layers.py 57.8 KB
Newer Older
1 2 3 4
from . import core
import proto.framework_pb2 as framework_pb2
from framework import OpProtoHolder, Variable, Program, Operator
from initializer import Constant, Normal, Xavier
Q
Qiao Longfei 已提交
5
from paddle.v2.fluid.layer_helper import LayerHelper, unique_name
Y
Yu Yang 已提交
6
import re
7
import cStringIO
Y
Yu Yang 已提交
8

Q
QI JUN 已提交
9
__all__ = [
Y
Yu Yang 已提交
10
    'fc', 'data', 'cross_entropy', 'conv2d', 'pool2d', 'embedding', 'concat',
D
dzhwinter 已提交
11
    'StaticRNN', 'cast', 'sequence_conv', 'sequence_pool', 'sums', 'cos_sim',
12
    'batch_norm', 'accuracy', 'split_lod_tensor'
Q
QI JUN 已提交
13
]
Y
Yu Yang 已提交
14 15


F
fengjiayi 已提交
16 17
def fc(input,
       size,
C
chengduoZH 已提交
18
       num_flatten_dims=1,
F
fengjiayi 已提交
19
       param_attr=None,
20
       param_initializer=None,
Q
QI JUN 已提交
21
       bias_attr=None,
22
       bias_initializer=None,
F
fengjiayi 已提交
23
       act=None,
C
chengduoZH 已提交
24
       name=None,
25 26
       main_program=None,
       startup_program=None):
27 28 29 30 31 32
    """
    Fully Connected Layer.

    Args:
       input: The input tensor to the function
       size: The size of the layer
C
chengduoZH 已提交
33
       num_flatten_dims: Number of columns in input
34
       param_attr: The parameters/weights to the FC Layer
35 36
       param_initializer: Initializer used for the weight/parameter.
       If None, XavierInitializer() is used
37
       bias_attr: The bias parameter for the FC layer
38 39
       bias_initializer: Initializer used for the bias.
       If None, then ConstantInitializer() is used
40
       act: Activation to be applied to the output of FC layer
C
chengduoZH 已提交
41
       name: Name/alias of the function
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
       main_program: Name of the main program that calls this
       startup_program: Name of the startup program

    This function can take in multiple inputs and performs the Fully Connected
    function (linear transformation) on top of each of them.
    So for input x, the output will be : Wx + b. Where W is the parameter,
    b the bias and x is the input.

    The function also applies an activation (non-linearity) on top of the
    output, if activation is passed in the input.

    All the input variables of this function are passed in as local variables
    to the LayerHelper constructor.

    """
57 58

    def _get_default_param_initializer():
59
        return Xavier()
60 61

    def _get_default_bias_initializer():
62
        return Constant()
63

Y
Yu Yang 已提交
64 65 66 67
    helper = LayerHelper('fc', **locals())

    dtype = helper.input_dtype()

68 69 70 71 72 73
    if param_initializer is None:
        param_initializer = _get_default_param_initializer()

    if bias_initializer is None:
        bias_initializer = _get_default_bias_initializer()

Y
Yu Yang 已提交
74 75 76
    mul_results = []
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
Y
Yu Yang 已提交
77 78 79
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
Y
Yu Yang 已提交
80
        w = helper.create_parameter(
81 82 83 84
            attr=param_attr,
            initializer=param_initializer,
            shape=param_shape,
            dtype=dtype)
Y
Yu Yang 已提交
85 86 87 88 89 90 91 92
        tmp = helper.create_tmp_variable(dtype)
        helper.append_op(
            type="mul",
            inputs={
                "X": input_var,
                "Y": w,
            },
            outputs={"Out": tmp},
Y
Yu Yang 已提交
93 94
            attrs={'x_num_col_dims': num_flatten_dims,
                   'y_num_col_dims': 1})
Y
Yu Yang 已提交
95 96 97 98 99 100 101 102 103 104
        mul_results.append(tmp)

    # sum
    if len(mul_results) == 1:
        pre_bias = mul_results[0]
    else:
        pre_bias = helper.create_tmp_variable(dtype)
        helper.append_op(
            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
    # add bias
105
    pre_activation = helper.append_bias_op(pre_bias, bias_initializer)
Y
Yu Yang 已提交
106 107 108 109
    # add activation
    return helper.append_activation(pre_activation)


Q
QI JUN 已提交
110 111
def embedding(input,
              size,
112
              is_sparse=False,
Q
Qiao Longfei 已提交
113
              param_initializer=None,
Q
QI JUN 已提交
114
              param_attr=None,
F
fengjiayi 已提交
115
              dtype='float32',
116 117
              main_program=None,
              startup_program=None):
118 119 120 121 122 123 124 125
    """
    Embedding Layer.

    Args:
       input: The input to the function
       size: The size of the layer
       is_sparse: A flag that decleares whether the input is sparse
       param_attr: Parameters for this layer
F
fengjiayi 已提交
126
       dtype: The type of data : float32, float_16, int etc
127 128 129 130 131 132 133 134 135 136 137
       main_program: Name of the main program that calls this
       startup_program: Name of the startup program

    This function can take in the input (which is a vector of IDs) and
    performs a lookup in the lookup_table using these IDs, to result into
    the embedding of each ID in the input.

    All the input variables of this function are passed in as local variables
    to the LayerHelper constructor.

    """
Q
Qiao Longfei 已提交
138 139

    def _get_default_param_initializer():
140
        return Xavier()
Q
Qiao Longfei 已提交
141

Q
QI JUN 已提交
142 143
    helper = LayerHelper('embedding', **locals())
    w = helper.create_parameter(
Q
Qiao Longfei 已提交
144 145
        attr=helper.param_attr,
        shape=size,
F
fengjiayi 已提交
146
        dtype=dtype,
Q
Qiao Longfei 已提交
147
        initializer=param_initializer or _get_default_param_initializer())
F
fengjiayi 已提交
148
    tmp = helper.create_tmp_variable(dtype)
Q
QI JUN 已提交
149 150 151 152
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
153 154
        outputs={'Out': tmp},
        attrs={'is_sparse': is_sparse})
Q
QI JUN 已提交
155 156 157
    return tmp


Q
QI JUN 已提交
158 159 160 161 162 163 164 165 166 167
# TODO(qijun): expose H0 and C0
def dynamic_lstm(input,
                 size,
                 param_attr=None,
                 bias_attr=None,
                 use_peepholes=True,
                 is_reverse=False,
                 gate_activation='sigmoid',
                 cell_activation='tanh',
                 candidate_activation='tanh',
F
fengjiayi 已提交
168
                 dtype='float32',
Q
QI JUN 已提交
169 170 171 172 173
                 main_program=None,
                 startup_program=None):
    helper = LayerHelper('lstm', **locals())
    size = size / 4
    weight = helper.create_parameter(
F
fengjiayi 已提交
174
        attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype)
Q
QI JUN 已提交
175 176 177 178
    bias_size = [1, 7 * size]
    if not use_peepholes:
        bias_size[1] = 4 * size
    bias = helper.create_parameter(
F
fengjiayi 已提交
179
        attr=helper.bias_attr, shape=bias_size, dtype=dtype, suffix='b')
Q
QI JUN 已提交
180

F
fengjiayi 已提交
181 182 183 184
    hidden = helper.create_tmp_variable(dtype)
    cell = helper.create_tmp_variable(dtype)
    batch_gate = helper.create_tmp_variable(dtype)
    batch_cell_pre_act = helper.create_tmp_variable(dtype)
Q
QI JUN 已提交
185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206

    helper.append_op(
        type='lstm',
        inputs={'Input': input,
                'Weight': weight,
                'Bias': bias},
        outputs={
            'Hidden': hidden,
            'Cell': cell,
            'BatchGate': batch_gate,
            'BatchCellPreAct': batch_cell_pre_act
        },
        attrs={
            'use_peepholes': use_peepholes,
            'is_reverse': is_reverse,
            'gate_activation': gate_activation,
            'cell_activation': cell_activation,
            'candidate_activation': candidate_activation
        })
    return hidden, cell


F
fengjiayi 已提交
207 208
def data(name,
         shape,
C
chengduoZH 已提交
209
         append_batch_size=True,
F
fengjiayi 已提交
210
         dtype='float32',
F
fengjiayi 已提交
211
         type=core.VarDesc.VarType.LOD_TENSOR,
212
         main_program=None,
213 214
         startup_program=None,
         stop_gradient=True):
215 216 217 218 219 220
    """
    Data Layer.

