nets.py 21.8 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
14 15

from __future__ import print_function
M
minqiyang 已提交
16
import six
17
from . import layers
F
fengjiayi 已提交
18

19 20 21
__all__ = [
    "simple_img_conv_pool",
    "sequence_conv_pool",
22
    "glu",
23
    "scaled_dot_product_attention",
24
]
D
dzhwinter 已提交
25

F
fengjiayi 已提交
26 27 28

def simple_img_conv_pool(input,
                         num_filters,
D
dzhwinter 已提交
29
                         filter_size,
F
fengjiayi 已提交
30 31
                         pool_size,
                         pool_stride,
C
chengduoZH 已提交
32
                         pool_padding=0,
C
chengduoZH 已提交
33
                         pool_type='max',
C
chengduoZH 已提交
34 35 36 37 38 39 40 41
                         global_pooling=False,
                         conv_stride=1,
                         conv_padding=0,
                         conv_dilation=1,
                         conv_groups=1,
                         param_attr=None,
                         bias_attr=None,
                         act=None,
42 43
                         use_cudnn=True,
                         use_mkldnn=False):
C
chengduoZH 已提交
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102
    """
    The simple_img_conv_pool is composed with one Convolution2d and one Pool2d.

    Args:
        input (Variable): The input image with [N, C, H, W] format.
        num_filters(int): The number of filter. It is as same as the output
            feature channel.
        filter_size (int|list|tuple): The filter size. If filter_size is a list or
            tuple, it must contain two integers, (filter_size_H, filter_size_W). Otherwise,
            the filter_size_H = filter_size_W = filter_size.
        pool_size (int|list|tuple): The pooling size of Pool2d layer. If pool_size
            is a list or tuple, it must contain two integers, (pool_size_H, pool_size_W).
            Otherwise, the pool_size_H = pool_size_W = pool_size.
        pool_stride (int|list|tuple): The pooling stride of Pool2d layer. If pool_stride
            is a list or tuple, it must contain two integers, (pooling_stride_H, pooling_stride_W).
            Otherwise, the pooling_stride_H = pooling_stride_W = pool_stride.
        pool_padding (int|list|tuple): The padding of Pool2d layer. If pool_padding is a list or
            tuple, it must contain two integers, (pool_padding_H, pool_padding_W).
            Otherwise, the pool_padding_H = pool_padding_W = pool_padding. Default 0.
        pool_type (str): Pooling type can be :math:`max` for max-pooling and :math:`avg` for
            average-pooling. Default :math:`max`.
        global_pooling (bool): Whether to use the global pooling. If global_pooling = true,
            pool_size and pool_padding while be ignored. Default False
        conv_stride (int|list|tuple): The stride size of the Conv2d Layer. If stride is a
            list or tuple, it must contain two integers, (conv_stride_H, conv_stride_W). Otherwise,
            the conv_stride_H = conv_stride_W = conv_stride. Default: conv_stride = 1.
        conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is
            a list or  tuple, it must contain two integers, (conv_padding_H, conv_padding_W).
            Otherwise, the conv_padding_H = conv_padding_W = conv_padding. Default: conv_padding = 0.
        conv_dilation (int|list|tuple): The dilation size of the Conv2d Layer. If dilation is
            a list or tuple, it must contain two integers, (conv_dilation_H, conv_dilation_W).
            Otherwise, the conv_dilation_H = conv_dilation_W = conv_dilation. Default: conv_dilation = 1.
        conv_groups (int): The groups number of the Conv2d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: groups=1
        param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None
        bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None
        act (str): Activation type for Conv2d. Default: None
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
        use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled
            with mkldnn library. Default: False

    Return:
        Variable: The result of input after Convolution2d and Pool2d.

