common.py 58.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
#   Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15
# TODO: define the common classes to build a neural network
16
import paddle
Z
zhiboniu 已提交
17
from ...fluid.dygraph import Flatten  # noqa: F401
T
tangwei12 已提交
18
from ...fluid.framework import in_dygraph_mode
19
from .. import functional as F
20
from ...fluid.framework import _dygraph_tracer
Z
zhiboniu 已提交
21
from paddle.nn import Layer
22

23 24
__all__ = []

25

26 27 28 29 30 31 32
def _npairs(x, n):
    if isinstance(x, (paddle.Tensor, list)):
        return x
    x = [x] * (n * 2)
    return x


Z
zhiboniu 已提交
33
class Linear(Layer):
34
    r"""
35 36 37

    Fully-connected linear transformation layer. For each input :math:`X` ,
    the equation is:
38 39 40

    .. math::

41
        Out = XW + b
42

43
    where :math:`W` is the weight and :math:`b` is the bias.
44

45 46 47 48 49 50 51
    Linear layer takes only one multi-dimensional tensor as input with the
    shape :math:`[batch\_size, *, in\_features]` , where :math:`*` means any
    number of additional dimensions. It multiplies input tensor with the weight
    (a 2-D tensor of shape :math:`[in\_features, out\_features]` ) and produces
    an output tensor of shape :math:`[batch\_size, *, out\_features]` .
    If :math:`bias\_attr` is not False, the bias (a 1-D tensor of
    shape :math:`[out\_features]` ) will be created and added to the output.
52 53

    Parameters:
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76
        in_features (int): The number of input units.
        out_features (int): The number of output units.
        weight_attr (ParamAttr, optional): The attribute for the learnable
            weight of this layer. The default value is None and the weight will be
            initialized to zero. For detailed information, please refer to
            paddle.ParamAttr.
        bias_attr (ParamAttr|bool, optional): The attribute for the learnable bias
            of this layer. If it is set to False, no bias will be added to the output.
            If it is set to None or one kind of ParamAttr, a bias parameter will
            be created according to ParamAttr. For detailed information, please refer
            to paddle.ParamAttr. The default value is None and the bias will be
            initialized to zero.
        name (str, optional): Normally there is no need for user to set this parameter.
            For detailed information, please refer to :ref:`api_guide_Name` .

    Attribute:
        **weight** (Parameter): the learnable weight of this layer.

        **bias** (Parameter): the learnable bias of this layer.

    Shape:
        - input: Multi-dimentional tensor with shape :math:`[batch\_size, *, in\_features]` .
        - output: Multi-dimentional tensor with shape :math:`[batch\_size, *, out\_features]` .
77 78 79 80 81

    Examples:
        .. code-block:: python

          import paddle
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102

          # Define the linear layer.
          weight_attr = paddle.ParamAttr(
              name="weight",
              initializer=paddle.nn.initializer.Constant(value=0.5))
          bias_attr = paddle.ParamAttr(
              name="bias",
              initializer=paddle.nn.initializer.Constant(value=1.0))
          linear = paddle.nn.Linear(2, 4, weight_attr=weight_attr, bias_attr=bias_attr)
          # linear.weight: [[0.5 0.5 0.5 0.5]
          #                 [0.5 0.5 0.5 0.5]]
          # linear.bias: [1. 1. 1. 1.]

          x = paddle.randn((3, 2), dtype="float32")
          # x: [[-0.32342386 -1.200079  ]
          #     [ 0.7979031  -0.90978354]
          #     [ 0.40597573  1.8095392 ]]
          y = linear(x)
          # y: [[0.23824859 0.23824859 0.23824859 0.23824859]
          #     [0.9440598  0.9440598  0.9440598  0.9440598 ]
          #     [2.1077576  2.1077576  2.1077576  2.1077576 ]]
103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131
    """

    def __init__(self,
                 in_features,
                 out_features,
                 weight_attr=None,
                 bias_attr=None,
                 name=None):
        super(Linear, self).__init__()
        self._dtype = self._helper.get_default_dtype()
        self._weight_attr = weight_attr
        self._bias_attr = bias_attr
        self.weight = self.create_parameter(
            shape=[in_features, out_features],
            attr=self._weight_attr,
            dtype=self._dtype,
            is_bias=False)
        self.bias = self.create_parameter(
            shape=[out_features],
            attr=self._bias_attr,
            dtype=self._dtype,
            is_bias=True)
        self.name = name

    def forward(self, input):
        out = F.linear(
            x=input, weight=self.weight, bias=self.bias, name=self.name)
        return out

132 133 134 135 136
    def extra_repr(self):
        name_str = ', name={}'.format(self.name) if self.name else ''
        return 'in_features={}, out_features={}, dtype={}{}'.format(
            self.weight.shape[0], self.weight.shape[1], self._dtype, name_str)

137

Z
zhiboniu 已提交
138
class Upsample(Layer):
139 140
    """
    This op resizes a batch of images.
141

142 143 144
    The input must be a 3-D Tensor of the shape (num_batches, channels, in_w)
    or 4-D (num_batches, channels, in_h, in_w), or a 5-D Tensor of the shape
    (num_batches, channels, in_d, in_h, in_w) or (num_batches, in_d, in_h, in_w, channels),
145 146
    Where in_w is width of the input tensor, in_h is the height of the input tensor,
    in_d is the depth of the intput tensor.
147
    and the resizing only applies on the three dimensions(depth, height and width).
X
xiaoting 已提交
148

149
    Supporting resample methods:
150 151 152 153 154 155
        'linear' : Linear interpolation
        'bilinear' : Bilinear interpolation
        'trilinear' : Trilinear interpolation
        'nearest' : Nearest neighbor interpolation
        'bicubic' : Bicubic interpolation

T
tangwei12 已提交
156 157 158
    Linear interpolation is the method of using a line connecting two known quantities
    to determine the value of an unknown quantity between the two known quantities.

159 160 161 162 163 164 165 166 167
    Nearest neighbor interpolation is to perform nearest neighbor interpolation
    in both the 3rd dimension(in height direction) and the 4th dimension(in width
    direction) on input tensor.

    Bilinear interpolation is an extension of linear interpolation for
    interpolating functions of two variables (e.g. H-direction and
    W-direction in this op) on a rectilinear 2D grid. The key idea is
    to perform linear interpolation first in one direction, and then
    again in the other direction.
T
tangwei12 已提交
168

169 170 171 172
    Bicubic interpolation is an extension of cubic interpolation for interpolating
    data points on a two-dimensional regular grid. The interpolated surface is
    smoother than corresponding surfaces obtained by bilinear interpolation or
    nearest-neighbor interpolation.
173 174 175 176 177

    Trilinear interpolation is an extension of linear interpolation for
    interpolating functions of three variables (e.g. D-direction,
    H-direction and W-direction in this op) on a rectilinear 3D grid.
    The linear interpolation is performed on three directions.
X
xiaoting 已提交
178
    align_corners and align_mode are optional parameters,the calculation method
179 180
    of interpolation can be selected by them.

181 182 183 184 185 186
    Area interpolation is to perform area interpolation
    in both the 3rd dimension(in height direction) , the 4th dimension(in width
    direction) and the 5th dimension(in depth direction) on input tensor. Set to
    area will directly call `paddle.nn.functional.adaptive_avg_pool1d` or
    `paddle.nn.functional.adaptive_avg_pool2d` or `paddle.nn.functional.adaptive_avg_pool3d`.

