framework.py 55.1 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.

Y
Yu Yang 已提交
15
import collections
Q
qiaolongfei 已提交
16
import contextlib
F
fengjiayi 已提交
17
import re
18

Y
Yu Yang 已提交
19
import numpy as np
Q
qiaolongfei 已提交
20

21
import proto.framework_pb2 as framework_pb2
Q
qiaolongfei 已提交
22
from . import core
Y
Yu Yang 已提交
23
import unique_name
Y
Yu Yang 已提交
24

25
__all__ = [
26 27 28 29 30 31 32 33 34
    'Block',
    'Variable',
    'Program',
    'Operator',
    'default_startup_program',
    'default_main_program',
    'program_guard',
    'switch_startup_program',
    'switch_main_program',
X
xuwei06 已提交
35
    'get_var',
36
]
Y
Yu Yang 已提交
37

Q
qiaolongfei 已提交
38 39 40 41 42 43 44 45 46 47 48 49
EMPTY_VAR_NAME = core.kEmptyVarName()
TEMP_VAR_NAME = core.kTempVarName()
GRAD_VAR_SUFFIX = core.kGradVarSuffix()
ZERO_VAR_SUFFIX = core.kZeroVarSuffix()


def grad_var_name(var_name):
    """
    return gradient name for a certain var name
    """
    return var_name + GRAD_VAR_SUFFIX

Y
Yu Yang 已提交
50

51
def convert_np_dtype_to_dtype_(np_dtype):
52 53 54 55 56
    """
    Convert the data type in numpy to the data type in Paddle
    Args:
        np_dtype(np.dtype): the data type in numpy

57
    Returns(core.VarDesc.VarType): the data type in Paddle
58 59

    """
60 61
    dtype = np.dtype(np_dtype)
    if dtype == np.float32:
62
        return core.VarDesc.VarType.FP32
63
    elif dtype == np.float64:
64
        return core.VarDesc.VarType.FP64
65
    elif dtype == np.float16:
66
        return core.VarDesc.VarType.FP16
67
    elif dtype == np.int32:
68
        return core.VarDesc.VarType.INT32
69
    elif dtype == np.int16:
70
        return core.VarDesc.VarType.INT16
71
    elif dtype == np.int64:
72
        return core.VarDesc.VarType.INT64
73
    elif dtype == np.bool:
74
        return core.VarDesc.VarType.BOOL
75 76
    elif dtype == np.uint16:
        return core.VarDesc.VarType.INT16
77 78
    elif dtype == np.uint8:
        return core.VarDesc.VarType.UINT8
79 80 81 82 83
    else:
        raise ValueError("Not supported numpy dtype " + str(dtype))


def dtype_is_floating(dtype):
84 85 86
    """
    Check the data type is floating or not.
    Args:
87
        dtype(np.dtype|core.VarDesc.VarType): data type.
88 89 90 91 92
            Could be numpy format or Paddle format

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

    """
93
    if not isinstance(dtype, core.VarDesc.VarType):
94 95
        dtype = convert_np_dtype_to_dtype_(dtype)

96 97 98 99
    return dtype in [
        core.VarDesc.VarType.FP16, core.VarDesc.VarType.FP32,
        core.VarDesc.VarType.FP64
    ]
100 101


Y
Yang Yang(Tony) 已提交
102
def _debug_string_(proto, throw_on_error=True):
103 104 105 106 107 108 109 110 111 112 113
    """
    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
114
    error_fields = list()
Y
Yang Yang(Tony) 已提交
115
    if not proto.IsInitialized(error_fields) and throw_on_error:
C
caoying03 已提交
116 117
        raise ValueError("{0} are not initialized.\nThe message is {1}:\n".
                         format(error_fields, proto))
Y
Yu Yang 已提交
118 119 120
    return proto.__str__()


Y
Yu Yang 已提交
121
class Variable(object):
122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144
    """
    Python variable. Every input and output of an operator is a variable. Every
    variable belongs to a block. The variable has a name and two variables in
    different blocks could have the same name.

    There are many kinds of variables. Please reference the framework.proto for
    details.

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

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

    Args:
        block(Block): The associated block. It will be passed by
            `Block.create_var` automatically.
        type(core.VarDesc.VarType): Variable type. Please reference the
            framework.proto for details.
        shape(tuple|list|None): The shape of variable. -1 means the batch size.
            Some kinds of variable do not contain shape, just set it to None.
145
        dtype(np.dtype|core.VarDesc.VarType|str): The data type of variable.
146
        lod_level(int): The level of lod tensor. 0 means it is not a time
147
            series data.
148 149
        capacity(int): The capacity of Channel variable. Ignored
            for other types.
150 151 152 153 154 155
        persistable(bool): True if the variable should be saved as check point.
            Defaults to False.
        stop_gradient(bool): True if the variable will stop to calculate
            gradients when backward. Defaults to False.
    """

Y
Yu Yang 已提交
156 157
    def __init__(self,
                 block,
Y
Yu Yang 已提交
158
                 type=core.VarDesc.VarType.LOD_TENSOR,
Y
Yu Yang 已提交
159 160 161 162
                 name=None,
                 shape=None,
                 dtype=None,
                 lod_level=None,
163
                 capacity=None,
Q
QI JUN 已提交
164
                 persistable=None,
F
fengjiayi 已提交
165
                 error_clip=None,
Y
Yu Yang 已提交
166
                 stop_gradient=False,
F
fengjiayi 已提交
167
                 is_data=False,
Y
Yu Yang 已提交
168
                 **kwargs):
Y
Yu Yang 已提交
169
        self.block = block
F
fengjiayi 已提交
170
        self.error_clip = error_clip
Y
Yu Yang 已提交
171 172

        if name is None:
Y
Yu Yang 已提交
173
            name = unique_name.generate('_generated_var')
D
Dong Zhihong 已提交
174 175 176 177
        is_new_var = False
        self.desc = self.block.desc.find_var(name)

        if self.desc is None:
D
dongzhihong 已提交
178
            self.desc = self.block.desc.var(name)
Y
Yu Yang 已提交
179
            is_new_var = True
Y
Yu Yang 已提交
180

Y
Yu Yang 已提交
181 182 183 184 185 186 187 188
        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))