    Examples:
        .. code-block:: python

            img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32')
            conv_pool = fluid.nets.simple_img_conv_pool(input=img,
                                                        filter_size=5,
                                                        num_filters=20,
                                                        pool_size=2,
                                                        pool_stride=2,
                                                        act="relu")
    """
F
fengjiayi 已提交
103 104 105 106
    conv_out = layers.conv2d(
        input=input,
        num_filters=num_filters,
        filter_size=filter_size,
C
chengduoZH 已提交
107 108 109 110
        stride=conv_stride,
        padding=conv_padding,
        dilation=conv_dilation,
        groups=conv_groups,
F
fengjiayi 已提交
111
        param_attr=param_attr,
C
chengduoZH 已提交
112
        bias_attr=bias_attr,
C
chengduoZH 已提交
113
        act=act,
114 115
        use_cudnn=use_cudnn,
        use_mkldnn=use_mkldnn)
F
fengjiayi 已提交
116 117 118 119

    pool_out = layers.pool2d(
        input=conv_out,
        pool_size=pool_size,
Q
Qiao Longfei 已提交
120
        pool_type=pool_type,
C
chengduoZH 已提交
121
        pool_stride=pool_stride,
C
chengduoZH 已提交
122 123
        pool_padding=pool_padding,
        global_pooling=global_pooling,
124 125
        use_cudnn=use_cudnn,
        use_mkldnn=use_mkldnn)
Q
Qiao Longfei 已提交
126 127 128 129 130 131 132 133 134
    return pool_out


def img_conv_group(input,
                   conv_num_filter,
                   pool_size,
                   conv_padding=1,
                   conv_filter_size=3,
                   conv_act=None,
F
fengjiayi 已提交
135
                   param_attr=None,
Q
Qiao Longfei 已提交
136
                   conv_with_batchnorm=False,
W
wanghaoshuang 已提交
137
                   conv_batchnorm_drop_rate=0.0,
Q
Qiao Longfei 已提交
138
                   pool_stride=1,
C
chengduoZH 已提交
139
                   pool_type="max",
140 141
                   use_cudnn=True,
                   use_mkldnn=False):
Q
Qiao Longfei 已提交
142
    """
C
chengduoZH 已提交
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197
    The Image Convolution Group is composed of Convolution2d, BatchNorm, DropOut,
    and Pool2d. According to the input arguments, img_conv_group will do serials of
    computation for Input using Convolution2d, BatchNorm, DropOut, and pass the last
    result to Pool2d.

    Args:
        input (Variable): The input image with [N, C, H, W] format.
        conv_num_filter(list|tuple): Indicates the numbers of filter of this group.
        pool_size (int|list|tuple): The pooling size of Pool2d Layer. If pool_size
            is a list or tuple, it must contain two integers, (pool_size_H, pool_size_W).
            Otherwise, the pool_size_H = pool_size_W = pool_size.
        conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is
            a list or tuple, its length must be equal to the length of conv_num_filter.
            Otherwise the conv_padding of all Conv2d Layers are the same. Default 1.
        conv_filter_size (int|list|tuple): The filter size. If filter_size is a list or
            tuple, its length must be equal to the length of conv_num_filter.
            Otherwise the conv_filter_size of all Conv2d Layers are the same. Default 3.
        conv_act (str): Activation type for Conv2d Layer that is not followed by BatchNorm.
            Default: None.
        param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None
        conv_with_batchnorm (bool|list): Indicates whether to use BatchNorm after Conv2d Layer.
            If conv_with_batchnorm is a list, its length must be equal to the length of
            conv_num_filter. Otherwise, conv_with_batchnorm indicates whether all the
            Conv2d Layer follows a BatchNorm. Default False.
        conv_batchnorm_drop_rate (float|list): Indicates the drop_rate of Dropout Layer
            after BatchNorm. If conv_batchnorm_drop_rate is a list, its length must be
            equal to the length of conv_num_filter. Otherwise, drop_rate of all Dropout
            Layers is conv_batchnorm_drop_rate. Default 0.0.
        pool_stride (int|list|tuple): The pooling stride of Pool2d layer. If pool_stride
            is a list or tuple, it must contain two integers, (pooling_stride_H,
            pooling_stride_W). Otherwise, the pooling_stride_H = pooling_stride_W = pool_stride.
            Default 1.
        pool_type (str): Pooling type can be :math:`max` for max-pooling and :math:`avg` for
            average-pooling. Default :math:`max`.
        use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn
            library is installed. Default: True
        use_mkldnn (bool): Use mkldnn kernels or not, it is valid only when compiled
            with mkldnn library. Default: False

    Return:
        Variable: The final result after serial computation using Convolution2d,
            BatchNorm, DropOut, and Pool2d.

