layers.py 20.8 KB
Newer Older
1
from paddle.v2.framework.layer_helper import LayerHelper, unique_name
2
import paddle.v2.framework.core as core
3
from paddle.v2.framework.framework import OpProtoHolder, Variable, Program
4
from paddle.v2.framework.initializer import ConstantInitializer
5 6
import re

7
__all__ = [
8
    'fc', 'data', 'cross_entropy', 'conv2d', 'pool2d', 'embedding', 'concat',
D
dzhwinter 已提交
9 10
    'StaticRNN', 'cast', 'sequence_conv', 'sequence_pool', 'sums', 'cos_sim',
    'batch_norm', 'accuracy'
11
]
12 13


F
fengjiayi 已提交
14 15 16 17 18 19 20
def fc(input,
       size,
       param_attr=None,
       bias_attr=True,
       name=None,
       act=None,
       num_flatten_dims=1,
Q
QI JUN 已提交
21 22
       program=None,
       init_program=None):
23 24 25 26 27 28 29 30 31
    # create helper
    helper = LayerHelper('fc', **locals())

    dtype = helper.input_dtype()

    # mul
    mul_results = []
    for input_var, param_attr in helper.iter_inputs_and_params():
        input_shape = input_var.shape
32 33 34
        param_shape = [
            reduce(lambda a, b: a * b, input_shape[num_flatten_dims:], 1)
        ] + [size]
35

36 37 38 39 40 41 42 43 44 45
        w = helper.create_parameter(
            attr=param_attr, shape=param_shape, dtype=dtype)
        tmp = helper.create_tmp_variable(dtype)
        helper.append_op(
            type="mul",
            inputs={
                "X": input_var,
                "Y": w,
            },
            outputs={"Out": tmp},
46 47
            attrs={'x_num_col_dims': num_flatten_dims,
                   'y_num_col_dims': 1})
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
        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
    pre_activation = helper.append_bias_op(pre_bias)
    # add activation
    return helper.append_activation(pre_activation)


63 64 65
def embedding(input,
              size,
              data_type='float32',
66
              is_sparse=False,
67 68 69 70 71 72 73 74 75 76 77
              param_attr=None,
              program=None,
              init_program=None):
    helper = LayerHelper('embedding', **locals())
    w = helper.create_parameter(
        attr=helper.param_attr, shape=size, dtype=data_type)
    tmp = helper.create_tmp_variable(data_type)
    helper.append_op(
        type='lookup_table',
        inputs={'Ids': input,
                'W': w},
78 79
        outputs={'Out': tmp},
        attrs={'is_sparse': is_sparse})
80 81 82
    return tmp


F
fengjiayi 已提交
83 84 85 86
def data(name,
         shape,
         data_type='float32',
         type=core.VarDesc.VarType.LOD_TENSOR,
87
         append_batch_size=True,
Q
QI JUN 已提交
88 89
         program=None,
         init_program=None):
90
    helper = LayerHelper('data', **locals())
91 92
    if append_batch_size:
        shape = [-1] + shape  # append batch size as -1
93 94 95 96 97 98 99 100 101 102 103
    return helper.create_global_variable(
        name=name, shape=shape, dtype=data_type, type=type)


def _convert_(name):
    s1 = re.sub('(.)([A-Z][a-z]+)', r'\1_\2', name)
    return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower()


def _create_op_func_(op_type):
    op_proto = OpProtoHolder.instance().get_op_proto(op_type)
104 105 106 107 108 109
    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:
110
        raise ValueError(
111 112
            "Only one not intermediate output operator can be automatically generated"
        )
113

114
    if not_intermediate_outputs[0].duplicable:
115 116 117
        raise ValueError(
            "Only not duplicable op can be automatically generated")

118 119 120 121 122 123 124 125
    for output in intermediate_outputs:
        if output.duplicable:
            raise ValueError(
                "Only when all intermediate ops are not duplicable, "
                "this op can be automatically generated")

    o_name = not_intermediate_outputs[0].name
    intermediate_output_names = [output.name for output in intermediate_outputs]
126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147

    def func(**kwargs):
        helper = LayerHelper(op_type, **kwargs)
        inputs = dict()
        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:
                    dtype = each.data_type
                elif dtype != each.data_type:
                    raise ValueError(
                        "operator {0} must input same dtype".format(op_type))
            inputs[ipt.name] = val

