framework.py 115.7 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15 16
from __future__ import print_function

Y
Yu Yang 已提交
17
import collections
X
Xin Pan 已提交
18
from collections import defaultdict
W
WangZhen 已提交
19
from collections import Iterable
Q
qiaolongfei 已提交
20
import contextlib
S
rename  
sneaxiy 已提交
21
from .wrapped_decorator import signature_safe_contextmanager
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
Q
qiaolongfei 已提交
30

M
minqiyang 已提交
31
from .. import compat as cpt
32
from .proto import framework_pb2
33
try:
P
peizhilin 已提交
34
    if os.name == 'nt':
P
peizhilin 已提交
35
        import sys
P
peizhilin 已提交
36 37 38 39 40
        third_lib_path = os.path.abspath(os.path.dirname(
            __file__)) + os.sep + '..' + os.sep + 'libs'
        os.environ['path'] += ';' + third_lib_path
        sys.path.append(third_lib_path)

41
    from . import core
42
except ImportError as e:
P
peizhilin 已提交
43
    if os.name == 'nt':
44
        executable_path = os.path.abspath(os.path.dirname(sys.executable))
P
peizhilin 已提交
45
        raise ImportError(
46 47 48 49 50
            """NOTE: You may need to run \"set PATH=%s;%%PATH%%\"
        if you encounters \"DLL load failed\" errors. If you have python
        installed in other directory, replace \"%s\" with your own
        directory. The original error is: \n %s""" %
            (executable_path, executable_path, cpt.get_exception_message(e)))
P
peizhilin 已提交
51 52 53 54 55 56
    else:
        raise ImportError(
            """NOTE: You may need to run \"export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH\"
        if you encounters \"libmkldnn.so not found\" errors. If you have python
        installed in other directory, replace \"/usr/local/lib\" with your own
        directory. The original error is: \n""" + cpt.get_exception_message(e))
57
except Exception as e:
58
    raise e
59
from . import unique_name
Y
Yu Yang 已提交
60

61
__all__ = [
62 63 64 65
    'Program',
    'default_startup_program',
    'default_main_program',
    'program_guard',
66
    'name_scope',
S
sneaxiy 已提交
67 68 69
    'cuda_places',
    'cpu_places',
    'cuda_pinned_places',
70
]
Y
Yu Yang 已提交
71

Q
qiaolongfei 已提交
72 73 74 75
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
76 77
CONTROL_DEP_VAR_PREFIX = core.kControlDepVarName()

L
lujun 已提交
78 79
_dygraph_tracer_ = None
_dygraph_current_expected_place_ = None
80 81


L
lujun 已提交
82 83
def _in_dygraph_mode():
    return _dygraph_tracer_ is not None
84 85


L
lujun 已提交
86 87
def _dygraph_tracer():
    return _dygraph_tracer_
88

W
Wu Yi 已提交
89

M
minqiyang 已提交
90
def _current_expected_place():
L
lujun 已提交
91
    return _dygraph_current_expected_place_
M
minqiyang 已提交
92 93


S
sneaxiy 已提交
94 95 96 97 98
def _cpu_num():
    return int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))


def cuda_places(device_ids=None):
S
add doc  
sneaxiy 已提交
99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
    '''
    Create a list of :code:`fluid.CUDAPlace` objects.

    If :code:`device_ids` is None, environment variable of
    :code:`FLAGS_selected_gpus` would be checked first. If
    :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
    gpu places would be returned.  

    If :code:`device_ids` is not None, it should be the device
    ids of gpus. For example, if :code:`device_ids=[0,1,2]`, 
    the returned list would be 
    [fluid.CUDAPlace(0), fluid.CUDAPlace(1), fluid.CUDAPlace(2)].
    
    Args: 
        device_ids (None|list(int)|tuple(int)): gpu device id list.

    Returns:
        out (list(fluid.CUDAPlace)): gpu place list.
    '''
S
sneaxiy 已提交
120 121 122 123 124 125 126 127 128 129 130 131 132 133
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_ids is None:
        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())
    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):
S
add doc  
sneaxiy 已提交
134 135 136 137 138 139 140 141 142 143 144 145 146 147
    '''
    Create a list of :code:`fluid.CPUPlace` objects.
    
    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
    If :code:`CPU_NUM` is not set, the device count would
    be determined by :code:`multiprocessing.cpu_count()`. 

    Args:
        device_count (None|int): device number.

    Returns:
        out (list(fluid.CPUPlace)): cpu place list.
    '''
S
sneaxiy 已提交
148 149 150 151 152 153
    if device_count is None:
        device_count = _cpu_num()
    return [core.CPUPlace()] * device_count


def cuda_pinned_places(device_count=None):
S
add doc  
sneaxiy 已提交
154 155 156 157 158 159 160 161 162 163 164 165 166 167
    '''
    Create a list of :code:`fluid.CUDAPinnedPlace` objects.

    If :code:`device_count` is None, the device count would
    be determined by environment variable :code:`CPU_NUM`. 
    If :code:`CPU_NUM` is not set, the device count would
    be determined by :code:`multiprocessing.cpu_count()`. 

    Args:
        device_count (None|int): device number.

    Returns:
        out (list(fluid.CUDAPinnedPlace)): cuda pinned place list.
    '''
S
sneaxiy 已提交
168 169 170 171 172 173 174
    assert core.is_compiled_with_cuda(), \
        "Not compiled with CUDA"
    if device_count is None:
        device_count = _cpu_num()
    return [core.cuda_pinned_places()] * device_count


175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200
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
201
@signature_safe_contextmanager
202 203 204 205 206 207 208 209 210 211 212 213
def name_scope(prefix=None):
    """
    Generate hierarchical name prefix for the operators.

    Note: This should only used for debugging and visualization purpose.
    Don't use it for serious analysis such as graph/program transformations.

    Args:
        prefix(str): prefix.

    Examples:
        .. code-block:: python
T
Tink_Y 已提交
214

215 216 217 218
          with name_scope("encoder"):
             ...
          with name_scope("decoder"):
             ...
T
Tink_Y 已提交
219 220
          with name_scope("attention"):
             ...
221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239
    """
    # TODO(panyx0718): Only [0-9a-z].
    assert prefix, "namescope prefix cannot be empty."
    global _name_scope
    _name_scope = _name_scope.child(prefix)
    yield
    _name_scope = _name_scope.parent()


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
240 241 242
def generate_control_dev_var_name():
    import random
    return CONTROL_DEP_VAR_PREFIX + "@" + str(random.random())
Q
qiaolongfei 已提交
243 244 245 246


def grad_var_name(var_name):
    """
247 248
    Returns:
        str: gradient name for a certain var name
Q
qiaolongfei 已提交
249 250 251
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
252

253
def convert_np_dtype_to_dtype_(np_dtype):
254 255
    """
    Convert the data type in numpy to the data type in Paddle
256

257
    Args:
258
        np_dtype(np.dtype): the data type in numpy.
259

260 261
    Returns:
        core.VarDesc.VarType: the data type in Paddle.
262 263

    """
264 265
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
266
        return core.VarDesc.VarType.FP32
267
    elif dtype == np.float64:
268
        return core.VarDesc.VarType.FP64
269
    elif dtype == np.float16:
270
        return core.VarDesc.VarType.FP16
271
    elif dtype == np.int32:
272
        return core.VarDesc.VarType.INT32
273
    elif dtype == np.int16:
274
        return core.VarDesc.VarType.INT16
275
    elif dtype == np.int64:
276
        return core.VarDesc.VarType.INT64
277
    elif dtype == np.bool:
278
        return core.VarDesc.VarType.BOOL
279 280
    elif dtype == np.uint16:
        return core.VarDesc.VarType.INT16
281 282
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
Q
qingqing01 已提交
283 284
    elif dtype == np.int8:
        return core.VarDesc.VarType.INT8
285
    else:
M
minqiyang 已提交
286
        raise ValueError("Not supported numpy dtype %s" % dtype)
287 288 289


def dtype_is_floating(dtype):
290 291 292
    """
    Check the data type is floating or not.
    Args:
293
        dtype(np.dtype|core.VarDesc.VarType): data type.
294 295 296 297 298
            Could be numpy format or Paddle format

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

    """
299
    if not isinstance(dtype, core.VarDesc.VarType):
300 301
        dtype = convert_np_dtype_to_dtype_(dtype)

302 303 304 305
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
306 307


Y
Yang Yang(Tony) 已提交
308
def _debug_string_(proto, throw_on_error=True):
309 310 311 312 313 314 315 316 317 318 319
    """
    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
320
    error_fields = list()
Y
Yang Yang(Tony) 已提交
321
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
322 323
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
324 325 326
    return proto.__str__()


X
Xin Pan 已提交
327
class Variable(object):
328
    """
329 330 331
    In Fluid, every input and output of an operator is a variable. In most
    cases, variables are used for holding different kinds of data or training
    labels. A variable belongs to a block. All variable has its own name and
332
    two variables in different blocks could have the same name.
333

334 335
    There are many kinds of variables. Each kind of them has its own attributes
    and usages. Please reference the framework.proto for details.
336

337
    Most of a Variable's member variables can be setted to be None. It mean
338
    it is not available or will be specified later.
339 340

    Args:
341
        block(Block): The block that the variable belongs to.
342 343
        type(core.VarDesc.VarType): Variable type. Please reference the
            framework.proto for details.
344 345
        name(str|None): The name of the variable. If setted None, it will be
            generated automatically. Default: None
346
        shape(tuple|list|None): The shape of the variable. -1 means the batch size.
347
            Some kinds of variable do not contain shape, just set it to None.
348 349 350
            Default: None
        dtype(np.dtype|core.VarDesc.VarType|str|None): The data type of variable.
            Default: None
351
        lod_level (int|None): The level of lod tensor. 0 means it is not a time
352
            series data.
353
            Default: None
354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375
        capacity (int|None): The capacity of Channel variable. Ignored for other
            types. Default: None
        persistable (bool|None): True if the variable is persistable. A persistable
            variable will not be deleted after an iteration ending. Defaults: None.
        error_clip (BaseErrorClipAttr|None): The error clip attributes of the
            corresponding gradient variable. Default: None
        stop_gradient (bool): True if the variable will stop to calculate its
            gradients when backward. Default: False.
        is_data (bool): True if the variable is an input data. Default: False

    Notes:
        The constructor of Variable should not be invoked directly. Please
        use `Block.create_var` to create a variable.

