307
          ::
L
Luo Tao 已提交
4308

4309 4310 4311 4312 4313 4314 4315 4316
            import paddle.fluid as fluid
            img = fluid.layers.data(name='image', shape=[784])
            pred = fluid.layers.fc(input=img, size=10, act='relu')
            loss = fluid.layers.mean(pred)
            # Here we use clone before Momentum
            test_program = fluid.default_main_program().clone(for_test=True)
            optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
            optimizer.minimize(loss)
4317

J
Jiabin Yang 已提交
4318
        Args:
4319

4320 4321
            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` .
4322

J
Jiabin Yang 已提交
4323
        Returns:
4324
            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``
4325

Y
yuyang18 已提交
4326 4327 4328

        Examples:

J
Jiabin Yang 已提交
4329
        **Notes: The Program's order maybe different after** :code:`clone` **and
4330
        this will not affect your training or testing progress. In the following
J
Jiabin Yang 已提交
4331
        example we give you an simple method** :code:`print_prog(program)` **to
4332
        print Program Descs inorder to make sure you have same print result
J
Jiabin Yang 已提交
4333
        after** :code:`clone`:
4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346 4347 4348 4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369
            .. code-block:: python

                import paddle.fluid as fluid
                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))


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

                    import paddle.fluid as fluid
                    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))

                    train_program = fluid.Program()
                    startup_program = fluid.Program()
J
Jiabin Yang 已提交
4370 4371 4372

                    # startup_program is used to do some parameter init work,
                    # and main program is used to hold the network
4373 4374 4375 4376 4377 4378 4379 4380 4381
                    with fluid.program_guard(train_program, startup_program):
                        with fluid.unique_name.guard():
                            img = fluid.layers.data(name='image', shape=[784])
                            hidden = fluid.layers.fc(input=img, size=200, act='relu')
                            hidden = fluid.layers.dropout(hidden, dropout_prob=0.5)
                            loss = fluid.layers.cross_entropy(
                                                      input=fluid.layers.fc(hidden, size=10, act='softmax'),
                                        label=fluid.layers.data(name='label', shape=[1], dtype='int64'))
                            avg_loss = fluid.layers.mean(loss)
4382
                            test_program = train_program.clone(for_test=True)
4383
                    print_prog(test_program)
J
Jiabin Yang 已提交
4384 4385 4386 4387 4388 4389 4390 4391 4392

                    # 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

                    # In Paddle Fluid we will share weights by using the same Variable name. In train and test program
                    # 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.

4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414
                    with fluid.program_guard(train_program, startup_program):
                        with fluid.unique_name.guard():
                            sgd = fluid.optimizer.SGD(learning_rate=1e-3)
                            sgd.minimize(avg_loss)


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

                    import paddle.fluid as fluid
                    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))
4415 4416
                    
                    def network():
4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430
                        img = fluid.layers.data(name='image', shape=[784])
                        hidden = fluid.layers.fc(input=img, size=200, act='relu')
                        hidden = fluid.layers.dropout(hidden, dropout_prob=0.5)
                        loss = fluid.layers.cross_entropy(
                            input=fluid.layers.fc(hidden, size=10, act='softmax'),
                            label=fluid.layers.data(name='label', shape=[1], dtype='int64'))
                        avg_loss = fluid.layers.mean(loss)
                        return avg_loss

                    train_program_2 = fluid.Program()
                    startup_program_2 = fluid.Program()
                    test_program_2 = fluid.Program()
                    with fluid.program_guard(train_program_2, startup_program_2):
                        with fluid.unique_name.guard():
4431 4432 4433
                            avg_loss = network()
                            sgd = fluid.optimizer.SGD(learning_rate=1e-3)
                            sgd.minimize(avg_loss)
4434
                    # the test startup program is not used.
4435
                    with fluid.program_guard(test_program_2, startup_program_2):
4436
                        with fluid.unique_name.guard():
4437 4438
                            avg_loss = network()
                    print_prog(test_program_2)
4439 4440

        The two code snippets above will generate and print same programs.
4441
        """
4442 4443 4444 4445 4446

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

4447
        pruned_origin_block_id_map = None
4448
        if for_test:
4449 4450 4451 4452 4453 4454 4455 4456 4457
            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)
4458
        else:
4459
            p = Program()
G
gongweibao 已提交
4460 4461
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
4462
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
4463 4464 4465
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
4466 4467

            p._current_role = self._current_role
4468
            p.__op_role_var = self.__op_role_var
4469
            p._appending_grad_times = self._appending_grad_times
4470 4471
            if hasattr(self, 'lr_sheduler'):
                p.lr_sheduler = self.lr_sheduler
G
gongweibao 已提交
4472