Y
Yu Yang 已提交
189
        if shape is not None:
Y
Yu Yang 已提交
190
            if is_new_var:
191
                self.desc.set_shape(shape)
Y
Yu Yang 已提交
192 193 194 195 196 197 198 199
            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))
Y
Yu Yang 已提交
200
        if dtype is not None:
201
            if not isinstance(dtype, core.VarDesc.VarType):
202
                dtype = convert_np_dtype_to_dtype_(dtype)
Y
Yu Yang 已提交
203
            if is_new_var:
F
fengjiayi 已提交
204
                self.desc.set_dtype(dtype)
Y
Yu Yang 已提交
205
            else:
F
fengjiayi 已提交
206
                old_dtype = self.dtype
Q
QI JUN 已提交
207
                if dtype != old_dtype:
Y
Yu Yang 已提交
208 209 210 211 212
                    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))
Y
Yu Yang 已提交
213 214

        if lod_level is not None:
Y
Yu Yang 已提交
215
            if is_new_var:
216
                self.desc.set_lod_level(lod_level)
Y
Yu Yang 已提交
217 218 219 220 221 222 223
            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))
224 225 226 227 228 229 230 231 232 233 234
        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))

235 236 237 238 239 240 241 242
        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

Y
Yu Yang 已提交
243
        self.block.vars[name] = self
Y
Yu Yang 已提交
244
        self.op = None
Y
Yu Yang 已提交
245
        self.stop_gradient = stop_gradient
F
fengjiayi 已提交
246
        self.is_data = is_data
Y
Yu Yang 已提交
247

248
    def __str__(self):
Y
Yang Yang(Tony) 已提交
249 250
        return self.to_string(True)

F
update  
fengjiayi 已提交
251
    def to_string(self, throw_on_error, with_details=False):
252 253 254 255 256 257
        """
        Get debug string.

        Args:
            throw_on_error(bool): True if raise an exception when self is not
                intialized.
F
update  
fengjiayi 已提交
258 259
            with_details(bool): more details about variables and parameters
                (e.g. trainable, optimize_attr, ...) will be printed when with_details is True
260 261 262 263

        Returns(str): The debug string.

        """
F
update  
fengjiayi 已提交
264 265
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
266 267
        protostr = self.desc.serialize_to_string()
        proto = framework_pb2.VarDesc.FromString(str(protostr))
F
update  
fengjiayi 已提交
268 269 270 271 272 273 274
        res_str = _debug_string_(proto, throw_on_error)
        if with_details:
            additional_attr = ("error_clip", "stop_gradient")
            for attr_name in additional_attr:
                res_str += "%s: %s\n" % (attr_name,
                                         str(getattr(self, attr_name)))
        return res_str
275 276 277

    __repr__ = __str__

278 279 280
    def set_desc(self, input):
        self.desc = input

281 282 283 284
    @property
    def persistable(self):
        return self.desc.persistable()

Y
Yu Yang 已提交
285 286 287 288
    @persistable.setter
    def persistable(self, p):
        self.desc.set_persistable(p)

Y
Yu Yang 已提交
289 290
    @property
    def name(self):
291
        return self.desc.name()
Y
Yu Yang 已提交
292

T
typhoonzero 已提交
293 294 295 296
    @name.setter
    def name(self, new_name):
        self.desc.set_name(new_name)

Y
Yu Yang 已提交
297 298 299
    @property
    def shape(self):
        # convert to tuple, make it as same as numpy API.
300
        return tuple(self.desc.shape())
Y
Yu Yang 已提交
301 302

    @property
F
fengjiayi 已提交
303 304
    def dtype(self):
        return self.desc.dtype()
Y
Yu Yang 已提交
305 306 307

    @property
    def lod_level(self):
308
        return self.desc.lod_level()
Y
Yu Yang 已提交
309

Y
Yu Yang 已提交
310 311 312 313
    @property
    def type(self):
        return self.desc.type()

314 315 316
    def set_error_clip(self, error_clip):
        self.error_clip = error_clip

Y
Yu Yang 已提交
317

F
fengjiayi 已提交
318 319 320
def get_all_op_protos():
    """
    Get all registered op proto from PaddlePaddle C++ end.
321 322 323

    Returns(list): list of OpProto

F
fengjiayi 已提交
324 325 326 327 328 329 330 331 332 333
    """
    protostrs = core.get_all_op_protos()
    ret_values = []
    for pbstr in protostrs:
        op_proto = framework_pb2.OpProto.FromString(str(pbstr))
        ret_values.append(op_proto)
    return ret_values


class OpProtoHolder(object):
334 335 336 337
    """
    A global variable to hold all OpProtos from C++ as a map
    """

F
fengjiayi 已提交
338 339 340 341 342 343 344 345 346
    @classmethod
    def instance(cls):
        if not hasattr(cls, '_instance'):
            cls._instance = cls()
        return cls._instance

    def __init__(self):
        assert not hasattr(
            self.__class__,
347
            '_instance'), 'Please use `instance()` to get OpProtoHolder object!'
F
fengjiayi 已提交
348 349 350 351 352 353
        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):
354 355 356 357 358 359 360 361
        """
        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
362 363
        if type not in self.op_proto_map:
            raise ValueError("Operator \"%s\" has not been registered." % type)
F
fengjiayi 已提交
364 365
        return self.op_proto_map[type]

366 367 368 369 370 371 372
    @staticmethod
    def generated_op_attr_names():
        return {
            core.op_proto_and_checker_maker.kOpRoleAttrName(),
            core.op_proto_and_checker_maker.kOpRoleVarAttrName()
        }

F
fengjiayi 已提交
373

Y
Yu Yang 已提交
374
class Operator(object):
375
    """
376 377
    Python Operator class. The operator represents the build in instructions in a
    Block. Users can use the build in instructions to describe their neural
378 379
    network.
    """
380 381 382 383 384
    OP_WITHOUT_KERNEL_SET = {
        'feed', 'fetch', 'save', 'load', 'recurrent', 'go',
        'rnn_memory_helper_grad', 'conditional_block', 'while', 'send', 'recv',
        'listen_and_serv', 'parallel_do', 'save_combine', 'load_combine',
        'ncclInit', 'channel_create', 'channel_close', 'channel_send',
T
typhoonzero 已提交
385
        'channel_recv', 'select', 'gen_nccl_id'
386
    }
387

Y
Yu Yang 已提交
388 389
    def __init__(self,
                 block,
Y
Yu Yang 已提交
390
                 desc,
Y
Yu Yang 已提交
391 392 393 394
                 type=None,
                 inputs=None,
                 outputs=None,
                 attrs=None):
395 396 397 398 399 400 401 402 403 404 405 406 407 408
        """
        Constructor.