    Examples:
        .. code-block:: python

            cur_program = Program()
            cur_block = cur_program.current_block()
            new_variable = cur_block.create_var(name="X",
                                                shape=[-1, 23, 48],
                                                dtype='float32')
376 377
    """

Y
Yu Yang 已提交
378 379
    def __init__(self,
                 block,
Y
Yu Yang 已提交
380
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
381 382 383 384
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
385
                 capacity=None,
Q
QI JUN 已提交
386
                 persistable=None,
F
fengjiayi 已提交
387
                 error_clip=None,
Y
Yu Yang 已提交
388
                 stop_gradient=False,
F
fengjiayi 已提交
389
                 is_data=False,
Y
Yu Yang 已提交
390
                 **kwargs):
Y
Yu Yang 已提交
391 392
        self.block = block
        if name is None:
Y
Yu Yang 已提交
393
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
394

Y
Yu Yang 已提交
395
        if dtype is not None:
396
            if not isinstance(dtype, core.VarDesc.VarType):
397
                dtype = convert_np_dtype_to_dtype_(dtype)
398

L
lujun 已提交
399
        if _in_dygraph_mode():
M
minqiyang 已提交
400
            # record vars in tracer rather than blocks
M
minqiyang 已提交
401 402
            self._ivar = kwargs.get("ivar", None)
            if not self._ivar:
403 404 405
                self._ivar = core.VarBase(
                    name, dtype if dtype else core.VarDesc.VarType.FP32,
                    list(shape) if shape else [],
X
fix  
Xin Pan 已提交
406 407
                    _current_expected_place(), stop_gradient, True
                    if persistable else False)
M
minqiyang 已提交
408
            if persistable:
L
lujun 已提交
409
                _dygraph_tracer().trace_var(name, self)
M
minqiyang 已提交
410
        else:
411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482
            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))

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

            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))

            if shape is not None:
                if is_new_var:
                    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))

            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

M
minqiyang 已提交
483
            self.block.vars[name] = self
484
            self.op = None
485
            self._stop_gradient = stop_gradient
486
            self.is_data = is_data
Y
Yu Yang 已提交
487

488
    def numpy(self):
M
minqiyang 已提交
489
        new_ivar = self._ivar._copy_to(core.CPUPlace(), True)
P
Paddle CI 已提交
490
        return np.array(new_ivar.value().get_tensor())
491

492
    def backward(self):
X
Xin Pan 已提交
493
        self._ivar._run_backward()
494

495
    def gradient(self):
496 497
        new_ivar = self._ivar._grad_ivar()._copy_to(core.CPUPlace(), True)
        return np.array(new_ivar.value().get_tensor())
498

499
    def clear_gradient(self):
X
Xin Pan 已提交
500
        self._ivar._clear_gradient()
X
Xin Pan 已提交
501

502
    def __str__(self):
Y
Yang Yang(Tony) 已提交
503 504
        return self.to_string(True)

F
update  
fengjiayi 已提交
505
    def to_string(self, throw_on_error, with_details=False):
506 507 508 509
        """
        Get debug string.

        Args:
510 511
            throw_on_error(bool): True if raise an exception when self is
                not initialized.
F
update  
fengjiayi 已提交
512
            with_details(bool): more details about variables and parameters
513 514
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False;
515

516 517
        Returns:
            str: The debug string.
518
        """
L
lujun 已提交
519 520
        if _in_dygraph_mode():
            # TODO(panyx0718): add more dygraph debug info.
521 522 523
            return 'name %s, dtype: %s shape: %s' % (self.name, self.dtype,
                                                     self.shape)

F
update  
fengjiayi 已提交
524 525
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
526
        protostr = self.desc.serialize_to_string()
527
        proto = framework_pb2.VarDesc.FromString(six.binary_type(protostr))
F
update  
fengjiayi 已提交
528 529 530 531
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
532 533
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
534
        return res_str
535 536 537

    __repr__ = __str__

538
    def set_desc(self, input):
539 540 541 542 543 544 545 546 547
        """
        Set the variable description.

        Args:
            input(core.VarDesc): The new VarDesc.

        Returns:
            None
        """
548 549
        self.desc = input

550
    @property
551
    def stop_gradient(self):
L
lujun 已提交
552
        if _in_dygraph_mode():
M
minqiyang 已提交
553 554
            return self._ivar.stop_gradient
        else:
555
            return self._stop_gradient
556

557 558
    @stop_gradient.setter
    def stop_gradient(self, s):
L
lujun 已提交
559
        if _in_dygraph_mode():
M
minqiyang 已提交
560
            self._ivar.stop_gradient = s
561
        else:
562
            self._stop_gradient = s
563

564 565
    @property
    def persistable(self):
L
lujun 已提交
566
        if _in_dygraph_mode():
567 568 569
            return self._ivar.persistable
        else:
            return self.desc.persistable()
570

Y
Yu Yang 已提交
571 572
    @persistable.setter
    def persistable(self, p):
L
lujun 已提交
573
        if _in_dygraph_mode():
574 575 576
            return self._ivar.persistable
        else:
            self.desc.set_persistable(p)
Y
Yu Yang 已提交
577

Y
Yu Yang 已提交
578 579
    @property
    def name(self):
L
lujun 已提交
580
        if _in_dygraph_mode():
581 582 583
            return self._ivar.name
        else:
            return cpt.to_text(self.desc.name())
Y
Yu Yang 已提交
584

T
typhoonzero 已提交
585 586
    @name.setter
    def name(self, new_name):
L
lujun 已提交
587
        if _in_dygraph_mode():
588 589 590
            self._ivar.name = new_name
        else:
            self.desc.set_name(new_name)
T
typhoonzero 已提交
591

Y
Yu Yang 已提交
592 593 594
    @property
    def shape(self):
        # convert to tuple, make it as same as numpy API.
L
lujun 已提交
595
        if _in_dygraph_mode():
596 597 598
            return self._ivar.shape
        else:
            return tuple(self.desc.shape())
Y
Yu Yang 已提交
599 600

    @property
F
fengjiayi 已提交
601
    def dtype(self):
L
lujun 已提交
602
        if _in_dygraph_mode():
603 604 605
            return self._ivar.dtype
        else:
            return self.desc.dtype()
Y
Yu Yang 已提交
606 607 608

    @property
    def lod_level(self):
L
lujun 已提交
609
        # TODO(minqiyang): Support lod_level in dygraph mode
610
        return self.desc.lod_level()
Y
Yu Yang 已提交
611

Y
Yu Yang 已提交
612 613
    @property
    def type(self):
L
lujun 已提交
614
        if _in_dygraph_mode():
615 616 617
            return self._ivar.dtype
        else:
            return self.desc.type()
Y
Yu Yang 已提交
618

W
Wu Yi 已提交
619
    def _set_error_clip(self, error_clip):
620 621 622 623 624 625 626 627 628
        """
        Set the error_clip.

        Args:
            error_clip(BaseErrorClipAttr) : The new error_clip.

        Returns:
            None
        """
629 630
        self.error_clip = error_clip

631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717
    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:
            raise ValueError("slice step cannot be zero")

        # 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
718
    def _cloneVar(self, copy=False):
719 720 721 722 723
        if not copy:
            return self.block.create_var(
                name=unique_name.generate(".".join(self.name)),
                dtype=self.dtype,
                persistable=self.persistable,
724
                stop_gradient=self.stop_gradient, )
725 726 727 728
        else:
            return self

    def _sliceVar(self, axes, starts, ends):
L
lujun 已提交
729
        new_var = self._cloneVar()
730 731 732 733 734 735 736 737 738 739
        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
740
        new_var = self._cloneVar()
741 742 743 744 745 746 747 748 749 750
        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
751
                return self._cloneVar(True)
752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769
            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
770
                return self._cloneVar(True)
771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792
            index = int(item)
            if (index > 0 and index >= self.shape[axis])\
                    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):
        """
        Slice the variable.

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

        Returns:
            Sliced variable
        """
        new_var = None
        if isinstance(item, tuple):
            if len(item) > len(self.shape):
                raise IndexError("Too many indexes")
W
wopeizl 已提交
793 794 795 796 797 798
            fixedSize = True
            for i in range(len(self.shape)):
                if self.shape[i] == -1:
                    fixedSize = False
                    break

799
            newitem = self._reconstructSliceinfo(item) or item
W
wopeizl 已提交
800 801
            if fixedSize:
                check, info = self._detectContinuesSlice(newitem)
802
                if check:
W
wopeizl 已提交
803 804 805 806 807 808 809 810
                    starts = info[0]
                    ends = info[1]
                    axes = [i for i in range(len(starts))]
                    return self._sliceVar(axes, starts, ends)
                else:
                    new_var = self
                    for index, o in enumerate(newitem):
                        new_var = new_var._sliceAndConcatVar(o, index)
811 812 813 814 815 816 817 818
            else:
                new_var = self
                for index, o in enumerate(newitem):
                    new_var = new_var._sliceAndConcatVar(o, index)
        else:
            new_var = self._sliceAndConcatVar(item, 0)
        return new_var

Y
Yu Yang 已提交
819

F
fengjiayi 已提交
820 821 822
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
823

824 825
    Returns:
       list: list of OpProto.
F
fengjiayi 已提交
826 827 828 829
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
830
        op_proto = framework_pb2.OpProto.FromString(six.binary_type(pbstr))
F
fengjiayi 已提交
831 832 833 834 835
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
836 837 838 839
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
840 841 842 843 844 845 846 847 848
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
849
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
850 851 852 853 854 855
        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):
856 857 858 859 860 861 862 863
        """
        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
864 865
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
866 867
        return self.op_proto_map[type]

868 869 870 871
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
S
sneaxiy 已提交
872
            core.op_proto_and_checker_maker.kOpRoleVarAttrName(),
873 874
            core.op_proto_and_checker_maker.kOpNameScopeAttrName(),
            core.op_proto_and_checker_maker.kOpCreationCallstackAttrName()
875 876
        }

F
fengjiayi 已提交
877

X
Xin Pan 已提交
878
class Operator(object):
879
    """
880 881 882 883 884 885 886
    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
887
        type(str): The type of operator. Default None.
888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907
        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
908
        Block.append_op or Block._prepend_op instead.
909 910 911 912 913 914 915 916 917 918