4473 4474
            #NOTE(zhiqiu): we sync the cloned program, to update its program by
            # its desc.
W
Wu Yi 已提交
4475
            p._sync_with_cpp()
4476

W
Wu Yi 已提交
4477
        p._copy_param_info_from(self)
4478
        p._copy_data_info_from(self, pruned_origin_block_id_map)
4479
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
4480
        return p
4481

4482
    def _prune(self, targets):
Y
yuyang18 已提交
4483 4484 4485 4486 4487 4488 4489 4490
        """
        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:
4491
            targets(list|Variable|Operator): A list of variables, operators, or variable names
Y
yuyang18 已提交
4492 4493 4494 4495
                need to be pruned

        Returns:
            Program:  A new, pruned program.
4496
        """
4497
        return self._prune_with_input([], targets)
4498 4499

    def _prune_with_input(self, feeded_var_names, targets):
Y
yuyang18 已提交
4500
        """
4501 4502 4503 4504 4505 4506 4507 4508 4509 4510
        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()
4511
            targets(list|Variable|Operator): A list of variables, operators, or variable names
4512 4513 4514 4515 4516 4517
                need to be pruned

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

4518 4519 4520 4521
        #NOTE(zhiqiu): we sync the original program first, since its program may diff with
        # its desc due to modifying desc in c++ space. E.g. save op will add kLookupTablePath in desc.
        self._sync_with_cpp()

4522 4523
        if not isinstance(feeded_var_names, list):
            feeded_var_names = [feeded_var_names]
4524 4525
        if not isinstance(targets, list):
            targets = [targets]
4526 4527 4528

        for var in feeded_var_names:
            if not isinstance(var, six.string_types):
4529 4530 4531
                raise ValueError(
                    "All feeded_var_names of Program._prune_with_input() can only be "
                    "str, but received %s." % type(var))
4532

4533 4534 4535 4536
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
4537 4538 4539
                    name = t.name
                elif isinstance(t, six.string_types):
                    name = str(t)
4540
                else:
4541 4542 4543
                    raise ValueError(
                        "All targets of Program._prune_with_input() can only be "
                        "Variable or Operator, but received %s." % type(t))
4544 4545 4546 4547 4548 4549 4550 4551

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

4552 4553 4554 4555 4556 4557 4558 4559 4560 4561 4562 4563 4564 4565 4566 4567
                # 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.
                        # Skip optimize op except for optimize op in targets, 
                        # since optimize op generates parameters.
                        if op._is_optimize_op() and op not in targets:
                            continue
                        else:
                            target_op = op
                            break
4568 4569 4570 4571 4572 4573 4574 4575
                if target_op is None:
                    raise ValueError(
                        "The target variable used for pruning should have an "
                        "associated operator that generates it.")
                else:
                    targets_idx.append([target_op.block.idx, target_op.idx])
            else:
                targets_idx.append([t.block.idx, t.idx])
4576

4577
        res = Program()
4578 4579 4580
        res.desc, pruned_origin_block_id_map = core.prune(self.desc,
                                                          set(feeded_var_names),
                                                          targets_idx)
M
minqiyang 已提交
4581 4582 4583
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
4584
        res._sync_with_cpp()
4585 4586 4587 4588 4589

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

4590 4591
        return res

X
Xin Pan 已提交
4592
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
4593
        """
F
fengjiayi 已提交
4594 4595 4596 4597 4598
        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.

4599
        3. change the :code:`is_test`
Y
yuyang18 已提交
4600 4601 4602
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

4603
        Args:
X
Xin Pan 已提交
4604 4605
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
4606

Y
yuyang18 已提交
4607 4608 4609 4610 4611 4612
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
4613
        res = Program()
4614
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
4615 4616 4617 4618

        # 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
4619
        if prune_read_op:
4620 4621 4622 4623 4624 4625 4626 4627 4628
            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
4629
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
4630 4631

        # change all `is_test` attributes to True
M
minqiyang 已提交
4632
        for i in six.moves.range(res.desc.num_blocks()):
4633
            block = res.desc.block(i)
M
minqiyang 已提交
4634
            for j in six.moves.range(block.op_size()):
4635 4636
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
4637
                    op._set_attr('is_test', True)
M
minqiyang 已提交
4638 4639 4640
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
4641
        res._sync_with_cpp()
4642 4643
        return res

4644 4645
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
4646
        """
J
Jiabin Yang 已提交
4647 4648 4649 4650
        **Notes**:
            **1. All information about parameters will be lost after serialization**

            **2. This API has no effect in Dygraph mode**
Y
yuyang18 已提交
4651

4652 4653
        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
4654

J
Jiabin Yang 已提交
4655
        Args:
Y
yuyang18 已提交
4656

J
Jiabin Yang 已提交
4657
            binary_str_type (str): the binary prootbuf string.
4658