    Examples:
        .. code-block:: python

            img = fluid.layers.data(name='img', shape=[1, 28, 28], dtype='float32')
            conv_pool = fluid.nets.img_conv_group(input=img,
                                                  num_channels=3,
                                                  conv_padding=1,
                                                  conv_num_filter=[3, 3],
                                                  conv_filter_size=3,
                                                  conv_act="relu",
                                                  pool_size=2,
                                                  pool_stride=2)
Q
Qiao Longfei 已提交
198 199 200
    """
    tmp = input
    assert isinstance(conv_num_filter, list) or \
201
        isinstance(conv_num_filter, tuple)
Q
Qiao Longfei 已提交
202 203 204 205 206

    def __extend_list__(obj):
        if not hasattr(obj, '__len__'):
            return [obj] * len(conv_num_filter)
        else:
C
chengduoZH 已提交
207
            assert len(obj) == len(conv_num_filter)
Q
Qiao Longfei 已提交
208 209 210 211
            return obj

    conv_padding = __extend_list__(conv_padding)
    conv_filter_size = __extend_list__(conv_filter_size)
F
fengjiayi 已提交
212
    param_attr = __extend_list__(param_attr)
Q
Qiao Longfei 已提交
213 214 215
    conv_with_batchnorm = __extend_list__(conv_with_batchnorm)
    conv_batchnorm_drop_rate = __extend_list__(conv_batchnorm_drop_rate)

M
minqiyang 已提交
216
    for i in six.moves.range(len(conv_num_filter)):
Q
Qiao Longfei 已提交
217 218 219 220 221 222 223 224 225
        local_conv_act = conv_act
        if conv_with_batchnorm[i]:
            local_conv_act = None

        tmp = layers.conv2d(
            input=tmp,
            num_filters=conv_num_filter[i],
            filter_size=conv_filter_size[i],
            padding=conv_padding[i],
F
fengjiayi 已提交
226
            param_attr=param_attr[i],
C
chengduoZH 已提交
227
            act=local_conv_act,
228 229
            use_cudnn=use_cudnn,
            use_mkldnn=use_mkldnn)
Q
Qiao Longfei 已提交
230 231

        if conv_with_batchnorm[i]:
232
            tmp = layers.batch_norm(input=tmp, act=conv_act)
Q
Qiao Longfei 已提交
233 234
            drop_rate = conv_batchnorm_drop_rate[i]
            if abs(drop_rate) > 1e-5:
235
                tmp = layers.dropout(x=tmp, dropout_prob=drop_rate)
Q
Qiao Longfei 已提交
236 237 238 239 240

    pool_out = layers.pool2d(
        input=tmp,
        pool_size=pool_size,
        pool_type=pool_type,
C
chengduoZH 已提交
241
        pool_stride=pool_stride,
242 243
        use_cudnn=use_cudnn,
        use_mkldnn=use_mkldnn)
F
fengjiayi 已提交
244
    return pool_out
D
dzhwinter 已提交
245 246 247 248 249


def sequence_conv_pool(input,
                       num_filters,
                       filter_size,
F
fengjiayi 已提交
250
                       param_attr=None,
251
                       act="sigmoid",
252
                       pool_type="max"):
C
chengduoZH 已提交
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285
    """
    The sequence_conv_pool is composed with Sequence Convolution and Pooling.