148
        outputs = dict()
149
        out = helper.create_tmp_variable(dtype=dtype)
150 151 152
        outputs[o_name] = [out]
        for name in intermediate_output_names:
            outputs[name] = [helper.create_tmp_variable(dtype=dtype)]
153
        helper.append_op(
154
            type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs)
Q
Qiao Longfei 已提交
155
        return helper.append_activation(out)
156 157 158 159 160 161 162 163

    func.__name__ = op_type
    globals()[op_type] = func
    global __all__
    __all__.append(op_type)


_create_op_func_('mean')
164
_create_op_func_('mul')
Q
Qiao Longfei 已提交
165
_create_op_func_('elementwise_add')
166
_create_op_func_('dropout')
167
_create_op_func_('reshape')
168 169


Y
Yu Yang 已提交
170 171 172 173 174 175 176 177 178 179 180 181
def cast(x, data_type, program=None):
    helper = LayerHelper('cast', **locals())
    out = helper.create_tmp_variable(dtype=data_type)
    helper.append_op(
        type='cast',
        inputs={'X': [x]},
        outputs={'Out': [out]},
        attrs={'in_data_type': x.data_type,
               'out_data_type': out.data_type})
    return out


182 183
def concat(input, axis, program=None, init_program=None):
    helper = LayerHelper('concat', **locals())
D
dzhwinter 已提交
184
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
185 186 187 188 189 190 191 192
    helper.append_op(
        type='concat',
        inputs={'X': input},
        outputs={'Out': [out]},
        attrs={'axis': axis})
    return out


D
dzhwinter 已提交
193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
def sums(input, program=None, init_program=None):
    helper = LayerHelper('sum', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype())
    helper.append_op(type='sum', inputs={'X': [input]}, outputs={'Out': out})
    return out


def cos_sim(X, Y, program=None, init_program=None):
    helper = LayerHelper('cos_sim', **locals())
    out = helper.create_tmp_variable(dtype=helper.input_dtype("X"))
    xnorm = helper.create_tmp_variable(dtype=helper.input_dtype("X"))
    ynorm = helper.create_tmp_variable(dtype=helper.input_dtype("X"))
    helper.append_op(
        type='cos_sim',
        inputs={'X': [X],
                'Y': [Y]},
        outputs={'Out': [out],
                 'XNorm': [xnorm],
                 'YNorm': [ynorm]})
    return out, xnorm, ynorm


215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
def cross_entropy(input, label, **kwargs):
    helper = LayerHelper('cross_entropy', **kwargs)
    out = helper.create_tmp_variable(dtype=input.data_type)
    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):
    helper = LayerHelper('square_error_cost', **kwargs)
    minus_out = helper.create_tmp_variable(dtype=input.data_type)
    helper.append_op(
        type='elementwise_sub',
        inputs={'X': [input],
                'Y': [label]},
        outputs={'Out': [minus_out]})

    square_out = helper.create_tmp_variable(dtype=input.data_type)
    helper.append_op(
238
        type='square', inputs={'X': [minus_out]}, outputs={'Y': [square_out]})
239
    return square_out
240 241


F
fengjiayi 已提交
242 243 244 245 246 247 248 249 250 251 252 253 254 255
def accuracy(input, label, k=1, **kwargs):
    helper = LayerHelper("accuracy", **kwargs)
    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})
    acc_out_dtype = kwargs.get("out_dtype", "float32")
    acc_out = helper.create_tmp_variable(dtype=acc_out_dtype)
    helper.append_op(
        type="accuracy",
256 257 258 259 260
        inputs={
            "Out": [topk_out],
            "Indices": [topk_indices],
            "Label": [label]
        },
F
fengjiayi 已提交
261 262 263 264
        outputs={"Accuracy": [acc_out]})
    return acc_out