        Notes: The constructor of operator should not be invoked directly. Use
        Block.append_op or Block.prepend_op instead.

        >>> 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]})

        Args:
C
caoying03 已提交
409 410
            block(Block): The block has the current operator.
            desc(core.OpDesc): The protobuf description.
411 412 413
            type(str): The type of operator.
            inputs(dict): The input dictionary. Key is the input parameter name.
                Value is a list of variables.
C
caoying03 已提交
414 415
            outputs(dict): The output dictionary which has the same format with
                           inputs.
416 417 418 419
            attrs(dict): The attributes dictionary. Key is attribute name. Value
                is the attribute value. The attribute type should be as same as
                the type registered in C++
        """
Y
Yu Yang 已提交
420
        self.block = block
Y
Yu Yang 已提交
421
        self.desc = desc
T
typhoonzero 已提交
422
        self.attrs = attrs
Y
yuyang18 已提交
423 424 425 426 427 428 429 430
        if self.attrs is None:
            self.attrs = dict()
        del attrs

        op_maker = core.op_proto_and_checker_maker

        if op_maker.kOpRoleAttrName() not in self.attrs:
            self.attrs[op_maker.kOpRoleAttrName()] = self.block.program.op_role
Y
yuyang18 已提交
431 432 433 434 435 436 437 438

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

        if role_var_name in self.attrs and len(self.attrs[role_var_name]) == 0:
            del self.attrs[role_var_name]
Y
yuyang18 已提交
439

F
fengjiayi 已提交
440 441 442 443 444
        if len(self.desc.type()) != 0:
            return
        if type is None:
            raise ValueError(
                "`type` to initilized an Operator can not be None.")
F
Update  
fengjiayi 已提交
445
        self.desc.set_type(type)
F
fengjiayi 已提交
446
        proto = OpProtoHolder.instance().get_op_proto(type)
447

Y
Yang Yang(Tony) 已提交
448 449
        def find_name(var_list, name):
            for var_name in var_list:
Q
Qiao Longfei 已提交
450
                if var_list[var_name] is not None and var_name == name:
Y
Yang Yang(Tony) 已提交
451 452
                    return True
            return False
Q
QI JUN 已提交
453

Y
Yang Yang(Tony) 已提交
454 455 456 457 458 459 460
        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:
461 462 463 464
                    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:
Y
Yang Yang(Tony) 已提交
465 466
                        raise ValueError(
                            "Input %s expects only one input, but %d are given."
467 468 469
                            % (in_proto.name, len(in_args)))
                    in_arg_names = []
                    for arg in in_args:
Y
Yang Yu 已提交
470 471 472 473
                        if isinstance(arg, basestring):
                            in_arg_names.append(arg)
                        else:
                            in_arg_names.append(arg.name)
474
                    self.desc.set_input(in_proto.name, in_arg_names)
Y
Yang Yang(Tony) 已提交
475 476
                else:
                    self.desc.set_input(in_proto.name, [])
F
Update  
fengjiayi 已提交
477

Y
Yu Yang 已提交
478
        if outputs is not None:
479 480 481 482 483 484 485
            given = set()
            need = set()
            for n in outputs:
                given.add(n)
            for m in proto.outputs:
                need.add(m.name)
            if not given == need:
C
caoying03 已提交
486 487
                raise ValueError(("Incorrect setting for output(s) of "
                                  "operator \"%s\". Need: [%s] Given: [%s]") %
488 489
                                 (type, ", ".join(str(e) for e in need),
                                  ", ".join(str(e) for e in given)))
490

F
fengjiayi 已提交
491
            for out_proto in proto.outputs:
492 493 494 495
                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:
F
Update  
fengjiayi 已提交
496 497
                    raise ValueError(
                        "Output %s expects only one output, but %d are given." %
498 499 500 501 502 503
                        (out_proto.name, len(out_args)))
                out_arg_names = []
                for arg in out_args:
                    out_arg_names.append(arg.name)
                    arg.op = self
                self.desc.set_output(out_proto.name, out_arg_names)
F
Update  
fengjiayi 已提交
504

Y
yuyang18 已提交
505 506
        if self.attrs is not None:
            if not isinstance(self.attrs, dict):
507
                raise TypeError("'attrs' should be a dict.")
F
fengjiayi 已提交
508
            for attr in proto.attrs:
F
Update  
fengjiayi 已提交
509
                attr_name = attr.name
Y
yuyang18 已提交
510 511
                if (attr_name not in self.attrs) or (
                        self.attrs[attr_name] is None):
F
Update  
fengjiayi 已提交
512
                    continue
Y
yuyang18 已提交
513 514 515 516 517
                if isinstance(self.attrs[attr_name], Block):
                    self.desc.set_block_attr(attr_name,
                                             self.attrs[attr_name].desc)
                elif isinstance(self.attrs[attr_name], core.BlockDesc) or \
                        isinstance(self.attrs[attr_name], core.ProgramDesc):
T
typhoonzero 已提交
518
                    self.desc.set_serialized_attr(
Y
yuyang18 已提交
519
                        attr_name, self.attrs[attr_name].serialize_to_string())
Y
Yang Yang(Tony) 已提交
520
                else:
Y
yuyang18 已提交
521
                    self.desc.set_attr(attr_name, self.attrs[attr_name])
522
        self.desc.check_attrs()
523
        if self.has_kernel(type):
Q
QI JUN 已提交
524
            self.desc.infer_var_type(self.block.desc)
Y
Yu Yang 已提交
525
            self.desc.infer_shape(self.block.desc)
F
fengjiayi 已提交
526

527 528 529
    def has_kernel(self, op_type):
        return op_type not in self.OP_WITHOUT_KERNEL_SET

Y
Yang Yang(Tony) 已提交
530
    def to_string(self, throw_on_error):
531 532 533 534 535 536 537 538 539
        """
        To debug string.
        Args:
            throw_on_error(bool): raise exception when self is not initialized
                when throw_on_error is True

        Returns(str): The debug string.