187 188 189 190
    Example:

    .. code-block:: text

191
        For scale_factor:
192 193 194 195 196
            if align_corners = True && out_size > 1 :
              scale_factor = (in_size-1.0)/(out_size-1.0)
            else:
              scale_factor = float(in_size/out_size)

197 198 199 200 201 202 203 204 205 206
        Linear interpolation:
            if:
                align_corners = False , align_mode = 0
                input : (N,C,W_in)
                output: (N,C,W_out) where:
                W_out = (W_{in}+0.5) * scale_{factor} - 0.5
            else:
                input : (N,C,W_in)
                output: (N,C,W_out) where:
                W_out = W_{in} * scale_{factor}
207 208 209 210 211 212 213 214 215 216 217 218 219 220

        Nearest neighbor interpolation:
          if:
              align_corners = False
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = floor (H_{in} * scale_{factor})
              W_out = floor (W_{in} * scale_{factor})
          else:
              align_corners = True
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = round(H_{in} * scale_{factor})
              W_out = round(W_{in} * scale_{factor})
T
tangwei12 已提交
221

222 223 224
        Bilinear interpolation:
          if:
              align_corners = False , align_mode = 0
225

226 227 228 229 230
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5
          else:
231

232 233 234 235 236 237 238 239 240 241 242 243
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}

        Bicubic interpolation:
          if:
              align_corners = False
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5
244

245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265
          else:
              input : (N,C,H_in,W_in)
              output: (N,C,H_out,W_out) where:
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}

        Trilinear interpolation:
          if:
              align_corners = False , align_mode = 0
              input : (N,C,D_in,H_in,W_in)
              output: (N,C,D_out,H_out,W_out) where:
              D_out = (D_{in}+0.5) * scale_{factor} - 0.5
              H_out = (H_{in}+0.5) * scale_{factor} - 0.5
              W_out = (W_{in}+0.5) * scale_{factor} - 0.5
          else:
              input : (N,C,D_in,H_in,W_in)
              output: (N,C,D_out,H_out,W_out) where:
              D_out = D_{in} * scale_{factor}
              H_out = H_{in} * scale_{factor}
              W_out = W_{in} * scale_{factor}

266 267
    https://en.wikipedia.org/wiki/Linear_interpolation.
    For details of linear interpolation, please refer to Wikipedia:
T
tangwei12 已提交
268

269 270
    For details of nearest neighbor interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation.
T
tangwei12 已提交
271

272 273
    For details of bilinear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Bilinear_interpolation.
T
tangwei12 已提交
274

275 276
    For details of bicubic interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Bicubic_interpolation
T
tangwei12 已提交
277

278 279
    For details of trilinear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Trilinear_interpolation.
T
tangwei12 已提交
280

281
    Parameters:
X
xiaoting 已提交
282
        x (Tensor): 3-D, 4-D or 5-D Tensor, its data type is float32, float64, or uint8,
283
                          its data format is specified by :attr:`data_format`.
X
xiaoting 已提交
284
        size (list|tuple|Tensor|None): Output shape of image resize
285 286
             layer, the shape is (out_w, ) when input is a 3-D Tensor, the shape is (out_h, out_w) 
             when input is a 4-D Tensor and is (out_d, out_h, out_w) when input is a 5-D Tensor. 
287
             Default: None. If a list/tuple, each element can be an integer or a Tensor of shape: [1].
X
xiaoting 已提交
288
             If a Tensor , its dimensions size should be a 1.
289 290 291
        scale_factor (float|Tensor|list|tuple|None): The multiplier for the input height or width. At
             least one of :attr:`size` or :attr:`scale_factor` must be set.
             And :attr:`size` has a higher priority than :attr:`scale_factor`. Has to match input size if it is either a list or a tuple or a Tensor.
292
             Default: None.
293 294
        mode (str): The resample method. It supports 'linear', 'nearst', 'bilinear',
                       'bicubic' and 'trilinear' currently. Default: 'nearest'
295 296 297
        align_corners(bool) :  An optional bool, If True, the centers of the 4 corner pixels of the
                               input and output tensors are aligned, preserving the values at the
                               corner pixels.
298 299 300 301
                               Default: False
        align_mode(int)  :  An optional for linear/bilinear/trilinear interpolation. Refer to the formula in the example above,
                            it can be \'0\' for src_idx = scale_factor*(dst_indx+0.5)-0.5 , can be \'1\' for
                            src_idx = scale_factor*dst_index.
302
        data_format (str, optional): Specify the data format of the input, and the data format of the output
303
            will be consistent with that of the input. An optional string from:`NCW`, `NWC`, `"NCHW"`, `"NHWC"`, `"NCDHW"`,
304 305 306
            `"NDHWC"`. The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`. When it is `"NCHW"`, the data is stored
            in the order of: `[batch_size, input_channels, input_depth, input_height, input_width]`.
307 308 309
        name(str, optional): The default value is None.
                             Normally there is no need for user to set this property.
                             For more information, please refer to :ref:`api_guide_Name`
310 311 312
    Returns:
        A 3-D Tensor of the shape (num_batches, channels, out_w) or (num_batches, out_w, channels),
        A 4-D Tensor of the shape (num_batches, channels, out_h, out_w) or (num_batches, out_h, out_w, channels),
313
        or 5-D Tensor of the shape (num_batches, channels, out_d, out_h, out_w) or (num_batches, out_d, out_h, out_w, channels).
314
    Raises:
X
xiaoting 已提交
315
        TypeError: size should be a list or tuple or Tensor.
316 317 318 319 320 321 322 323 324 325
        ValueError: The 'mode' of image_resize can only be 'linear', 'bilinear',
                    'trilinear', 'bicubic', or 'nearest' currently.
        ValueError: 'linear' only support 3-D tensor.
        ValueError: 'bilinear', 'bicubic' and 'nearest' only support 4-D tensor.
        ValueError: 'trilinear' only support 5-D tensor.
        ValueError: One of size and scale_factor must not be None.
        ValueError: size length should be 1 for input 3-D tensor.
        ValueError: size length should be 2 for input 4-D tensor.
        ValueError: size length should be 3 for input 5-D tensor.
        ValueError: scale_factor should be greater than zero.
326 327
        TypeError: align_corners should be a bool value
        ValueError: align_mode can only be '0' or '1'
328
        ValueError: data_format can only be 'NCW', 'NWC', 'NCHW', 'NHWC', 'NCDHW' or 'NDHWC'.
329 330 331

    Examples:
        .. code-block:: python
C
Chen Long 已提交
332
            
333
            import paddle
X
xiaoting 已提交
334
            import paddle.nn as nn
335
            import numpy as np
X
xiaoting 已提交
336

337
            input_data = np.random.rand(2,3,6,10).astype("float32")
338
            upsample_out  = paddle.nn.Upsample(size=[12,12])
X
xiaoting 已提交
339 340 341 342 343 344

            input = paddle.to_tensor(input_data)
            output = upsample_out(x=input)
            print(output.shape)
            # [2L, 3L, 12L, 12L]

345 346 347
    """

    def __init__(self,
348 349 350 351
                 size=None,
                 scale_factor=None,
                 mode='nearest',
                 align_corners=False,
X
xiaoting 已提交
352 353 354
                 align_mode=0,
                 data_format='NCHW',
                 name=None):
355
        super(Upsample, self).__init__()
356 357 358
        self.size = size
        self.scale_factor = scale_factor
        self.mode = mode.lower()
359 360 361
        self.align_corners = align_corners
        self.align_mode = align_mode
        self.data_format = data_format
X
xiaoting 已提交
362
        self.name = name
363

X
xiaoting 已提交
364
    def forward(self, x):
365
        out = F.interpolate(
X
xiaoting 已提交
366
            x,
367 368 369
            size=self.size,
            scale_factor=self.scale_factor,
            mode=self.mode,
370 371
            align_corners=self.align_corners,
            align_mode=self.align_mode,
X
xiaoting 已提交
372 373
            data_format=self.data_format,
            name=self.name)
X
xiaoting 已提交
374 375 376

        return out

377 378 379 380 381 382 383 384 385 386
    def extra_repr(self):
        if self.scale_factor is not None:
            main_str = 'scale_factor={}'.format(self.scale_factor)
        else:
            main_str = 'size={}'.format(self.size)
        name_str = ', name={}'.format(self.name) if self.name else ''
        return '{}, mode={}, align_corners={}, align_mode={}, data_format={}{}'.format(
            main_str, self.mode, self.align_corners, self.align_mode,
            self.data_format, name_str)

X
xiaoting 已提交
387

Z
zhiboniu 已提交
388
class UpsamplingNearest2D(Layer):
X
xiaoting 已提交
389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405
    """
    This op upsamples a batch of images, using nearest neighbours' pixel values.
    The input must be a 4-D Tensor of the shape (num_batches, channels, in_h, in_w),
    where in_w is width of the input tensor, in_h is the height of the input tensor.
    And the upsampling only applies on the two dimensions(height and width).
    Nearest neighbor interpolation is to perform nearest neighbor interpolation
    in both the 3rd dimension(in height direction) and the 4th dimension(in width
    direction) on input tensor.

