common.py 67.3 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.

X
xiaoting 已提交
15
import warnings
16 17
import paddle
from ...fluid.framework import in_dygraph_mode, default_main_program
X
xiaoting 已提交
18
from paddle.fluid.layer_helper import LayerHelper
L
littletomatodonkey 已提交
19
from paddle.fluid.layers.tensor import Variable, fill_constant, zeros, concat
20 21
from ...fluid.layers import core
from ...fluid import dygraph_utils
22
# TODO: define the common functions to build a neural network  
Z
zhiboniu 已提交
23 24 25
from ...fluid.layers import unfold  # noqa: F401
from ...fluid.layers import squeeze
from ...fluid.layers import unsqueeze
Y
Yang Zhang 已提交
26 27 28
from ...tensor import clip
from ...tensor import sum
from ...tensor import sqrt
29 30
from ...fluid.data_feeder import check_variable_and_dtype, check_dtype
from ...fluid.framework import Variable, in_dygraph_mode, _varbase_creator
X
xiaoting 已提交
31

32 33
from ...fluid.framework import in_dygraph_mode
from ...fluid import core, dygraph_utils
34 35
from ...fluid import core, layers
from ...fluid.data_feeder import check_variable_and_dtype
36

37 38
__all__ = []

X
xiaoting 已提交
39

X
xiaoting 已提交
40
def interpolate(x,
41 42 43 44
                size=None,
                scale_factor=None,
                mode='nearest',
                align_corners=False,
X
xiaoting 已提交
45
                align_mode=0,
46 47
                data_format='NCHW',
                name=None):
X
xiaoting 已提交
48
    """
S
swtkiwi 已提交
49

X
xiaoting 已提交
50
    This op resizes a batch of images.
51 52
    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
X
xiaoting 已提交
53
    (num_batches, channels, in_d, in_h, in_w) or (num_batches, in_d, in_h, in_w, channels),
54 55
    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.
X
xiaoting 已提交
56
    and the resizing only applies on the three dimensions(depth, height and width).
X
xiaoting 已提交
57

X
xiaoting 已提交
58
    Supporting resample methods:
59 60 61 62 63
        'linear' : Linear interpolation
        'bilinear' : Bilinear interpolation
        'trilinear' : Trilinear interpolation
        'nearest' : Nearest neighbor interpolation
        'bicubic' : Bicubic interpolation
64
        'area': Area interpolation
65 66 67 68

    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. 
    
X
xiaoting 已提交
69 70 71 72 73 74 75 76 77 78 79 80 81 82
    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.

    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
83
    align_corners and align_mode are optional parameters,the calculation method
X
xiaoting 已提交
84 85 86 87 88 89 90
    of interpolation can be selected by them.

    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.

91 92 93 94 95 96
    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`.

X
xiaoting 已提交
97 98 99 100
    Example:

    .. code-block:: text

101
        For scale_factor:
X
xiaoting 已提交
102 103 104 105 106
            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)

107 108 109 110 111 112 113 114 115 116 117
        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}
        
X
xiaoting 已提交
118
        Nearest neighbor interpolation:
X
xiaoting 已提交
119

X
xiaoting 已提交
120 121 122 123 124
              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})
125

X
xiaoting 已提交
126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166
        Bilinear interpolation:
          if:
              align_corners = False , align_mode = 0
              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:
              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
          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}

167 168 169
    For details of linear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Linear_interpolation.
    
X
xiaoting 已提交
170 171
    For details of nearest neighbor interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation.
172
    
X
xiaoting 已提交
173 174
    For details of bilinear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Bilinear_interpolation.
175
    
X
xiaoting 已提交
176 177
    For details of trilinear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Trilinear_interpolation.
178
    
X
xiaoting 已提交
179 180
    For details of bicubic interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Bicubic_interpolation
181
    