    Examples:
        .. code-block:: python

            cur_program = Program()
            cur_block = cur_program.current_block()
            # var1 += var2 + var3
            cur_block.append_op(type="sum",
                                inputs={"X": [var1, var2, var3]},
                                outputs={"Out": [var1]})
919
    """
920
    OP_WITHOUT_KERNEL_SET = {
921 922 923
        'feed', 'fetch', 'recurrent', 'go', 'rnn_memory_helper_grad',
        'conditional_block', 'while', 'send', 'recv', 'listen_and_serv',
        'ncclInit', 'select', 'checkpoint_notify', 'gen_nccl_id'
924
    }
925

Y
Yu Yang 已提交
926 927
    def __init__(self,
                 block,
Y
Yu Yang 已提交
928
                 desc,
Y
Yu Yang 已提交
929 930 931
                 type=None,
                 inputs=None,
                 outputs=None,
M
minqiyang 已提交
932
                 attrs=None):
L
lujun 已提交
933
        if _in_dygraph_mode():
934 935
            if type is None:
                raise ValueError(
936
                    "`type` to initialized an Operator can not be None.")
937
            self.iop = core.OpBase(type)
M
minqiyang 已提交
938

939 940
            # TODO(minqiyang): remove these lines after we take apart all
            # backward grads and forward variables
X
Xin Pan 已提交
941
            self.inputs = defaultdict(list)
X
Xin Pan 已提交
942
            if inputs is not None:
X
Xin Pan 已提交
943 944 945 946 947
                for k, v in six.iteritems(inputs):
                    if isinstance(v, Variable):
                        self.inputs[k].append(v._ivar)
                    elif isinstance(v, list) or isinstance(v, tuple):
                        self.inputs[k].extend([var._ivar for var in v])
M
minqiyang 已提交
948

X
Xin Pan 已提交
949
            self.outputs = defaultdict(list)
X
Xin Pan 已提交
950
            if outputs is not None:
X
Xin Pan 已提交
951 952 953 954 955
                for k, v in six.iteritems(outputs):
                    if isinstance(v, Variable):
                        self.outputs[k].append(v._ivar)
                    elif isinstance(v, list) or isinstance(v, tuple):
                        self.outputs[k].extend([var._ivar for var in v])
F
fengjiayi 已提交
956

957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051
            self.attrs = attrs if attrs else {}
        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(
                )] = self.block.program.op_role

            role_var_name = op_maker.kOpRoleVarAttrName()
            if len(self.block.program.
                   op_role_var) != 0 and role_var_name not in op_attrs:
                op_attrs[role_var_name] = self.block.program.op_role_var

            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(
                    "`type` to initilized an Operator can not be None.")
            else:
                callstack_var_name = op_maker.kOpCreationCallstackAttrName()
                op_attrs[callstack_var_name] = list(
                    reversed(traceback.format_stack()))[1:]

            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()

            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]
                        if not isinstance(in_args, list):
                            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 = []
                        for arg in in_args:
                            if isinstance(arg, six.string_types):
                                in_arg_names.append(arg)
                            elif isinstance(arg, six.binary_type):
                                in_arg_names.append(arg.decode())
                            else:
                                in_arg_names.append(cpt.to_text(arg.name))
                        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
1052
                        if not _in_dygraph_mode():
1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
                            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
1072
    def _has_kernel(self, op_type):
1073 1074
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
1075
    def to_string(self, throw_on_error):
1076
        """
1077 1078
        Get debug string.

1079
        Args:
1080 1081
            throw_on_error(bool): Whether to raise exception if self is not
                initialized.
1082

1083 1084
        Returns:
            str: The debug string.
1085 1086

        """
1087
        protostr = self.desc.serialize_to_string()
1088
        proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr))
Y
Yang Yang(Tony) 已提交
1089 1090 1091 1092
        return _debug_string_(proto, throw_on_error)

    def __str__(self):
        return self.to_string(True)
1093 1094 1095

    __repr__ = __str__

F
fengjiayi 已提交
1096 1097
    @property
    def type(self):
L
lujun 已提交
1098
        if _in_dygraph_mode():
1099 1100 1101
            return self.iop.type
        else:
            return self.desc.type()
F
fengjiayi 已提交
1102 1103

    def input(self, name):
1104
        """
1105
        Get the input arguments according to the input parameter name.
1106

1107 1108
        Args:
            name(str): The input parameter name.
1109

1110 1111 1112
        Returns:
            list: return the list of argument names that associated with \
                the specific parameter name.
1113
        """
F
fengjiayi 已提交
1114 1115
        return self.desc.input(name)

W
Wu Yi 已提交
1116
    def _rename_input(self, old_name, new_name):
1117 1118 1119 1120 1121 1122 1123 1124 1125 1126
        """
        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
1127
        self.desc._rename_input(old_name, new_name)
T
typhoonzero 已提交
1128

W
Wu Yi 已提交
1129
    def _rename_output(self, old_name, new_name):
1130 1131 1132 1133 1134 1135 1136 1137 1138 1139
        """
        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
1140
        self.desc._rename_output(old_name, new_name)
T
typhoonzero 已提交
1141

F
fengjiayi 已提交
1142 1143 1144 1145
    @property
    def input_names(self):
        return self.desc.input_names()

T
typhoonzero 已提交
1146 1147 1148 1149 1150 1151 1152 1153
    @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
1154
    def output(self, name):
1155
        """
1156
        Get output arguments by the output parameter name.
1157

1158 1159
        Args:
            name(str): The output parameter name.
1160

1161 1162 1163
        Returns:
            list: return the list of argument names associated with \
                the specific parameter name.
1164
        """
F
fengjiayi 已提交
1165 1166 1167 1168 1169 1170
        return self.desc.output(name)

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

1171 1172 1173 1174 1175 1176 1177 1178
    @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
1179
    def has_attr(self, name):
1180
        """
1181 1182
        Whether this Operator has the attribute with name or not.

1183
        Args:
1184
            name(str): the attribute name.
1185

1186 1187
        Returns:
            bool: True if has this attribute.
1188 1189

        """
F
fengjiayi 已提交
1190 1191 1192
        return self.desc.has_attr(name)

    def attr_type(self, name):
1193
        """
1194
        Get the type of attribute by attribute's name.
1195

1196 1197
        Args:
            name(str): the attribute name.
1198

1199 1200
        Returns:
            core.AttrType: the attribute type.
1201
        """
F
fengjiayi 已提交
1202 1203
        return self.desc.attr_type(name)

W
Wu Yi 已提交
1204
    def _set_attr(self, name, val):
1205 1206 1207 1208 1209 1210 1211 1212 1213 1214
        """
        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
1215 1216
        self._update_desc_attr(name, val)

1217 1218 1219
    def _remove_attr(self, name):
        self.desc.remove_attr(name)

G
gongweibao 已提交
1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230
    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
1231 1232
        if isinstance(val, Block):
            self.desc.set_block_attr(name, val.desc)
Y
Yancey1989 已提交
1233 1234
        elif isinstance(val, list) and val and all(
                isinstance(v, Block) for v in val):
1235
            self.desc.set_blocks_attr(name, [v.desc for v in val])
Q
Qiyang Min 已提交
1236 1237 1238 1239
        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
1240
            self.desc._set_attr(name, val)
Y
yuyang18 已提交
1241

F
fengjiayi 已提交
1242 1243 1244 1245 1246
    @property
    def attr_names(self):
        return self.desc.attr_names()

    def attr(self, name):
1247
        """
1248 1249
        Get the attribute by name.

1250
        Args:
1251
            name(str): the attribute name.
1252

1253 1254
        Returns:
            bool|int|str|float|list: The attribute value. The return value
1255 1256
            can be any valid attribute type.
        """
F
fengjiayi 已提交
1257
        return self.desc.attr(name)
Y
Yu Yang 已提交
1258

W
Wu Yi 已提交
1259
    def _block_attr_id(self, name):
1260
        """
G
gongweibao 已提交
1261
        Get the block attribute's id by name.
1262

1263 1264
        Args:
            name(str): the attribute name.
1265

1266 1267
        Returns:
            int: the block index.
1268
        """
W
Wu Yi 已提交
1269
        return self.desc._block_attr_id(name)
G
gongweibao 已提交
1270

W
Wu Yi 已提交
1271
    def _block_attr(self, name):
G
gongweibao 已提交
1272 1273 1274 1275 1276 1277 1278 1279 1280 1281
        """
        Get the block attribute  by name.

        Args:
            name(str): the attribute name.

        Returns:
            block: the block attribute.
        """

W
Wu Yi 已提交
1282
        id = self._block_attr_id(name)
G
gongweibao 已提交
1283 1284 1285
        assert (id >= 0 and id < len(self.block.program.blocks))
        return self.block.program.blocks[id]

W
Wu Yi 已提交
1286
    def _blocks_attr(self, name):
G
gongweibao 已提交
1287 1288 1289 1290 1291 1292 1293 1294 1295 1296
        """
        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
1297
        for i in self._blocks_attr_ids(name):
G
gongweibao 已提交
1298 1299 1300 1301 1302
            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
1303
    def _blocks_attr_ids(self, name):
G
gongweibao 已提交
1304 1305 1306 1307 1308 1309 1310 1311 1312 1313
        """
        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
1314
        return self.desc._blocks_attr_ids(name)
Y
Yu Yang 已提交
1315

J
JiayiFeng 已提交
1316
    def all_attrs(self):
F
fengjiayi 已提交
1317
        """
1318 1319 1320
        Get the attribute dict.

        Returns:
G
gongweibao 已提交
1321
            dict: The Operator's attribute dict, name->attr.
F
fengjiayi 已提交
1322 1323 1324 1325
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
G
gongweibao 已提交
1326 1327
            attr_type = self.desc.attr_type(n)
            if attr_type == core.AttrType.BLOCK:
W
Wu Yi 已提交
1328
                attr_map[n] = self._block_attr(n)
G
gongweibao 已提交
1329 1330 1331
                continue

            if attr_type == core.AttrType.BLOCKS:
W
Wu Yi 已提交
1332
                attr_map[n] = self._blocks_attr(n)
G
gongweibao 已提交
1333 1334 1335 1336
                continue

            attr_map[n] = self.attr(n)

F
fengjiayi 已提交
1337 1338
        return attr_map

Y
Yu Yang 已提交
1339

Y
Yu Yang 已提交
1340
class Block(object):
1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354
    """
    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
1355
        use `Program._create_block()` to create a block.
1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369

    Examples:
        .. code-block:: python

            cur_program = Program()
            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
1370
    def __init__(self, program, idx):
Y
Yu Yang 已提交
1371
        self.desc = program.desc.block(idx)
1372
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
1373
        self.ops = list()  # operator list
Y
Yu Yang 已提交
1374
        self.program = program
1375
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
1376

1377
    def __str__(self):
Y
Yang Yang(Tony) 已提交
1378 1379
        return self.to_string(True)

F
fengjiayi 已提交
1380 1381
    def to_string(self, throw_on_error, with_details=False):
        """
1382 1383
        Get debug string.