        """
540 541
        protostr = self.desc.serialize_to_string()
        proto = framework_pb2.OpDesc.FromString(str(protostr))
Y
Yang Yang(Tony) 已提交
542 543 544 545
        return _debug_string_(proto, throw_on_error)

    def __str__(self):
        return self.to_string(True)
546 547 548

    __repr__ = __str__

F
fengjiayi 已提交
549 550 551 552 553
    @property
    def type(self):
        return self.desc.type()

    def input(self, name):
554 555 556 557 558 559 560 561 562
        """
        Get input arguments by the input parameter name
        Args:
            name(str): The input parameter name

        Returns(list): return the list of argument names associated with the
            specific parameter name.

        """
F
fengjiayi 已提交
563 564
        return self.desc.input(name)

T
typhoonzero 已提交
565 566 567 568 569 570
    def rename_input(self, old_name, new_name):
        self.desc.rename_input(old_name, new_name)

    def rename_output(self, old_name, new_name):
        self.desc.rename_output(old_name, new_name)

F
fengjiayi 已提交
571 572
    @property
    def input_names(self):
573 574 575 576 577
        """
        Get all input parameter names
        Returns(list): return a list of input parameter names

        """
F
fengjiayi 已提交
578 579
        return self.desc.input_names()

T
typhoonzero 已提交
580 581 582 583 584 585 586 587
    @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
588
    def output(self, name):
589 590 591 592 593 594 595 596 597
        """
        Get output arguments by the output parameter name
        Args:
            name(str): The output parameter name

        Returns(list): return the list of argument names associated with the
            specific parameter name.

        """
F
fengjiayi 已提交
598 599 600 601
        return self.desc.output(name)

    @property
    def output_names(self):
602 603 604 605 606
        """
        Get all output parameter names
        Returns(list): return a list of output parameter names

        """
F
fengjiayi 已提交
607 608
        return self.desc.output_names()

609 610
    @property
    def idx(self):
611 612 613 614 615 616
        """
        Return the array index of current operator.
        Returns(int): The array index in block.ops array
        Raises:
            ValueError: when the operator is not found.
        """
617 618 619 620 621 622
        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
623
    def has_attr(self, name):
624 625 626 627 628 629 630 631
        """
        operator has the attribute with name or not.
        Args:
            name(str): the attribute name

        Returns(bool): True if has this attribute.

        """
F
fengjiayi 已提交
632 633 634
        return self.desc.has_attr(name)

    def attr_type(self, name):
635 636 637 638 639 640 641 642
        """
        Get the type of attribute by attribute name
        Args:
            name(str): the attribute name

        Returns(core.AttrType): the attribute type

        """
F
fengjiayi 已提交
643 644
        return self.desc.attr_type(name)

Y
yuyang18 已提交
645 646 647 648
    def set_attr(self, name, val):
        self.attrs[name] = val
        self.desc.set_attr(name, val)

F
fengjiayi 已提交
649 650
    @property
    def attr_names(self):
651 652 653 654 655
        """
        Get all attribute names
        Returns(list): The list of attribute name

        """
F
fengjiayi 已提交
656 657 658
        return self.desc.attr_names()

    def attr(self, name):
659 660 661 662 663 664 665 666 667
        """
        Get attribute by name
        Args:
            name(str): the attribute name

        Returns(bool|int|str|float|list): The attribute value. The return value
            can be any valid attribute type.

        """
F
fengjiayi 已提交
668
        return self.desc.attr(name)
Y
Yu Yang 已提交
669

F
fengjiayi 已提交
670
    def block_attr(self, name):
671 672 673 674 675 676 677 678
        """
        Get the block attribute by name
        Args:
            name(str): the attribute name

        Returns(int): the block index

        """
F
fengjiayi 已提交
679
        return self.desc.block_attr(name)
Y
Yu Yang 已提交
680

J
JiayiFeng 已提交
681
    def all_attrs(self):
F
fengjiayi 已提交
682 683 684 685 686 687 688 689 690 691 692 693 694
        """
        Get the attribute dict
        Returns(dict): The Operator's attribute dict
        """
        attr_names = self.attr_names
        attr_map = {}
        for n in attr_names:
            if n == 'sub_block':
                attr_map[n] = self.block_attr(n)
            else:
                attr_map[n] = self.attr(n)
        return attr_map

Y
Yu Yang 已提交
695

Y
Yu Yang 已提交
696 697
class Block(object):
    def __init__(self, program, idx):
Y
Yu Yang 已提交
698
        self.desc = program.desc.block(idx)
699
        self.vars = collections.OrderedDict()  # var_name --> var
Q
qiaolongfei 已提交
700
        self.ops = list()  # operator list
Y
Yu Yang 已提交
701
        self.program = program
702
        self.removed_vars = collections.OrderedDict()
Y
Yu Yang 已提交
703

704
    def __str__(self):
Y
Yang Yang(Tony) 已提交
705 706
        return self.to_string(True)

F
fengjiayi 已提交
707 708 709 710 711 712
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
        Args:
            throw_on_error(bool): raise exception when self is not initialized
                when throw_on_error is True
F
update  
fengjiayi 已提交
713 714
            with_details(bool): more details about variables and parameters
                (e.g. trainable, optimize_attr, ...) will be printed when with_details is True
F
fengjiayi 已提交
715 716 717 718 719 720 721

        Returns(str): The debug string.