X
xiaoting 已提交
182
    Parameters:
X
xiaoting 已提交
183
        x (Tensor): 3-D, 4-D or 5-D Tensor, its data type is float32, float64, or uint8,
X
xiaoting 已提交
184
                          its data format is specified by :attr:`data_format`.
X
xiaoting 已提交
185
        size (list|tuple|Tensor|None): Output shape of image resize
186 187
             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. 
188
             Default: None. If a list/tuple, each element can be an integer or a Tensor of shape: [1].
X
xiaoting 已提交
189
             If a Tensor, its dimensions size should be a 1.
190 191 192
        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.
X
xiaoting 已提交
193
             Default: None.
194
        mode (str): The resample method. It supports 'linear', 'area', 'nearest', 'bilinear',
195
                       'bicubic' and 'trilinear' currently. Default: 'nearest'
X
xiaoting 已提交
196 197
        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
X
xiaoting 已提交
198
                               corner pixels.This only has an effect when 'linear', 'bilinear', 'bicubic' or 'trilinear'.
199 200 201 202
                               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.
X
xiaoting 已提交
203
        data_format (str, optional): Specify the data format of the input, and the data format of the output
204
            will be consistent with that of the input. An optional string from:`NCW`, `NWC`,  `"NCHW"`, `"NHWC"`, `"NCDHW"`,
X
xiaoting 已提交
205 206 207
            `"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]`.
208 209 210
        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`
X
xiaoting 已提交
211
    Returns:
212
        A 3-D Tensor of the shape (num_batches, channels, out_w) or (num_batches, out_w, channels),
X
xiaoting 已提交
213 214 215
        A 4-D Tensor of the shape (num_batches, channels, out_h, out_w) or (num_batches, out_h, out_w, channels),
        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).
    Raises:
X
xiaoting 已提交
216
        TypeError: size should be a list or tuple or Tensor.
217
        ValueError: The 'mode' of image_resize can only be 'linear', 'bilinear',
218
                    'trilinear', 'bicubic', 'area' or 'nearest' currently.
219 220 221 222 223 224 225 226
        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.
X
xiaoting 已提交
227 228
        TypeError: align_corners should be a bool value
        ValueError: align_mode can only be '0' or '1'
229 230
        ValueError: data_format can only be 'NCW', 'NWC', 'NCHW', 'NHWC', 'NCDHW' or 'NDHWC'.

X
xiaoting 已提交
231 232 233 234 235
    Examples:
        .. code-block:: python

	    import paddle
	    import numpy as np
X
xiaoting 已提交
236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
            import paddle.nn.functional as F
            
            # given out size
            input_data = np.random.rand(2,3,6,10).astype("float32")
            x = paddle.to_tensor(input_data)
            output_1 = F.interpolate(x=x, size=[12,12])
    	    print(output_1.shape)
	    # [2L, 3L, 12L, 12L]
            
            # given scale
            output_2 = F.interpolate(x=x, scale_factor=[2,1])
            print(output_2.shape)
            # [2L, 3L, 12L, 10L]
            
            # bilinear interp
            output_3 = F.interpolate(x=x, scale_factor=[2,1], mode="bilinear")
            print(output_2.shape)
            # [2L, 3L, 12L, 10L]
X
xiaoting 已提交
254
    """
255 256 257 258 259 260 261 262 263 264
    data_format = data_format.upper()
    resample = mode.upper()
    resample_type = mode.lower()

    resample_methods = [
        'LINEAR',
        'BILINEAR',
        'TRILINEAR',
        'NEAREST',
        'BICUBIC',
265
        'AREA',
266
    ]
X
xiaoting 已提交
267 268
    if resample not in resample_methods:
        raise ValueError(
269
            "The 'resample' of image_resize can only be 'area', 'linear', 'bilinear', 'trilinear', "
270
            " 'bicubic' or 'nearest' currently.")
X
xiaoting 已提交
271

X
xiaoting 已提交
272
    if resample in ['LINEAR'] and len(x.shape) != 3:
273
        raise ValueError("'linear' only support 3-D tensor.")
274

X
xiaoting 已提交
275
    if resample in ['BILINEAR', 'NEAREST', 'BICUBIC'] and len(x.shape) != 4:
X
xiaoting 已提交
276
        raise ValueError(
277
            "'bilinear', 'bicubic' and 'nearest' only support 4-D tensor.")
X
xiaoting 已提交
278
    if resample == 'TRILINEAR' and len(x.shape) != 5:
279 280 281 282
        raise ValueError("'trilinear'only support 5-D tensor.")