J
Jiabin Yang 已提交
4659 4660
        Returns:
            Program: A deserialized Program.
4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681 4682

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                startup_prog = fluid.Program()
                main_prog = fluid.Program()
                with fluid.program_guard(startup_prog, main_prog):
                    x = fluid.layers.data(
                        name='X', shape=[1000, 784], dtype='float32', append_batch_size=False)

                    y = fluid.layers.data(
                        name='Y', shape=[784, 100], dtype='float32', append_batch_size=False)

                    z = fluid.layers.mul(x=x, y=y)

                    binary_str = fluid.default_main_program().desc.serialize_to_string()
                    prog_restored = fluid.default_main_program().parse_from_string(binary_str)

                    print(fluid.default_main_program())
                    print(prog_restored)
Y
yuyang18 已提交
4683
        """
4684 4685
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
4686
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
4687
        p._sync_with_cpp()
4688
        return p
Y
Yu Yang 已提交
4689

4690
    @staticmethod
4691
    def _construct_from_desc(desc):
4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706
        """
        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
4707 4708
    @property
    def random_seed(self):
Y
yuyang18 已提交
4709
        """
J
Jiabin Yang 已提交
4710
        The default random seed for random operators in Program. ``0`` means get
Y
yuyang18 已提交
4711 4712
        the random seed from random device.

J
Jiabin Yang 已提交
4713 4714 4715 4716
        **Notes: It must be set before the operators have been added.**

        Returns:
            int64: Random seed in current Program
4717

4718 4719 4720 4721 4722 4723 4724 4725

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
                random_seed = prog.random_seed
4726
                x_var = fluid.layers.data(name="X", shape=[3,3], dtype="float32", append_batch_size=False)
4727 4728 4729
                print(random_seed)
                ## 0
                ## the default random seed is 0
4730 4731

                # Here we need to set random seed before we use fluid.layers.dropout
4732
                prog.random_seed = 1
4733 4734
                z_var = fluid.layers.dropout(x_var, 0.7)

4735
                print(prog.random_seed)
4736 4737
                ## 1
                ## the random seed is change to 1
Y
yuyang18 已提交
4738
        """
D
dzhwinter 已提交
4739 4740
        return self._seed

Q
qiaolongfei 已提交
4741 4742
    @property
    def num_blocks(self):
Y
yuyang18 已提交
4743
        """
4744 4745
        The number of :ref:`api_guide_Block_en`  in this Program.

J
Jiabin Yang 已提交
4746 4747 4748 4749
        **Notes: This API has no effect in Dygraph mode**

        Returns:
            int(Platform-dependent size): num of :ref:`api_guide_Block_en`  in current Program
4750

4751 4752 4753 4754 4755 4756 4757 4758 4759

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
                num_blocks = prog.num_blocks
                print(num_blocks)
4760 4761


Y
yuyang18 已提交
4762
        """
Q
qiaolongfei 已提交
4763 4764
        return self.desc.num_blocks()

D
dzhwinter 已提交
4765 4766 4767
    @random_seed.setter
    def random_seed(self, seed):
        if not isinstance(seed, int):
4768 4769 4770
            raise ValueError(
                "Program.random_seed's input seed must be an integer, but received %s."
                % type(seed))
D
dzhwinter 已提交
4771 4772
        self._seed = seed

Y
Yu Yang 已提交
4773
    def __repr__(self):
4774
        return self.__str__()
4775

Y
Yu Yang 已提交
4776
    def global_block(self):
Y
yuyang18 已提交
4777
        """
J
Jiabin Yang 已提交
4778 4779
        **Notes**:
            **This API has no effect in Dygraph mode**
4780 4781 4782

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

J
Jiabin Yang 已提交
4783 4784
        Returns:
            :ref:`api_guide_Block_en`: The first :ref:`api_guide_Block_en`  of this Program.
4785

4786 4787 4788 4789 4790 4791 4792 4793 4794

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
                gb_block = prog.global_block()
                print(gb_block)
4795

Y
yuyang18 已提交
4796
        """
Y
Yu Yang 已提交
4797 4798
        return self.blocks[0]

Q
Qiao Longfei 已提交
4799
    def block(self, index):
Y
yuyang18 已提交
4800
        """
J
Jiabin Yang 已提交
4801 4802
        **Notes**:
            **This API has no effect in Dygraph mode**
Y
yuyang18 已提交
4803

4804 4805
        Get the :code:`index`  :ref:`api_guide_Block_en`  of this Program

J
Jiabin Yang 已提交
4806 4807
        Args:
            index (int) - The index of  :ref:`api_guide_Block_en`  to get
4808

J
Jiabin Yang 已提交
4809 4810
        Returns:
            :ref:`api_guide_Block_en`: The :code:`index` block
4811 4812 4813 4814 4815 4816 4817 4818 4819