        """
        assert isinstance(throw_on_error, bool) and isinstance(with_details,
                                                               bool)
        if with_details:
F
fengjiayi 已提交
722
            re_add_indent = re.compile(r"\n(.)")
F
fengjiayi 已提交
723 724 725
            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
                self.idx, self.parent_idx)
            for var in self.vars.itervalues():
F
fengjiayi 已提交
726
                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
F
update  
fengjiayi 已提交
727
                    r"\n    \1", var.to_string(throw_on_error, with_details))
F
fengjiayi 已提交
728
            for op in self.ops:
F
fengjiayi 已提交
729 730
                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
                    r"\n    \1", op.to_string(throw_on_error))
F
fengjiayi 已提交
731 732 733 734 735 736
            res_str += "\n}"
        else:
            protostr = self.desc.serialize_to_string()
            proto = framework_pb2.BlockDesc.FromString(str(protostr))
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
737 738 739

    __repr__ = __str__

Y
Yu Yang 已提交
740 741
    @property
    def parent_idx(self):
Y
Yu Yang 已提交
742
        return self.desc.parent
Y
Yu Yang 已提交
743

Y
Yu Yang 已提交
744 745 746 747 748 749 750
    @property
    def forward_block_idx(self):
        return self.desc.get_forward_block_idx()

    def set_forward_block_idx(self, idx):
        self.desc.set_forward_block_idx(idx)

Y
Yu Yang 已提交
751 752
    @property
    def idx(self):
Y
Yu Yang 已提交
753
        return self.desc.id
Y
Yu Yang 已提交
754

Q
Qiao Longfei 已提交
755
    def var(self, name):
Y
Yu Yang 已提交
756
        if not isinstance(name, basestring):
757 758 759
            raise TypeError(
                "var require string as parameter, but get %s instead." %
                (type(name)))
Y
Yu Yang 已提交
760 761
        v = self.vars.get(name, None)
        if v is None:
Q
Qiao Longfei 已提交
762
            raise ValueError("var %s not in this block" % name)
Y
Yu Yang 已提交
763
        return v
Q
Qiao Longfei 已提交
764

F
fengjiayi 已提交
765
    def var_recursive(self, name):
Y
Yu Yang 已提交
766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791
        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))

        raise ValueError("Var {0} is not found recursively".format(name))
F
fengjiayi 已提交
792

Q
Qiao Longfei 已提交
793
    def all_parameters(self):
794 795 796 797 798
        return list(self.iter_parameters())

    def iter_parameters(self):
        return (item[1] for item in self.vars.iteritems()
                if isinstance(item[1], Parameter))
Q
Qiao Longfei 已提交
799

Y
Yu Yang 已提交
800
    def create_var(self, *args, **kwargs):
801
        var = Variable(block=self, *args, **kwargs)
802 803
        if 'initializer' in kwargs:
            kwargs['initializer'](var, self)
Q
Qiao Longfei 已提交
804
        return var
Y
Yu Yang 已提交
805

Q
Qiao Longfei 已提交
806 807 808
    def has_var(self, name):
        return name in self.vars

T
typhoonzero 已提交
809 810 811 812 813
    def rename_var(self, name, new_name):
        """
        Rename variable in vars and ops' inputs and outputs
        """
        if not self.has_var(name):
814
            raise ValueError("var %s is not in current block" % name)
T
wip  
typhoonzero 已提交
815 816
        v = self.var(name)
        if type(v) == Parameter:
T
typhoonzero 已提交
817
            var_type = "Parameter"
T
wip  
typhoonzero 已提交
818 819 820 821 822 823 824
            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
825
            var_type = "Variable"
T
wip  
typhoonzero 已提交
826 827 828 829
            error_clip = v.error_clip
            stop_gradient = v.stop_gradient
        else:
            raise ValueError("unsupported var type: %s", type(v))
T
typhoonzero 已提交
830
        orig_var_type = v.type
T
typhoonzero 已提交
831
        self.desc.rename_var(name, new_name)
T
typhoonzero 已提交
832
        # NOTE: v is destroyed by C++ after calling rename_var.
T
wip  
typhoonzero 已提交
833
        d = self.desc.find_var(new_name)
T
typhoonzero 已提交
834
        if var_type == "Parameter":
T
wip  
typhoonzero 已提交
835 836 837 838
            var = Parameter(
                self,
                d.shape(),
                d.dtype(),
T
typhoonzero 已提交
839
                type=orig_var_type,
T
wip  
typhoonzero 已提交
840 841 842 843 844 845 846
                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
847
        elif var_type == "Variable":
T
wip  
typhoonzero 已提交
848 849
            var = Variable(
                self,
T
typhoonzero 已提交
850
                type=orig_var_type,
T
wip  
typhoonzero 已提交
851 852 853 854 855 856 857 858
                name=new_name,
                error_clip=error_clip,
                stop_gradient=stop_gradient)

        # rename the python side, sync_with_cpp will only add
        # new vars/ops to python side.
        self.vars[new_name] = var
        del self.vars[name]
T
typhoonzero 已提交
859
        self.sync_with_cpp()
860
        return var
T
typhoonzero 已提交
861

862 863 864 865 866
    def remove_var(self, name):
        self.sync_with_cpp()
        self.desc.remove_var(name)
        del self.vars[name]

Y
Yu Yang 已提交
867 868
    def create_parameter(self, *args, **kwargs):
        global_block = self.program.global_block()
Q
Qiao Longfei 已提交
869
        param = Parameter(global_block, *args, **kwargs)
870 871
        if 'initializer' in kwargs:
            kwargs['initializer'](param, self)
Q
Qiao Longfei 已提交
872
        return param
Y
Yu Yang 已提交
873

Y
Yu Yang 已提交
874
    def append_op(self, *args, **kwargs):
Y
Yu Yang 已提交
875
        op_desc = self.desc.append_op()
876
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
Y
Yu Yang 已提交
877 878 879
        self.ops.append(op)
        return op

Q
qiaolongfei 已提交
880 881 882 883 884 885 886
    def insert_op(self, index, *args, **kwargs):
        self.sync_with_cpp()
        op_desc = self.desc.insert_op(index)
        op = Operator(block=self, desc=op_desc, *args, **kwargs)
        self.ops.insert(index, op)
        return op

887 888 889 890 891
    def remove_op(self, index):
        self.sync_with_cpp()
        self.desc.remove_op(index, index + 1)
        del self.ops[index]

Y
Yancey1989 已提交
892
    def slice_ops(self, start, end):
Q
qiaolongfei 已提交
893
        return self.ops[start:end]
Y
Yancey1989 已提交
894