    if size is None and scale_factor is None:
        raise ValueError("One of size and scale_factor must not be None.")
X
xiaoting 已提交
283 284 285

    if not isinstance(align_corners, bool):
        raise TypeError("Attr align_corners should be a bool value")
286

X
xiaoting 已提交
287 288
    if align_mode != 0 and align_mode != 1:
        raise ValueError("align_mode can only be 0 or 1")
X
xiaoting 已提交
289 290 291 292
    if align_corners != 0 and resample == 'NEAREST':
        raise ValueError(
            "align_corners option can only be set with the interpolating modes: linear | bilinear | bicubic | trilinear"
        )
293 294 295 296 297 298 299 300 301

    if resample == 'AREA' and len(x.shape) == 3:
        return paddle.nn.functional.adaptive_avg_pool1d(x, size)

    if resample == 'AREA' and len(x.shape) == 4:
        return paddle.nn.functional.adaptive_avg_pool2d(x, size)
    if resample == 'AREA' and len(x.shape) == 5:
        return paddle.nn.functional.adaptive_avg_pool3d(x, size)

X
xiaoting 已提交
302
    helper = LayerHelper('{}_interp_v2'.format(resample_type), **locals())
303
    dtype = helper.input_dtype(input_param_name='x')
X
xiaoting 已提交
304
    if len(x.shape) == 3 and data_format not in ['NCW', 'NWC']:
305 306
        raise ValueError(
            "Got wrong value for param `data_format`: " + data_format +
307
            " received but only `NCW` or `NWC` supported for 3-D input.")
X
xiaoting 已提交
308
    elif len(x.shape) == 4 and data_format not in ['NCHW', 'NHWC']:
X
xiaoting 已提交
309 310 311
        raise ValueError(
            "Got wrong value for param `data_format`: " + data_format +
            " received but only `NCHW` or `NHWC` supported for 4-D input.")
X
xiaoting 已提交
312
    elif len(x.shape) == 5 and data_format not in ['NCDHW', 'NDHWC']:
X
xiaoting 已提交
313 314 315 316 317 318 319
        raise ValueError(
            "Got wrong value for param `data_format`: " + data_format +
            " received but only `NCDHW` or `NDHWC` supported for 5-D input.")

    def _is_list_or_turple_(data):
        return (isinstance(data, list) or isinstance(data, tuple))

320
    if data_format == 'NCHW' or data_format == 'NCDHW' or data_format == 'NCW':
X
xiaoting 已提交
321
        data_layout = 'NCHW'
322
    if data_format == 'NHWC' or data_format == 'NDHWC' or data_format == 'NWC':
X
xiaoting 已提交
323 324
        data_layout = 'NHWC'

X
xiaoting 已提交
325 326 327 328
    if resample == 'NEAREST':
        align_corners = False

    inputs = {"X": x}
X
xiaoting 已提交
329 330 331 332 333 334 335 336 337 338
    attrs = {
        "out_d": -1,
        "out_h": -1,
        "out_w": -1,
        "interp_method": resample_type,
        "align_corners": align_corners,
        "align_mode": align_mode,
        "data_layout": data_layout
    }

339 340
    out_shape = size
    scale = scale_factor
341 342
    if out_shape is not None and scale is not None:
        raise ValueError("Only one of size or scale_factor should be defined.")
X
xiaoting 已提交
343
    if out_shape is not None:
344 345

        if isinstance(out_shape, Variable) and not in_dygraph_mode():
X
xiaoting 已提交
346 347
            out_shape.stop_gradient = True
            inputs['OutSize'] = out_shape
348