F
fengjiayi 已提交
1384 1385
        Args:
            throw_on_error(bool): raise exception when self is not initialized
1386
                when throw_on_error is True.
F
update  
fengjiayi 已提交
1387
            with_details(bool): more details about variables and parameters
1388 1389
                (e.g. trainable, optimize_attr, ...) will be printed when
                with_details is True. Default False.
F
fengjiayi 已提交
1390

1391 1392
        Returns:
            str: The debug string.
F
fengjiayi 已提交
1393 1394 1395 1396
        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
1397
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
1398 1399
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
1400
            for var in list(self.vars.values()):
F
fengjiayi 已提交
1401
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
1402
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
1403
            for op in self.ops:
F
fengjiayi 已提交
1404 1405
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
1406 1407 1408
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
1409 1410
            proto = framework_pb2.BlockDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
1411 1412
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1413 1414 1415

    __repr__ = __str__

Y
Yu Yang 已提交
1416 1417
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
1418
        return self.desc.parent
Y
Yu Yang 已提交
1419

Y
Yu Yang 已提交
1420 1421 1422 1423
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

W
Wu Yi 已提交
1424
    def _set_forward_block_idx(self, idx):
1425 1426 1427 1428 1429 1430 1431 1432 1433
        """
        Set the forward block Idx.

        Args:
            idx(int): the block index.

        Returns:
            None
        """
W
Wu Yi 已提交
1434
        self.desc._set_forward_block_idx(idx)
Y
Yu Yang 已提交
1435

Y
Yu Yang 已提交
1436 1437
    @property
    def idx(self):
Y
Yu Yang 已提交
1438
        return self.desc.id
Y
Yu Yang 已提交
1439

Q
Qiao Longfei 已提交
1440
    def var(self, name):
1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453
        """
        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.
        """
1454
        if not isinstance(name, six.string_types):
M
minqiyang 已提交
1455 1456 1457
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
1458 1459
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
1460
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
1461
        return v
Q
Qiao Longfei 已提交
1462

X
Xin Pan 已提交
1463
    def _find_var_recursive(self, name):
1464 1465 1466 1467 1468 1469 1470
        """
        Get a Variable by name from this block recursively.

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

        Returns:
X
Xin Pan 已提交
1471
            Variable: the Variable with the giving name. Or None if not found.
1472
        """
Y
Yu Yang 已提交
1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496
        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
1497
        return None
Y
Yu Yang 已提交
1498

X
Xin Pan 已提交
1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517
    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
1518

Q
Qiao Longfei 已提交
1519
    def all_parameters(self):
1520
        return list(self.iter_parameters())
1521

1522
    def iter_parameters(self):
M
minqiyang 已提交
1523
        return (item[1] for item in six.iteritems(self.vars)
1524
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
1525

Y
Yu Yang 已提交
1526
    def create_var(self, *args, **kwargs):
1527
        var = Variable(block=self, *args, **kwargs)
1528 1529
        if 'initializer' in kwargs:
            kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
1530
        return var
Y
Yu Yang 已提交
1531

Q
Qiao Longfei 已提交
1532 1533 1534
    def has_var(self, name):
        return name in self.vars

W
Wu Yi 已提交
1535
    def _rename_var(self, name, new_name):
T
typhoonzero 已提交
1536 1537
        """
        Rename variable in vars and ops' inputs and outputs
1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549

        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
1550
        """
M
minqiyang 已提交
1551 1552
        name = cpt.to_text(name)
        new_name = cpt.to_text(new_name)
M
minqiyang 已提交
1553

T
typhoonzero 已提交
1554
        if not self.has_var(name):
1555
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
1556 1557
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
1558
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
1559 1560 1561 1562 1563 1564 1565
            stop_gradient = v.stop_gradient
            trainable = v.trainable
            optimize_attr = v.optimize_attr
            regularizer = v.regularizer
            gradient_clip_attr = v.gradient_clip_attr
            error_clip = v.error_clip
        elif type(v) == Variable:
T
typhoonzero 已提交
1566
            var_type = "Variable"
T
wip  
typhoonzero 已提交
1567 1568 1569 1570
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
1571
        orig_var_type = v.type
M
minqiyang 已提交
1572
        self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name))
W
Wu Yi 已提交
1573
        # NOTE: v is destroyed by C++ after calling _rename_var.
M
minqiyang 已提交
1574
        d = self.desc.find_var(cpt.to_bytes(new_name))
T
typhoonzero 已提交
1575
        if var_type == "Parameter":
T
wip  
typhoonzero 已提交
1576 1577 1578 1579
            var = Parameter(
                self,
                d.shape(),
                d.dtype(),
T
typhoonzero 已提交
1580
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1581 1582 1583 1584 1585 1586 1587
                name=new_name,
                stop_gradient=stop_gradient,
                trainable=trainable,
                optimize_attr=optimize_attr,
                regularizer=regularizer,
                gradient_clip_attr=gradient_clip_attr,
                error_clip=error_clip)
T
typhoonzero 已提交
1588
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
1589 1590
            var = Variable(
                self,
T
typhoonzero 已提交
1591
                type=orig_var_type,
T
wip  
typhoonzero 已提交
1592 1593 1594 1595
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

W
Wu Yi 已提交
1596
        # rename the python side, _sync_with_cpp will only add
T
wip  
typhoonzero 已提交
1597 1598 1599
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
W
Wu Yi 已提交
1600
        self._sync_with_cpp()
1601
        return var
T
typhoonzero 已提交
1602

W
Wu Yi 已提交
1603 1604
    def _remove_var(self, name):
        self._sync_with_cpp()
M
minqiyang 已提交
1605
        self.desc._remove_var(cpt.to_bytes(name))
1606 1607
        del self.vars[name]

Y
Yu Yang 已提交
1608 1609
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
Q
Qiao Longfei 已提交
1610
        param = Parameter(global_block, *args, **kwargs)
1611
        if 'initializer' in kwargs:
1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631

            def _is_inited_by(block, var):
                init_ops = []
                for op in block.ops:
                    if var.name in op.output_arg_names:
                        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:
                #TODO already inited, do nothing, should log a warning
                pass
            else:
                initializer(param, self)
Q
Qiao Longfei 已提交
1632
        return param
Y
Yu Yang 已提交
1633

Y
Yu Yang 已提交
1634
    def append_op(self, *args, **kwargs):
1635 1636 1637 1638 1639 1640
        """
        Appends a new Operator according to the giving arguments.

        Returns:
            Operator: the append Operator.
        """
L
lujun 已提交
1641
        if _in_dygraph_mode():
1642 1643 1644 1645 1646 1647 1648 1649
            op = Operator(
                block=self,
                desc=None,
                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
1650 1651 1652
            # record ops in tracer rather than blocks
            #
            # TODO(minqiyang): add op stop_gradient support in static mode too.
L
lujun 已提交
1653 1654
            # currently, we only support stop_gradient in dygraph mode.
            _dygraph_tracer().trace_op(op, kwargs.get("stop_gradient", False))
M
minqiyang 已提交
1655
        else:
1656 1657 1658 1659 1660 1661 1662 1663 1664
            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
1665
            self.ops.append(op)
M
minqiyang 已提交
1666

1667 1668
        return op

W
Wu Yi 已提交
1669
    def _insert_op(self, index, *args, **kwargs):
1670 1671 1672 1673 1674 1675 1676 1677 1678
        """
        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
1679 1680
        self._sync_with_cpp()
        op_desc = self.desc._insert_op(index)
Q
qiaolongfei 已提交
1681 1682 1683 1684
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

W
Wu Yi 已提交
1685
    def _remove_op(self, index):
1686 1687 1688 1689 1690 1691 1692 1693 1694
        """
        Remove the specific position operator.

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

        Returns:
            None
        """
W
Wu Yi 已提交
1695 1696
        self._sync_with_cpp()
        self.desc._remove_op(index, index + 1)
1697 1698
        del self.ops[index]

W
Wu Yi 已提交
1699
    def _slice_ops(self, start, end):
1700 1701 1702 1703 1704 1705 1706 1707 1708 1709
        """
        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
1710
        return self.ops[start:end]
Y
Yancey1989 已提交
1711

W
Wu Yi 已提交
1712
    def _prepend_op(self, *args, **kwargs):
L
lujun 已提交
1713
        if _in_dygraph_mode():
1714 1715 1716 1717 1718 1719 1720
            op = Operator(
                self,
                None,
                type=kwargs.get("type", None),
                inputs=kwargs.get("inputs", None),
                outputs=kwargs.get("outputs", None),
                attrs=kwargs.get("attrs", None))
L
lujun 已提交
1721
            _dygraph_tracer().trace_op(op, kwargs.get("stop_gradient", False))
M
minqiyang 已提交
1722
        else:
1723 1724 1725 1726 1727 1728 1729 1730
            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
1731
            self.ops.insert(0, op)
1732

Y
Yu Yang 已提交
1733 1734
        return op

W
Wu Yi 已提交
1735
    def _sync_with_cpp(self):
1736
        """
1737 1738
        Sync from the desc on the c++ end. This method is used to synchronize
        the c++ desc instance generated by backward.
1739
        """
Q
Qiao Longfei 已提交
1740 1741 1742 1743 1744
        # 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())

1745
        # sync variables removed from c++ end
1746
        for var in list(self.vars.keys()):
M
minqiyang 已提交
1747
            if not self.desc.find_var(cpt.to_bytes(var)):
1748 1749
                self.vars.pop(var)

Q
Qiao Longfei 已提交
1750
        # sync operators from cpp
1751 1752 1753 1754
        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
1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770
        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
1771 1772 1773 1774 1775

        # 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
1776
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
1777 1778 1779 1780 1781 1782 1783

        # 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)

1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796
        # 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
1797 1798 1799 1800
        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
1801
    def _copy_param_info_from(self, other):
1802
        """
1803 1804
        Copy the information of parameters from the other block.