    Args:
       name: The name/alias of the function
       shape: Tuple declaring the shape.
C
chengduoZH 已提交
221
       append_batch_size: Whether or not to append the data as a batch.
F
fengjiayi 已提交
222
       dtype: The type of data : float32, float_16, int etc
223 224 225 226 227 228 229 230 231 232 233 234 235 236
       type: The output type. By default it is LOD_TENSOR.
       main_program: Name of the main program that calls this
       startup_program: Name of the startup program
       stop_gradient: A boolean that mentions whether gradient should flow.

    This function takes in input and based on whether data has
    to be returned back as a minibatch, it creates the global variable using
    the helper functions. The global variables can be accessed by all the
    following operations and layers in the graph.

    All the input variables of this function are passed in as local variables
    to the LayerHelper constructor.

    """
Y
Yu Yang 已提交
237
    helper = LayerHelper('data', **locals())
Y
Yu Yang 已提交
238 239 240 241 242 243 244 245
    shape = list(shape)
    for i in xrange(len(shape)):
        if shape[i] is None:
            shape[i] = -1
            append_batch_size = False
        elif shape[i] < 0:
            append_batch_size = False

Y
Yu Yang 已提交
246 247
    if append_batch_size:
        shape = [-1] + shape  # append batch size as -1
Y
Yu Yang 已提交
248

Y
Yu Yang 已提交
249
    return helper.create_global_variable(
250 251
        name=name,
        shape=shape,
F
fengjiayi 已提交
252
        dtype=dtype,
253 254
        type=type,
        stop_gradient=stop_gradient)
Y
Yu Yang 已提交
255 256


Y
Yu Yang 已提交
257
def create_tensor(dtype, name=None, main_program=None, startup_program=None):
Y
Yu Yang 已提交
258 259
    helper = LayerHelper("create_tensor", **locals())
    return helper.create_variable(name=helper.name, dtype=dtype)
Y
Yu Yang 已提交
260 261 262


def _convert_(name):
263 264 265 266 267 268 269 270 271 272 273
    """
    Formatting.

    Args:
       name: The name/alias

    This function takes in a name and converts it to a standard format of
    group1_group2. Where as per the regular expression, group1 can have
    alphabets and numbers and group2 has capital alphabets.

    """
Y
Yu Yang 已提交
274 275 276 277
    s1 = re.sub('(.)([A-Z][a-z]+)', r'\1_\2', name)
    return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower()


278 279 280
def _generate_doc_string_(op_proto):
    """
    Generate docstring by OpProto
X
xuwei06 已提交
281

282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329
    Args:
        op_proto (framework_pb2.OpProto): a protobuf message typed OpProto

    Returns:
        str: the document string
    """

    def _type_to_str_(tp):
        return framework_pb2.AttrType.Name(tp)

    if not isinstance(op_proto, framework_pb2.OpProto):
        raise TypeError("OpProto should be `framework_pb2.OpProto`")

    buf = cStringIO.StringIO()
    buf.write(op_proto.comment)
    buf.write('\nArgs:\n')
    for each_input in op_proto.inputs:
        line_begin = '    {0}: '.format(_convert_(each_input.name))
        buf.write(line_begin)
        buf.write(each_input.comment)
        buf.write('\n')
        buf.write(' ' * len(line_begin))
        buf.write('Duplicable: ')
        buf.write(str(each_input.duplicable))
        buf.write('  Optional: ')
        buf.write(str(each_input.dispensable))
        buf.write('\n')

    for each_attr in op_proto.attrs:
        buf.write('    ')
        buf.write(each_attr.name)
        buf.write(' (')
        buf.write(_type_to_str_(each_attr.type))
        buf.write('): ')
        buf.write(each_attr.comment)
        buf.write('\n')

    if len(op_proto.outputs) != 0:
        buf.write('\nReturns:\n')
        buf.write('    ')
        for each_opt in op_proto.outputs:
            if not each_opt.intermediate:
                break
        buf.write(each_opt.comment)

    return buf.getvalue()


Y
Yu Yang 已提交
330
def _create_op_func_(op_type):
331 332 333 334 335 336 337 338 339 340
    """
    Create an Operator for a Function.

    Args:
       op_type: The name of the operator to be created

    This function takes in the operator type (sigmoid, mean , average etc) and
    creates the operator functionality.

    """
Y
Yu Yang 已提交
341
    op_proto = OpProtoHolder.instance().get_op_proto(op_type)
342 343 344 345 346 347
    not_intermediate_outputs = \
        filter(lambda output: not output.intermediate, op_proto.outputs)
    intermediate_outputs = \
        filter(lambda output: output.intermediate, op_proto.outputs)

    if len(not_intermediate_outputs) != 1:
348 349
        raise ValueError("Only one non intermediate output operator can be",
                         "automatically generated")
Y
Yu Yang 已提交
350

351
    if not_intermediate_outputs[0].duplicable:
Y
Yu Yang 已提交
352
        raise ValueError(
353
            "Only non duplicable op can be automatically generated")
Y
Yu Yang 已提交
354

355 356
    for output in intermediate_outputs:
        if output.duplicable:
357 358
            raise ValueError("The op can be automatically generated only when ",
                             "all intermediate ops are not duplicable")
359 360 361

    o_name = not_intermediate_outputs[0].name
    intermediate_output_names = [output.name for output in intermediate_outputs]
Y
Yu Yang 已提交
362

F
fengjiayi 已提交
363
    def infer_and_check_dtype(op_proto, **kwargs):
364
        """
F
fengjiayi 已提交
365
        This function performs the sanity check for dtype and
366 367
        instance type.
        """
Y
Yu Yang 已提交
368 369 370 371 372 373 374 375 376 377 378 379
        dtype = None
        for ipt in op_proto.inputs:
            name = _convert_(ipt.name)
            val = kwargs.pop(name, [])
            if not isinstance(val, list) and not isinstance(val, tuple):
                val = [val]
            for each in val:
                if not isinstance(each, Variable):
                    raise ValueError("input of {0} must be variable".format(
                        op_type))

                if dtype is None:
F
fengjiayi 已提交
380 381
                    dtype = each.dtype
                elif dtype != each.dtype:
Y
Yu Yang 已提交
382 383
                    raise ValueError(
                        "operator {0} must input same dtype".format(op_type))
Y
Yang Yang(Tony) 已提交
384 385 386 387 388 389

        return dtype

    def func(**kwargs):
        helper = LayerHelper(op_type, **kwargs)

F
fengjiayi 已提交
390
        dtype = infer_and_check_dtype(op_proto, **kwargs)
Y
Yang Yang(Tony) 已提交
391 392 393 394 395 396 397

        inputs = dict()
        for ipt in op_proto.inputs:
            name = _convert_(ipt.name)
            val = kwargs.pop(name, [])
            if not isinstance(val, list) and not isinstance(val, tuple):
                val = [val]
Y
Yu Yang 已提交
398 399
            inputs[ipt.name] = val

400
        outputs = dict()
Y
Yu Yang 已提交
401
        out = helper.create_tmp_variable(dtype=dtype)
402 403 404
        outputs[o_name] = [out]
        for name in intermediate_output_names:
            outputs[name] = [helper.create_tmp_variable(dtype=dtype)]
Y
Yu Yang 已提交
405
        helper.append_op(
406
            type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs)
Q
Qiao Longfei 已提交
407
        return helper.append_activation(out)
Y
Yu Yang 已提交
408 409 410

    func.__name__ = op_type
    globals()[op_type] = func
411
    func.__doc__ = _generate_doc_string_(op_proto)
Y
Yu Yang 已提交
412 413 414 415 416
    global __all__
    __all__.append(op_type)


_create_op_func_('mean')
Y
Yu Yang 已提交
417
_create_op_func_('mul')
Q
Qiao Longfei 已提交
418
_create_op_func_('elementwise_add')
Y
Yu Yang 已提交
419
_create_op_func_('elementwise_div')
420
_create_op_func_('dropout')
Q
Qiao Longfei 已提交
421
_create_op_func_('reshape')
Y
Yu Yang 已提交
422 423
_create_op_func_('sigmoid')
_create_op_func_('scale')
Y
Yang Yang(Tony) 已提交
424 425 426 427
_create_op_func_('reshape')
_create_op_func_('transpose')


F
fengjiayi 已提交
428
def cast(x, dtype, main_program=None):
429
    """
F
fengjiayi 已提交
430 431
    This function takes in the input with input_dtype
    and casts it to the output_dtype as the output.
432
    """
Y
Yu Yang 已提交
433
    helper = LayerHelper('cast', **locals())
F
fengjiayi 已提交
434
    out = helper.create_tmp_variable(dtype=dtype)
Y
Yu Yang 已提交
435 436 437 438
    helper.append_op(
        type='cast',
        inputs={'X': [x]},
        outputs={'Out': [out]},
F
fengjiayi 已提交
439 440
        attrs={'in_dtype': x.dtype,
               'out_dtype': out.dtype})
Y
Yu Yang 已提交
441 442 443
    return out