    For details of nearest neighbor interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation.

    Parameters:
        x (Tensor): 4-D Tensor, its data type is float32, float64, or uint8,
                          its data format is specified by :attr:`data_format`.
        size (list|tuple|Tensor|None): Output shape of image resize
             layer, the shape is (out_h, out_w) when input is a 4-D Tensor.
406
             Default: None. If a list/tuple, each element can be an integer or a Tensor of shape: [1].
X
xiaoting 已提交
407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430
             If a Tensor , its dimensions size should be a 1.
        scale_factor (float|int|list|tuple|Tensor|None): The multiplier for the input height or width. At
             least one of :attr:`size` or :attr:`scale_factor` must be set.
             And :attr:`size` has a higher priority than :attr:`scale_factor`.
             Has to match input size if it is either a list or a tuple or a Tensor.
             Default: None.
        data_format (str, optional): Specify the data format of the input, and the data format of the output
            will be consistent with that of the input. An optional string from:`NCW`, `NWC`, `"NCHW"`, `"NHWC"`, `"NCDHW"`,
            `"NDHWC"`. The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`. When it is `"NCHW"`, the data is stored
            in the order of: `[batch_size, input_channels, input_depth, input_height, input_width]`.
        name(str, optional): The default value is None.
                             Normally there is no need for user to set this property.
                             For more information, please refer to :ref:`api_guide_Name`
    Returns:
        A 4-D Tensor of the shape (num_batches, channels, out_h, out_w) or (num_batches, out_h, out_w, channels),


    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn as nn

X
xiaoting 已提交
431
            input_data = paddle.rand(shape=(2,3,6,10)).astype("float32")
X
xiaoting 已提交
432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462
            upsample_out  = paddle.nn.UpsamplingNearest2D(size=[12,12])
            input = paddle.to_tensor(input_data)
            output = upsample_out(x=input)
            print(output.shape)
            # [2L, 3L, 12L, 12L]
    """

    def __init__(self,
                 size=None,
                 scale_factor=None,
                 data_format='NCHW',
                 name=None):
        super(UpsamplingNearest2D, self).__init__()
        self.size = size
        self.scale_factor = scale_factor
        self.data_format = data_format
        self.name = name

    def forward(self, x):
        out = F.interpolate(
            x,
            size=self.size,
            scale_factor=self.scale_factor,
            mode='nearest',
            align_corners=False,
            align_mode=0,
            data_format=self.data_format,
            name=self.name)

        return out

463 464 465 466 467 468 469 470 471
    def extra_repr(self):
        if self.scale_factor is not None:
            main_str = 'scale_factor={}'.format(self.scale_factor)
        else:
            main_str = 'size={}'.format(self.size)
        name_str = ', name={}'.format(self.name) if self.name else ''
        return '{}, data_format={}{}'.format(main_str, self.data_format,
                                             name_str)

X
xiaoting 已提交
472

Z
zhiboniu 已提交
473
class UpsamplingBilinear2D(Layer):
X
xiaoting 已提交
474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492
    """
    This op upsamples a batch of images, using bilinear' pixel values.
    The input must be a 4-D Tensor of the shape (num_batches, channels, in_h, in_w),
    where in_w is width of the input tensor, in_h is the height of the input tensor.
    And the upsampling only applies on the two dimensions(height and width).
    Bilinear interpolation is an extension of linear interpolation for
    interpolating functions of two variables (e.g. H-direction and
    W-direction in this op) on a rectilinear 2D grid. The key idea is
    to perform linear interpolation first in one direction, and then
    again in the other direction.

    For details of bilinear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Bilinear_interpolation.

    Parameters:
        x (Tensor): 4-D Tensor, its data type is float32, float64, or uint8,
                          its data format is specified by :attr:`data_format`.
        size (list|tuple|Tensor|None): Output shape of image resize
             layer, the shape is (out_h, out_w) when input is a 4-D Tensor.
493
             Default: None. If a list/tuple, each element can be an integer or a Tensor  of shape: [1].
X
xiaoting 已提交
494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516
             If a Tensor , its dimensions size should be a 1.
        scale_factor (float|int|list|tuple|Tensor|None): The multiplier for the input height or width. At
             least one of :attr:`size` or :attr:`scale_factor` must be set.
             And :attr:`size` has a higher priority than :attr:`scale_factor`.
             Has to match input size if it is either a list or a tuple or a Tensor.
             Default: None.
        data_format (str, optional): Specify the data format of the input, and the data format of the output
            will be consistent with that of the input. An optional string from:`NCW`, `NWC`, `"NCHW"`, `"NHWC"`, `"NCDHW"`,
            `"NDHWC"`. The default is `"NCHW"`. When it is `"NCHW"`, the data is stored in the order of:
            `[batch_size, input_channels, input_height, input_width]`. When it is `"NCHW"`, the data is stored
            in the order of: `[batch_size, input_channels, input_depth, input_height, input_width]`.
        name(str, optional): The default value is None.
                             Normally there is no need for user to set this property.
                             For more information, please refer to :ref:`api_guide_Name`
    Returns:
        A 4-D Tensor of the shape (num_batches, channels, out_h, out_w) or (num_batches, out_h, out_w, channels),

    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn as nn

X
xiaoting 已提交
517
            input_data = paddle.rand(shape=(2,3,6,10)).astype("float32")
X
xiaoting 已提交
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
            upsample_out  = paddle.nn.UpsamplingBilinear2D(size=[12,12])
            input = paddle.to_tensor(input_data)
            output = upsample_out(x=input)
            print(output.shape)
            # [2L, 3L, 12L, 12L]
    """

    def __init__(self,
                 size=None,
                 scale_factor=None,
                 data_format='NCHW',
                 name=None):
        super(UpsamplingBilinear2D, self).__init__()
        self.size = size
        self.scale_factor = scale_factor
        self.data_format = data_format
        self.name = name

    def forward(self, x):
        out = F.interpolate(
            x,
            size=self.size,
            scale_factor=self.scale_factor,
            mode='bilinear',
            align_corners=True,
            align_mode=0,
            data_format=self.data_format,
            name=self.name)
X
xiaoting 已提交
546 547 548

        return out

549 550 551 552 553 554 555 556 557
    def extra_repr(self):
        if self.scale_factor is not None:
            main_str = 'scale_factor={}'.format(self.scale_factor)
        else:
            main_str = 'size={}'.format(self.size)
        name_str = ', name={}'.format(self.name) if self.name else ''
        return '{}, data_format={}{}'.format(main_str, self.data_format,
                                             name_str)

X
xiaoting 已提交
558

Z
zhiboniu 已提交
559
class Bilinear(Layer):
560
    r"""
561 562 563 564

    This layer performs bilinear on two inputs.