1805
        Args:
1806 1807 1808 1809 1810
            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.
1811 1812 1813 1814 1815

        Returns:
            None
        """
        if not isinstance(other, Block):
W
Wu Yi 已提交
1816 1817
            raise TypeError(
                "_copy_param_info_from should be invoked with Block")
1818
        for p in other.iter_parameters():
1819 1820 1821
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
W
Wu Yi 已提交
1822
                raise ValueError("_copy_param_info_from should be invoked with "
1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834
                                 "same topology")
            assert isinstance(v, Variable)
            new_p = Parameter(
                block=self,
                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,
F
fengjiayi 已提交
1835
                gradient_clip_attr=p.gradient_clip_attr,
F
fengjiayi 已提交
1836
                error_clip=p.error_clip,
1837 1838 1839
                name=v.name)
            self.vars[new_p.name] = new_p

1840
    def _clone_variable(self, var, force_persistable=True):
1841 1842
        """
        Clone a variable into current block.
1843

1844 1845
        Args:
            var: the variable to be cloned.
1846 1847 1848
            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.
1849 1850

        Returns:
1851
            Variable: the new  variable cloned from 'var' in current block.
1852 1853
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
1854 1855 1856 1857 1858
        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
1859 1860
        elif var.type == core.VarDesc.VarType.RAW:
            ret_var = self.create_var(
T
tangwei12 已提交
1861
                name=var.name, persistable=var.persistable, type=var.type)
T
typhoonzero 已提交
1862 1863 1864 1865 1866 1867
        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,
1868
                persistable=True if force_persistable else var.persistable,
F
fengjiayi 已提交
1869
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1870 1871 1872 1873 1874 1875 1876
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
1877
                persistable=True if force_persistable else var.persistable,
F
fengjiayi 已提交
1878
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1879
        return ret_var
1880

Y
Yu Yang 已提交
1881

1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976
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()

1977
    def remove_input_by_id(self, node_id):
1978 1979 1980 1981 1982 1983
        """
        Remove a node from inputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
1984
        self.node.remove_input(node_id)
1985

1986
    def remove_input(self, node):
1987 1988 1989 1990
        """
        Remove a node from inputs.

        Args:
1991
            node(IrNode): the node being removed.
1992
        """
1993
        self.node.remove_input(node.node)
1994

1995
    def append_input(self, node):
1996 1997 1998 1999
        """
        Append a node in inputs.

        Args:
2000
            node(IrNode): the node being appended.
2001
        """
2002
        self.node.append_input(node.node)
2003 2004 2005 2006 2007 2008 2009 2010

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

2011
    def remove_output_by_id(self, node_id):
2012 2013 2014 2015 2016 2017
        """
        Remove a node from outputs by the given node id.

        Args:
            node_id(int): the given node id.
        """
2018
        self.node.remove_output(node_id)
2019

2020
    def remove_output(self, node):
2021 2022 2023 2024
        """
        Remove a node from outputs.

        Args:
2025
            node(IrNode): the node being removed.
2026
        """
2027
        self.node.remove_output(node.node)
2028

2029
    def append_output(self, node):
2030 2031 2032 2033
        """
        Append a node in outputs.

        Args:
2034
            node(IrNode): the node being appended.
2035
        """
2036
        self.node.append_output(node.node)
2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097

    @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, \
            "The node variable description cannot be None."
        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, \
            "The node variable description cannot be None."
        return self.node.var().persistable()

2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130
    def type(self):
        """
        Return the variable type.

        Returns:
            core.VarDesc.VarType: the variable type.
        """
        assert self.node.var() is not None, \
            "The node variable description cannot be None."
        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, \
            "The node variable description cannot be None."
        return self.node.var().dtype()

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

        Returns:
            list: the variable shape.
        """
        assert self.node.var() is not None, \
            "The node variable description cannot be None."
        return self.node.var().shape()

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
    @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, \
            "The node operator description cannot be None."
        self.node.op()._rename_input(old_input_name, new_input_name)

2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219
    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, \
            "The node operator description cannot be None."
        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, \
            "The node operator description cannot be None."
        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, \
            "The node operator description cannot be None."
        return self.node.op().set_type(new_type)

2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247
    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, \
            "The node operator description cannot be None."
        desc = self.node.op()
        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)

2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269
    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, \
            "The node operator description cannot be None."
        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, \
            "The node operator description cannot be None."
        return self.node.op().output_arg_names()

2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290
    @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]


2291 2292
class IrGraph(object):
    """
2293
    Python IrGraph. Beneath it is a core.Graph, which is used for
2294
    creating a c++ Ir Pass Graph. An IrGraph is just a graph view of
2295 2296
    a Program. In an IrGraph, both Variables and Operators are graph
    nodes.
2297 2298 2299 2300
    """

    def __init__(self, graph, for_test=False):
        """
2301 2302
        Construct an IrGraph using core.Graph.

2303 2304 2305 2306 2307 2308 2309 2310 2311
        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

2312 2313 2314 2315
    def clone(self):
        """
        Create a new and duplicated IrGraph.

2316 2317 2318
        Warns:
            The method only clones the graph structure, not its attributes.

2319 2320 2321
        Returns:
            IrGraph: A new and duplicated graph.
        """
2322
        g = self.graph.clone()
2323 2324
        return IrGraph(g, self._for_test)

2325
    def is_test(self):
2326 2327 2328
        """
        If the graph is used for testing, the function returns true. Otherwise, returns false.
        """
2329 2330
        return self._for_test

W
WangZhen 已提交
2331
    def all_nodes(self):
2332 2333 2334
        """
        Return all nodes included in the graph as a set.
        """
2335
        return {IrNode(node) for node in self.graph.nodes()}
2336

2337
    def all_var_nodes(self):
2338 2339 2340
        """
        Return all variable nodes included in the graph as a set.
        """
2341
        return {IrVarNode(node) for node in self.graph.nodes() if node.is_var()}
2342

2343
    def all_persistable_nodes(self):
2344 2345 2346
        """
        Return all persistable variable nodes included in the graph as a set.
        """
W
WangZhen 已提交
2347 2348 2349 2350 2351
        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)
2352
        return {IrVarNode(p) for p in persistable_nodes}
W
WangZhen 已提交
2353

2354
    def all_op_nodes(self):
2355 2356 2357
        """
        Return all operator nodes included in the graph as a set.
        """
2358
        return {IrOpNode(node) for node in self.graph.nodes() if node.is_op()}
2359

2360
    def create_persistable_node(self, name, var_type, shape, var_dtype):
2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371
        """
        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:
2372
            IrVarNode: the created persistable variable node.
2373
        """
2374 2375 2376 2377 2378
        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)
2379
        return IrVarNode(self.graph.create_var_node(var_desc))
2380 2381

    def create_var_node(self, name, var_type, shape, var_dtype):
2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392
        """
        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:
2393
            IrVarNode: the created variable node.
2394 2395
        """

2396 2397 2398 2399
        var_desc = core.VarDesc(name)
        var_desc.set_type(var_type)
        var_desc.set_shape(shape)
        var_desc.set_dtype(var_dtype)
2400
        return IrVarNode(self.graph.create_var_node(var_desc))
2401 2402

    def create_var_node_from_desc(self, var_desc):
2403 2404 2405 2406 2407 2408 2409 2410
        """
        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:
2411
            IrVarNode: the created variable node.
2412
        """
2413
        return IrVarNode(self.graph.create_var_node(var_desc))
2414 2415

    def create_op_node(self, op_type, attrs, inputs, outputs):
2416 2417 2418 2419 2420 2421 2422 2423 2424 2425
        """
        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.
            outputs(dict): the outpus of the operator node.

        Returns:
2426
            IrOpNode: the created operator node.
2427
        """
2428 2429
        op_desc = core.OpDesc()
        op_desc.set_type(op_type)
2430
        for attr, value in six.iteritems(attrs):
2431
            self._update_desc_attr(op_desc, attr, value)
2432
        for input_name, var_nodes in six.iteritems(inputs):
2433 2434 2435 2436
            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])
2437
        for output_name, var_nodes in six.iteritems(outputs):
2438 2439 2440 2441
            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])
2442
        return IrOpNode(self.graph.create_op_node(op_desc))
2443 2444

    def create_op_node_from_desc(self, op_desc):
2445 2446 2447 2448 2449 2450 2451
        """
        Create a operator node by using an existing OpDesc in the graph.

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

        Returns:
2452
            IrOpNode: the created operator node.
2453
        """
2454
        return IrOpNode(self.graph.create_op_node(op_desc))
2455 2456

    def update_input_link(self, old_input_node, new_input_node, op_node):
2457 2458 2459 2460
        """
        Update the input's link of a operator node.

        Args:
2461 2462 2463
            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.
2464
        """
2465 2466
        assert old_input_node.node in self.graph.nodes() and new_input_node.node in \
        self.graph.nodes() and op_node.node in self.graph.nodes(), \
W
WangZhen 已提交
2467
        'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.'
2468 2469 2470 2471
        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)
2472
        op_node.rename_input(old_input_node.name(), new_input_node.name())
2473 2474

    def link_to(self, node_in, node_out):
2475 2476 2477 2478
        """
        Connect two nodes.

        Args:
2479 2480
            node_in(IrNode): the input node.
            node_out(IrNode): the output node.
2481
        """
2482
        assert node_in.node in self.graph.nodes() and node_out.node in self.graph.nodes(), \
W
WangZhen 已提交
2483
            'The two arguments(node_in&node_out) must be in the graph nodes.'
2484 2485
        node_in.append_output(node_out)
        node_out.append_input(node_in)
2486 2487

    def safe_remove_nodes(self, remove_nodes):
2488 2489 2490 2491 2492 2493 2494
        """
        Remove nodes safely since links connected to these removed nodes are
        also removed.

        Args:
            remove_nodes(set): the nodes prepared to be removed.
        """
2495
        if not isinstance(remove_nodes, set):
W
WangZhen 已提交
2496 2497 2498 2499
            if isinstance(remove_nodes, Iterable):
                remove_nodes = set(remove_nodes)
            else:
                remove_nodes = {remove_nodes}
2500 2501
        original_nodes = {n.node for n in remove_nodes}
        core.graph_safe_remove_nodes(self.graph, original_nodes)
2502

Z
Zhen Wang 已提交
2503 2504 2505 2506 2507 2508 2509 2510
    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] = [
2511
                            self._find_node_by_name(node.inputs, each_var_name)
Z
Zhen Wang 已提交
2512 2513 2514 2515
                        ]
                for each_var_name in node.op().output_arg_names():
                    if each_var_name not in var_nodes:
                        var_nodes[each_var_name] = [
2516
                            self._find_node_by_name(node.outputs, each_var_name)
Z
Zhen Wang 已提交
2517 2518 2519
                        ]
                    else:
                        var_nodes[each_var_name].append(
2520 2521
                            self._find_node_by_name(node.outputs,
                                                    each_var_name))
Z
Zhen Wang 已提交
2522 2523
        self.graph.resolve_hazard(var_nodes)

W
WangZhen 已提交
2524
    def has_circle(self):
2525 2526 2527 2528 2529 2530
        """
        Check if the graph has a circle.