444
def concat(input, axis, main_program=None, startup_program=None):
445 446 447 448
    """
    This function concats the input along the axis mentioned
    and returns that as the output.
    """
Q
QI JUN 已提交
449
    helper = LayerHelper('concat', **locals())
D
dzhwinter 已提交
450
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
Q
QI JUN 已提交
451 452 453 454 455 456 457 458
    helper.append_op(
        type='concat',
        inputs={'X': input},
        outputs={'Out': [out]},
        attrs={'axis': axis})
    return out


Y
Yu Yang 已提交
459
def sums(input, out=None, main_program=None, startup_program=None):
460 461 462 463
    """
    This function takes in the input and performs the sum operation on it
    and returns that as the output.
    """
D
dzhwinter 已提交
464
    helper = LayerHelper('sum', **locals())
Y
Yu Yang 已提交
465 466
    if out is None:
        out = helper.create_tmp_variable(dtype=helper.input_dtype())
Y
Yu Yang 已提交
467
    helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out})
D
dzhwinter 已提交
468 469 470
    return out


Q
Qiao Longfei 已提交
471 472 473 474 475 476 477
def linear_chain_crf(input,
                     label,
                     param_attr=None,
                     param_initializer=None,
                     main_program=None,
                     startup_program=None):
    def _get_default_param_initializer():
478
        return Xavier()
Q
Qiao Longfei 已提交
479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505

    helper = LayerHelper('linear_chain_crf', **locals())
    size = input.shape[1]
    transition = helper.create_parameter(
        attr=helper.param_attr,
        shape=[size + 2, size],
        dtype=helper.input_dtype(),
        initializer=param_initializer or _get_default_param_initializer())
    alpha = helper.create_tmp_variable(dtype=helper.input_dtype())
    emission_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
    transition_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
    log_likelihood = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(
        type='linear_chain_crf',
        inputs={"Emission": [input],
                "Transition": transition,
                "Label": label},
        outputs={
            "Alpha": [alpha],
            "EmissionExps": [emission_exps],
            "TransitionExps": transition_exps,
            "LogLikelihood": log_likelihood
        })

    return log_likelihood


Y
Yu Yang 已提交
506
def assign(input, output, main_program=None, startup_program=None):
Y
Yu Yang 已提交
507 508 509 510 511 512 513 514 515
    helper = LayerHelper('assign', **locals())
    helper.append_op(
        type='scale',
        inputs={'X': [input]},
        outputs={'Out': [output]},
        attrs={'scale': 1.0})
    return output


516 517
def split_lod_tensor(input,
                     mask,
Y
Yu Yang 已提交
518
                     level=0,
519 520 521
                     main_program=None,
                     startup_program=None):
    helper = LayerHelper('split_lod_tensor', **locals())
F
fengjiayi 已提交
522 523
    out_true = helper.create_tmp_variable(dtype=input.dtype)
    out_false = helper.create_tmp_variable(dtype=input.dtype)
524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539
    helper.append_op(
        type='split_lod_tensor',
        inputs={
            'X': input,
            'Mask': mask,
        },
        outputs={'OutTrue': out_true,
                 'OutFalse': out_false},
        attrs={'level': level})
    return out_true, out_false


def merge_lod_tensor(in_true,
                     in_false,
                     x,
                     mask,
Y
Yu Yang 已提交
540
                     level=0,
541 542 543
                     main_program=None,
                     startup_program=None):
    helper = LayerHelper('merge_lod_tensor', **locals())
F
fengjiayi 已提交
544
    out = helper.create_tmp_variable(dtype=in_true.dtype)
545 546 547 548 549 550 551 552 553 554 555
    helper.append_op(
        type='merge_lod_tensor',
        inputs={'X': x,
                'Mask': mask,
                'InTrue': in_true,
                'InFalse': in_false},
        outputs={'Out': out},
        attrs={'level': level})
    return out


556
def cos_sim(X, Y, **kwargs):
557 558 559 560
    """
    This function performs the cosine similarity between two tensors
    X and Y and returns that as the output.
    """
561
    helper = LayerHelper('cos_sim', **kwargs)
F
fengjiayi 已提交
562 563 564
    out = helper.create_tmp_variable(dtype=X.dtype)
    xnorm = helper.create_tmp_variable(dtype=X.dtype)
    ynorm = helper.create_tmp_variable(dtype=X.dtype)
D
dzhwinter 已提交
565 566 567 568 569 570 571
    helper.append_op(
        type='cos_sim',
        inputs={'X': [X],
                'Y': [Y]},
        outputs={'Out': [out],
                 'XNorm': [xnorm],
                 'YNorm': [ynorm]})
572
    return out
D
dzhwinter 已提交
573 574


Y
Yu Yang 已提交
575
def cross_entropy(input, label, **kwargs):
576 577 578
    """
    This function computes cross_entropy using the input and label.
    """
Y
Yu Yang 已提交
579
    helper = LayerHelper('cross_entropy', **kwargs)
F
fengjiayi 已提交
580
    out = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
581 582 583 584 585 586 587 588 589 590
    helper.append_op(
        type='cross_entropy',
        inputs={'X': [input],
                'Label': [label]},
        outputs={'Y': [out]},
        attrs=kwargs)
    return out


def square_error_cost(input, label, **kwargs):
591 592 593 594
    """
    This functions returns the squared error cost using the input and label.
    The output is appending the op to do the above.
    """
Y
Yu Yang 已提交
595
    helper = LayerHelper('square_error_cost', **kwargs)
F
fengjiayi 已提交
596
    minus_out = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
597 598 599 600 601 602
    helper.append_op(
        type='elementwise_sub',
        inputs={'X': [input],
                'Y': [label]},
        outputs={'Out': [minus_out]})

F
fengjiayi 已提交
603
    square_out = helper.create_tmp_variable(dtype=input.dtype)
Y
Yu Yang 已提交
604
    helper.append_op(
Q
QI JUN 已提交
605
        type='square', inputs={'X': [minus_out]}, outputs={'Y': [square_out]})
Y
Yu Yang 已提交
606
    return square_out
607 608


Y
Yu Yang 已提交
609
def accuracy(input, label, k=1, correct=None, total=None, **kwargs):
610 611 612 613
    """
    This function computes the accuracy using the input and label.
    The output is the top_k inputs and their indices.
    """
F
fengjiayi 已提交
614
    helper = LayerHelper("accuracy", **kwargs)
F
fengjiayi 已提交
615
    topk_out = helper.create_tmp_variable(dtype=input.dtype)
F
fengjiayi 已提交
616 617 618 619 620 621 622
    topk_indices = helper.create_tmp_variable(dtype="int64")
    helper.append_op(
        type="top_k",
        inputs={"X": [input]},
        outputs={"Out": [topk_out],
                 "Indices": [topk_indices]},
        attrs={"k": k})
D
Dong Zhihong 已提交
623
    acc_out = helper.create_tmp_variable(dtype="float32")
Y
Yu Yang 已提交
624 625 626 627
    if correct is None:
        correct = helper.create_tmp_variable(dtype="int64")
    if total is None:
        total = helper.create_tmp_variable(dtype="int64")
F
fengjiayi 已提交
628 629
    helper.append_op(
        type="accuracy",
武毅 已提交
630 631 632 633 634
        inputs={
            "Out": [topk_out],
            "Indices": [topk_indices],
            "Label": [label]
        },
D
Dong Zhihong 已提交
635 636 637 638 639
        outputs={
            "Accuracy": [acc_out],
            "Correct": [correct],
            "Total": [total],
        })
F
fengjiayi 已提交
640 641 642
    return acc_out


D
dzhwinter 已提交
643 644 645
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
646
                  filter_stride=1,
D
dzhwinter 已提交
647 648
                  padding=None,
                  bias_attr=None,
649
                  bias_initializer=None,
D
dzhwinter 已提交
650
                  param_attr=None,
651
                  param_initializer=None,
C
chengduoZH 已提交
652
                  act=None,
653 654
                  main_program=None,
                  startup_program=None):
655 656 657 658 659
    """
    This function creates the op for sequence_conv, using the inputs and
    other convolutional configurations for the filters and stride as given
    in the input parameters to the function.
    """
660 661

    def _get_default_bias_initializer():
662
        return Constant()
663 664

    def _get_default_param_initializer():
665
        return Xavier()
666

D
dzhwinter 已提交
667 668 669 670 671 672 673
    # FIXME(dzh) : want to unify the argument of python layer
    # function. So we ignore some unecessary attributes.
    # such as, padding_trainable, context_start.