    Args:
        input (Variable): The input of sequence_conv, which supports variable-time
            length input sequence. The underlying of input is a matrix with shape
            (T, N), where T is the total time steps in this mini-batch and N is
            the input_hidden_size
        num_filters(int): The number of filter.
        filter_size (int): The filter size.
        param_attr (ParamAttr): The parameters to the Sequence_conv Layer. Default: None.
        act (str): Activation type for Sequence_conv Layer. Default: "sigmoid".
        pool_type (str): Pooling type can be :math:`max` for max-pooling, :math:`average` for
            average-pooling, :math:`sum` for sum-pooling, :math:`sqrt` for sqrt-pooling.
            Default :math:`max`.

    Return:
        Variable: The final result after Sequence Convolution and Pooling.

    Examples:
        .. code-block:: python

            input_dim = len(word_dict)
            emb_dim = 128
            hid_dim = 512
            data = fluid.layers.data( ame="words", shape=[1], dtype="int64", lod_level=1)
            emb = fluid.layers.embedding(input=data, size=[input_dim, emb_dim], is_sparse=True)
            seq_conv = fluid.nets.sequence_conv_pool(input=emb,
                                                     num_filters=hid_dim,
                                                     filter_size=3,
                                                     act="tanh",
                                                     pool_type="sqrt")
    """
D
dzhwinter 已提交
286 287 288 289
    conv_out = layers.sequence_conv(
        input=input,
        num_filters=num_filters,
        filter_size=filter_size,
F
fengjiayi 已提交
290
        param_attr=param_attr,
291
        act=act)
D
dzhwinter 已提交
292

293
    pool_out = layers.sequence_pool(input=conv_out, pool_type=pool_type)
D
dzhwinter 已提交
294
    return pool_out
G
guosheng 已提交
295 296 297 298


def glu(input, dim=-1):
    """
C
chengduoZH 已提交
299 300 301
    The Gated Linear Units(GLU) composed by split, sigmoid activation and element-wise
    multiplication. Specifically, Split the input into two equal sized parts,
    :math:`a` and :math:`b`, along the given dimension and then compute as
G
guosheng 已提交
302
    following:
G
guosheng 已提交
303 304 305 306 307

        .. math::

            {GLU}(a, b)= a \otimes \sigma(b)

Y
ying 已提交
308
    Refer to `Language Modeling with Gated Convolutional Networks
G
guosheng 已提交
309
    <https://arxiv.org/pdf/1612.08083.pdf>`_.
Y
ying 已提交
310

G
guosheng 已提交
311 312
    Args:
        input (Variable): The input variable which is a Tensor or LoDTensor.
Y
ying 已提交
313
        dim (int): The dimension along which to split. If :math:`dim < 0`, the
C
chengduoZH 已提交
314
            dimension to split along is :math:`rank(input) + dim`. Default -1.
G
guosheng 已提交
315 316

    Returns:
C
chengduoZH 已提交
317
        Variable: Variable with half the size of input.
G
guosheng 已提交
318 319 320 321

    Examples:
        .. code-block:: python

C
chengduoZH 已提交
322 323
            data = fluid.layers.data(name="words", shape=[3, 6, 9], dtype="float32")
            output = fluid.nets.glu(input=data, dim=1)  # shape of output: [3, 3, 9]
G
guosheng 已提交
324 325 326
    """

    a, b = layers.split(input, num_or_sections=2, dim=dim)
G
guosheng 已提交
327 328
    act_b = layers.sigmoid(x=b)
    out = layers.elementwise_mul(x=a, y=act_b)
G
guosheng 已提交
329
    return out
330 331


Y
ying 已提交
332 333 334
def scaled_dot_product_attention(queries,
                                 keys,
                                 values,
Y
ying 已提交
335
                                 num_heads=1,
Y
ying 已提交
336
                                 dropout_rate=0.):
337 338 339
    """
    The dot-product attention.