X
xiaoting 已提交
349
        else:
350 351 352 353 354 355
            if in_dygraph_mode():
                if isinstance(out_shape, Variable):
                    out_shape = list(out_shape.numpy())
                for i, dim in enumerate(out_shape):
                    if isinstance(dim, Variable):
                        out_shape[i] = dim.numpy()[0]
X
xiaoting 已提交
356
            if not (_is_list_or_turple_(out_shape)):
357
                raise TypeError("size should be a list or tuple or Variable.")
X
xiaoting 已提交
358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385
            # Validate the shape
            contain_var = False
            for dim_idx, dim_size in enumerate(out_shape):
                if isinstance(dim_size, Variable):
                    contain_var = True
                    continue
                assert dim_size > 0, (
                    "Each dimension size given in out_shape must be greater than 0."
                )

            if contain_var:
                new_size_tensor = []
                size_list = []
                for dim in out_shape:
                    if isinstance(dim, Variable):
                        dim.stop_gradient = True
                        new_size_tensor.append(dim)
                        size_list.append(-1)
                    else:
                        assert (isinstance(dim, int))
                        temp_out = helper.create_variable_for_type_inference(
                            'int32')
                        fill_constant(
                            [1], 'int32', dim, force_cpu=True, out=temp_out)
                        new_size_tensor.append(temp_out)
                        size_list.append(dim)
                inputs['SizeTensor'] = new_size_tensor

X
xiaoting 已提交
386
            if len(x.shape) == 3:
387 388
                if len(out_shape) != 1:
                    raise ValueError(
389
                        "size length should be 2 for input 3-D tensor")
390 391 392 393 394
                if contain_var:
                    attrs['out_w'] = size_list[0]
                else:
                    out_shape = list(map(int, out_shape))
                    attrs['out_w'] = out_shape[0]
X
xiaoting 已提交
395
            if len(x.shape) == 4:
X
xiaoting 已提交
396
                if len(out_shape) != 2:
397
                    raise ValueError("size length should be 2 for "
X
xiaoting 已提交
398 399 400 401 402 403 404 405
                                     "input 4-D tensor.")
                if contain_var:
                    attrs['out_h'] = size_list[0]
                    attrs['out_w'] = size_list[1]
                else:
                    out_shape = list(map(int, out_shape))
                    attrs['out_h'] = out_shape[0]
                    attrs['out_w'] = out_shape[1]
X
xiaoting 已提交
406
            if len(x.shape) == 5:
X
xiaoting 已提交
407
                if len(out_shape) != 3:
408
                    raise ValueError("size length should be 3 for "
X
xiaoting 已提交
409 410 411 412 413 414 415 416 417 418 419 420
                                     "input 5-D tensor.")
                if contain_var:
                    attrs['out_d'] = size_list[0]
                    attrs['out_h'] = size_list[1]
                    attrs['out_w'] = size_list[2]
                else:
                    out_shape = list(map(int, out_shape))
                    attrs['out_d'] = out_shape[0]
                    attrs['out_h'] = out_shape[1]
                    attrs['out_w'] = out_shape[2]

    else:
421 422
        if in_dygraph_mode() and isinstance(scale, Variable):
            scale = list(scale.numpy())
X
xiaoting 已提交
423 424 425 426 427 428
        if isinstance(scale, Variable):
            scale.stop_gradient = True
            inputs["Scale"] = scale
        elif isinstance(scale, float) or isinstance(scale, int):
            if scale <= 0:
                raise ValueError("Attr(scale) should be greater than zero.")
X
xiaoting 已提交
429 430 431 432
            scale_list = []
            for i in range(len(x.shape) - 2):
                scale_list.append(scale)
            attrs['scale'] = list(map(float, scale_list))
X
xiaoting 已提交
433
        elif isinstance(scale, list) or isinstance(scale, tuple):
X
xiaoting 已提交
434 435 436 437 438 439 440 441
            if len(scale) != len(x.shape) - 2:
                raise ValueError("scale_shape length should be {} for "
                                 "input {}-D tensor.".format(
                                     len(x.shape) - 2, len(x.shape)))
            for value in scale:
                if value <= 0:
                    raise ValueError("Attr(scale) should be greater than zero.")
            attrs['scale'] = list(map(float, scale))
X
xiaoting 已提交
442 443
        else:
            raise TypeError(
444 445
                "Attr(scale)'s type should be float, int, list, tuple, or Tensor."
            )
X
xiaoting 已提交
446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464