    helper = LayerHelper('sequence_conv', **locals())
    dtype = helper.input_dtype()

674 675 676 677 678
    if param_initializer is None:
        param_initializer = _get_default_param_initializer()
    if bias_initializer is None:
        bias_initializer = _get_default_bias_initializer()

D
dzhwinter 已提交
679
    filter_shape = [filter_size * input.shape[1], num_filters]
D
dzhwinter 已提交
680
    filter = helper.create_parameter(
681 682 683 684
        attr=helper.param_attr,
        shape=filter_shape,
        dtype=dtype,
        initializer=param_initializer)
D
dzhwinter 已提交
685 686 687 688 689 690
    pre_bias = helper.create_tmp_variable(dtype)

    helper.append_op(
        type='sequence_conv',
        inputs={
            'X': [input],
D
dzhwinter 已提交
691
            'Filter': [filter],
D
dzhwinter 已提交
692 693 694
        },
        outputs={"Out": pre_bias},
        attrs={
695
            'contextStride': filter_stride,
696
            'contextStart': -int(filter_size / 2),
697
            'contextLength': filter_size
D
dzhwinter 已提交
698
        })
699
    pre_act = helper.append_bias_op(pre_bias, bias_initializer)
D
dzhwinter 已提交
700 701 702
    return helper.append_activation(pre_act)


F
fengjiayi 已提交
703 704
def conv2d(input,
           num_filters,
C
chengduoZH 已提交
705
           filter_size,
F
fengjiayi 已提交
706 707
           stride=[1, 1],
           padding=None,
C
chengduoZH 已提交
708
           groups=None,
F
fengjiayi 已提交
709
           param_attr=None,
710
           param_initializer=None,
C
chengduoZH 已提交
711 712 713 714
           bias_attr=None,
           bias_initializer=None,
           act=None,
           name=None,
715 716
           main_program=None,
           startup_program=None):
717 718 719 720 721 722 723
    """
    This function creates the op for a 2-dimensional Convolution.
    This is performed using the parameters of filters(size, dimensionality etc)
    , stride and other configurations for a Convolution operation.
    This funciton can also append an activation on top of the
    conv-2d output, if mentioned in the input parameters.
    """
724 725

    def _get_default_bias_initializer():
726
        return Constant()
727 728 729

    def _get_default_param_initializer(filter_size, num_channels):
        std = (2.0 / (filter_size[0]**2 * num_channels))**0.5
730
        return Normal(0.0, std, 0)
731

732 733 734 735 736 737 738
    helper = LayerHelper('conv2d', **locals())
    dtype = helper.input_dtype()

    num_channels = input.shape[1]
    if groups is None:
        num_filter_channels = num_channels
    else:
C
chengduoZH 已提交
739
        if num_channels % groups != 0:
740 741 742
            raise ValueError("num_channels must be divisible by groups.")
        num_filter_channels = num_channels / groups

F
fengjiayi 已提交
743 744 745 746 747 748 749
    if isinstance(filter_size, int):
        filter_size = [filter_size, filter_size]
    if isinstance(stride, int):
        stride = [stride, stride]
    if isinstance(padding, int):
        padding = [padding, padding]

750 751
    input_shape = input.shape
    filter_shape = [num_filters, num_filter_channels] + filter_size
752

753 754 755 756 757 758
    if param_initializer is None:
        param_initializer = _get_default_param_initializer(filter_size,
                                                           num_channels)
    if bias_initializer is None:
        bias_initializer = _get_default_bias_initializer()

759
    filter = helper.create_parameter(
760 761 762
        attr=helper.param_attr,
        shape=filter_shape,
        dtype=dtype,
763
        initializer=param_initializer)
764 765 766 767 768 769 770 771 772 773 774 775 776
    pre_bias = helper.create_tmp_variable(dtype)

    helper.append_op(
        type='conv2d',
        inputs={
            'Input': input,
            'Filter': filter,
        },
        outputs={"Output": pre_bias},
        attrs={'strides': stride,
               'paddings': padding,
               'groups': groups})

777 778
    pre_act = helper.append_bias_op(
        pre_bias, bias_initializer, dim_start=1, dim_end=2)
779 780

    return helper.append_activation(pre_act)
F
fengjiayi 已提交
781 782


D
dzhwinter 已提交
783
def sequence_pool(input, pool_type, **kwargs):
784 785 786 787 788
    """
    This function add the operator for sequence pooling.
    This is applied on top of the input using pool_type mentioned
    in the parameters.
    """
789
    helper = LayerHelper('sequence_pool', input=input, **kwargs)
D
dzhwinter 已提交
790 791
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)
D
dangqingqing 已提交
792
    max_index = helper.create_tmp_variable(dtype)
D
dzhwinter 已提交
793 794 795

    helper.append_op(
        type="sequence_pool",
D
dangqingqing 已提交
796 797 798
        inputs={"X": input},
        outputs={"Out": pool_out,
                 "MaxIndex": max_index},
D
dzhwinter 已提交
799
        attrs={"pooltype": pool_type.upper()})
D
dzhwinter 已提交
800 801 802 803

    return pool_out


F
fengjiayi 已提交
804 805 806 807 808 809
def pool2d(input,
           pool_size,
           pool_type,
           pool_stride=[1, 1],
           pool_padding=[0, 0],
           global_pooling=False,
810 811
           main_program=None,
           startup_program=None):
812 813 814 815
    """
    This function adds the operator for pooling in 2 dimensions, using the
    pooling configurations mentioned in input parameters.
    """
F
fengjiayi 已提交
816 817 818 819 820 821 822 823 824 825 826
    if pool_type not in ["max", "avg"]:
        raise ValueError(
            "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",
            str(pool_type))
    if isinstance(pool_size, int):
        pool_size = [pool_size, pool_size]
    if isinstance(pool_stride, int):
        pool_stride = [pool_stride, pool_stride]
    if isinstance(pool_padding, int):
        pool_padding = [pool_padding, pool_padding]

D
dzhwinter 已提交
827
    helper = LayerHelper('pool2d', **locals())
F
fengjiayi 已提交
828 829 830 831 832 833 834 835
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

    helper.append_op(
        type="pool2d",
        inputs={"X": input},
        outputs={"Out": pool_out},
        attrs={
C
chengduoZH 已提交
836
            "pooling_type": pool_type,
F
fengjiayi 已提交
837
            "ksize": pool_size,
C
chengduoZH 已提交
838
            "global_pooling": global_pooling,
F
fengjiayi 已提交
839 840 841 842 843
            "strides": pool_stride,
            "paddings": pool_padding
        })

    return pool_out
Y
Yu Yang 已提交
844 845


Q
Qiao Longfei 已提交
846 847 848 849
def batch_norm(input,
               act=None,
               is_test=False,
               momentum=0.9,
850
               epsilon=1e-05,
Q
Qiao Longfei 已提交
851 852 853
               param_attr=None,
               bias_attr=None,
               data_layout='NCHW',
854 855
               main_program=None,
               startup_program=None):
856 857 858 859
    """
    This function helps create an operator to implement
    the BatchNorm layer using the configurations from the input parameters.
    """
Q
Qiao Longfei 已提交
860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875
    helper = LayerHelper('batch_norm', **locals())
    dtype = helper.input_dtype()

    input_shape = input.shape
    if data_layout == 'NCHW':
        channel_num = input_shape[1]
    else:
        if data_layout == 'NHWC':
            channel_num = input_shape[-1]
        else:
            raise ValueError("unsupported data layout:" + data_layout)

    param_shape = [channel_num]

    # create parameter
    scale = helper.create_parameter(
876 877 878
        attr=helper.param_attr,
        shape=param_shape,
        dtype=dtype,
879
        initializer=Constant(1.0))
Q
Qiao Longfei 已提交
880
    bias = helper.create_parameter(
881 882 883
        attr=helper.param_attr,
        shape=param_shape,
        dtype=dtype,
884
        initializer=Constant(0.0))
885 886

    mean = helper.create_global_variable(
F
fengjiayi 已提交
887
        dtype=input.dtype, shape=param_shape, persistable=True)
888
    helper.set_variable_initializer(var=mean, initializer=Constant(0.0))
889 890

    variance = helper.create_global_variable(
F
fengjiayi 已提交
891
        dtype=input.dtype, shape=param_shape, persistable=True)
892
    helper.set_variable_initializer(var=variance, initializer=Constant(1.0))
Q
Qiao Longfei 已提交
893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926