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
                block_0 = prog.block(0)
                print(block_0)
Y
yuyang18 已提交
4820
        """
Q
Qiao Longfei 已提交
4821 4822
        return self.blocks[index]

Y
Yu Yang 已提交
4823
    def current_block(self):
Y
yuyang18 已提交
4824
        """
J
Jiabin Yang 已提交
4825 4826
        **Notes**:
            **This API has no effect in Dygraph mode**
4827

J
Jiabin Yang 已提交
4828 4829
        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.
4830

J
Jiabin Yang 已提交
4831 4832
        Returns:
             :ref:`api_guide_Block_en`: The :code:`index`  :ref:`api_guide_Block_en`
4833

4834 4835 4836 4837 4838 4839 4840 4841
        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
                current_blk = prog.current_block()
                print(current_blk)
Y
yuyang18 已提交
4842
        """
Y
Yu Yang 已提交
4843 4844
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
4845
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
4846 4847 4848 4849 4850
        """
        Create a new block with the :code:`parent_idx` and change the current block
        to new block.

        Args:
J
Jiabin Yang 已提交
4851

Y
yuyang18 已提交
4852 4853 4854 4855 4856
            parent_idx(int): The parent block index.

        Returns:
            Block: The new block.
        """
Y
Yu Yang 已提交
4857
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
4858 4859 4860
        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
4861 4862 4863 4864
        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
4865
    def _rollback(self):
Y
yuyang18 已提交
4866 4867 4868 4869 4870
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
4871 4872
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
4873
    def _sync_with_cpp(self):
Y
yuyang18 已提交
4874 4875 4876 4877 4878 4879 4880 4881 4882 4883
        """
        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
4884 4885 4886
        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
4887
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
4888

W
Wu Yi 已提交
4889
    def _copy_param_info_from(self, other):
4890
        """
4891
        Copy the information of parameters from other program.
D
dzhwinter 已提交
4892

Y
yuyang18 已提交
4893 4894 4895
        Notes: This is a very low level API. Users should not invoke it
        directly.

4896 4897 4898 4899 4900 4901 4902
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
4903 4904 4905
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
4906

W
Wu Yi 已提交
4907
        self.global_block()._copy_param_info_from(other.global_block())
4908

4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919
    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):
4920 4921 4922
            raise TypeError(
                "Function Program._copy_param_info_from() needs to pass in a source Program, but received %s"
                % type(other))
4923 4924
        self._is_distributed = other._is_distributed
        self._is_chief = other._is_chief
4925
        self._parameters_on_pservers = other._parameters_on_pservers
4926
        self._endpoints = other._endpoints
4927
        self._ps_endpoint = other._ps_endpoint
4928 4929
        self._distributed_lookup_table = other._distributed_lookup_table

4930
    def _copy_data_info_from(self, other, pruned_origin_block_id_map=None):
F
fengjiayi 已提交
4931 4932
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
4933

Y
yuyang18 已提交
4934 4935 4936
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
4937 4938
        Args:
            other(Program): Other program
4939 4940 4941 4942
            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
4943 4944 4945 4946 4947

        Returns:
            None
        """
        if not isinstance(other, Program):
4948 4949 4950
            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
4951

4952 4953 4954 4955 4956
        if not pruned_origin_block_id_map:
            pruned_origin_block_id_map = {
                i: i
                for i in six.moves.range(self.desc.num_blocks())
            }
4957 4958 4959

        # NOTE(zhiqiu): All vars in cloned program exist in original program.
        # The reverse is not true, due to backward pruning.
4960 4961
        for i, block in enumerate(self.blocks):
            other_block = other.blocks[pruned_origin_block_id_map[i]]
4962
            for var in list(block.vars.values()):
4963 4964 4965 4966 4967 4968 4969
                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
4970

4971
    def list_vars(self):
Y
yuyang18 已提交
4972
        """
J
Jiabin Yang 已提交
4973
        Get all :ref:`api_guide_Variable_en` from this Program. A iterable object is returned.
Y
yuyang18 已提交
4974

J
Jiabin Yang 已提交
4975 4976
        Returns:
            iterable :ref:`api_guide_Variable_en`: The Generator will yield every variable in this program.
4977 4978 4979 4980 4981 4982 4983 4984 4985 4986 4987

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid

                prog = fluid.default_main_program()
                img = fluid.layers.data(name='img', shape=[1,28,28], dtype='float32')
                label = fluid.layers.data(name='label', shape=[128,1], dtype='int64')
                for var in prog.list_vars():
                    print(var)
Y
yuyang18 已提交
4988
        """
4989
        for each_block in self.blocks:
4990
            for each_var in list(each_block.vars.values()):
4991 4992
                yield each_var