    .. math::
565

566
      out_{i} = x1 * W_{i} * {x2^\mathrm{T}}, i=0,1,...,outfeatures-1
567

568 569 570 571 572 573
      out = out + b

    In this formula:
     - :math:`x1`: the first input contains in1_features elements, shape is [batch_size, in1_features].
     - :math:`x2`: the second input contains in2_features elements, shape is [batch_size, in2_features].
     - :math:`W_{i}`: the i-th learned weight, shape is [in1_features, in2_features], and learned weight's shape is [out_features, in1_features, in2_features].
574
     - :math:`out_{i}`: the i-th element of out, shape is [batch_size], and out's shape is [batch_size, out_features].
575 576 577 578 579 580 581
     - :math:`b`: the learned bias, shape is [1, out_features].
     - :math:`x2^\mathrm{T}`: the transpose of :math:`x2`.

    Parameters:
       in1_features (int): The dimension of each first input(`x1`).
       in2_features (int): The dimension of each second input(`x2`).
       out_features (int): The dimension of output of this layer.
T
tangwei12 已提交
582
       weight_attr (ParamAttr, optional): The parameter attribute for the learnable w, parameters/weights of
583 584 585
       this layer. The default value is None.
       bias_attr (ParamAttr, optional): The parameter attribute for the bias
           of this layer. If it is set to False, no bias will be added to the output units.
T
tangwei12 已提交
586
           If it is set to None, the bias is initialized zero. The default value is None.
587 588 589 590 591 592 593 594 595
       name (str, optional): The default value is None. Normally there is no need for user
           to set this property. For more information, please refer to :ref:`api_guide_Name`. Default: None.

    Attribute:
        **weight** (Parameter): the learnable weights of this layer.

        **bias** (Parameter): the learnable bias of this layer.

    Returns:
596
       Tensor: A 2-D Tensor of shape [batch_size, out_features].
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

    Examples:
       .. code-block:: python

        import paddle
        import numpy

        layer1 = numpy.random.random((5, 5)).astype('float32')
        layer2 = numpy.random.random((5, 4)).astype('float32')
        bilinear = paddle.nn.Bilinear(
            in1_features=5, in2_features=4, out_features=1000)
        result = bilinear(paddle.to_tensor(layer1),
                        paddle.to_tensor(layer2))     # result shape [5, 1000]

    """

    def __init__(self,
                 in1_features,
                 in2_features,
                 out_features,
                 weight_attr=None,
                 bias_attr=None,
                 name=None):
        super(Bilinear, self).__init__()
        self._weight_attr = weight_attr
        self._bias_attr = bias_attr
        self._name = name
        self._in1_features = in1_features
        self._in2_features = in2_features
        self._out_features = out_features
        self._dtype = self._helper.get_default_dtype()

        weight_shape = [
            self._out_features, self._in1_features, self._in2_features
        ]
        self.weight = self.create_parameter(
            attr=self._weight_attr,
            shape=weight_shape,
            dtype=self._dtype,
            is_bias=False)
        bias_shape = [1, self._out_features]
        self.bias = self.create_parameter(
            attr=self._bias_attr,
            shape=bias_shape,
            dtype=self._dtype,
            is_bias=True)

    def forward(self, x1, x2):
        return F.bilinear(x1, x2, self.weight, self.bias, self._name)

647 648 649 650 651 652
    def extra_repr(self):
        name_str = ', name={}'.format(self._name) if self._name else ''
        return 'in1_features={}, in2_features={}, out_features={}, dtype={}{}'.format(
            self._in1_features, self._in2_features, self._out_features,
            self._dtype, name_str)

653

Z
zhiboniu 已提交
654
class Dropout(Layer):
655 656 657
    """
    Dropout is a regularization technique for reducing overfitting by preventing
    neuron co-adaption during training as described in the paper:
T
tangwei12 已提交
658
    `Improving neural networks by preventing co-adaptation of feature detectors <https://arxiv.org/abs/1207.0580>`_
659 660 661 662
    The dropout operator randomly sets the outputs of some units to zero, while upscale others
    according to the given dropout probability.

    See ``paddle.nn.functional.dropout`` for more details.
663 664

    In dygraph mode, please use ``eval()`` to switch to evaluation mode, where dropout is disabled.
665 666

    Parameters:
667 668
        p (float|int): Probability of setting units to zero. Default: 0.5
        axis (int|list|tuple): The axis along which the dropout is performed. Default None.
669 670 671 672 673 674 675 676 677 678 679
        mode(str, optional): ['upscale_in_train'(default) | 'downscale_in_infer']

                               1. upscale_in_train(default), upscale the output at training time

                                  - train: out = input * mask / ( 1.0 - p )
                                  - inference: out = input

                               2. downscale_in_infer, downscale the output at inference

                                  - train: out = input * mask
                                  - inference: out = input * (1.0 - p)
680
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
681 682 683 684 685

    Shape:
        - input: N-D tensor.
        - output: N-D tensor, the same shape as input.

686

687 688
    Examples:
        .. code-block:: python
689

690 691 692 693 694 695 696 697 698
            import paddle
            import numpy as np

            x = np.array([[1,2,3], [4,5,6]]).astype('float32')
            x = paddle.to_tensor(x)
            m = paddle.nn.Dropout(p=0.5)
            y_train = m(x)
            m.eval()  # switch the model to test phase
            y_test = m(x)
699 700 701
            print(x)
            print(y_train)
            print(y_test)
702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721
   """

    def __init__(self, p=0.5, axis=None, mode="upscale_in_train", name=None):
        super(Dropout, self).__init__()

        self.p = p
        self.axis = axis
        self.mode = mode
        self.name = name

    def forward(self, input):
        out = F.dropout(
            input,
            p=self.p,
            axis=self.axis,
            training=self.training,
            mode=self.mode,
            name=self.name)
        return out

722 723 724 725 726
    def extra_repr(self):
        name_str = ', name={}'.format(self.name) if self.name else ''
        return 'p={}, axis={}, mode={}{}'.format(self.p, self.axis, self.mode,
                                                 name_str)

727

Z
zhiboniu 已提交
728
class Dropout2D(Layer):
729 730 731 732
    """
    Randomly zero out entire channels (in the batched input 4d tensor with the shape `NCHW` ,
    a channel is a 2D feature map with the shape `HW`). Each channel will be zeroed out independently
    on every forward call with probability `p` using samples from a Bernoulli distribution.
C
cnn 已提交
733
    Dropout2D will help promote independence between feature maps as described in the paper:
T
tangwei12 已提交
734
    `Efficient Object Localization Using Convolutional Networks <https://arxiv.org/abs/1411.4280>`_
735 736 737

    See ``paddle.nn.functional.dropout2d`` for more details.

738 739
    In dygraph mode, please use ``eval()`` to switch to evaluation mode, where dropout is disabled.

740 741
    Parameters:
        p (float, optional): Probability of setting units to zero. Default: 0.5
742
        data_format (str, optional): Specify the data format of the input, and the data format of the output will be consistent with that of the input. An optional string from `NCHW` or `NHWC`. The default is `NCHW`. When it is `NCHW`, the data is stored in the order of: [batch_size, input_channels, input_height, input_width].
743 744 745 746 747 748
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Shape:
        - input: 4-D tensor.
        - output: 4-D tensor, the same shape as input.