        Returns:
            bool: True if the graph has a circle else False.
        """
W
WangZhen 已提交
2531 2532 2533
        return core.has_circle(self.graph)

    def graph_num(self):
2534 2535 2536 2537 2538 2539
        """
        Count the number of unconnected graphs in this graph.

        Returns:
            int: the number of unconnected graphs.
        """
W
WangZhen 已提交
2540 2541 2542
        return core.graph_num(self.graph)

    def topology_sort(self):
2543 2544 2545 2546 2547 2548
        """
        Perform the topology sort operation on the graph.

        Notes: the `graph` cannot contain a circle.

        Returns:
Z
Zhen Wang 已提交
2549
            list(IrNode): nodes in topology order.
2550
        """
2551
        ordered_nodes = core.topology_sort(self.graph)
Z
Zhen Wang 已提交
2552
        return [IrNode(n) for n in ordered_nodes]
W
WangZhen 已提交
2553 2554

    def build_adjacency_list(self):
2555 2556 2557 2558
        """
        Build an adjacency list of operations for the `graph`.

        Returns:
2559
            dict{IrNode: set(IrNode)}: the adjacency list.
2560
        """
2561 2562 2563 2564 2565
        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
2566

2567 2568 2569 2570 2571 2572 2573 2574
    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.
2575
            marked_nodes(set(IrNode)): nodes that are needed to be marked.
2576 2577 2578 2579 2580
            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.
        """

2581 2582 2583 2584 2585 2586 2587 2588 2589
        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 \
                            + ' -o ' + pdf_save_path, shell=True)
            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))

2590
        remove_ctr_vars = set()
2591
        if remove_ctr_var:
2592
            for node in self.all_var_nodes():
2593 2594 2595
                if node.is_ctrl_var():
                    remove_ctr_vars.add(node)
            self.safe_remove_nodes(remove_ctr_vars)
2596 2597
        print('Total ops num = {}.'.format(len(self.all_op_nodes())))

2598 2599
        if marked_nodes is not None:
            if not isinstance(marked_nodes, set):
2600 2601 2602 2603 2604 2605
                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}
2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616
            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)
        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):
2617 2618 2619
        """
        Convert the graph into a Program.

Z
Zhen Wang 已提交
2620
        WARN: When the graph includes backward operator nodes, the
2621 2622 2623 2624 2625 2626
        conversion process may be failed. Usually, this function is
        only used to convert a test graph.

        Returns:
            Program: a program converted from the graph.
        """
2627
        convert_pass = core.get_pass('graph_to_program_pass')
2628 2629
        desc = core.ProgramDesc()
        convert_pass.set_not_owned('program', desc)
2630 2631 2632 2633
        convert_pass.apply(self.graph)
        program = Program._construct_from_desc(desc)
        return program

2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644
    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

2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660
    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
2661
class Program(object):
D
dzhwinter 已提交
2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672
    """
    Python Program. Beneath it is a ProgramDesc, which is used for
    create c++ Program. A program is a self-contained programing
    language like container. It has at least one Block, when the
    control flow op like conditional_block, while_op is included,
    it will contains nested block.
    Please reference the framework.proto for details.

    Notes: we have default_startup_program and default_main_program
    by default, a pair of them will shared the parameters.
    The default_startup_program only run once to initialize parameters,
Y
yuyang18 已提交
2673
    default_main_program run in every mini batch and adjust the weights.
D
dzhwinter 已提交
2674 2675

    Returns:
Y
yuyang18 已提交
2676
        A empty program.
D
dzhwinter 已提交
2677 2678

    Examples:
Y
yuyang18 已提交
2679 2680 2681 2682 2683 2684
        >>> main_program = fluid.Program()
        >>> startup_program = fluid.Program()
        >>> with fluid.program_guard(main_program=main_program, startup_program=startup_program):
        >>>     fluid.layers.data(name="x", shape=[-1, 784], dtype='float32')
        >>>     fluid.layers.data(name="y", shape=[-1, 1], dtype='int32')
        >>>     fluid.layers.fc(name="fc", shape=[10], dtype='float32', act="relu")
D
dzhwinter 已提交
2685 2686 2687

    """

2688 2689
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
2690 2691
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
D
dzhwinter 已提交
2692
        self._seed = 0
Y
yuyang18 已提交
2693
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
Y
yuyang18 已提交
2694
        self._op_role_var = []
T
tangwei12 已提交
2695

2696 2697
        # for distribute training
        # _is_distributed = True if under distributed training
T
tangwei12 已提交
2698
        self._is_distributed = False
2699
        # _is_chief = True if the trainer is the first one, usually No.0
T
tangwei12 已提交
2700
        self._is_chief = False
2701 2702 2703
        # _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
2704
        self._endpoints = []
2705 2706 2707
        # if current role is parameter server, the _ps_endpoint is its "ip:port"
        self._ps_endpoint = None
        # trainers_endpoints, it is used for distribution.
2708
        self._trainers_endpoints = []
2709
        # the distributed lookup table names
T
tangwei12 已提交
2710
        self._distributed_lookup_table = None
2711 2712 2713 2714

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

D
dzhwinter 已提交
2715
        # @deprecated(the python memory optimize transpiler is deprecated)
D
dzhwinter 已提交
2716
        # whether the program is optimized by memory_optimize_transpiler
D
dzhwinter 已提交
2717
        self.__is_mem_optimized = False
D
dzhwinter 已提交
2718

2719 2720 2721
        # 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
2722
        self._program_config = None
2723

D
dzhwinter 已提交
2724
    @property
D
dzhwinter 已提交
2725
    def _is_mem_optimized(self):
D
dzhwinter 已提交
2726 2727
        # if the program is optimized, operator input/outputs
        # maybe same, which conflict with save_inference_model.
D
dzhwinter 已提交
2728
        return self.__is_mem_optimized
D
dzhwinter 已提交
2729

D
dzhwinter 已提交
2730 2731 2732
    @_is_mem_optimized.setter
    def _is_mem_optimized(self, target):
        self.__is_mem_optimized = target
Y
yuyang18 已提交
2733 2734 2735

    @property
    def op_role(self):
Y
yuyang18 已提交
2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748
        """
        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
        parameter gradient of backward (use :code:`op_role_var` to get this
        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
2749 2750 2751
        return self._current_role

    @op_role.setter
D
dzhwinter 已提交
2752
    def op_role(self, role):
Y
yuyang18 已提交
2753 2754 2755 2756
        self._current_role = role

    @property
    def op_role_var(self):
Y
yuyang18 已提交
2757 2758 2759 2760 2761 2762 2763
        """
        The auxiliary variables for :code:`op_role` property.

        See Also: :code:`Program.op_role`'s documentation for details.

        Notes: This is a very low-level API. Users should not use it directly.
        """
Y
yuyang18 已提交
2764 2765 2766 2767
        return self._op_role_var

    @op_role_var.setter
    def set_op_role_var(self, var_name):
Y
yuyang18 已提交
2768
        self._op_role_var = [var_name]
Y
yuyang18 已提交
2769

2770 2771 2772 2773 2774 2775 2776 2777 2778
    @contextlib.contextmanager
    def _backward_role_guard(self):
        tmp_role = self._current_role

        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Backward
        yield
        self._current_role = tmp_role

S
rename  
sneaxiy 已提交
2779
    @signature_safe_contextmanager
W
Wu Yi 已提交
2780
    def _optimized_guard(self, param_and_grads):
Y
yuyang18 已提交
2781 2782 2783 2784 2785 2786 2787
        """
        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:
2788
            param_and_grads(list): The variables (names) to be optimized.
Y
yuyang18 已提交
2789 2790 2791 2792

        Examples:

            >>> p, g = backward(...)
W
Wu Yi 已提交
2793
            >>> with program._optimized_guard([p,g]):
Y
yuyang18 已提交
2794 2795
            >>>     p = p - 0.001 * g
        """
X
Xin Pan 已提交
2796 2797 2798
        tmp_role = self._current_role
        tmp_var = self._op_role_var

Y
yuyang18 已提交
2799 2800
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
2801 2802 2803 2804
        self._op_role_var = [
            var.name if isinstance(var, Variable) else var
            for var in param_and_grads
        ]
Y
yuyang18 已提交
2805
        yield
X
Xin Pan 已提交
2806 2807
        self._op_role_var = tmp_var
        self._current_role = tmp_role
Y
Yu Yang 已提交
2808

S
rename  
sneaxiy 已提交
2809
    @signature_safe_contextmanager
X
Xin Pan 已提交
2810
    def _lr_schedule_guard(self, is_with_opt=False):
2811 2812 2813 2814 2815 2816 2817
        """
        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
2818 2819 2820 2821
        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.
2822 2823 2824 2825 2826 2827 2828

        Examples:

            >>> p, g = backward(...)
            >>> with program.lr_schedule_guard():
            >>>     lr = lr * decay
        """
2829 2830 2831 2832

        tmp_role = self._current_role
        tmp_var = self._op_role_var

2833 2834
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.LRSched
X
Xin Pan 已提交
2835 2836
        if is_with_opt:
            self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize)
2837 2838 2839
        # TODO(typhoonzero): how to set target learning rate var
        self._op_role_var = []
        yield
2840 2841
        self._op_role_var = tmp_var
        self._current_role = tmp_role
2842

2843
    def __str__(self):
Y
yuyang18 已提交
2844 2845 2846 2847 2848 2849 2850 2851 2852
        """
        Get the protobuf debug string of this Program.

        Returns:
            (str): The protobuf debug string.

        Raises:
            ValueError: If any of required fields is not set.
        """
Y
Yang Yang(Tony) 已提交
2853 2854
        return self.to_string(True)

F
fengjiayi 已提交
2855 2856 2857
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
2858

F
fengjiayi 已提交
2859
        Args:
Y
yuyang18 已提交
2860 2861
            throw_on_error(bool): raise Value error when any of required fields
                is not set.
F
fengjiayi 已提交
2862

Y
yuyang18 已提交
2863 2864 2865 2866
            with_details(bool): True if more details about variables and
                parameters, e.g., :code:`trainable`, :code:`optimize_attr`, need
                to print.