4993 4994 4995 4996 4997 4998 4999 5000 5001 5002 5003 5004 5005 5006 5007 5008 5009 5010 5011 5012 5013 5014 5015 5016 5017 5018 5019 5020 5021 5022 5023 5024 5025 5026 5027 5028 5029 5030 5031 5032 5033 5034 5035 5036 5037 5038 5039 5040 5041 5042 5043 5044 5045 5046 5047 5048 5049 5050
    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

                import paddle.fluid as fluid

                program = fluid.default_main_program()
                data = fluid.data(name='x', shape=[None, 13], dtype='float32')
                hidden = fluid.layers.fc(input=data, size=10)
                loss = fluid.layers.mean(hidden)
                fluid.optimizer.SGD(learning_rate=0.01).minimize(loss)

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

                # Here will print all parameters in current program, in this example,
                # the result is like:
                #
                # name: "fc_0.w_0"
                # type {
                #   type: LOD_TENSOR
                #   lod_tensor {
                #     tensor {
                #       data_type: FP32
                #       dims: 13
                #       dims: 10
                #     }
                #   }
                # }
                # persistable: true
                #
                # name: "fc_0.b_0"
                # type {
                # type: LOD_TENSOR
                # lod_tensor {
                #     tensor {
                #       data_type: FP32
                #       dims: 10
                #     }
                #   }
                # }
                # persistable: true
                #
                # Here print(param) will print out all the properties of a parameter,
                # including name, type and persistable, you can access to specific
                # property of a parameter, such as param.name, param.type
        """
        parameters = []
        for each_block in self.blocks:
            parameters.extend(each_block.all_parameters())
        return parameters

Y
Yu Yang 已提交
5051

5052
@six.add_metaclass(ParameterMetaClass)
Y
Yu Yang 已提交
5053
class Parameter(Variable):
5054
    """
5055
    Parameter is derived from Variable. A parameter is a persistable
5056
    Variable, and will be updated by optimizers after each iteration.
5057
    The training of a neural network is essentially the updating of
5058 5059
    its parameters.

5060
    Relative to a general Variable, a Parameter has several its own
5061 5062
    member variables:

5063 5064 5065 5066 5067 5068 5069 5070 5071 5072
    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.
5073 5074
    """

5075 5076 5077 5078 5079 5080
    def __init__(self,
                 block,
                 shape,
                 dtype,
                 type=core.VarDesc.VarType.LOD_TENSOR,
                 **kwargs):
5081 5082 5083 5084 5085
        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
5086
        if len(shape) == 0:
5087 5088
            raise ValueError(
                "The dimensions of shape for Parameter must be greater than 0")
Y
Yu Yang 已提交
5089 5090 5091

        for each in shape:
            if each < 0:
5092 5093 5094
                raise ValueError(
                    "Each dimension of shape for Parameter must be greater than 0, but received %s"
                    % list(shape))
5095 5096

        Variable.__init__(
5097 5098 5099 5100 5101 5102 5103
            self,
            block,
            persistable=True,
            shape=shape,
            dtype=dtype,
            type=type,
            **kwargs)
Y
Yu Yang 已提交
5104 5105 5106 5107
        self.trainable = kwargs.get('trainable', True)

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

5108 5109
        self.regularizer = kwargs.get('regularizer', None)

W
wanghaoshuang 已提交
5110
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
5111

5112 5113
        self.is_distributed = False

F
fengjiayi 已提交
5114
    def __str__(self):
5115
        return self._to_readable_code()
F
fengjiayi 已提交
5116

F
update  
fengjiayi 已提交
5117 5118 5119
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
5120

F
update  
fengjiayi 已提交
5121 5122 5123 5124 5125 5126 5127 5128
        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.

5129 5130 5131 5132 5133 5134 5135 5136 5137
        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
5138 5139 5140 5141 5142 5143
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
            res_str = Variable.to_string(self, throw_on_error, True)
            additional_attr = ("trainable", "optimize_attr", "regularizer",
5144
                               "do_model_average")
F
update  
fengjiayi 已提交
5145
            for attr_name in additional_attr:
5146 5147
                res_str += "%s: %s\n" % (attr_name,
                                         cpt.to_text(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
5148 5149
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
5150 5151 5152 5153
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
5154

5155 5156
class ParamBase(core.VarBase):
    """
5157 5158 5159
    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.
5160 5161 5162
    The training of a neural network is essentially the updating of
    its ParamBase.