265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
def sequence_conv(input,
                  num_filters,
                  filter_size=3,
                  stride=1,
                  padding=None,
                  bias_attr=None,
                  param_attr=None,
                  program=None,
                  init_program=None):
    # 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()

D
dzhwinter 已提交
281
    filter_shape = [filter_size * input.shape[1], num_filters]
282 283 284 285 286 287 288 289
    filter = helper.create_parameter(
        attr=helper.param_attr, shape=filter_shape, dtype=dtype)
    pre_bias = helper.create_tmp_variable(dtype)

    helper.append_op(
        type='sequence_conv',
        inputs={
            'X': [input],
D
dzhwinter 已提交
290
            'Filter': [filter],
291 292 293 294 295 296 297 298 299 300 301
        },
        outputs={"Out": pre_bias},
        attrs={
            'context_stride': stride,
            'context_start': 0,
            'context_length': filter_size
        })
    pre_act = helper.append_bias_op(pre_bias)
    return helper.append_activation(pre_act)


F
fengjiayi 已提交
302 303 304 305 306 307 308 309 310 311
def conv2d(input,
           num_filters,
           name=None,
           filter_size=[1, 1],
           act=None,
           groups=None,
           stride=[1, 1],
           padding=None,
           bias_attr=None,
           param_attr=None,
Q
QI JUN 已提交
312 313
           program=None,
           init_program=None):
314 315 316 317 318 319 320 321 322 323 324
    helper = LayerHelper('conv2d', **locals())
    dtype = helper.input_dtype()

    num_channels = input.shape[1]
    if groups is None:
        num_filter_channels = num_channels
    else:
        if num_channels % groups is not 0:
            raise ValueError("num_channels must be divisible by groups.")
        num_filter_channels = num_channels / groups

325 326 327 328 329 330 331
    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]

332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351
    input_shape = input.shape
    filter_shape = [num_filters, num_filter_channels] + filter_size
    filter = helper.create_parameter(
        attr=helper.param_attr, shape=filter_shape, dtype=dtype)
    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})

    pre_act = helper.append_bias_op(pre_bias)

    return helper.append_activation(pre_act)
352 353


D
dzhwinter 已提交
354
def sequence_pool(input, pool_type, program=None, init_program=None):
355 356 357
    # FIXME(dzh) : want to unify the argument of python layer
    # function. So we ignore some unecessary attributes

D
dzhwinter 已提交
358 359 360 361 362 363 364 365 366
    ENUM_POOL_TYPE = dict({
        "AVERAGE": 0,
        "SUM": 1,
        "SQRT": 2,
        "MAX": 3,
        "LAST": 4,
        "FIRST": 5
    })
    if pool_type.upper() not in ENUM_POOL_TYPE:
367
        raise ValueError("Unknown pool_type: '%s'. It can only be %s.",
D
dzhwinter 已提交
368
                         str(pool_type), " ".join(ENUM_POOL_TYPE.keys()))
369 370 371 372 373

    helper = LayerHelper('sequence_pool', **locals())
    dtype = helper.input_dtype()
    pool_out = helper.create_tmp_variable(dtype)

D
dzhwinter 已提交
374
    # FIXME(dzh): strategy
375 376 377
    helper.append_op(
        type="sequence_pool",
        inputs={"X": [input]},
D
dzhwinter 已提交
378 379
        outputs={"Out": [pool_out]},
        attrs={"strategy": ENUM_POOL_TYPE[pool_type.upper()]})
380 381 382 383

    return pool_out


384 385 386 387 388 389
def pool2d(input,
           pool_size,
           pool_type,
           pool_stride=[1, 1],
           pool_padding=[0, 0],
           global_pooling=False,
Q
QI JUN 已提交
390 391
           program=None,
           init_program=None):
392 393 394 395 396 397 398 399 400 401 402
    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]

403
    helper = LayerHelper('pool2d', **locals())
404 405 406 407 408 409 410 411
    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
412
            "poolingType": pool_type,
413
            "ksize": pool_size,
C
chengduoZH 已提交
414
            "globalPooling": global_pooling,
415 416 417 418 419
            "strides": pool_stride,
            "paddings": pool_padding
        })

    return pool_out
420 421


422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443
def batch_norm(input,
               act=None,
               is_test=False,
               momentum=0.9,
               epsilon=1e05,
               param_attr=None,
               bias_attr=None,
               data_layout='NCHW',
               program=None,
               init_program=None):
    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)