Y
Yu Yang 已提交
895
    def prepend_op(self, *args, **kwargs):
Y
Yu Yang 已提交
896 897
        op_desc = self.desc.prepend_op()
        op = Operator(self, op_desc, *args, **kwargs)
Q
qiaolongfei 已提交
898
        self.ops.insert(0, op)
Y
Yu Yang 已提交
899 900
        return op

Q
Qiao Longfei 已提交
901
    def sync_with_cpp(self):
902
        """
G
gongweibao 已提交
903
        Sync from the desc on the c++ end.
904 905 906

        This method is used to synchronize the c++ desc instance generated by backward.
        """
Q
Qiao Longfei 已提交
907 908 909 910 911
        # 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())

912 913 914 915 916
        # sync variables removed from c++ end
        for var in self.vars.keys():
            if not self.desc.find_var(var):
                self.vars.pop(var)

Q
Qiao Longfei 已提交
917
        # sync operators from cpp
918 919 920 921
        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
922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937
        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
938 939 940 941 942

        # 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
943
            self.ops.insert(0, op)
Q
Qiao Longfei 已提交
944 945 946 947 948 949 950

        # 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)

951 952 953 954 955 956 957 958 959 960 961 962 963
        # 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
964 965 966 967
        assert len(self.ops) == len(ops_in_cpp)
        for index in range(len(self.ops)):
            assert self.ops[index].desc == ops_in_cpp[index]

968 969
    def copy_param_info_from(self, other):
        """
970
        Copy the information of parameters from the other block
971
        Args:
972
            other(Block): the other block
973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995

        Returns:
            None
        """
        if not isinstance(other, Block):
            raise TypeError("copy_param_info_from should be invoked with Block")
        for p in other.iter_parameters():
            assert isinstance(p, Parameter)
            v = self.vars.get(p.name, None)
            if v is None:
                raise ValueError("copy_param_info_from should be invoked with "
                                 "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
996
                gradient_clip_attr=p.gradient_clip_attr,
F
fengjiayi 已提交
997
                error_clip=p.error_clip,
998 999 1000
                name=v.name)
            self.vars[new_p.name] = new_p

1001 1002 1003 1004 1005 1006 1007 1008 1009 1010
    def clone_variable(self, var):
        """
        Clone a variable into current block.
        Args:
            var: the variable to be cloned.

        Returns:
            The new  variable cloned from 'var' in current block.
        """
        assert isinstance(var, Variable)
T
update  
typhoonzero 已提交
1011 1012 1013 1014 1015
        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
typhoonzero 已提交
1016 1017 1018 1019 1020 1021
        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,
F
fengjiayi 已提交
1022 1023
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1024 1025 1026 1027 1028 1029 1030
        else:
            ret_var = self.create_var(
                name=var.name,
                shape=var.shape,
                dtype=var.dtype,
                type=var.type,
                lod_level=var.lod_level,
F
fengjiayi 已提交
1031 1032
                persistable=True,
                is_data=var.is_data)
T
update  
typhoonzero 已提交
1033
        return ret_var
1034

Y
Yu Yang 已提交
1035 1036

class Program(object):
D
dzhwinter 已提交
1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047
    """
    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
1048
    default_main_program run in every mini batch and adjust the weights.
D
dzhwinter 已提交
1049 1050

    Returns:
Y
yuyang18 已提交
1051
        A empty program.
D
dzhwinter 已提交
1052 1053

    Examples:
Y
yuyang18 已提交
1054 1055 1056 1057 1058 1059
        >>> 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
1060 1061 1062

    """

1063 1064
    def __init__(self):
        self.desc = core.ProgramDesc()
Y
Yu Yang 已提交
1065 1066
        self.blocks = [Block(self, 0)]
        self.current_block_idx = 0
D
dzhwinter 已提交
1067
        self._seed = 0
Y
yuyang18 已提交
1068
        self._current_role = core.op_proto_and_checker_maker.OpRole.Forward
Y
yuyang18 已提交
1069
        self._op_role_var = []
Y
yuyang18 已提交
1070 1071 1072

    @property
    def op_role(self):
Y
yuyang18 已提交
1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085
        """
        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
1086 1087 1088 1089 1090 1091 1092 1093
        return self._current_role

    @op_role.setter
    def set_op_role(self, role):
        self._current_role = role

    @property
    def op_role_var(self):
Y
yuyang18 已提交
1094 1095 1096 1097 1098 1099 1100
        """
        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
1101 1102 1103 1104
        return self._op_role_var

    @op_role_var.setter
    def set_op_role_var(self, var_name):
Y
yuyang18 已提交
1105
        self._op_role_var = [var_name]
Y
yuyang18 已提交
1106 1107

    @contextlib.contextmanager
Y
yuyang18 已提交
1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123
    def optimization_guard(self, var):
        """
        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:
            var(Variable|str): The variable (name) to be optimized.

        Examples:

            >>> p, g = backward(...)
            >>> with program.optimization_guard(p):
            >>>     p = p - 0.001 * g
        """
Y
yuyang18 已提交
1124 1125
        OpRole = core.op_proto_and_checker_maker.OpRole
        self._current_role = OpRole.Optimize
Y
yuyang18 已提交
1126
        self._op_role_var = [var.name if isinstance(var, Variable) else var]
Y
yuyang18 已提交
1127
        yield
Y
yuyang18 已提交
1128
        self._op_role_var = []
Y
yuyang18 已提交
1129
        self._current_role = OpRole.Forward
Y
Yu Yang 已提交
1130

1131
    def __str__(self):
Y
yuyang18 已提交
1132 1133 1134 1135 1136 1137 1138 1139 1140
        """
        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
1141 1142
        return self.to_string(True)

F
fengjiayi 已提交
1143 1144 1145
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
Y
yuyang18 已提交
1146

F
fengjiayi 已提交
1147
        Args:
Y
yuyang18 已提交
1148 1149
            throw_on_error(bool): raise Value error when any of required fields
                is not set.
F
fengjiayi 已提交
1150

Y
yuyang18 已提交
1151 1152 1153 1154 1155 1156 1157 1158 1159 1160
            with_details(bool): True if more details about variables and
                parameters, e.g., :code:`trainable`, :code:`optimize_attr`, need
                to print.