749

750 751
    Examples:
        .. code-block:: python
752

753 754 755 756 757
            import paddle
            import numpy as np

            x = np.random.random(size=(2, 3, 4, 5)).astype('float32')
            x = paddle.to_tensor(x)
C
cnn 已提交
758
            m = paddle.nn.Dropout2D(p=0.5)
759 760 761
            y_train = m(x)
            m.eval()  # switch the model to test phase
            y_test = m(x)
762 763 764
            print(x)
            print(y_train)
            print(y_test)
765 766 767
   """

    def __init__(self, p=0.5, data_format='NCHW', name=None):
C
cnn 已提交
768
        super(Dropout2D, self).__init__()
769 770 771 772 773 774 775 776 777 778 779 780 781 782

        self.p = p
        self.data_format = data_format
        self.name = name

    def forward(self, input):
        out = F.dropout2d(
            input,
            p=self.p,
            training=self.training,
            data_format=self.data_format,
            name=self.name)
        return out

783 784 785 786 787
    def extra_repr(self):
        name_str = ', name={}'.format(self.name) if self.name else ''
        return 'p={}, data_format={}{}'.format(self.p, self.data_format,
                                               name_str)

788

Z
zhiboniu 已提交
789
class Dropout3D(Layer):
790 791 792 793
    """
    Randomly zero out entire channels (in the batched input 5d tensor with the shape `NCDHW` ,
    a channel is a 3D feature map with the shape `DHW` ). Each channel will be zeroed out independently
    on every forward call with probability `p` using samples from a Bernoulli distribution.
C
cnn 已提交
794
    Dropout3D will help promote independence between feature maps as described in the paper:
T
tangwei12 已提交
795
    `Efficient Object Localization Using Convolutional Networks <https://arxiv.org/abs/1411.4280>`_
796 797 798

    See ``paddle.nn.functional.dropout3d`` for more details.

799 800
    In dygraph mode, please use ``eval()`` to switch to evaluation mode, where dropout is disabled.

801 802
    Parameters:
        p (float | int): Probability of setting units to zero. Default: 0.5
803
        data_format (str, optional): Specify the data format of the input, and the data format of the output will be consistent with that of the input. An optional string from `NCDHW` or `NDHWC`. The default is `NCDHW`. When it is `NCDHW`, the data is stored in the order of: [batch_size, input_channels, input_depth, input_height, input_width].
804 805 806 807 808 809
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Shape:
        - input: 5-D tensor.
        - output: 5-D tensor, the same shape as input.

810

811 812
    Examples:
        .. code-block:: python
813

814 815 816 817 818
            import paddle
            import numpy as np

            x = np.random.random(size=(2, 3, 4, 5, 6)).astype('float32')
            x = paddle.to_tensor(x)
C
cnn 已提交
819
            m = paddle.nn.Dropout3D(p=0.5)
820 821 822
            y_train = m(x)
            m.eval()  # switch the model to test phase
            y_test = m(x)
823 824 825
            print(x)
            print(y_train)
            print(y_test)
826 827 828
   """

    def __init__(self, p=0.5, data_format='NCDHW', name=None):
C
cnn 已提交
829
        super(Dropout3D, self).__init__()
830 831 832 833 834 835 836 837 838 839 840 841 842 843

        self.p = p
        self.data_format = data_format
        self.name = name

    def forward(self, input):
        out = F.dropout3d(
            input,
            p=self.p,
            training=self.training,
            data_format=self.data_format,
            name=self.name)
        return out

844 845 846 847 848
    def extra_repr(self):
        name_str = ', name={}'.format(self.name) if self.name else ''
        return 'p={}, data_format={}{}'.format(self.p, self.data_format,
                                               name_str)

849

Z
zhiboniu 已提交
850
class AlphaDropout(Layer):
851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871
    """
    Alpha Dropout is a type of Dropout that maintains the self-normalizing property. For an input with
    zero mean and unit standard deviation, the output of Alpha Dropout maintains the original mean and
    standard deviation of the input. Alpha Dropout fits well to SELU activate function by randomly setting
    activations to the negative saturation value.

    For more information, please refer to:
    `Self-Normalizing Neural Networks <https://arxiv.org/abs/1706.02515>`_

    In dygraph mode, please use ``eval()`` to switch to evaluation mode, where dropout is disabled.

    Parameters:
        p (float | int): Probability of setting units to zero. Default: 0.5
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Shape:
        - input: N-D tensor.
        - output: N-D tensor, the same shape as input.

    Examples:
        .. code-block:: python
872

873 874 875 876 877 878 879 880 881
            import paddle
            import numpy as np

            x = np.array([[-1, 1], [-1, 1]]).astype('float32')
            x = paddle.to_tensor(x)
            m = paddle.nn.AlphaDropout(p=0.5)
            y_train = m(x)
            m.eval()  # switch the model to test phase
            y_test = m(x)
882 883
            print(x)
            print(y_train)
884
            # [[-0.10721093, 1.6655989 ], [-0.7791938, -0.7791938]] (randomly)
885
            print(y_test)
886 887 888 889 890 891 892 893 894 895 896 897
   """

    def __init__(self, p=0.5, name=None):
        super(AlphaDropout, self).__init__()
        self.p = p
        self.name = name

    def forward(self, input):
        out = F.alpha_dropout(
            input, p=self.p, training=self.training, name=self.name)
        return out

898 899 900 901
    def extra_repr(self):
        name_str = ', name={}'.format(self.name) if self.name else ''
        return 'p={}{}'.format(self.p, name_str)

902

Z
zhiboniu 已提交
903
class Pad1D(Layer):
L
littletomatodonkey 已提交
904
    """
L
littletomatodonkey 已提交
905 906 907
    This interface is used to construct a callable object of the ``Pad1D`` class.
    Pad tensor according to 'pad', 'mode' and 'value'.
    If mode is 'reflect', pad[0] and pad[1] must be no greater than width-1.
L
littletomatodonkey 已提交
908 909

    Parameters:
910 911
        padding (Tensor | List[int] | int): The padding size with data type int. If is int, use the
            same padding in both dimensions. Else [len(padding)/2] dimensions
L
littletomatodonkey 已提交
912
            of input will be padded. The pad has the form (pad_left, pad_right).
L
littletomatodonkey 已提交
913 914 915 916 917 918
        mode (str): Four modes: 'constant' (default), 'reflect', 'replicate', 'circular'.
            When in 'constant' mode, this op uses a constant value to pad the input tensor.
            When in 'reflect' mode, uses reflection of the input boundaries to pad the input tensor.
            When in 'replicate' mode, uses input boundaries to pad the input tensor.
            When in 'circular' mode, uses circular input to pad the input tensor.
            Default is 'constant'.
L
littletomatodonkey 已提交
919 920 921 922 923
        value (float32): The value to fill the padded areas. Default is 0.0
        data_format (str): An string from: "NCL", "NLC". Specify the data format of the input data.
           Default is  "NCL"
        name (str, optional) : The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
924 925

    Returns:
L
littletomatodonkey 已提交
926 927 928 929 930 931 932 933
        None

    Examples:
        .. code-block:: text

            x = [[[1., 2., 3.],
                  [4., 5., 6.]]]
            padding = [1, 2],
L
littletomatodonkey 已提交
934
            mode = "constant"
L
littletomatodonkey 已提交
935 936 937 938 939 940
            value = 0.0
            Out = [[[0. 1. 2. 3. 0. 0.]
                    [0. 4. 5. 6. 0. 0.]]]