    # create output
    # mean and mean_out share the same memory
    mean_out = mean
    # variance and variance out share the same memory
    variance_out = variance
    saved_mean = helper.create_tmp_variable(dtype)
    saved_variance = helper.create_tmp_variable(dtype)

    batch_norm_out = helper.create_tmp_variable(dtype)

    helper.append_op(
        type="batch_norm",
        inputs={
            "X": input,
            "Scale": scale,
            "Bias": bias,
            "Mean": mean,
            "Variance": variance
        },
        outputs={
            "Y": batch_norm_out,
            "MeanOut": mean_out,
            "VarianceOut": variance_out,
            "SavedMean": saved_mean,
            "SavedVariance": saved_variance
        },
        attrs={"momentum": momentum,
               "epsilon": epsilon,
               "is_test": is_test})

    return helper.append_activation(batch_norm_out)


927 928
def beam_search_decode(ids, scores, main_program=None, startup_program=None):
    helper = LayerHelper('beam_search_decode', **locals())
F
fengjiayi 已提交
929 930
    sentence_ids = helper.create_tmp_variable(dtype=ids.dtype)
    sentence_scores = helper.create_tmp_variable(dtype=ids.dtype)
931 932 933 934 935 936 937 938 939 940 941 942 943

    helper.append_op(
        type="beam_search_decode",
        inputs={"Ids": ids,
                "Scores": scores},
        outputs={
            "SentenceIds": sentence_ids,
            "SentenceScores": sentence_scores
        })

    return sentence_ids, sentence_scores


Y
Yu Yang 已提交
944 945
class BlockGuard(object):
    """
946 947 948 949
    BlockGuard class.

    BlockGuard class is used to create a sub-block in a program by
    using the Python `with` keyword.
Y
Yu Yang 已提交
950 951
    """

952 953
    def __init__(self, main_program):
        if not isinstance(main_program, Program):
Y
Yu Yang 已提交
954
            raise TypeError("BlockGuard takes a program")
955
        self.main_program = main_program
Y
Yu Yang 已提交
956 957

    def __enter__(self):
958
        self.main_program.create_block()
Y
Yu Yang 已提交
959 960

    def __exit__(self, exc_type, exc_val, exc_tb):
961
        self.main_program.rollback()
Y
Yu Yang 已提交
962 963 964 965 966 967
        if exc_type is not None:
            return False  # re-raise exception
        return True


class StaticRNNGuard(BlockGuard):
968 969 970 971 972 973
    """
    StaticRNNGuard class.

    StaticRNNGuard class is used to create a StaticRNN block in a program.
    """

Y
Yu Yang 已提交
974 975
    def __init__(self, rnn):
        if not isinstance(rnn, StaticRNN):
Y
Yang Yang(Tony) 已提交
976
            raise TypeError("StaticRNNGuard takes a StaticRNN")
977
        super(StaticRNNGuard, self).__init__(rnn.helper.main_program)
Y
Yu Yang 已提交
978 979 980 981 982 983 984
        self.rnn = rnn

    def __enter__(self):
        self.rnn.status = StaticRNN.IN_RNN_BLOCK
        return super(StaticRNNGuard, self).__enter__()

    def __exit__(self, exc_type, exc_val, exc_tb):
Y
Yu Yang 已提交
985 986
        if exc_type is not None:
            return False
Y
Yu Yang 已提交
987 988 989 990 991 992 993
        self.rnn.status = StaticRNN.AFTER_RNN_BLOCK
        self.rnn.complete_rnn_op()
        return super(StaticRNNGuard, self).__exit__(exc_type, exc_val, exc_tb)


class StaticRNNMemoryLink(object):
    """
994 995 996 997 998 999 1000 1001 1002 1003 1004 1005
    StaticRNNMemoryLink class.

    Args:
        init: the initial variable for Memory
        init: Variable
        pre_mem: the memory variable in previous time step
        pre_mem: Variable
        mem: the memory variable in current time step
        mem: Variable

    StaticRNNMemoryLink class is used to create a link between two
    memory cells of a StaticRNN.
Y
Yu Yang 已提交
1006 1007 1008 1009 1010 1011 1012 1013 1014
    """

    def __init__(self, init, pre_mem, mem=None):
        self.init = init
        self.pre_mem = pre_mem
        self.mem = mem


class StaticRNN(object):
1015 1016 1017 1018 1019 1020
    """
    StaticRNN class.

    StaticRNN class is used to create a StaticRNN. The RNN will have its
    own parameters like inputs, outputs, memories, status and length.
    """
Y
Yu Yang 已提交
1021 1022 1023 1024
    BEFORE_RNN_BLOCK = 0
    IN_RNN_BLOCK = 1
    AFTER_RNN_BLOCK = 2

1025 1026 1027
    def __init__(self, name=None, main_program=None):
        self.helper = LayerHelper(
            "static_rnn", name=name, main_program=main_program)
Y
Yu Yang 已提交
1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041
        self.memories = {}  # memory map, from pre_mem.name --> MemoryLink
        self.inputs = []  # input variable list in current block
        self.outputs = []  # output variable list in parent block
        self.status = StaticRNN.BEFORE_RNN_BLOCK  # status flag.
        # sequence length, since it is a static RNN, sequence length are fixed.
        self.seq_len = None

    def step(self):
        return StaticRNNGuard(self)

    def _assert_in_rnn_block_(self, method):
        if self.status != StaticRNN.IN_RNN_BLOCK:
            raise ValueError("You must invoke {0} in rnn block".format(method))

1042 1043 1044 1045 1046 1047 1048
    def memory(self,
               init=None,
               shape=None,
               batch_ref=None,
               init_value=0.0,
               init_batch_dim_idx=0,
               ref_batch_dim_idx=1):
1049 1050 1051 1052 1053 1054 1055 1056 1057
        """
        Args:
            init: boot memory, if not set, a shape, batch_ref must be provided
            shape: shape of the boot memory
            batch_ref: batch size reference variable
            init_value: the init value of boot memory
            init_batch_dim_idx: the index of batch size in init's dimension
            ref_batch_dim_idx: the index of batch size in batch_ref's dimension
        """
Y
Yu Yang 已提交
1058 1059
        self._assert_in_rnn_block_('memory')
        if init is None:
1060
            if shape is None or batch_ref is None:
Y
Yu Yang 已提交
1061
                raise ValueError(
1062
                    "if init is None, memory at least need shape and batch_ref")
Y
Yu Yang 已提交
1063 1064 1065
            parent_block = self.parent_block()
            var_name = unique_name("@".join([self.helper.name, "memory_boot"]))
            boot_var = parent_block.create_var(
1066 1067
                name=var_name,
                shape=shape,
F
fengjiayi 已提交
1068
                dtype=batch_ref.dtype,
1069
                persistable=False)
Y
Yu Yang 已提交
1070 1071

            parent_block.append_op(
1072 1073
                type="fill_constant_batch_size_like",
                inputs={'Input': [batch_ref]},
Y
Yu Yang 已提交
1074 1075 1076
                outputs={'Out': [boot_var]},
                attrs={
                    'value': init_value,
1077
                    'shape': boot_var.shape,
F
fengjiayi 已提交
1078
                    'dtype': boot_var.dtype,
1079 1080
                    'input_dim_idx': ref_batch_dim_idx,
                    'output_dim_idx': init_batch_dim_idx
Y
Yu Yang 已提交
1081 1082 1083 1084 1085 1086
                })

            return self.memory(init=boot_var)
        else:
            pre_mem = self.helper.create_variable(
                name=unique_name("@".join([self.helper.name, "mem"])),
F
fengjiayi 已提交
1087
                dtype=init.dtype,
Y
Yu Yang 已提交
1088 1089 1090 1091 1092 1093 1094 1095 1096 1097
                shape=init.shape)
            self.memories[pre_mem.name] = StaticRNNMemoryLink(
                init=init, pre_mem=pre_mem)
            return pre_mem

    def step_input(self, x):
        self._assert_in_rnn_block_('step_input')
        if not isinstance(x, Variable):
            raise TypeError("step input takes a Variable")
        if self.seq_len is None:
Y
Yu Yang 已提交
1098 1099
            self.seq_len = x.shape[0]
        elif self.seq_len != x.shape[0]:
Y
Yu Yang 已提交
1100 1101 1102
            raise ValueError("Static RNN only take fix seq_len input")

        ipt = self.helper.create_variable(
F
fengjiayi 已提交
1103
            name=x.name, dtype=x.dtype, shape=list(x.shape[1:]), type=x.type)
Y
Yu Yang 已提交
1104 1105 1106 1107 1108 1109 1110 1111
        self.inputs.append(ipt)
        return ipt

    def step_output(self, o):
        self._assert_in_rnn_block_('step_output')
        if not isinstance(o, Variable):
            raise TypeError("step output takes a Variable")