5163
    Relative to a general Tensor, a ParamBase has several its own
5164 5165 5166 5167 5168 5169 5170 5171 5172 5173 5174 5175 5176 5177 5178 5179 5180 5181 5182 5183 5184 5185 5186 5187 5188 5189 5190 5191 5192 5193 5194 5195 5196 5197 5198 5199 5200 5201 5202 5203 5204 5205
    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.
    """

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

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

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

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

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

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

5206 5207
        trainable = kwargs.get('trainable', True)
        self.stop_gradient = not trainable
5208 5209 5210 5211 5212 5213 5214 5215

        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.is_distributed = False
5216
        # self.block = default_main_program().global_block()
5217

5218 5219 5220 5221 5222 5223 5224 5225 5226 5227 5228 5229 5230
    @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))

5231
    def __str__(self):
5232
        """
5233
        Convert a ParamBase object to a readable string.
5234

5235
        Returns(str): A readable string.
5236 5237 5238 5239

        Examples:
            .. code-block:: python

5240
                import paddle
5241
                paddle.disable_static()
5242 5243 5244 5245 5246 5247 5248 5249
                conv = paddle.nn.Conv2D(3, 3, 5)
                print(conv.weight)
                # Parameter: conv2d_0.w_0
                #   - place: CUDAPlace(0)
                #   - shape: [3, 3, 5, 5]
                #   - layout: NCHW
                #   - dtype: float
                #   - data: [...] 
5250
                paddle.enable_static()
5251
        """
5252 5253
        return "Parameter containing:\n  {}\n  - stop_gradient: {}".format(
            super(ParamBase, self).__str__(), self.stop_gradient)
5254 5255 5256 5257

    __repr__ = __str__


Y
Yu Yang 已提交
5258
# program is a global instance.
Y
Yu Yang 已提交
5259 5260
_main_program_ = Program()
_startup_program_ = Program()
5261

5262

5263
def default_startup_program():
Y
Yu Yang 已提交
5264
    """
Y
yuyang18 已提交
5265 5266
    Get default/global startup program.

J
Jiabin Yang 已提交
5267 5268 5269
    The layer function in :ref:`api_fluid_layers` will create parameters, :ref:`api_paddle_data_reader_reader` ,
    `NCCL <https://developer.nvidia.com/nccl>`_ handles as global variables. The :code:`startup_program` will
    initialize them by the OPs in startup  :ref:`api_fluid_Program` . The  :ref:`api_fluid_layers`  function will
Y
yuyang18 已提交
5270 5271 5272
    append these initialization operators into startup program.

    This method will return the :code:`default` or the :code:`current` startup
J
Jiabin Yang 已提交
5273
    program. Users can use  :ref:`api_fluid_program_guard`  to switch :ref:`api_fluid_Program` .
5274

J
Jiabin Yang 已提交
5275
    Returns: current default startup :ref:`api_fluid_Program`
5276

J
Jiabin Yang 已提交
5277
    Returns type: :ref:`api_fluid_Program`
5278 5279 5280 5281 5282 5283 5284 5285 5286 5287 5288 5289 5290 5291 5292

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            main_program = fluid.Program()
            startup_program = fluid.Program()
            with fluid.program_guard(main_program=main_program, startup_program=startup_program):
                x = fluid.layers.data(name="x", shape=[-1, 784], dtype='float32')
                y = fluid.layers.data(name="y", shape=[-1, 1], dtype='int32')
                z = fluid.layers.fc(name="fc", input=x, size=10, act="relu")

                print("main program is: {}".format(fluid.default_main_program()))
                print("start up program is: {}".format(fluid.default_startup_program()))
Y
Yu Yang 已提交
5293
    """
Y
Yu Yang 已提交
5294
    return _startup_program_
5295

5296

5297
def default_main_program():
Y
Yu Yang 已提交
5298
    """
5299 5300 5301 5302 5303
    This API can be used to get ``default main program`` which store the 
    descriptions of ``op`` and ``variable``.
    
    For example ``z = fluid.layers.elementwise_add(x, y)`` will create a new ``elementwise_add`` 
    ``op`` and a new ``z`` ``variable``, and they will be recorded in ``default main program`` 
Y
yuyang18 已提交
5304

5305 5306
    The ``default_main_program`` is the default value for ``Program`` parameter in 
    a lot of ``fluid`` APIs. For example, the :code:`Executor.run()` will execute the
Y
yuyang18 已提交
5307
    :code:`default_main_program` when the program is not specified.
5308

5309 5310
    If you want to replace the ``default main program``, you can use :ref:`api_fluid_program_guard`
    
Y
Yu Yang 已提交
5311
    Returns:
5312
        :ref:`api_fluid_Program`: a ``Program`` which holding the descriptions of ops and variables in the network.
5313 5314 5315 5316 5317