H
haowang101779990 已提交
2867 2868
        Returns:
            str : The debug string.
Y
yuyang18 已提交
2869 2870 2871 2872

        Raises:
            ValueError: If any of required fields is not set and throw_on_error is
                True.
F
fengjiayi 已提交
2873 2874 2875 2876 2877 2878 2879 2880 2881 2882

        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        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()
2883 2884
            proto = framework_pb2.ProgramDesc.FromString(
                six.binary_type(protostr))
F
fengjiayi 已提交
2885 2886
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
2887

W
Wu Yi 已提交
2888
    def _get_desc(self):
Y
yuyang18 已提交
2889 2890 2891 2892 2893 2894 2895
        """
        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.
        """
2896 2897
        return self.desc

X
version  
Xin Pan 已提交
2898 2899 2900
    def _version(self):
        return self.desc._version()

2901
    def clone(self, for_test=False):
Y
yuyang18 已提交
2902 2903 2904
        """
        Create a new, duplicated program.

2905

Y
yuyang18 已提交
2906 2907 2908 2909
        Some operators, e.g., :code:`batch_norm`, behave differently between
        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`.
2910

Y
yuyang18 已提交
2911 2912 2913 2914
        * Set for_test to False when we want to clone the program for training.
        * Set for_test to True when we want to clone the program for testing.

        Notes: This API DOES NOT prune any operator. Use
L
Luo Tao 已提交
2915 2916 2917 2918 2919
        :code:`clone(for_test=True)` before backward and optimization please. e.g.

            >>> test_program = fluid.default_main_program().clone(for_test=True)
            >>> optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
            >>> optimizer.minimize()
2920 2921

        Args:
Y
yuyang18 已提交
2922 2923
            for_test(bool): True if change the :code:`is_test` attribute of
                operators to :code:`True`.
2924

D
dzhwinter 已提交
2925
        Returns:
Y
yuyang18 已提交
2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978
            Program: The new, duplicated Program object.

        Examples:

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

            >>> import paddle.fluid as fluid
            >>> train_program = fluid.Program()
            >>> startup_program = fluid.Program()
            >>> with fluid.program_guard(train_program, startup_program):
            >>>     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'))
            >>>
            >>> test_program = train_program.clone(for_test=True)
            >>>
            >>> sgd = fluid.optimizer.SGD(learning_rate=1e-3)
            >>> with fluid.program_guard(train_program, startup_program):
            >>>     sgd.minimize(loss)

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

            >>> import paddle.fluid as fluid
            >>>
            >>> def network(is_test):
            >>>     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, is_test=is_test)
            >>>     loss = fluid.layers.cross_entropy(
            >>>                 input=fluid.layers.fc(hidden, size=10, act='softmax'),
            >>>                 label=fluid.layers.data(name='label', shape=[1], dtype='int64'))
            >>>     return loss
            >>>
            >>> train_program = fluid.Program()
            >>> startup_program = fluid.Program()
            >>> test_program = fluid.Program()
            >>>
            >>> with fluid.program_guard(train_program, startup_program):
            >>>     with fluid.unique_name.guard():
            >>>         loss = network(is_test=False)
            >>>         sgd = fluid.optimizer.SGD(learning_rate=1e-3)
            >>>         sgd.minimize(loss)
            >>>
            >>> # the test startup program is not used.
            >>> with fluid.program_guard(test_program, fluid.Program()):
            >>>     with fluid.unique_name.guard():
            >>>         loss = network(is_test=True)

            The two code snippets above will generate same programs.
2979 2980
        """
        if for_test:
X
Xin Pan 已提交
2981
            p = self._inference_optimize(prune_read_op=False)
2982
        else:
2983
            p = Program()
G
gongweibao 已提交
2984 2985
            p.current_block_idx = self.current_block_idx
            p._seed = self._seed
2986
            p.desc = core.ProgramDesc(self.desc)
M
minqiyang 已提交
2987 2988 2989
            p.blocks = [
                Block(p, i) for i in six.moves.range(self.desc.num_blocks())
            ]
G
gongweibao 已提交
2990 2991 2992 2993

            p._current_role = self._current_role
            p._op_role_var = self._op_role_var

W
Wu Yi 已提交
2994
            p._sync_with_cpp()
2995

W
Wu Yi 已提交
2996
        p._copy_param_info_from(self)
W
Wu Yi 已提交
2997
        p._copy_data_info_from(self)
2998
        p._copy_dist_param_info_from(self)
Y
Yu Yang 已提交
2999
        return p
3000

W
Wu Yi 已提交
3001
    def _prune(self, targets):
Y
yuyang18 已提交
3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016
        """
        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:
            targets(list|Variable|Operator): A list of variables or operators
                need to be pruned

        Returns:
            Program:  A new, pruned program.

        """
3017 3018 3019 3020 3021 3022
        if not isinstance(targets, list):
            targets = [targets]
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
3023 3024
                    # After transpiler processing, the op that output this
                    # variable maybe has been changed, so t.op is not reliable
3025
                    # and we need to find the current op that generate this
3026 3027 3028 3029 3030 3031 3032 3033
                    # variable here.
                    t.op = None
                    global_block = self.global_block()
                    for idx, op in enumerate(global_block.ops):
                        if t.name in op.output_arg_names:
                            t.op = op
                            break

3034
                    t = t.op
3035 3036 3037 3038
                    if t is None:
                        raise ValueError(
                            "The target variable must have an "
                            "associated operator that generates it.")
3039
                else:
3040 3041
                    raise ValueError("All targets of prune() can only be "
                                     "Variable or Operator.")
3042 3043 3044 3045

            targets_idx.append([t.block.idx, t.idx])
        res = Program()
        res.desc = core.prune(self.desc, targets_idx)
M
minqiyang 已提交
3046 3047 3048
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
3049
        res._sync_with_cpp()
3050 3051
        return res

X
Xin Pan 已提交
3052
    def _inference_optimize(self, prune_read_op=True):
Y
yuyang18 已提交
3053
        """
F
fengjiayi 已提交
3054 3055 3056 3057 3058
        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.

3059
        3. change the :code:`is_test`
Y
yuyang18 已提交
3060 3061 3062
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

3063
        Args:
X
Xin Pan 已提交
3064 3065
            prune_read_op(bool): remove the read ops that are added by py_reader
                                 for cpp inference library
3066

Y
yuyang18 已提交
3067 3068 3069 3070 3071 3072
        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
3073
        res = Program()
3074
        res.desc = core.ProgramDesc(self.desc)
F
fengjiayi 已提交
3075 3076 3077 3078

        # 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
3079
        if prune_read_op:
3080 3081 3082 3083 3084 3085 3086 3087 3088
            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
3089
                    root_block._remove_var(cpt.to_bytes(var.name()))
F
fengjiayi 已提交
3090 3091

        # change all `is_test` attributes to True
M
minqiyang 已提交
3092
        for i in six.moves.range(res.desc.num_blocks()):
3093
            block = res.desc.block(i)
M
minqiyang 已提交
3094
            for j in six.moves.range(block.op_size()):
3095 3096
                op = block.op(j)
                if op.has_attr('is_test'):
W
Wu Yi 已提交
3097
                    op._set_attr('is_test', True)
M
minqiyang 已提交
3098 3099 3100
        res.blocks = [
            Block(res, i) for i in six.moves.range(res.desc.num_blocks())
        ]
W
Wu Yi 已提交
3101
        res._sync_with_cpp()
3102 3103
        return res

3104 3105
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
3106 3107 3108 3109 3110 3111 3112
        """
        Deserialize a program desc from protobuf binary string.

        Notes: All information about parameters will be lost after serialization
        and deserialization.

        Args:
3113
            binary_str_type(str): The binary prootbuf string.
Y
yuyang18 已提交
3114 3115 3116 3117

        Returns:
            Program: A deserialized program desc.
        """
3118 3119
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
M
minqiyang 已提交
3120
        p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())]
W
Wu Yi 已提交
3121
        p._sync_with_cpp()
3122
        return p
Y
Yu Yang 已提交
3123

3124
    @staticmethod
3125
    def _construct_from_desc(desc):
3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140
        """
        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
3141 3142
    @property
    def random_seed(self):
Y
yuyang18 已提交
3143 3144 3145 3146 3147 3148
        """
        The default random seed for random operators in Program. Zero means get
        the random seed from random device.

        Notes: It must be set before the operators have been added.
        """
D
dzhwinter 已提交
3149 3150
        return self._seed

Q
qiaolongfei 已提交
3151 3152
    @property
    def num_blocks(self):
Y
yuyang18 已提交
3153 3154 3155
        """
        The number of blocks in this program.
        """
Q
qiaolongfei 已提交
3156 3157
        return self.desc.num_blocks()

D
dzhwinter 已提交
3158 3159 3160 3161 3162 3163
    @random_seed.setter
    def random_seed(self, seed):
        if not isinstance(seed, int):
            raise ValueError("Seed must be a integer.")
        self._seed = seed

Y
Yu Yang 已提交
3164
    def __repr__(self):
3165
        return self.__str__()
3166

Y
Yu Yang 已提交
3167
    def global_block(self):
Y
yuyang18 已提交
3168 3169 3170
        """
        Get the first block of this program.
        """
Y
Yu Yang 已提交
3171 3172
        return self.blocks[0]

Q
Qiao Longfei 已提交
3173
    def block(self, index):
Y
yuyang18 已提交
3174 3175 3176 3177 3178 3179 3180 3181
        """
        Get the :code:`index` block of this program
        Args:
            index(int): The index of block to get

        Returns:
            Block: The :code:`index` block
        """
Q
Qiao Longfei 已提交
3182 3183
        return self.blocks[index]

Y
Yu Yang 已提交
3184
    def current_block(self):
Y
yuyang18 已提交
3185 3186 3187 3188
        """
        Get the current block. The :code:`current` block is the block to append
        operators.
        """
Y
Yu Yang 已提交
3189 3190
        return self.blocks[self.current_block_idx]

W
Wu Yi 已提交
3191
    def _create_block(self, parent_idx=None):
Y
yuyang18 已提交
3192 3193 3194 3195 3196 3197 3198 3199 3200 3201
        """
        Create a new block with the :code:`parent_idx` and change the current block
        to new block.

        Args:
            parent_idx(int): The parent block index.