    Code Examples:
        .. code-block:: python
941

L
littletomatodonkey 已提交
942 943 944 945 946 947
            import paddle
            import paddle.nn as nn
            import numpy as np

            input_shape = (1, 2, 3)
            pad = [1, 2]
L
littletomatodonkey 已提交
948 949 950
            mode = "constant"
            data = paddle.arange(np.prod(input_shape), dtype="float32").reshape(input_shape) + 1
            my_pad = nn.Pad1D(padding=pad, mode=mode)
L
littletomatodonkey 已提交
951
            result = my_pad(data)
L
littletomatodonkey 已提交
952
            print(result)
L
littletomatodonkey 已提交
953 954 955 956
            # [[[0. 1. 2. 3. 0. 0.]
            #   [0. 4. 5. 6. 0. 0.]]]
    """

L
littletomatodonkey 已提交
957 958 959 960 961 962 963
    def __init__(self,
                 padding,
                 mode='constant',
                 value=0.0,
                 data_format="NCL",
                 name=None):
        super(Pad1D, self).__init__()
964
        self._pad = _npairs(padding, 1)
L
littletomatodonkey 已提交
965
        self._mode = mode
L
littletomatodonkey 已提交
966
        self._value = value
L
littletomatodonkey 已提交
967
        self._data_format = data_format
L
littletomatodonkey 已提交
968 969 970 971 972 973 974 975 976 977
        self._name = name

    def forward(self, x):
        return F.pad(x,
                     pad=self._pad,
                     mode=self._mode,
                     value=self._value,
                     data_format=self._data_format,
                     name=self._name)

978 979 980 981 982
    def extra_repr(self):
        name_str = ', name={}'.format(self._name) if self._name else ''
        return 'padding={}, mode={}, value={}, data_format={}{}'.format(
            self._pad, self._mode, self._value, self._data_format, name_str)

L
littletomatodonkey 已提交
983

Z
zhiboniu 已提交
984
class Pad2D(Layer):
L
littletomatodonkey 已提交
985
    """
L
littletomatodonkey 已提交
986 987 988 989
    This interface is used to construct a callable object of the ``Pad2D`` class.
    Pad tensor according to 'pad', 'mode' and 'value'.
    If mode is 'reflect', pad[0] and pad[1] must be no greater
    than width-1. The height dimension has the same condition.
L
littletomatodonkey 已提交
990 991

    Parameters:
992 993 994
        padding (Tensor | List[int] | int): The padding size with data type int. If is int, use the
            same padding in all dimensions. Else [len(padding)/2] dimensions of input will be padded. 
            The pad has the form (pad_left, pad_right, pad_top, pad_bottom). 
L
littletomatodonkey 已提交
995 996 997 998 999 1000
        mode (str): Four modes: 'constant' (default), 'reflect', 'replicate', 'circular'.
            When in 'constant' mode, this op uses a constant value to pad the input tensor.
            When in 'reflect' mode, uses reflection of the input boundaries to pad the input tensor.
            When in 'replicate' mode, uses input boundaries to pad the input tensor.
            When in 'circular' mode, uses circular input to pad the input tensor.
            Default is 'constant'.
L
littletomatodonkey 已提交
1001 1002 1003 1004 1005
        value (float32): The value to fill the padded areas. Default is 0.0
        data_format (str): An string from: "NCHW", "NHWC". Specify the data format of the input data.
           Default is  "NCHW"
        name (str, optional) : The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
1006 1007

    Returns:
L
littletomatodonkey 已提交
1008 1009 1010 1011 1012 1013 1014 1015
        None

    Examples:
        .. code-block:: text

            x = [[[[1., 2., 3.],
                   [4., 5., 6.]]]]
            padding = [1, 1, 0, 0]
L
littletomatodonkey 已提交
1016
            mode = "constant"
L
littletomatodonkey 已提交
1017 1018 1019 1020 1021 1022
            value = 0.0
            Out = [[[[0. 1. 2. 3. 0.]
                     [0. 4. 5. 6. 0.]]]]

    Code Examples:
        .. code-block:: python
1023

L
littletomatodonkey 已提交
1024 1025 1026 1027 1028
            import paddle
            import paddle.nn as nn
            import numpy as np
            input_shape = (1, 1, 2, 3)
            pad = [1, 0, 1, 2]
L
littletomatodonkey 已提交
1029 1030 1031
            mode = "constant"
            data = paddle.arange(np.prod(input_shape), dtype="float32").reshape(input_shape) + 1
            my_pad = nn.Pad2D(padding=pad, mode=mode)
L
littletomatodonkey 已提交
1032
            result = my_pad(data)
L
littletomatodonkey 已提交
1033
            print(result)
L
littletomatodonkey 已提交
1034 1035 1036 1037 1038 1039 1040
            # [[[[0. 0. 0. 0.]
            #    [0. 1. 2. 3.]
            #    [0. 4. 5. 6.]
            #    [0. 0. 0. 0.]
            #    [0. 0. 0. 0.]]]]
    """

L
littletomatodonkey 已提交
1041 1042 1043 1044 1045 1046 1047
    def __init__(self,
                 padding,
                 mode='constant',
                 value=0.0,
                 data_format="NCHW",
                 name=None):
        super(Pad2D, self).__init__()
1048
        self._pad = _npairs(padding, 2)
L
littletomatodonkey 已提交
1049
        self._mode = mode
L
littletomatodonkey 已提交
1050 1051 1052 1053 1054 1055 1056 1057
        self._value = value
        self._data_format = data_format
        self._name = name

    def forward(self, x):
        return F.pad(x,
                     pad=self._pad,
                     mode=self._mode,
L
littletomatodonkey 已提交
1058
                     value=self._value,
L
littletomatodonkey 已提交
1059 1060 1061
                     data_format=self._data_format,
                     name=self._name)

1062 1063 1064 1065 1066
    def extra_repr(self):
        name_str = ', name={}'.format(self._name) if self._name else ''
        return 'padding={}, mode={}, value={}, data_format={}{}'.format(
            self._pad, self._mode, self._value, self._data_format, name_str)

L
littletomatodonkey 已提交
1067

Z
zhiboniu 已提交
1068
class Pad3D(Layer):
L
littletomatodonkey 已提交
1069
    """
L
littletomatodonkey 已提交
1070 1071 1072 1073
    This interface is used to construct a callable object of the ``Pad3D`` class.
    Pad tensor according to 'pad', 'mode' and 'value'.
    If mode is 'reflect', pad[0] and pad[1] must be no greater
    than width-1. The height and depth dimension has the same condition.
L
littletomatodonkey 已提交
1074 1075

    Parameters:
1076 1077
        padding (Tensor | List[int] | int): The padding size with data type int. If is int, use the
            same padding in all dimensions. Else [len(padding)/2] dimensions
L
littletomatodonkey 已提交
1078
            of input will be padded. The pad has the form (pad_left, pad_right, pad_top, pad_bottom, pad_front, pad_back).
L
littletomatodonkey 已提交
1079 1080 1081 1082 1083 1084
        mode (str): Four modes: 'constant' (default), 'reflect', 'replicate', 'circular'.
            When in 'constant' mode, this op uses a constant value to pad the input tensor.
            When in 'reflect' mode, uses reflection of the input boundaries to pad the input tensor.
            When in 'replicate' mode, uses input boundaries to pad the input tensor.
            When in 'circular' mode, uses circular input to pad the input tensor.
            Default is 'constant'.
L
littletomatodonkey 已提交
1085 1086 1087 1088 1089
        value (float32): The value to fill the padded areas. Default is 0.0
        data_format (str): An string from: "NCDHW", "NDHWC". Specify the data format of the input data.
           Default is  "NCDHW"
        name (str, optional) : The default value is None.  Normally there is no need for
            user to set this property.  For more information, please refer to :ref:`api_guide_Name`.
1090 1091

    Returns:
L
littletomatodonkey 已提交
1092 1093 1094 1095 1096 1097 1098 1099
        None

    Examples:
        .. code-block:: text

            x = [[[[[1., 2., 3.],
                    [4., 5., 6.]]]]]
            padding = [1, 2, 0, 0, 0, 0]
L
littletomatodonkey 已提交
1100
            mode = "constant"
L
littletomatodonkey 已提交
1101 1102 1103 1104 1105 1106
            value = 0.0
            Out = [[[[[0. 1. 2. 3. 0. 0.]
                      [0. 4. 5. 6. 0. 0.]]]]]