    Examples:
        ..  code-block:: python

            import paddle.fluid as fluid
5318

5319
            # Sample Network:
5320 5321
            data = fluid.data(name='image', shape=[None, 3, 224, 224], dtype='float32')
            label = fluid.data(name='label', shape=[None, 1], dtype='int64')
5322 5323 5324 5325 5326 5327 5328 5329 5330 5331 5332 5333 5334 5335 5336 5337 5338 5339 5340
            
            conv1 = fluid.layers.conv2d(data, 4, 5, 1, act=None)
            bn1 = fluid.layers.batch_norm(conv1, act='relu')
            pool1 = fluid.layers.pool2d(bn1, 2, 'max', 2)
            conv2 = fluid.layers.conv2d(pool1, 16, 5, 1, act=None)
            bn2 = fluid.layers.batch_norm(conv2, act='relu')
            pool2 = fluid.layers.pool2d(bn2, 2, 'max', 2)
            
            fc1 = fluid.layers.fc(pool2, size=50, act='relu')
            fc2 = fluid.layers.fc(fc1, size=102, act='softmax')
            
            loss = fluid.layers.cross_entropy(input=fc2, label=label)
            loss = fluid.layers.mean(loss)
            opt = fluid.optimizer.Momentum(
                learning_rate=0.1,
                momentum=0.9,
                regularization=fluid.regularizer.L2Decay(1e-4))
            opt.minimize(loss)
            
5341
            #print the number of blocks in the program, 1 in this case
5342
            print(fluid.default_main_program().num_blocks)
5343 5344

            #print the description of variable 'image'
5345
            print(fluid.default_main_program().blocks[0].var('image'))
5346

Y
Yu Yang 已提交
5347
    """
Y
Yu Yang 已提交
5348
    return _main_program_
Y
Yu Yang 已提交
5349 5350 5351 5352 5353


def switch_main_program(program):
    """
    Switch the main program to a new program.
5354

Y
Yu Yang 已提交
5355 5356 5357 5358 5359 5360 5361 5362 5363 5364 5365 5366 5367 5368
    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):
    """
5369
    Switch the startup program to a new program
Y
Yu Yang 已提交
5370 5371 5372 5373 5374 5375 5376 5377 5378 5379 5380 5381
    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
5382
@signature_safe_contextmanager
Y
Yu Yang 已提交
5383 5384
def program_guard(main_program, startup_program=None):
    """
5385 5386
    :api_attr: Static Graph

5387 5388
    Change the global main program and startup program with `"with"` statement.
    Layer functions in the Python `"with"` block will append operators and
Y
yuyang18 已提交
5389
    variables to the new main programs.
5390

G
guofei 已提交
5391 5392 5393 5394 5395 5396 5397
    Args:
        main_program(Program): New main program inside `"with"` statement.
        startup_program(Program, optional): New startup program inside `"with"` 
            statement. :code:`None` means not changing startup program, 
            default_startup_program is still used.
            Default: None.

Y
Yu Yang 已提交
5398
    Examples:
5399 5400 5401
       .. code-block:: python
       
         import paddle.fluid as fluid
Y
yuyang18 已提交
5402

5403 5404 5405
         main_program = fluid.Program()
         startup_program = fluid.Program()
         with fluid.program_guard(main_program, startup_program):
G
guofei 已提交
5406
             data = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
5407
             hidden = fluid.layers.fc(input=data, size=10, act='relu')
Y
yuyang18 已提交
5408 5409 5410

    Notes: The temporary :code:`Program` can be used if the user does not need
    to construct either of startup program or main program.
5411

Y
Yu Yang 已提交
5412
    Examples:
5413
       .. code-block:: python
Y
yuyang18 已提交
5414

5415 5416 5417 5418 5419
         import paddle.fluid as fluid

         main_program = fluid.Program()
         # does not care about startup program. Just pass a temporary value.
         with fluid.program_guard(main_program, fluid.Program()):
G
guofei 已提交
5420 5421
             data = fluid.data(name='image', shape=[None, 784, 784], dtype='float32')
    
Y
Yu Yang 已提交
5422
    """
5423 5424
    from .data_feeder import check_type
    check_type(main_program, 'main_program', Program, 'fluid.program_guard')
Y
Yu Yang 已提交
5425 5426
    main_program = switch_main_program(main_program)
    if startup_program is not None:
5427 5428
        check_type(startup_program, 'startup_program', Program,
                   'fluid.program_guard')
Y
Yu Yang 已提交
5429
        startup_program = switch_startup_program(startup_program)
5430 5431 5432 5433 5434 5435
    try:
        yield
    finally:
        switch_main_program(main_program)
        if startup_program is not None:
            switch_startup_program(startup_program)
X
xuwei06 已提交
5436 5437


W
Wu Yi 已提交
5438
def _get_var(name, program=None):
X
xuwei06 已提交
5439
    """
Y
yuyang18 已提交
5440
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
5441