    if in_dygraph_mode():
        attr_list = []
        for k, v in attrs.items():
            attr_list.append(k)
            attr_list.append(v)
        dy_attr = tuple(attr_list)

        if resample_type == "linear":
            out = core.ops.linear_interp_v2(x, *dy_attr)
        if resample_type == "bilinear":
            out = core.ops.bilinear_interp_v2(x, *dy_attr)
        if resample_type == "trilinear":
            out = core.ops.trilinear_interp_v2(x, *dy_attr)
        if resample_type == "nearest":
            out = core.ops.nearest_interp_v2(x, *dy_attr)
        if resample_type == "bicubic":
            out = core.ops.bicubic_interp_v2(x, *dy_attr)
        return out
X
xiaoting 已提交
465 466
    out = helper.create_variable_for_type_inference(dtype)
    helper.append_op(
X
xiaoting 已提交
467
        type='{}_interp_v2'.format(resample_type),
X
xiaoting 已提交
468 469 470 471
        inputs=inputs,
        outputs={"Out": out},
        attrs=attrs)
    return out
L
littletomatodonkey 已提交
472 473


X
xiaoting 已提交
474 475 476 477 478 479 480 481 482 483
def upsample(x,
             size=None,
             scale_factor=None,
             mode='nearest',
             align_corners=False,
             align_mode=0,
             data_format='NCHW',
             name=None):
    """
    This op resizes a batch of images.
484

X
xiaoting 已提交
485 486 487
    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),
488 489
    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.
X
xiaoting 已提交
490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513
    and the resizing only applies on the three dimensions(depth, height and width).

    Supporting resample methods:
        'linear' : Linear interpolation
        'bilinear' : Bilinear interpolation
        'trilinear' : Trilinear interpolation
        'nearest' : Nearest neighbor interpolation
        'bicubic' : Bicubic interpolation
    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. 
    
    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.
    
    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.
514

X
xiaoting 已提交
515 516 517
    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.
518

X
xiaoting 已提交
519 520 521
    The linear interpolation is performed on three directions.
    align_corners and align_mode are optional parameters,the calculation method
    of interpolation can be selected by them.
522 523 524 525 526 527 528

    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`.

X
xiaoting 已提交
529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618
    Example:
    .. code-block:: text
        For scale_factor:
            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)
        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}
        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})
        
        Bilinear interpolation:
          if:
              align_corners = False , align_mode = 0
              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:
              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
          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}
    https://en.wikipedia.org/wiki/Linear_interpolation.
    For details of linear interpolation, please refer to Wikipedia:
    
    For details of nearest neighbor interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation.
    
    For details of bilinear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Bilinear_interpolation.
    
    For details of bicubic interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Bicubic_interpolation
    
    For details of trilinear interpolation, please refer to Wikipedia:
    https://en.wikipedia.org/wiki/Trilinear_interpolation.
    
    Parameters:
        x (Tensor): 3-D, 4-D or 5-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_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. 
619
             Default: None. If a list/tuple, each element can be an integer or a Tensor of shape: [1].
X
xiaoting 已提交
620
             If a Tensor , its dimensions size should be a 1.
621 622 623 624
        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.
X
xiaoting 已提交
625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667
             Default: None.
        mode (str): The resample method. It supports 'linear', 'nearest', 'bilinear',
                       'bicubic' and 'trilinear' currently. Default: 'nearest'
        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.
                               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.
        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 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),
        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).
    Raises:
        TypeError: size should be a list or tuple or Tensor.
        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.
        TypeError: align_corners should be a bool value
        ValueError: align_mode can only be '0' or '1'
        ValueError: data_format can only be 'NCW', 'NWC', 'NCHW', 'NHWC', 'NCDHW' or 'NDHWC'.
        Examples:
        .. code-block:: python
            import paddle
            import numpy as np
            import paddle.nn.functional as F

X
xiaoting 已提交
668
            input_data = np.random.rand(2,3,6,10).astype("float32")
X
xiaoting 已提交
669
            input = paddle.to_tensor(input_data)
X
xiaoting 已提交
670
            output = F.upsample(x=input, size=[12,12])
X
xiaoting 已提交
671 672 673 674 675 676 677 678
            print(output.shape)
            # [2L, 3L, 12L, 12L]