F
fengjiayi 已提交
1112
        tmp_o = self.helper.create_tmp_variable(dtype=o.dtype)
Y
Yu Yang 已提交
1113 1114 1115 1116
        self.helper.append_op(
            type='rnn_memory_helper',
            inputs={'X': [o]},
            outputs={'Out': tmp_o},
F
fengjiayi 已提交
1117
            attrs={'dtype': o.dtype})
Y
Yu Yang 已提交
1118

Y
Yu Yang 已提交
1119
        out_var = self.parent_block().create_var(
Y
Yu Yang 已提交
1120 1121
            name=tmp_o.name,
            shape=[self.seq_len] + list(tmp_o.shape),
F
fengjiayi 已提交
1122
            dtype=tmp_o.dtype)
Y
Yu Yang 已提交
1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135

        self.outputs.append(out_var)

    def output(self, *outputs):
        for each in outputs:
            self.step_output(each)

    def update_memory(self, mem, var):
        if not isinstance(mem, Variable) or not isinstance(var, Variable):
            raise TypeError("update memory should take variables")
        self.memories[mem.name].mem = var

    def parent_block(self):
1136
        prog = self.helper.main_program
Y
Yu Yang 已提交
1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152
        parent_idx = prog.current_block().parent_idx
        assert parent_idx >= 0
        parent_block = prog.block(parent_idx)
        return parent_block

    def __call__(self, *args, **kwargs):
        if self.status != StaticRNN.AFTER_RNN_BLOCK:
            raise ValueError("RNN output can only be retrieved after rnn block")
        if len(self.outputs) == 0:
            raise ValueError("RNN has no output")
        elif len(self.outputs) == 1:
            return self.outputs[0]
        else:
            return self.outputs

    def complete_rnn_op(self):
1153 1154
        main_program = self.helper.main_program
        rnn_block = main_program.current_block()
Y
Yu Yang 已提交
1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193
        parent_block = self.parent_block()

        local_inputs = set()

        for op in rnn_block.ops:
            assert isinstance(op, Operator)
            for oname in op.output_names:
                for out_var_name in op.output(oname):
                    local_inputs.add(out_var_name)

        for var in self.inputs:
            local_inputs.add(var.name)
        for m in self.memories:
            local_inputs.add(m)

        params = list()
        for op in rnn_block.ops:
            assert isinstance(op, Operator)
            for iname in op.input_names:
                for in_var_name in op.input(iname):
                    if in_var_name not in local_inputs:
                        params.append(in_var_name)

        parameters = [parent_block.var(name) for name in params]

        step_scope = parent_block.create_var(
            type=core.VarDesc.VarType.STEP_SCOPES)

        inlinks = [parent_block.var(i.name) for i in self.inputs]
        outlinks = self.outputs

        boot_memories = []
        pre_memories = []
        memories = []
        for _, mem in self.memories.iteritems():
            boot_memories.append(mem.init)
            pre_memories.append(mem.pre_mem.name)
            mem_var = rnn_block.var(mem.mem.name)
            assert isinstance(mem_var, Variable)
F
fengjiayi 已提交
1194
            new_mem = self.helper.create_tmp_variable(dtype=mem_var.dtype)
Y
Yu Yang 已提交
1195 1196 1197 1198 1199

            rnn_block.append_op(
                type='rnn_memory_helper',
                inputs={'X': [mem_var]},
                outputs={'Out': [new_mem]},
F
fengjiayi 已提交
1200
                attrs={'dtype': mem_var.dtype})
Y
Yu Yang 已提交
1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217

            memories.append(new_mem.name)

        parent_block.append_op(
            type='recurrent',
            inputs={
                'inputs': inlinks,
                'initial_states': boot_memories,
                'parameters': parameters
            },
            outputs={'outputs': outlinks,
                     'step_scopes': [step_scope]},
            attrs={
                'ex_states': pre_memories,
                'states': memories,
                'step_block': rnn_block
            })
Y
Yu Yang 已提交
1218 1219


Y
Yang Yang(Tony) 已提交
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
class WhileGuard(BlockGuard):
    def __init__(self, while_op):
        if not isinstance(while_op, While):
            raise TypeError("WhileGuard takes a while op")
        super(WhileGuard, self).__init__(while_op.helper.main_program)
        self.while_op = while_op

    def __enter__(self):
        self.while_op.status = While.IN_WHILE_BLOCK
        return super(WhileGuard, self).__enter__()

    def __exit__(self, exc_type, exc_val, exc_tb):
        if exc_type is not None:
            return False
        self.while_op.status = While.AFTER_WHILE_BLOCK
        self.while_op.complete()
        return super(WhileGuard, self).__exit__(exc_type, exc_val, exc_tb)


class While(object):
    BEFORE_WHILE_BLOCK = 0
    IN_WHILE_BLOCK = 1
    AFTER_WHILE_BLOCK = 2

    def __init__(self, cond, name=None, main_program=None):
        self.helper = LayerHelper("while", name=name, main_program=main_program)
        self.status = While.BEFORE_WHILE_BLOCK
        if not isinstance(cond, Variable):
            raise TypeError("condition should be a variable")
        assert isinstance(cond, Variable)
F
fengjiayi 已提交
1250
        if cond.dtype != core.DataType.BOOL:
Y
Yang Yang(Tony) 已提交
1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295
            raise TypeError("condition should be a bool variable")
        if reduce(lambda a, b: a * b, cond.shape, 1) != 1:
            raise TypeError("condition should be a bool scalar")
        self.cond_var = cond

    def block(self):
        return WhileGuard(self)

    def complete(self):
        main_program = self.helper.main_program
        while_block = main_program.current_block()
        parent_block = main_program.block(main_program.current_block()
                                          .parent_idx)

        inner_outputs = {self.cond_var.name}
        x_name_list = set()
        for op in while_block.ops:
            for iname in op.input_names:
                for in_var_name in op.input(iname):
                    if in_var_name not in inner_outputs:
                        x_name_list.add(in_var_name)

            for oname in op.output_names:
                for out_var_name in op.output(oname):
                    inner_outputs.add(out_var_name)

        out_vars = []
        for inner_out_name in inner_outputs:
            if inner_out_name in parent_block.vars:
                out_vars.append(parent_block.var(inner_out_name))

        step_scope = parent_block.create_var(
            type=core.VarDesc.VarType.STEP_SCOPES)

        parent_block.append_op(
            type='while',
            inputs={
                'X': [parent_block.var(x_name) for x_name in x_name_list],
                'Condition': [self.cond_var]
            },
            outputs={'Out': out_vars,
                     'StepScopes': [step_scope]},
            attrs={'step_block': while_block})


Y
Yang Yang(Tony) 已提交
1296 1297 1298 1299 1300 1301
def lstm(x,
         c_pre_init,
         hidden_dim,
         forget_bias=None,
         main_program=None,
         startup_program=None):
1302 1303 1304 1305
    """
    This function helps create an operator for the LSTM (Long Short Term
    Memory) cell that can be used inside an RNN.
    """
Y
Yang Yang(Tony) 已提交
1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321
    helper = LayerHelper('lstm_unit', **locals())
    rnn = StaticRNN()
    with rnn.step():
        c_pre = rnn.memory(init=c_pre_init)
        x_t = rnn.step_input(x)

        before_fc = concat(
            input=[x_t, c_pre],
            axis=1,
            main_program=main_program,
            startup_program=startup_program)
        after_fc = fc(input=before_fc,
                      size=hidden_dim * 4,
                      main_program=main_program,
                      startup_program=startup_program)

F
fengjiayi 已提交
1322 1323 1324
        dtype = x.dtype
        c = helper.create_tmp_variable(dtype)
        h = helper.create_tmp_variable(dtype)
Y
Yang Yang(Tony) 已提交
1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339

        helper.append_op(
            type='lstm_unit',
            inputs={"X": after_fc,
                    "C_prev": c_pre},
            outputs={"C": c,
                     "H": h},
            attrs={"forget_bias": forget_bias})

        rnn.update_memory(c_pre, c)
        rnn.output(h)

    return rnn()


1340
def lod_rank_table(x, level=0, main_program=None):
1341 1342 1343 1344
    """
    This function creates an operator for creating a LOD_RANK_TABLE
    using the input x.
    """
Y
Yu Yang 已提交
1345 1346 1347 1348 1349 1350 1351 1352 1353 1354
    helper = LayerHelper("lod_rank_table", **locals())
    table = helper.create_variable(
        type=core.VarDesc.VarType.LOD_RANK_TABLE,
        name=unique_name("lod_rank_table"))
    helper.append_op(
        type='lod_rank_table',
        inputs={'X': x},
        outputs={'Out': table},
        attrs={'level': level})
    return table
Y
Yu Yang 已提交
1355 1356