        Returns
            (str): The debug string.

        Raises:
            ValueError: If any of required fields is not set and throw_on_error is
                True.
F
fengjiayi 已提交
1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173

        """
        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()
            proto = framework_pb2.ProgramDesc.FromString(str(protostr))
            res_str = _debug_string_(proto, throw_on_error)
        return res_str
1174

1175
    def get_desc(self):
Y
yuyang18 已提交
1176 1177 1178 1179 1180 1181 1182
        """
        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.
        """
1183 1184
        return self.desc

1185
    def clone(self, for_test=False):
Y
yuyang18 已提交
1186 1187 1188 1189 1190 1191 1192 1193
        """
        Create a new, duplicated program.


        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`.
1194

Y
yuyang18 已提交
1195 1196 1197 1198 1199
        * 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
        :code:`clone(for_test=True)` before backward and optimization please.
1200 1201

        Args:
Y
yuyang18 已提交
1202 1203
            for_test(bool): True if change the :code:`is_test` attribute of
                operators to :code:`True`.
1204

D
dzhwinter 已提交
1205
        Returns:
Y
yuyang18 已提交
1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258
            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.
1259 1260
        """
        if for_test:
1261
            p = self.inference_optimize()
1262
        else:
1263
            p = Program()
1264
            p.desc = core.ProgramDesc(self.desc)
1265 1266 1267
            p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())]
            p.sync_with_cpp()

1268
        p.copy_param_info_from(self)
F
fengjiayi 已提交
1269
        p.copy_data_info_from(self)
Y
Yu Yang 已提交
1270
        return p
1271

1272
    def prune(self, targets):
Y
yuyang18 已提交
1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287
        """
        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.

        """
1288 1289 1290 1291 1292 1293
        if not isinstance(targets, list):
            targets = [targets]
        targets_idx = []
        for t in targets:
            if not isinstance(t, Operator):
                if isinstance(t, Variable):
1294 1295
                    # After transpiler processing, the op that output this
                    # variable maybe has been changed, so t.op is not reliable
1296
                    # and we need to find the current op that generate this
1297 1298 1299 1300 1301 1302 1303 1304
                    # 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

1305
                    t = t.op
1306 1307 1308 1309
                    if t is None:
                        raise ValueError(
                            "The target variable must have an "
                            "associated operator that generates it.")
1310
                else:
1311 1312
                    raise ValueError("All targets of prune() can only be "
                                     "Variable or Operator.")
1313 1314 1315 1316 1317 1318 1319 1320

            targets_idx.append([t.block.idx, t.idx])
        res = Program()
        res.desc = core.prune(self.desc, targets_idx)
        res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())]
        res.sync_with_cpp()
        return res

1321
    def inference_optimize(self):
Y
yuyang18 已提交
1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332
        """
        This method will create a new program and change the :code:`is_test`
        attribute of operators to :code:`True`. All the :code:`Parameter`
        information will be lost.

        Notes: This API is a very low level API. Use
        :code:`Program.clone(for_test=True)` instead.

        Returns:
            Program: The new program.
        """
1333 1334
        # this is an alternative implement before
        # core.inference_optimize being fixed.
1335
        res = Program()
1336 1337 1338 1339 1340 1341 1342
        res.desc = core.ProgramDesc(self.desc)
        for i in xrange(res.desc.num_blocks()):
            block = res.desc.block(i)
            for j in xrange(block.op_size()):
                op = block.op(j)
                if op.has_attr('is_test'):
                    op.set_attr('is_test', True)
1343 1344 1345 1346
        res.blocks = [Block(res, i) for i in xrange(res.desc.num_blocks())]
        res.sync_with_cpp()
        return res

1347 1348
    @staticmethod
    def parse_from_string(binary_str):
Y
yuyang18 已提交
1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360
        """
        Deserialize a program desc from protobuf binary string.

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

        Args:
            binary_str(str): The binary prootbuf string.

        Returns:
            Program: A deserialized program desc.
        """
1361 1362
        p = Program()
        p.desc = core.ProgramDesc(binary_str)
1363
        p.blocks = [Block(p, i) for i in xrange(p.desc.num_blocks())]
1364 1365
        p.sync_with_cpp()
        return p
Y
Yu Yang 已提交
1366

D
dzhwinter 已提交
1367 1368
    @property
    def random_seed(self):
Y
yuyang18 已提交
1369 1370 1371 1372 1373 1374
        """
        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
1375 1376
        return self._seed

Q
qiaolongfei 已提交
1377 1378
    @property
    def num_blocks(self):
Y
yuyang18 已提交
1379 1380 1381
        """
        The number of blocks in this program.
        """
Q
qiaolongfei 已提交
1382 1383
        return self.desc.num_blocks()

D
dzhwinter 已提交
1384 1385 1386 1387 1388 1389
    @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
1390 1391
    def __repr__(self):
        return str(self)
1392

Y
Yu Yang 已提交
1393
    def global_block(self):
Y
yuyang18 已提交
1394 1395 1396
        """
        Get the first block of this program.
        """
Y
Yu Yang 已提交
1397 1398
        return self.blocks[0]

Q
Qiao Longfei 已提交
1399
    def block(self, index):
Y
yuyang18 已提交
1400 1401 1402 1403 1404 1405 1406 1407
        """
        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
1408 1409
        return self.blocks[index]

Y
Yu Yang 已提交
1410
    def current_block(self):
Y
yuyang18 已提交
1411 1412 1413 1414
        """
        Get the current block. The :code:`current` block is the block to append
        operators.
        """
Y
Yu Yang 已提交
1415 1416
        return self.blocks[self.current_block_idx]

F
update  
fengjiayi 已提交
1417
    def create_block(self, parent_idx=None):
Y
yuyang18 已提交
1418 1419 1420 1421 1422 1423 1424 1425 1426 1427
        """
        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
1428
        new_block_idx = len(self.blocks)
F
update  
fengjiayi 已提交
1429 1430 1431
        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
1432 1433 1434 1435 1436
        self.current_block_idx = new_block_idx
        self.blocks.append(Block(self, self.current_block_idx))
        return self.current_block()

    def rollback(self):
Y
yuyang18 已提交
1437 1438 1439 1440 1441
        """
        Exit a code block, i.e., roll back to the parent block.
        Returns:
            None
        """
Y
Yu Yang 已提交
1442 1443
        self.current_block_idx = self.current_block().parent_idx

Q
Qiao Longfei 已提交
1444
    def sync_with_cpp(self):
Y
yuyang18 已提交
1445 1446 1447 1448 1449 1450 1451 1452 1453 1454
        """
        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
1455 1456 1457 1458 1459
        for block_idx in range(len(self.blocks), self.desc.num_blocks()):
            self.blocks.append(Block(self, block_idx))
        for block in self.blocks:
            block.sync_with_cpp()

1460 1461
    def copy_param_info_from(self, other):
        """
1462
        Copy the information of parameters from other program.
D
dzhwinter 已提交
1463

Y
yuyang18 已提交
1464 1465 1466
        Notes: This is a very low level API. Users should not invoke it
        directly.