444
    def create_persistable_var(dtype, shape, initializer=None):
445 446 447
        name = unique_name(".".join([helper.name, "xxxx"]))
        var = init_program.global_block().create_var(
            dtype=dtype, shape=shape, name=name, persistable=True)
448 449
        if initializer is not None:
            initializer(var, var.block)
450 451 452 453 454 455 456 457 458 459 460 461
        return program.global_block().create_var(
            name=name, dtype=dtype, shape=shape, persistable=True)

    param_shape = [channel_num]

    # create parameter
    scale = helper.create_parameter(
        attr=helper.param_attr, shape=param_shape, dtype=dtype)
    bias = helper.create_parameter(
        attr=helper.param_attr, shape=param_shape, dtype=dtype)

    # create input
462 463 464
    mean = create_persistable_var(dtype, param_shape, ConstantInitializer(0.0))
    variance = create_persistable_var(dtype, param_shape,
                                      ConstantInitializer(1.0))
465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498

    # 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)


499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663
class BlockGuard(object):
    """
    BlockGuard used to create sub-block in program by using Python `with` 
    keyword.
    """

    def __init__(self, program):
        if not isinstance(program, Program):
            raise TypeError("BlockGuard takes a program")
        self.program = program

    def __enter__(self):
        self.program.create_block()

    def __exit__(self, exc_type, exc_val, exc_tb):
        self.program.rollback()
        if exc_type is not None:
            return False  # re-raise exception
        return True


class StaticRNNGuard(BlockGuard):
    def __init__(self, rnn):
        if not isinstance(rnn, StaticRNN):
            raise TypeError("StaticRNNGuard takes an StaticRNN")
        super(StaticRNNGuard, self).__init__(rnn.helper.program)
        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):
        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):
    """
    :param init: the initial variable for Memory
    :type init: Variable
    :param pre_mem: the memory variable in previous time step
    :type pre_mem: Variable
    :param mem: the memory variable in current time step
    :type mem: Variable
    """

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


class StaticRNN(object):
    BEFORE_RNN_BLOCK = 0
    IN_RNN_BLOCK = 1
    AFTER_RNN_BLOCK = 2

    def __init__(self, name=None, program=None):
        self.helper = LayerHelper("static_rnn", name=name, program=program)
        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))

    def memory(self, init=None, shape=None, dtype=None, init_value=0):
        self._assert_in_rnn_block_('memory')
        if init is None:
            if shape is None or dtype is None:
                raise ValueError(
                    "if init is None, memory at least need shape and dtype")
            parent_block = self.parent_block()
            var_name = unique_name("@".join([self.helper.name, "memory_boot"]))
            boot_var = parent_block.create_var(
                name=var_name, shape=shape, dtype=dtype, persistable=False)

            parent_block.append_op(
                type="fill_constant",
                inputs={},
                outputs={'Out': [boot_var]},
                attrs={
                    'value': init_value,
                    'shape': boot_var.shape,
                    'data_type': boot_var.data_type
                })

            return self.memory(init=boot_var)
        else:
            pre_mem = self.helper.create_variable(
                name=unique_name("@".join([self.helper.name, "mem"])),
                dtype=init.data_type,
                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:
            self.seq_len = x.shape[1]
        elif self.seq_len != x.shape[1]:
            raise ValueError("Static RNN only take fix seq_len input")

        ipt = self.helper.create_variable(
            name=x.name,
            dtype=x.data_type,
            shape=[-1] + list(x.shape[2:]),
            type=x.type)
        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")

        out_var = self.parent_block().create_var(
            name=o.name,
            shape=[-1, self.seq_len] + list(o.shape[1:]),
            dtype=o.data_type)

        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):
        prog = self.helper.program
        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):
        # TODO(yuyang18): Create RNN Op here.
        # Implement this method after RNN op complete.
        pass
反馈
建议
客服 返回
顶部