    Code Examples:
        .. code-block:: python
1107

L
littletomatodonkey 已提交
1108 1109 1110 1111 1112
            import paddle
            import paddle.nn as nn
            import numpy as np
            input_shape = (1, 1, 1, 2, 3)
            pad = [1, 0, 1, 2, 0, 0]
L
littletomatodonkey 已提交
1113 1114 1115
            mode = "constant"
            data = paddle.arange(np.prod(input_shape), dtype="float32").reshape(input_shape) + 1
            my_pad = nn.Pad3D(padding=pad, mode=mode)
L
littletomatodonkey 已提交
1116
            result = my_pad(data)
L
littletomatodonkey 已提交
1117
            print(result)
L
littletomatodonkey 已提交
1118 1119 1120 1121 1122 1123 1124
            # [[[[[0. 0. 0. 0.]
            #     [0. 1. 2. 3.]
            #     [0. 4. 5. 6.]
            #     [0. 0. 0. 0.]
            #     [0. 0. 0. 0.]]]]]
    """

L
littletomatodonkey 已提交
1125 1126 1127 1128 1129 1130 1131
    def __init__(self,
                 padding,
                 mode='constant',
                 value=0.0,
                 data_format="NCDHW",
                 name=None):
        super(Pad3D, self).__init__()
1132
        self._pad = _npairs(padding, 3)
L
littletomatodonkey 已提交
1133
        self._mode = mode
L
littletomatodonkey 已提交
1134 1135 1136 1137 1138 1139 1140 1141
        self._value = value
        self._data_format = data_format
        self._name = name

    def forward(self, x):
        return F.pad(x,
                     pad=self._pad,
                     mode=self._mode,
L
littletomatodonkey 已提交
1142
                     value=self._value,
L
littletomatodonkey 已提交
1143 1144 1145
                     data_format=self._data_format,
                     name=self._name)

1146 1147 1148 1149 1150
    def extra_repr(self):
        name_str = ', name={}'.format(self._name) if self._name else ''
        return 'padding={}, mode={}, value={}, data_format={}{}'.format(
            self._pad, self._mode, self._value, self._data_format, name_str)

L
littletomatodonkey 已提交
1151

Z
zhiboniu 已提交
1152
class CosineSimilarity(Layer):
L
littletomatodonkey 已提交
1153
    """
1154
    This interface is used to compute cosine similarity between x1 and x2 along axis.
L
littletomatodonkey 已提交
1155 1156

    Parameters:
1157
        axis (int): Dimension of vectors to compute cosine similarity. Default is 1.
L
littletomatodonkey 已提交
1158
        eps(float): Small value to avoid division by zero. Default is 1e-8.
1159
    Returns:
L
littletomatodonkey 已提交
1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173
        None

    Examples:
        .. code-block:: text

            Case 0:
                x1 = [[0.8024077  0.9927354  0.27238318 0.8344984 ]
                     [0.48949873 0.5797396  0.65444374 0.66510963]
                     [0.1031398  0.9614342  0.08365563 0.6796464 ]
                     [0.10760343 0.7461209  0.7726148  0.5801006 ]]
                x2 = [[0.62913156 0.1536727  0.9847992  0.04591406]
                     [0.9098952  0.15715368 0.8671125  0.3156102 ]
                     [0.4427798  0.54136837 0.5276275  0.32394758]
                     [0.3769419  0.8535014  0.48041078 0.9256797 ]]
1174
                axis = 1
L
littletomatodonkey 已提交
1175 1176 1177 1178 1179
                eps = 1e-8
                Out: [0.5275037  0.8368967  0.75037485 0.9245899]

    Code Examples:
        .. code-block:: python
1180

L
littletomatodonkey 已提交
1181 1182 1183 1184 1185 1186 1187 1188 1189 1190
            import paddle
            import paddle.nn as nn
            import numpy as np

            np.random.seed(0)
            x1 = np.random.rand(2,3)
            x2 = np.random.rand(2,3)
            x1 = paddle.to_tensor(x1)
            x2 = paddle.to_tensor(x2)

1191
            cos_sim_func = nn.CosineSimilarity(axis=0)
L
littletomatodonkey 已提交
1192
            result = cos_sim_func(x1, x2)
L
littletomatodonkey 已提交
1193
            print(result)
L
littletomatodonkey 已提交
1194 1195 1196
            # [0.99806249 0.9817672  0.94987036]
    """

1197
    def __init__(self, axis=1, eps=1e-8):
L
littletomatodonkey 已提交
1198
        super(CosineSimilarity, self).__init__()
1199
        self._axis = axis
L
littletomatodonkey 已提交
1200 1201 1202
        self._eps = eps

    def forward(self, x1, x2):
1203
        return F.cosine_similarity(x1, x2, axis=self._axis, eps=self._eps)
T
tangwei12 已提交
1204

1205 1206 1207
    def extra_repr(self):
        return 'axis={_axis}, eps={_eps}'.format(**self.__dict__)

T
tangwei12 已提交
1208

Z
zhiboniu 已提交
1209
class Embedding(Layer):
1210
    r"""
T
tangwei12 已提交
1211 1212 1213 1214
    **Embedding Layer**

    This interface is used to construct a callable object of the ``Embedding`` class.
    For specific usage, refer to code examples. It implements the function of the Embedding Layer.
T
tangwei12 已提交
1215
    This layer is used to lookup embeddings vector of ids provided by :attr:`x` .
T
tangwei12 已提交
1216
    It automatically constructs a 2D embedding matrix based on the
T
tangwei12 已提交
1217
    input :attr:`num_embeddings` and :attr:`embedding_dim`.
T
tangwei12 已提交
1218 1219 1220 1221

    The shape of output Tensor is generated by appending an emb_size dimension to the
    last dimension of the input Tensor shape.

T
tangwei12 已提交
1222
    **Note:** The id in :attr:`x` must satisfy :math:`0 =< id < num_embeddings` ,
T
tangwei12 已提交
1223 1224 1225 1226 1227 1228
    otherwise the program will throw an exception and exit.

    .. code-block:: text

        Case 1:

T
tangwei12 已提交
1229 1230 1231
        x is a Tensor. padding_idx = -1
            x.data = [[1, 3], [2, 4], [4, 127]
            x.shape = [3, 2]
T
tangwei12 已提交
1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248
        Given size = [128, 16]
        output is a Tensor:
            out.shape = [3, 2, 16]
            out.data = [[[0.129435295, 0.244512452, ..., 0.436322452],
                        [0.345421456, 0.524563927, ..., 0.144534654]],

                        [[0.345249859, 0.124939536, ..., 0.194353745],
                        [0.945345345, 0.435394634, ..., 0.435345365]],

                        [[0.945345345, 0.435394634, ..., 0.435345365],
                        [0.0,         0.0,         ..., 0.0        ]]]  # padding data
        The input padding_idx is less than 0, it is automatically converted to padding_idx = -1 + 128 = 127
        It will pad all-zero data when ids is 127.