F
fengjiayi 已提交
1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370
def max_sequence_len(rank_table, main_program=None):
    """
    This function creates an operator to calculate the length of 
    max seqence through input rank_table(should be a lod_rank_table)
    """
    helper = LayerHelper("max_seqence_len", **locals())
    res = helper.create_tmp_variable(dtype="int64")
    helper.append_op(
        type="max_sequence_len",
        inputs={"RankTable": rank_table},
        outputs={"Out": res})
    return res


Y
Yu Yang 已提交
1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383
def topk(input, k, main_program=None, startup_program=None):
    helper = LayerHelper('topk', **locals())
    topk_out = helper.create_tmp_variable(dtype=input.data_type)
    topk_indices = helper.create_tmp_variable(dtype='int64')
    helper.append_op(
        type='top_k',
        inputs={'X': [input]},
        outputs={'Out': [topk_out],
                 'Indices': [topk_indices]},
        attrs={'k': k})
    return topk_out, topk_indices


1384
def lod_tensor_to_array(x, table, main_program=None):
1385 1386 1387 1388
    """
    This function creates an operator to convert an LOD_Tensor to
    an array.
    """
1389 1390 1391
    helper = LayerHelper("lod_tensor_to_array", **locals())
    array = helper.create_variable(
        name=unique_name("lod_tensor_to_array"),
1392
        type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
F
fengjiayi 已提交
1393
        dtype=x.dtype)
1394 1395 1396 1397 1398 1399 1400 1401 1402
    helper.append_op(
        type='lod_tensor_to_array',
        inputs={'X': x,
                'RankTable': table},
        outputs={'Out': array})
    return array


def array_to_lod_tensor(x, table, main_program=None):
1403 1404 1405 1406
    """
    This function creates an operator to convert an array to a
    LOD_Tensor.
    """
1407
    helper = LayerHelper("array_to_lod_tensor", **locals())
F
fengjiayi 已提交
1408
    tmp = helper.create_tmp_variable(dtype=x.dtype)
1409 1410 1411 1412 1413 1414 1415 1416
    helper.append_op(
        type="array_to_lod_tensor",
        inputs={'X': x,
                'RankTable': table},
        outputs={'Out': tmp})
    return tmp


Y
Yu Yang 已提交
1417 1418 1419 1420 1421 1422
def fill_constant(shape,
                  dtype,
                  value,
                  out=None,
                  main_program=None,
                  startup_program=None):
1423 1424
    """
    This function creates a tensor , with shape as mentioned in the input and
F
fengjiayi 已提交
1425
    specified dtype and fills this up with a constant value that
1426 1427
    comes in the input. It also sets the stop_gradient to be True.
    """
Y
Yang Yu 已提交
1428
    helper = LayerHelper("fill_constant", **locals())
Y
Yu Yang 已提交
1429 1430
    if out is None:
        out = helper.create_tmp_variable(dtype=dtype)
Y
Yu Yang 已提交
1431 1432 1433 1434
    helper.append_op(
        type='fill_constant',
        inputs={},
        outputs={'Out': [out]},
F
fengjiayi 已提交
1435 1436 1437
        attrs={'shape': shape,
               'dtype': out.dtype,
               'value': float(value)})
Y
Yu Yang 已提交
1438 1439 1440 1441
    out.stop_gradient = True
    return out


Y
Yu Yang 已提交
1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457
def fill_constant_batch_size_like(input,
                                  shape,
                                  dtype,
                                  value,
                                  input_dim_idx=0,
                                  output_dim_idx=0,
                                  main_program=None,
                                  startup_program=None):
    helper = LayerHelper("fill_constant_batch_size_like", **locals())
    out = helper.create_tmp_variable(dtype=dtype)
    helper.append_op(
        type='fill_constant_batch_size_like',
        inputs={'Input': input},
        outputs={'Out': [out]},
        attrs={
            'shape': shape,
F
fengjiayi 已提交
1458
            'dtype': out.dtype,
Y
Yu Yang 已提交
1459 1460 1461 1462 1463 1464 1465 1466
            'value': float(value),
            'input_dim_idx': input_dim_idx,
            'output_dim_idx': output_dim_idx
        })
    out.stop_gradient = True
    return out


Y
Yu Yang 已提交
1467
def ones(shape, dtype, main_program=None):
1468 1469 1470 1471
    """
    This function performs the same function as fill_constant() declared above
    with the constant value being 1.0.
    """
Y
Yu Yang 已提交
1472 1473 1474 1475
    return fill_constant(value=1.0, **locals())


def zeros(shape, dtype, main_program=None):
1476 1477 1478 1479
    """
    This function performs the same function as fill_constant() declared above
    with the constant value being 0.0.
    """
Y
Yu Yang 已提交
1480 1481 1482
    return fill_constant(value=0.0, **locals())


1483
def increment(x, value=1.0, in_place=True, main_program=None):
1484 1485 1486 1487 1488
    """
    This function creates an operator to increment each value in the input
    `x` by an amount: `value` as mentioned in the input parameter. This
    operation is performed in-place by default.
    """
Y
Yu Yang 已提交
1489
    helper = LayerHelper("increment", **locals())
Y
Yang Yang(Tony) 已提交
1490
    if not in_place:
F
fengjiayi 已提交
1491
        out = helper.create_tmp_variable(dtype=x.dtype)
Y
Yang Yang(Tony) 已提交
1492 1493
    else:
        out = x
Y
Yu Yang 已提交
1494 1495 1496
    helper.append_op(
        type='increment',
        inputs={'X': [x]},
Y
Yang Yu 已提交
1497
        outputs={'Out': [out]},
Y
Yu Yang 已提交
1498
        attrs={'step': value})
Y
Yang Yu 已提交
1499
    return out
Y
Yu Yang 已提交
1500 1501 1502


def array_write(x, i, array=None, main_program=None):
1503 1504 1505 1506
    """
    This function creates an operator to write the data out as a
    LOD_TENSOR_ARRAY.
    """
Y
Yu Yang 已提交
1507 1508 1509 1510 1511
    helper = LayerHelper('array_write', **locals())
    if array is None:
        array = helper.create_variable(
            name="{0}.out".format(helper.name),
            type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
F
fengjiayi 已提交
1512
            dtype=x.dtype)
Y
Yu Yang 已提交
1513 1514 1515 1516 1517 1518 1519 1520
    helper.append_op(
        type='write_to_array',
        inputs={'X': [x],
                'I': [i]},
        outputs={'Out': [array]})
    return array


Y
Yang Yang(Tony) 已提交
1521 1522 1523 1524 1525 1526 1527 1528
def create_array(dtype, main_program=None):
    helper = LayerHelper("array", **locals())
    return helper.create_variable(
        name="{0}.out".format(helper.name),
        type=core.VarDesc.VarType.LOD_TENSOR_ARRAY,
        dtype=dtype)


Y
Yu Yang 已提交
1529
def less_than(x, y, cond=None, main_program=None, **ignored):
Y
Yang Yang(Tony) 已提交
1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540
    helper = LayerHelper("less_than", **locals())
    if cond is None:
        cond = helper.create_tmp_variable(dtype='bool')
        cond.stop_gradient = True

    helper.append_op(
        type='less_than', inputs={'X': [x],
                                  'Y': [y]}, outputs={'Out': [cond]})
    return cond


Y
Yu Yang 已提交
1541
def array_read(array, i, main_program=None):
1542 1543 1544 1545
    """
    This function creates an operator to read the data in as a
    LOD_TENSOR_ARRAY.
    """
Y
Yu Yang 已提交
1546 1547 1548 1549 1550
    helper = LayerHelper('array_read', **locals())
    if not isinstance(
            array,
            Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY:
        raise TypeError("array should be tensor array vairable")
F
fengjiayi 已提交
1551
    out = helper.create_tmp_variable(dtype=array.dtype)
Y
Yu Yang 已提交
1552 1553 1554 1555 1556 1557
    helper.append_op(
        type='read_from_array',
        inputs={'X': [array],
                'I': [i]},
        outputs={'Out': [out]})
    return out
Y
Yang Yu 已提交
1558 1559 1560


def shrink_memory(x, i, table, main_program=None):
1561 1562 1563 1564
    """
    This function creates an operator to shrink_rnn_memory using the RankTable
    as mentioned in the input parameter.
    """
Y
Yang Yu 已提交
1565
    helper = LayerHelper('shrink_memory', **locals())
F
fengjiayi 已提交
1566
    out = helper.create_tmp_variable(dtype=x.dtype)
Y
Yang Yu 已提交
1567
    helper.append_op(
Y
Yang Yu 已提交
1568
        type='shrink_rnn_memory',
Y
Yang Yu 已提交
1569 1570 1571 1572 1573 1574
        inputs={'X': [x],
                'I': [i],
                'RankTable': [table]},
        outputs={'Out': [out]},
        attrs={})
    return out
Y
Yang Yu 已提交
1575 1576 1577