    """
    return interpolate(x, size, scale_factor, mode, align_corners, align_mode,
                       data_format)


679 680 681 682
def bilinear(x1, x2, weight, bias=None, name=None):
    """

    This layer performs bilinear on two inputs.
683
    See :ref:`api_nn_Bilinear` for details and output shape.
684 685 686 687 688 689 690 691 692 693

    Parameters:
       x1 (Tensor): the first input tensor, it's data type should be float32, float64.
       x2 (Tensor): the second input tensor, it's data type should be float32, float64.
       weight (Parameter): The learnable weights of this layer, shape is [out_features, in1_features, in2_features].
       bias (Parameter, optional): The learnable bias(Bias) of this layer, shape is [1, out_features]. If it is set to None, no bias will be added to the output units. The default value is None.
       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.

    Returns:
694
       Tensor: A 2-D Tensor of shape [batch_size, out_features].
695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730

    Examples:
       .. code-block:: python

        import paddle
        import numpy
        import paddle.nn.functional as F

        x1 = numpy.random.random((5, 5)).astype('float32')
        x2 = numpy.random.random((5, 4)).astype('float32')
        w = numpy.random.random((1000, 5, 4)).astype('float32')
        b = numpy.random.random((1, 1000)).astype('float32')

        result = F.bilinear(paddle.to_tensor(x1), paddle.to_tensor(x2), paddle.to_tensor(w), paddle.to_tensor(b))           # result shape [5, 1000]

    """

    if in_dygraph_mode():
        return core.ops.bilinear_tensor_product(x1, x2, weight, bias)

    check_variable_and_dtype(x1, 'x1', ['float32', 'float64'], 'bilinear')
    check_variable_and_dtype(x2, 'x2', ['float32', 'float64'], 'bilinear')

    inputs = {"X": x1, "Y": x2, "Weight": weight}
    if bias is not None:
        inputs["Bias"] = bias

    helper = LayerHelper("bilinear", **locals())
    out = helper.create_variable_for_type_inference(dtype=x1.dtype)

    helper.append_op(
        type="bilinear_tensor_product", inputs=inputs, outputs={"Out": out})

    return out


731 732 733 734 735 736 737 738 739 740 741 742 743 744
def dropout(x,
            p=0.5,
            axis=None,
            training=True,
            mode="upscale_in_train",
            name=None):
    """
    Dropout is a regularization technique for reducing overfitting by preventing
    neuron co-adaption during training. The dropout operator randomly sets the
    outputs of some units to zero, while upscale others according to the given
    dropout probability.

    Args:
        x (Tensor): The input tensor. The data type is float32 or float64.
745 746
        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.
747
        training (bool): A flag indicating whether it is in train phrase or not. Default True.
748
        mode(str): ['upscale_in_train'(default) | 'downscale_in_infer'].
749 750 751 752 753 754 755 756 757 758

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

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

                           2. downscale_in_infer, downscale the output at inference

                              - train: out = input * mask
                              - inference: out = input * (1.0 - dropout_prob)
759
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.
760 761 762 763

    Returns:
        A Tensor representing the dropout, has same shape and data type as `x` .

764

765 766
    Examples:
        We use ``p=0.5`` in the following description for simplicity.
767

768
        1. When ``axis=None`` , this is commonly used dropout, which dropout each element of x randomly.
769 770 771

        ..  code-block:: text

772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796
            Let's see a simple case when x is a 2d tensor with shape 2*3:
            [[1 2 3]
             [4 5 6]]
            we generate mask with the same shape as x, which is 2*3. The value of mask is
            sampled from a Bernoulli distribution randomly. For example, we may get such mask:
            [[0 1 0]
             [1 0 1]]
            So the output is obtained from elementwise multiply of x and mask:
            [[0 2 0]
             [4 0 6]]
            Using default setting, i.e. ``mode='upscale_in_train'`` ,
            if in training phase, the final upscale output is:
            [[0 4 0 ]
             [8 0 12]]
            if in test phase, the output is the same as input:
            [[1 2 3]
             [4 5 6]]
            we can also set ``mode='downscale_in_infer'`` , then
            if in training phase, the final output is:
            [[0 2 0]
             [4 0 6]]
            if in test phase, the scale output is:
            [[0.5 1.  1.5]
             [2.  2.5 3. ]]