340 341 342
    Attention mechanism can be seen as mapping a query and a set of key-value
    pairs to an output. The output is computed as a weighted sum of the values,
    where the weight assigned to each value is computed by a compatibility
343
    function (dot-product here) of the query with the corresponding key.
Y
ying 已提交
344 345

    The dot-product attention can be implemented through (batch) matrix
346 347 348 349
    multipication as follows:

        .. math::

350
            Attention(Q, K, V)= softmax(QK^\mathrm{T})V
351

Y
ying 已提交
352
    Refer to `Attention Is All You Need
353 354
    <https://arxiv.org/pdf/1706.03762.pdf>`_.

Y
ying 已提交
355 356 357 358 359
    Args:
        queries (Variable): The input variable which should be a 3-D Tensor.
        keys (Variable): The input variable which should be a 3-D Tensor.
        values (Variable): The input variable which should be a 3-D Tensor.
        num_heads (int): Head number to compute the scaled dot product
C
chengduoZH 已提交
360
            attention. Default: 1.
Y
ying 已提交
361
        dropout_rate (float): The dropout rate to drop the attention weight.
C
chengduoZH 已提交
362
            Default: 0.0.
363 364

    Returns:
C
chengduoZH 已提交
365 366
        Variable: A 3-D Tensor computed by multi-head scaled dot product\
            attention.
367

Y
ying 已提交
368 369 370
    Raises:
        ValueError: If input queries, keys, values are not 3-D Tensors.

C
chengduoZH 已提交
371
    NOTES:
Y
ying 已提交
372
        1. When num_heads > 1, three linear projections are learned respectively
C
chengduoZH 已提交
373 374 375 376 377
           to map input queries, keys and values into queries', keys' and values'.
           queries', keys' and values' have the same shapes with queries, keys
           and values.
        2. When num_heads == 1, scaled_dot_product_attention has no learnable
           parameters.
Y
ying 已提交
378

379 380 381
    Examples:
        .. code-block:: python

C
chengduoZH 已提交
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397
            queries = fluid.layers.data(name="queries",
                                        shape=[3, 5, 9],
                                        dtype="float32",
                                        append_batch_size=False)
            queries.stop_gradient = False
            keys = fluid.layers.data(name="keys",
                                     shape=[3, 6, 9],
                                     dtype="float32",
                                     append_batch_size=False)
            keys.stop_gradient = False
            values = fluid.layers.data(name="values",
                                       shape=[3, 6, 10],
                                       dtype="float32",
                                       append_batch_size=False)
            values.stop_gradient = False
            contexts = fluid.nets.scaled_dot_product_attention(queries, keys, values)
Y
ying 已提交
398
            contexts.shape  # [3, 5, 10]
399
    """
Y
ying 已提交
400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
    if not (len(queries.shape) == len(keys.shape) == len(values.shape) == 3):
        raise ValueError(
            "Inputs quries, keys and values should all be 3-D tensors.")

    if queries.shape[-1] != keys.shape[-1]:
        raise ValueError(
            "The hidden size of queries and keys should be the same.")
    if keys.shape[-2] != values.shape[-2]:
        raise ValueError(
            "The max sequence length in query batch and in key batch "
            "should be the same.")
    if keys.shape[-1] % num_heads != 0:
        raise ValueError("The hidden size of keys (%d) must be divisible "
                         "by the number of attention heads (%d)." %
                         (keys.shape[-1], num_heads))
    if values.shape[-1] % num_heads != 0:
        raise ValueError("The hidden size of values (%d) must be divisible "
                         "by the number of attention heads (%d)." %
                         (values.shape[-1], num_heads))

Y
ying 已提交
420
    def __compute_qkv(queries, keys, values, num_heads):
Y
ying 已提交
421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436
        """
        Add linear projection to queries, keys, and values.

        Args:
            queries(Tensor): a 3-D input Tensor.
            keys(Tensor): a 3-D input Tensor.
            values(Tensor): a 3-D input Tensor.
            num_heads(int): The number of heads. Linearly project the inputs
                            ONLY when num_heads > 1.