        Returns:
            Block: The new block.
        """
Y
Yu Yang 已提交
3202
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
3203 3204 3205
        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
3206 3207 3208 3209
        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
3210
    def _rollback(self):
Y
yuyang18 已提交
3211 3212 3213 3214 3215
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
3216 3217
        self.current_block_idx = self.current_block().parent_idx

W
Wu Yi 已提交
3218
    def _sync_with_cpp(self):
Y
yuyang18 已提交
3219 3220 3221 3222 3223 3224 3225 3226 3227 3228
        """
        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
3229 3230 3231
        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
3232
            block._sync_with_cpp()
Q
Qiao Longfei 已提交
3233

W
Wu Yi 已提交
3234
    def _copy_param_info_from(self, other):
3235
        """
3236
        Copy the information of parameters from other program.
D
dzhwinter 已提交
3237

Y
yuyang18 已提交
3238 3239 3240
        Notes: This is a very low level API. Users should not invoke it
        directly.

3241 3242 3243 3244 3245 3246 3247
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
3248
            raise TypeError("_copy_param_info_from should be invoked with "
3249 3250 3251
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
3252
            raise ValueError("_copy_param_info_from should be invoked with two "
3253
                             "program, with represent the same topology")
W
Wu Yi 已提交
3254
        self.global_block()._copy_param_info_from(other.global_block())
3255

3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270
    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):
            raise TypeError("_copy_dist_param_info_from should be invoked with "
                            "Program")
        self._is_distributed = other._is_distributed
        self._is_chief = other._is_chief
3271
        self._parameters_on_pservers = other._parameters_on_pservers
3272
        self._endpoints = other._endpoints
3273
        self._ps_endpoint = other._ps_endpoint
3274 3275
        self._distributed_lookup_table = other._distributed_lookup_table

W
Wu Yi 已提交
3276
    def _copy_data_info_from(self, other):
F
fengjiayi 已提交
3277 3278
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
3279

Y
yuyang18 已提交
3280 3281 3282
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
3283 3284 3285 3286 3287 3288 3289
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
W
Wu Yi 已提交
3290
            raise TypeError("_copy_param_info_from should be invoked with "
F
fengjiayi 已提交
3291 3292 3293
                            "Program")

        if len(self.blocks) != len(other.blocks):
W
Wu Yi 已提交
3294
            raise ValueError("_copy_param_info_from should be invoked with two "
F
fengjiayi 已提交
3295
                             "program, with represent the same topology")
3296
        for var in list(other.global_block().vars.values()):
F
fengjiayi 已提交
3297 3298 3299
            if var.is_data:
                self.global_block().var(var.name).is_data = True

3300
    def list_vars(self):
Y
yuyang18 已提交
3301 3302 3303 3304 3305 3306
        """
        Get all variables from this Program. A iterable object is returned.

        Returns:
            iterable: The generator will yield every variable in this program.
        """
3307
        for each_block in self.blocks:
3308
            for each_var in list(each_block.vars.values()):
3309 3310
                yield each_var

Y
Yu Yang 已提交
3311

Y
Yu Yang 已提交
3312
class Parameter(Variable):
3313
    """
3314
    Parameter is derived from Variable. A parameter is a persistable
3315
    Variable, and will be updated by optimizers after each iteration.
3316
    The training of a neural network is essentially the updating of
3317 3318
    its parameters.

3319
    Relative to a general Variable, a Parameter has several its own
3320 3321
    member variables:

3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333
    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
        gradient_clip_attr(BaseGradientClipAttr): The gradint clip strategy
            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.
3334 3335
    """

Y
Yu Yang 已提交
3336 3337 3338 3339 3340 3341 3342 3343 3344 3345
    def __init__(self, block, shape, dtype, **kwargs):
        if shape is None or dtype is None:
            raise ValueError("Parameter must set shape and dtype")
        if len(shape) == 0:
            raise ValueError("Parameter shape cannot be empty")

        for each in shape:
            if each < 0:
                raise ValueError("Parameter shape should not be related with "
                                 "batch-size")
3346 3347 3348

        Variable.__init__(
            self, block, persistable=True, shape=shape, dtype=dtype, **kwargs)
Y
Yu Yang 已提交
3349 3350 3351 3352
        self.trainable = kwargs.get('trainable', True)

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

3353 3354
        self.regularizer = kwargs.get('regularizer', None)

F
fengjiayi 已提交
3355
        self.gradient_clip_attr = kwargs.get('gradient_clip_attr', None)
Y
Yu Yang 已提交
3356

W
wanghaoshuang 已提交
3357
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
3358

F
fengjiayi 已提交
3359 3360 3361
    def __str__(self):
        return self.to_string(True)

F
update  
fengjiayi 已提交
3362 3363 3364
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
3365

F
update  
fengjiayi 已提交
3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379
        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.

        """
        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",
W
wanghaoshuang 已提交
3380
                               "gradient_clip_attr", "do_model_average")
F
update  
fengjiayi 已提交
3381
            for attr_name in additional_attr:
3382 3383
                res_str += "%s: %s\n" % (
                    attr_name, six.binary_type(getattr(self, attr_name)))
F
update  
fengjiayi 已提交
3384 3385
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
3386 3387 3388 3389
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
3390

Y
Yu Yang 已提交
3391
# program is a global instance.
Y
Yu Yang 已提交
3392 3393
_main_program_ = Program()
_startup_program_ = Program()
3394

3395

3396
def default_startup_program():
Y
Yu Yang 已提交
3397
    """
Y
yuyang18 已提交
3398 3399 3400 3401 3402 3403 3404 3405 3406
    Get default/global startup program.

    The layer function in :code:`fluid.layers` will create parameters, readers,
    NCCL handles as global variables. The :code:`startup_program` will
    initialize them by the operators in startup program. The layer function will
    append these initialization operators into startup program.

    This method will return the :code:`default` or the :code:`current` startup
    program. Users can use :code:`fluid.program_guard` to switch program.
3407

Y
Yu Yang 已提交
3408 3409 3410
    Returns:
        Program: startup program
    """
Y
Yu Yang 已提交
3411
    return _startup_program_
3412

3413

3414
def default_main_program():
Y
Yu Yang 已提交
3415
    """
Y
yuyang18 已提交
3416 3417 3418 3419 3420 3421 3422 3423 3424
    Get default/global main program. The main program is used for training or
    testing.

    All layer function in :code:`fluid.layers` will append operators and
    variables to the :code:`default_main_program`.

    The :code:`default_main_program` is the default program in a lot of APIs.
    For example, the :code:`Executor.run()` will execute the
    :code:`default_main_program` when the program is not specified.
3425

Y
Yu Yang 已提交
3426 3427 3428
    Returns:
        Program: main program
    """
Y
Yu Yang 已提交
3429
    return _main_program_
Y
Yu Yang 已提交
3430 3431 3432 3433 3434


def switch_main_program(program):
    """
    Switch the main program to a new program.
3435

Y
Yu Yang 已提交
3436 3437 3438 3439 3440 3441 3442 3443 3444 3445 3446 3447 3448 3449
    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):
    """
3450
    Switch the startup program to a new program
Y
Yu Yang 已提交
3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462
    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
3463
@signature_safe_contextmanager
Y
Yu Yang 已提交
3464 3465
def program_guard(main_program, startup_program=None):
    """
Y
yuyang18 已提交
3466 3467 3468
    Change the global main program and startup program with `with` statement.
    Layer functions in the Python `with` block will append operators and
    variables to the new main programs.
3469

Y
Yu Yang 已提交
3470
    Examples:
Y
yuyang18 已提交
3471 3472 3473 3474 3475 3476 3477 3478 3479 3480

        >>> import paddle.fluid as fluid
        >>> main_program = fluid.Program()
        >>> startup_program = fluid.Program()
        >>> with fluid.program_guard(main_program, startup_program):
        >>>     data = fluid.layers.data(...)
        >>>     hidden = fluid.layers.fc(...)

    Notes: The temporary :code:`Program` can be used if the user does not need
    to construct either of startup program or main program.
3481

Y
Yu Yang 已提交
3482
    Examples:
Y
yuyang18 已提交
3483 3484 3485 3486 3487 3488

        >>> 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()):
        >>>     data = ...
3489

Y
Yu Yang 已提交
3490
    Args:
Y
yuyang18 已提交
3491
        main_program(Program): New main program inside `with` statement.
3492
        startup_program(Program): New startup program inside `with` statement.
Y
Yu Yang 已提交
3493 3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505
            None means do not change startup program.
    """
    if not isinstance(main_program, Program):
        raise TypeError("main_program should be Program")
    main_program = switch_main_program(main_program)
    if startup_program is not None:
        if not isinstance(startup_program, Program):
            raise TypeError("startup_program should be Program")
        startup_program = switch_startup_program(startup_program)
    yield
    switch_main_program(main_program)
    if startup_program is not None:
        switch_startup_program(startup_program)
X
xuwei06 已提交
3506 3507


W
Wu Yi 已提交
3508
def _get_var(name, program=None):
X
xuwei06 已提交
3509
    """
Y
yuyang18 已提交
3510
    Get a variable by name from the global block of a program.
F
fengjiayi 已提交
3511

X
xuwei06 已提交
3512 3513 3514
    Args:
        name(str): name of the variable
        program(Program|None): program object.
T
tangwei12 已提交
3515
        If None, default_global_program() will be used.
X
xuwei06 已提交
3516 3517 3518 3519 3520 3521 3522

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
3523
    assert isinstance(program, Program)
X
xuwei06 已提交
3524 3525

    return program.global_block().var(name)
3526 3527


S
rename  
sneaxiy 已提交
3528
@signature_safe_contextmanager
L
lujun 已提交
3529 3530 3531 3532
def _dygraph_guard(tracer):
    global _dygraph_tracer_
    tmp_trace = _dygraph_tracer_
    _dygraph_tracer_ = tracer
M
minqiyang 已提交
3533

3534
    yield
P
Paddle CI 已提交
3535

L
lujun 已提交
3536
    _dygraph_tracer_ = tmp_trace
P
Paddle CI 已提交
3537 3538


S
rename  
sneaxiy 已提交
3539
@signature_safe_contextmanager
L
lujun 已提交
3540 3541 3542 3543
def _dygraph_place_guard(place):
    global _dygraph_current_expected_place_
    tmp_place = _dygraph_current_expected_place_
    _dygraph_current_expected_place_ = place
M
minqiyang 已提交
3544

3545
    yield
M
minqiyang 已提交
3546

L
lujun 已提交
3547
    _dygraph_current_expected_place_ = tmp_place