X
xuwei06 已提交
5442 5443 5444
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
5445
        If None, default_global_program() will be used.
X
xuwei06 已提交
5446 5447 5448 5449 5450 5451 5452

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
5453
    assert isinstance(program, Program)
X
xuwei06 已提交
5454 5455

    return program.global_block().var(name)
5456 5457


S
rename  
sneaxiy 已提交
5458
@signature_safe_contextmanager
L
lujun 已提交
5459 5460 5461 5462
def _dygraph_guard(tracer):
    global _dygraph_tracer_
    tmp_trace = _dygraph_tracer_
    _dygraph_tracer_ = tracer
5463
    core._switch_tracer(tracer)
M
minqiyang 已提交
5464

5465 5466 5467 5468 5469
    try:
        yield
    finally:
        core._switch_tracer(tmp_trace)
        _dygraph_tracer_ = tmp_trace
P
Paddle CI 已提交
5470 5471


S
rename  
sneaxiy 已提交
5472
@signature_safe_contextmanager
L
lujun 已提交
5473
def _dygraph_place_guard(place):
5474 5475 5476
    global _global_expected_place_
    tmp_place = _global_expected_place_
    _global_expected_place_ = place
M
minqiyang 已提交
5477

5478 5479 5480
    try:
        yield
    finally:
5481
        _global_expected_place_ = tmp_place
5482 5483 5484 5485


def load_op_library(lib_filename):
    """
5486 5487
    :api_attr: Static Graph
    
5488 5489 5490
    Load a dynamic library, including custom operators and kernels.
    When library is loaded, ops and kernels registered in the library
    will be available in PaddlePaddle main process.
T
tianshuo78520a 已提交
5491
    Please note, the type of custom operators can't have the same type
5492 5493 5494 5495 5496 5497 5498 5499 5500 5501 5502 5503 5504 5505
    with the existing operators in the framework.

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

    Examples:
        .. code-block:: python

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

    """
    core.load_op_library(lib_filename)
    OpProtoHolder.instance().update_op_proto()
5506 5507 5508 5509 5510 5511 5512 5513 5514 5515 5516 5517 5518 5519 5520 5521 5522 5523 5524 5525 5526 5527 5528 5529 5530 5531 5532 5533 5534 5535 5536 5537 5538 5539 5540 5541 5542 5543 5544 5545 5546 5547 5548 5549 5550 5551 5552 5553 5554 5555 5556 5557


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


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

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

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

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

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

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

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

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

5558 5559 5560 5561 5562
    index = None
    if device and ':' in device:
        device, index = device.split(':')
        if device == 'cpu':
            raise ValueError("Should not set device id for cpu.")
5563 5564 5565 5566
    if device not in ['cpu', 'gpu', '', None]:
        raise ValueError(
            "The Attr(device) should be 'cpu' or 'gpu', and it can also be empty string or None "
            "when there is no need to specify device. But received %s" % device)
5567 5568
    if index:
        device = ":".join([device, index])
5569
    pre_device = switch_device(device)
5570 5571 5572 5573
    try:
        yield
    finally:
        switch_device(pre_device)
G
guofei 已提交
5574 5575 5576 5577 5578 5579 5580 5581 5582 5583 5584 5585 5586 5587 5588 5589 5590 5591 5592 5593 5594 5595 5596 5597 5598 5599 5600 5601 5602 5603 5604 5605 5606 5607 5608 5609 5610 5611 5612 5613 5614 5615 5616 5617 5618 5619 5620 5621 5622 5623 5624 5625 5626 5627 5628 5629 5630 5631 5632 5633 5634 5635 5636 5637 5638 5639 5640


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

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

    Examples:
            .. code-block:: python

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


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

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

    Returns:
        flag's value in Paddle.

    Examples:
        .. code-block:: python

            import paddle.fluid as fluid

            flags = ['FLAGS_eager_delete_tensor_gb', 'FLAGS_check_nan_inf']
            res = fluid.get_flags(flags)
            print(res)
            # {'FLAGS_eager_delete_tensor_gb': 0.0, 'FLAGS_check_nan_inf': False}
    """
    flags_value = {}
    if isinstance(flags, (list, tuple)):
        for key in flags:
            if (core.globals().is_public(key)):
                value = core.globals()[key]
                temp = {key: value}
                flags_value.update(temp)
            else:
                raise ValueError(
                    'Flag %s cannot get its value through this function.' %
                    (key))
    elif isinstance(flags, str):
        if (core.globals().is_public(flags)):
            value = core.globals()[flags]
            temp = {flags: value}
            flags_value.update(temp)
        else:
            raise ValueError(
                'Flag %s cannot get its value through this function.' % (flags))
    else:
        raise TypeError('Flags in get_flags should be a list, tuple or string.')
    return flags_value