        Returns:
            Tensor: linearly projected output Tensors: queries', keys' and
                    values'. They have the same shapes with queries, keys and
                    values.
        """

Y
ying 已提交
437 438 439 440 441 442 443 444
        if num_heads == 1:
            return queries, keys, values

        q = layers.fc(input=queries, size=queries.shape[-1], num_flatten_dims=2)
        k = layers.fc(input=keys, size=keys.shape[-1], num_flatten_dims=2)
        v = layers.fc(input=values, size=values.shape[-1], num_flatten_dims=2)
        return q, k, v

Y
ying 已提交
445 446 447 448 449 450
    def __split_heads(x, num_heads):
        """
        Reshape the last dimension of inpunt tensor x so that it becomes two
        dimensions.

        Args:
Y
ying 已提交
451 452
            x(Tensor): a 3-D input Tensor.
            num_heads(int): The number of heads.
Y
ying 已提交
453 454

        Returns:
Y
ying 已提交
455 456
            Tensor: a Tensor with shape [..., n, m/num_heads], where m is size
                    of the last dimension of x.
Y
ying 已提交
457
        """
Y
ying 已提交
458 459
        if num_heads == 1:
            return x
460

Y
ying 已提交
461
        hidden_size = x.shape[-1]
462 463 464
        # reshape the 3-D input: [batch_size, max_sequence_length, hidden_dim]
        # into a 4-D output:
        # [batch_size, max_sequence_length, num_heads, hidden_size_per_head].
Y
ying 已提交
465
        reshaped = layers.reshape(
466 467
            x=x,
            shape=list(x.shape[:-1]) + [num_heads, hidden_size // num_heads])
468 469

        # permuate the dimensions into:
470 471 472 473
        # [batch_size, num_heads, max_sequence_len, hidden_size_per_head]
        return layers.transpose(x=reshaped, perm=[0, 2, 1, 3])

    def __combine_heads(x):
Y
ying 已提交
474 475 476 477 478 479 480 481 482 483 484 485 486
        """
        Reshape the last two dimensions of inpunt tensor x so that it becomes
        one dimension.

        Args:
            x(Tensor): a 4-D input Tensor with shape
                       [bs, num_heads, max_sequence_length, hidden_dim].

        Returns:
            Tensor: a Tensor with shape
                    [bs, max_sequence_length, num_heads * hidden_dim].
        """

Y
ying 已提交
487
        if len(x.shape) == 3: return x
488 489 490
        if len(x.shape) != 4:
            raise ValueError("Input(x) should be a 4-D Tensor.")

Y
ying 已提交
491
        trans_x = layers.transpose(x, perm=[0, 2, 1, 3])
Y
ying 已提交
492
        return layers.reshape(
493
            x=trans_x,
494 495 496 497 498
            shape=list(
                map(int, [
                    trans_x.shape[0], trans_x.shape[1], trans_x.shape[2] *
                    trans_x.shape[3]
                ])))
499

Y
ying 已提交
500 501 502 503 504
    q, k, v = __compute_qkv(queries, keys, values, num_heads)

    q = __split_heads(q, num_heads)
    k = __split_heads(k, num_heads)
    v = __split_heads(v, num_heads)
Y
ying 已提交
505 506

    key_dim_per_head = keys.shape[-1] // num_heads
507 508
    scaled_q = layers.scale(x=q, scale=key_dim_per_head**-0.5)
    product = layers.matmul(x=k, y=scaled_q, transpose_y=True)
Y
ying 已提交
509

Y
ying 已提交
510
    weights = layers.reshape(
511
        x=layers.reshape(
Y
ying 已提交
512
            x=product, shape=[-1, product.shape[-1]], act="softmax"),
513
        shape=product.shape)
Y
ying 已提交
514
    if dropout_rate:
G
guosheng 已提交
515 516
        weights = layers.dropout(
            weights, dropout_prob=dropout_rate, is_test=False)
Y
ying 已提交
517 518
    ctx_multiheads = layers.matmul(weights, v)
    return __combine_heads(ctx_multiheads)