1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
            raise TypeError("copy_param_info_from should be invoked with "
                            "Program")

        if len(self.blocks) != len(other.blocks):
            raise ValueError("copy_param_info_from should be invoked with two "
                             "program, with represent the same topology")
        self.global_block().copy_param_info_from(other.global_block())

F
fengjiayi 已提交
1482 1483 1484
    def copy_data_info_from(self, other):
        """
        Copy the information of data variables from other program.
D
dzhwinter 已提交
1485

Y
yuyang18 已提交
1486 1487 1488
        Notes: This is a very low level API. Users should not invoke it
        directly.

F
fengjiayi 已提交
1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505
        Args:
            other(Program): Other program

        Returns:
            None
        """
        if not isinstance(other, Program):
            raise TypeError("copy_param_info_from should be invoked with "
                            "Program")

        if len(self.blocks) != len(other.blocks):
            raise ValueError("copy_param_info_from should be invoked with two "
                             "program, with represent the same topology")
        for var in other.global_block().vars.itervalues():
            if var.is_data:
                self.global_block().var(var.name).is_data = True

1506
    def list_vars(self):
Y
yuyang18 已提交
1507 1508 1509 1510 1511 1512
        """
        Get all variables from this Program. A iterable object is returned.

        Returns:
            iterable: The generator will yield every variable in this program.
        """
1513 1514 1515 1516
        for each_block in self.blocks:
            for each_var in each_block.vars.itervalues():
                yield each_var

Y
Yu Yang 已提交
1517

Y
Yu Yang 已提交
1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528
class Parameter(Variable):
    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")
1529 1530 1531

        Variable.__init__(
            self, block, persistable=True, shape=shape, dtype=dtype, **kwargs)
Y
Yu Yang 已提交
1532 1533 1534 1535
        self.trainable = kwargs.get('trainable', True)

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

1536 1537
        self.regularizer = kwargs.get('regularizer', None)

F
fengjiayi 已提交
1538
        self.gradient_clip_attr = kwargs.get('gradient_clip_attr', None)
Y
Yu Yang 已提交
1539

W
wanghaoshuang 已提交
1540
        self.do_model_average = kwargs.get('do_model_average', None)
W
wanghaoshuang 已提交
1541

F
fengjiayi 已提交
1542 1543 1544
    def __str__(self):
        return self.to_string(True)

F
update  
fengjiayi 已提交
1545 1546 1547
    def to_string(self, throw_on_error, with_details=False):
        """
        To debug string.
D
dzhwinter 已提交
1548

F
update  
fengjiayi 已提交
1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562
        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
1563
                               "gradient_clip_attr", "do_model_average")
F
update  
fengjiayi 已提交
1564 1565 1566 1567 1568
            for attr_name in additional_attr:
                res_str += "%s: %s\n" % (attr_name,
                                         str(getattr(self, attr_name)))
        else:
            res_str = Variable.to_string(self, throw_on_error, False)
F
fengjiayi 已提交
1569 1570 1571 1572
        return res_str

    __repr__ = __str__

Y
Yu Yang 已提交
1573

Y
Yu Yang 已提交
1574
# program is a global instance.
Y
Yu Yang 已提交
1575 1576
_main_program_ = Program()
_startup_program_ = Program()
1577

1578

1579
def default_startup_program():
Y
Yu Yang 已提交
1580 1581 1582
    """
    Get default startup program. In startup program, Paddle will initialize
    parameters, initialize nccl handle, etc.
1583

Y
Yu Yang 已提交
1584 1585 1586
    Returns:
        Program: startup program
    """
Y
Yu Yang 已提交
1587
    return _startup_program_
1588

1589

1590
def default_main_program():
Y
Yu Yang 已提交
1591 1592
    """
    Get default main program. The main program is used for training or testing.
1593

Y
Yu Yang 已提交
1594 1595 1596
    Returns:
        Program: main program
    """
Y
Yu Yang 已提交
1597
    return _main_program_
Y
Yu Yang 已提交
1598 1599 1600 1601 1602


def switch_main_program(program):
    """
    Switch the main program to a new program.
1603

Y
Yu Yang 已提交
1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617
    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):
    """
1618
    Switch the startup program to a new program
Y
Yu Yang 已提交
1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634
    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


@contextlib.contextmanager
def program_guard(main_program, startup_program=None):
    """
    Switch program with `with` statement
1635

Y
Yu Yang 已提交
1636 1637 1638 1639
    Examples:
        >>> with program_guard(Program()):
        >>>   data = fluid.layers.data(...)
        >>>   hidden = fluid.layers.fc(...)
1640

Y
Yu Yang 已提交
1641 1642
    Args:
        main_program(Program): New main program inside `with` statement
1643
        startup_program(Program): New startup program inside `with` statement.
Y
Yu Yang 已提交
1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659
            None means do not change startup program.

    Returns:
        None
    """
    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
1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675


def get_var(name, program=None):
    """
    Get a variable by name from the global block of a program
    Args:
        name(str): name of the variable
        program(Program|None): program object.
             If None, default_global_program() will be used.

    Returns:
        Variable
    """
    if program is None:
        program = default_main_program()
    assert isinstance(name, str)
1676
    assert isinstance(program, Program)
X
xuwei06 已提交
1677 1678

    return program.global_block().var(name)