    Parameters:
        num_embeddings (int): Just one element which indicate the size
            of the dictionary of embeddings.
T
tangwei12 已提交
1249
        embedding_dim (int):  Just one element which indicate the size of each embedding vector respectively.
T
tangwei12 已提交
1250
        padding_idx(int|long|None): padding_idx needs to be in the interval [-num_embeddings, num_embeddings).
T
tangwei12 已提交
1251 1252 1253 1254 1255 1256 1257
            If :math:`padding\_idx < 0`, the :math:`padding\_idx` will automatically be converted
            to :math:`vocab\_size + padding\_idx` . It will output all-zero padding data whenever lookup
            encounters :math:`padding\_idx` in id. And the padding data will not be updated while training.
            If set None, it makes no effect to output. Default: None.
        sparse(bool): The flag indicating whether to use sparse update. This parameter only
            affects the performance of the backwards gradient update. It is recommended to set
            True because sparse update is faster. But some optimizer does not support sparse update,
T
tangwei12 已提交
1258
            such as :ref:`api_paddle_optimizer_adadelta_Adadelta` , :ref:`api_paddle_optimizer_adamax_Adamax` , :ref:`api_paddle_optimizer_lamb_Lamb`.
T
tangwei12 已提交
1259
            In these case, sparse must be False. Default: False.
T
tangwei12 已提交
1260
        weight_attr(ParamAttr): To specify the weight parameter property. Default: None, which means the
T
tangwei12 已提交
1261
            default weight parameter property is used. See usage for details in :ref:`api_ParamAttr` . In addition,
T
tangwei12 已提交
1262 1263
            user-defined or pre-trained word vectors can be loaded with the :attr:`param_attr` parameter.
            The local word vector needs to be transformed into numpy format, and the shape of local word
T
tangwei12 已提交
1264 1265
            vector should be consistent with :attr:`num_embeddings` . Then :ref:`api_initializer_NumpyArrayInitializer`
            is used to load custom or pre-trained word vectors. See code example for details.
T
tangwei12 已提交
1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279
        name(str|None): For detailed information, please refer
               to :ref:`api_guide_Name`. Usually name is no need to set and
               None by default.

    Attribute:
        **weight** (Parameter): the learnable weights of this layer.

    Returns:
        None

    Examples:

        .. code-block:: python

T
tangwei12 已提交
1280 1281 1282 1283 1284
            import paddle
            import numpy as np

            x_data = np.arange(3, 6).reshape((3, 1)).astype(np.int64)
            y_data = np.arange(6, 12).reshape((3, 2)).astype(np.float32)
T
tangwei12 已提交
1285

T
tangwei12 已提交
1286 1287 1288 1289 1290 1291 1292
            x = paddle.to_tensor(x_data, stop_gradient=False)
            y = paddle.to_tensor(y_data, stop_gradient=False)

            embedding = paddle.nn.Embedding(10, 3, sparse=True)

            w0=np.full(shape=(10, 3), fill_value=2).astype(np.float32)
            embedding.weight.set_value(w0)
T
tangwei12 已提交
1293

T
tangwei12 已提交
1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306
            adam = paddle.optimizer.Adam(parameters=[embedding.weight], learning_rate=0.01)
            adam.clear_grad()

            # weight.shape = [10, 3]

            # x.data = [[3],[4],[5]]
            # x.shape = [3, 1]

            # out.data = [[2,2,2], [2,2,2], [2,2,2]]
            # out.shape = [3, 1, 3]
            out=embedding(x)
            out.backward()
            adam.step()
T
tangwei12 已提交
1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321

    """

    def __init__(self,
                 num_embeddings,
                 embedding_dim,
                 padding_idx=None,
                 sparse=False,
                 weight_attr=None,
                 name=None):
        super(Embedding, self).__init__()
        self._num_embeddings = num_embeddings
        self._embedding_dim = embedding_dim
        self._sparse = sparse
        self._is_distributed = False
1322
        self._padding_idx = padding_idx
T
tangwei12 已提交
1323 1324 1325 1326 1327 1328 1329

        if self._num_embeddings <= 0:
            raise ValueError("num_embeddings must be gather than 0")

        if self._embedding_dim <= 0:
            raise ValueError("embedding_dim must be gather than 0")

1330 1331 1332 1333
        padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (
            num_embeddings + padding_idx)

        if padding_idx >= num_embeddings or padding_idx < -num_embeddings:
T
tangwei12 已提交
1334 1335 1336
            raise ValueError("padding_idx must be within [-{}, {})".format(
                num_embeddings, num_embeddings))

T
tangwei12 已提交
1337 1338 1339 1340 1341 1342
        self._dtype = self._helper.get_default_dtype()
        self._size = [self._num_embeddings, self._embedding_dim]

        self._weight_attr = weight_attr
        self._remote_prefetch = False
        self._name = name
T
tangwei12 已提交
1343
        self.weight = self.create_parameter(
T
tangwei12 已提交
1344 1345 1346 1347 1348
            attr=self._weight_attr,
            shape=self._size,
            dtype=self._dtype,
            is_bias=False)

T
tangwei12 已提交
1349 1350 1351
        if in_dygraph_mode() and padding_idx != -1:
            self.weight[padding_idx] = 0.0

T
tangwei12 已提交
1352 1353 1354
    def forward(self, x):
        return F.embedding(
            x,
T
tangwei12 已提交
1355
            weight=self.weight,
T
tangwei12 已提交
1356 1357 1358
            padding_idx=self._padding_idx,
            sparse=self._sparse,
            name=self._name)
1359 1360 1361 1362 1363 1364 1365 1366 1367

    def extra_repr(self):
        main_str = '{_num_embeddings}, {_embedding_dim}'
        if self._padding_idx is not None:
            main_str += ', padding_idx={_padding_idx}'
        main_str += ', sparse={_sparse}'
        if self._name is not None:
            main_str += ', name={_name}'
        return main_str.format(**self.__dict__)
F
FNRE 已提交
1368 1369


Z
zhiboniu 已提交
1370
class Unfold(Layer):
F
FNRE 已提交
1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430
    """
    This op returns a col buffer of sliding local blocks of input x, also known
    as im2col for batched 2D image tensors. For each block under the convolution filter,
    all element will be rearranged as a column. While the convolution filter sliding over
    the input feature map, a series of such columns will be formed.

    For each input :math:`x` with shape [N, C, H, W], the output shape [N, Cout, Lout]
    can be calculated as following.

    See ``paddle.nn.functional.unfold`` for more details.

    
    Parameters:
        kernel_sizes(int|list):   The size of convolution kernel, should be [k_h, k_w]
                                  or an integer k treated as [k, k].
        strides(int|list):        The strides, should be [stride_h, stride_w]
                                  or an integer stride treated as [sride, stride].
                                  For default, strides will be [1, 1].
        paddings(int|list):       The paddings of each dimension, should be
                                  [padding_top, padding_left, padding_bottom, padding_right]
                                  or [padding_h, padding_w] or an integer padding.
                                  If [padding_h, padding_w] was given, it will expanded to
                                  [padding_h, padding_w, padding_h, padding_w]. If an integer
                                  padding was given, [padding, padding, padding, padding] will
                                  be used. For default, paddings will be [0, 0, 0, 0]
        dilations(int|list):      the dilations of convolution kernel, should be
                                  [dilation_h, dilation_w], or an integer dilation treated as
                                  [dilation, dilation]. For default, it will be [1, 1].
        name(str, optional): The default value is None.
                             Normally there is no need for user to set this property.
                             For more information, please refer to :ref:`api_guide_Name`


    Examples:
        .. code-block:: python

            import paddle
            import paddle.nn as nn

            x = paddle.randn((100,3,224,224))
            unfold = nn.Unfold(kernel_sizes=[3, 3])
            result = unfold(x)
            print(result)
   """

    def __init__(self,
                 kernel_sizes,
                 dilations=1,
                 paddings=0,
                 strides=1,
                 name=None):
        super(Unfold, self).__init__()

        self.kernel_sizes = kernel_sizes
        self.dilations = dilations
        self.paddings = paddings
        self.strides = strides
        self.name = name

    def forward(self, input):
1431 1432 1433 1434 1435 1436 1437
        return F.unfold(
            input,
            kernel_sizes=self.kernel_sizes,
            strides=self.strides,
            paddings=self.paddings,
            dilations=self.dilations,
            name=self.name)
F
FNRE 已提交
1438 1439 1440 1441 1442

    def extra_repr(self):
        name_str = ', name={}'.format(self.name) if self.name else ''
        return 'kernel_size={}, dilation={}, padding={}, stride={}{}'.\
                format(self.kernel_sizes, self.dilations, self.paddings, self.strides, name_str)