def array_length(array, main_program=None):
1578 1579 1580 1581
    """
    This function creates an operator to find the length of the
    LOD_TENSOR_ARRAY.
    """
Y
Yang Yu 已提交
1582 1583 1584 1585 1586 1587
    helper = LayerHelper('array_length', **locals())
    tmp = helper.create_tmp_variable(dtype='int64')
    tmp.stop_gradient = True
    helper.append_op(
        type='lod_array_length', inputs={'X': [array]}, outputs={'Out': [tmp]})
    return tmp
Y
Yu Yang 已提交
1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606


class ConditionalBlockGuard(BlockGuard):
    def __init__(self, block):
        if not isinstance(block, ConditionalBlock):
            raise TypeError("block should be conditional block")
        super(ConditionalBlockGuard, self).__init__(block.helper.main_program)
        self.block = block

    def __enter__(self):
        return super(ConditionalBlockGuard, self).__enter__()

    def __exit__(self, exc_type, exc_val, exc_tb):
        self.block.complete()
        return super(ConditionalBlockGuard, self).__exit__(exc_type, exc_val,
                                                           exc_tb)


class ConditionalBlock(object):
Y
Yu Yang 已提交
1607 1608 1609 1610 1611
    def __init__(self,
                 inputs,
                 name=None,
                 main_program=None,
                 startup_program=None):
Y
Yu Yang 已提交
1612 1613 1614 1615 1616
        for each_input in inputs:
            if not isinstance(each_input, Variable):
                raise TypeError("Each input should be variable")
        self.inputs = inputs
        self.helper = LayerHelper(
Y
Yu Yang 已提交
1617 1618 1619 1620
            'conditional_block',
            name=name,
            main_program=main_program,
            startup_program=startup_program)
Y
Yu Yang 已提交
1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664

    def block(self):
        return ConditionalBlockGuard(self)

    def complete(self):
        inside_block = self.helper.main_program.current_block()
        parent_block = self.helper.main_program.block(inside_block.parent_idx)

        intermediate = set()
        params = set()

        for each_op in inside_block.ops:
            assert isinstance(each_op, Operator)
            for iname in each_op.input_names:
                for in_var_name in each_op.input(iname):
                    if in_var_name not in intermediate:
                        params.add(in_var_name)

            for oname in each_op.output_names:
                for out_var_name in each_op.output(oname):
                    intermediate.add(out_var_name)
        input_set = set([ipt.name for ipt in self.inputs])

        param_list = [
            parent_block.var(each_name) for each_name in params
            if each_name not in input_set
        ]

        out_list = [
            parent_block.var(var_name) for var_name in parent_block.vars
            if var_name not in intermediate
        ]

        step_scope = parent_block.create_var(
            type=core.VarDesc.VarType.STEP_SCOPES)
        parent_block.append_op(
            type='conditional_block',
            inputs={
                'X': self.inputs,
                'Params': param_list,
            },
            outputs={'Out': out_list,
                     'Scope': [step_scope]},
            attrs={'block': inside_block})
Y
Yu Yang 已提交
1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727


class IfElseBlockGuard(object):
    def __init__(self, is_true, ifelse):
        if not isinstance(ifelse, IfElse):
            raise TypeError("ifelse must be an instance of IfElse class")

        if ifelse.status != IfElse.OUT_IF_ELSE_BLOCKS:
            raise ValueError("You cannot invoke IfElse.block() inside a block")

        self.is_true = is_true
        self.ie = ifelse
        if is_true:
            self.cond_block = ifelse.conditional_true_block
        else:
            self.cond_block = ifelse.conditional_false_block

        if not isinstance(self.cond_block, ConditionalBlock):
            raise TypeError("Unexpected situation")

        self.cond_block = self.cond_block.block()

    def __enter__(self):
        self.ie.status = IfElse.IN_IF_ELSE_TRUE_BLOCKS if self.is_true else IfElse.IN_IF_ELSE_FALSE_BLOCKS
        self.cond_block.__enter__()

    def __exit__(self, exc_type, exc_val, exc_tb):
        if not self.cond_block.__exit__(exc_type, exc_val, exc_tb):
            # re-raise inside exception
            return False
        if len(self.ie.output_table[1 if self.is_true else 0]) == 0:
            raise ValueError("Must set output inside block")
        self.ie.status = IfElse.OUT_IF_ELSE_BLOCKS


class IfElse(object):
    OUT_IF_ELSE_BLOCKS = 0
    IN_IF_ELSE_TRUE_BLOCKS = 1
    IN_IF_ELSE_FALSE_BLOCKS = 2

    def __init__(self, cond, name=None, main_program=None,
                 startup_program=None):
        if not isinstance(cond, Variable):
            raise TypeError("cond must be a Variable")
        self.helper = LayerHelper(
            'ifelse',
            name=name,
            main_program=main_program,
            startup_program=startup_program)
        self.cond = cond
        self.input_table = {}
        self.status = IfElse.OUT_IF_ELSE_BLOCKS
        self.conditional_true_block = ConditionalBlock(inputs=[self.cond])
        self.conditional_false_block = ConditionalBlock(inputs=[self.cond])
        self.output_table = ([], [])  # (true_outs, false_outs)

    def input(self, x):
        if self.status == IfElse.OUT_IF_ELSE_BLOCKS:
            raise ValueError("input must in true/false blocks")
        if id(x) not in self.input_table:
            parent_block = self.parent_block()
            out_true = parent_block.create_var(
                name=unique_name('ifelse_input' + self.helper.name),
F
fengjiayi 已提交
1728
                dtype=x.dtype)
Y
Yu Yang 已提交
1729 1730 1731

            out_false = parent_block.create_var(
                name=unique_name('ifelse_input' + self.helper.name),
F
fengjiayi 已提交
1732
                dtype=x.dtype)
Y
Yu Yang 已提交
1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773
            parent_block.append_op(
                type='split_lod_tensor',
                inputs={
                    'X': x,
                    'Mask': self.cond,
                },
                outputs={'OutTrue': out_true,
                         'OutFalse': out_false},
                attrs={'level': 0})
            self.input_table[id(x)] = (out_true, out_false)
        else:
            out_true, out_false = self.input_table[id(x)]

        if self.status == IfElse.IN_IF_ELSE_TRUE_BLOCKS:
            return out_true
        else:
            return out_false

    def parent_block(self):
        current_block = self.helper.main_program.current_block()
        return self.helper.main_program.block(current_block.parent_idx)

    def true_block(self):
        return IfElseBlockGuard(True, self)

    def false_block(self):
        return IfElseBlockGuard(False, self)

    def output(self, *outs):
        if self.status == self.OUT_IF_ELSE_BLOCKS:
            raise ValueError("output can only be invoked in the sub-block")

        out_table = self.output_table[1 if self.status ==
                                      self.IN_IF_ELSE_TRUE_BLOCKS else 0]
        parent_block = self.parent_block()
        for each_out in outs:
            if not isinstance(each_out, Variable):
                raise TypeError("Each output should be a variable")
            # create outside tensor
            outside_out = parent_block.create_var(
                name=unique_name("_".join([self.helper.name, 'output'])),
F
fengjiayi 已提交
1774
                dtype=each_out.dtype)
Y
Yu Yang 已提交
1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809
            out_table.append(outside_out)

            # assign local var to outside
            assign(
                input=each_out,
                output=outside_out,
                main_program=self.helper.main_program,
                startup_program=self.helper.startup_program)

    def __call__(self):
        if self.status != self.OUT_IF_ELSE_BLOCKS:
            raise ValueError("IfElse::__call__ must be out of sub-block")
        false_len, true_len = map(len, self.output_table)
        if false_len == 0 and true_len == 0:
            raise ValueError("Must invoke true_block/false_block before "
                             "__call__")
        elif false_len != true_len and false_len != 0 and true_len != 0:
            raise ValueError("The output side must be same")
        elif false_len == 0 or true_len == 0:
            return self.output_table[0 if false_len != 0 else 1]

        # else none of false_len/true_len is zero
        # merge together
        rlist = []
        for false_var, true_var in zip(*self.output_table):
            rlist.append(
                merge_lod_tensor(
                    in_true=true_var,
                    in_false=false_var,
                    mask=self.cond,
                    x=self.cond,
                    level=0,
                    main_program=self.helper.main_program,
                    startup_program=self.helper.startup_program))
        return rlist