797 798


799
        2. When ``axis!=None`` , this is useful for dropping whole channels from an image or sequence.
800 801 802

        ..  code-block:: text

803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830
            Let's see the simple case when x is a 2d tensor with shape 2*3 again:
            [[1 2 3]
             [4 5 6]]
            (1) If ``axis=0`` , this means the dropout is only performed in axis `0` .
                we generate mask with the shape 2*1. Only in axis `0` the value is randomly selected.
                For example, we may get such mask:
                [[1]
                 [0]]
                The output is obtained from elementwise multiply of x and mask. Doing that the mask will be
                broadcast from 2*1 to 2*3:
                [[1 1 1]
                 [0 0 0]]
                and the result after elementwise multiply is:
                [[1 2 3]
                 [0 0 0]]
                then we can do upscale or downscale according to the setting of other arguments.
            (2) If ``axis=1`` , this means the dropout is only performed in axis `1` .
                we generate mask with the shape 1*3. Only in axis `1` the value is randomly selected.
                For example, we may get such mask:
                [[1 0 1]]
                Doing elementwise multiply the mask will be broadcast from 1*3 to 2*3:
                [[1 0 1]
                 [1 0 1]]
                and the result after elementwise multiply is:
                [[1 0 3]
                 [4 0 6]]
            (3) What about ``axis=[0, 1]`` ? This means the dropout is performed in all axes of x,
                which is the same case as default setting ``axis=None`` .
831
            (4) You may note that logically `axis=None` means the dropout is performed in none axis of x,
832 833 834 835 836 837 838 839 840 841
                We generate mask with the shape 1*1. Whole input is randomly selected or dropped.
                For example, we may get such mask:
                [[0]]
                Doing elementwise multiply the mask will be broadcast from 1*1 to 2*3:
                [[0 0 0]
                 [0 0 0]]
                and the result after elementwise multiply is:
                [[0 0 0]
                 [0 0 0]]
                Actually this is not what we want because all elements may set to zero~
842 843 844

        When x is a 4d tensor with shape `NCHW`, we can set ``axis=[0,1]`` and the dropout will be performed in channel `N` and `C`, `H` and `W` is tied, i.e. paddle.nn.dropout(x, p, axis=[0,1]) . Please refer to ``paddle.nn.functional.dropout2d`` for more details.
        Similarly, when x is a 5d tensor with shape `NCDHW`, we can set ``axis=[0,1]`` to perform dropout3d. Please refer to ``paddle.nn.functional.dropout3d`` for more details.
845 846

        .. code-block:: python
847

848 849 850 851 852 853 854 855 856 857
            import paddle
            import numpy as np

            x = np.array([[1,2,3], [4,5,6]]).astype('float32')
            x = paddle.to_tensor(x)
            y_train = paddle.nn.functional.dropout(x, 0.5)
            y_test = paddle.nn.functional.dropout(x, 0.5, training=False) 
            y_0 = paddle.nn.functional.dropout(x, axis=0)
            y_1 = paddle.nn.functional.dropout(x, axis=1)
            y_01 = paddle.nn.functional.dropout(x, axis=[0,1])
858 859 860 861 862 863
            print(x)
            print(y_train)
            print(y_test)
            print(y_0)
            print(y_1)
            print(y_01)
864 865

    """
866 867 868 869
    # fast return for p == 0
    if p == 0:
        return x

870 871 872 873 874 875 876
    if not isinstance(p, (float, int)):
        raise TypeError("p argument should be a number")
    if p < 0 or p > 1:
        raise ValueError("p argument should between 0 and 1")
    if mode not in ('downscale_in_infer', 'upscale_in_train'):
        raise ValueError(
            "mode argument should be 'downscale_in_infer' or 'upscale_in_train'")
877
    if axis and not isinstance(axis, (int, list, tuple)):
878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935
        raise TypeError("datatype of axis argument should be int or list")

    if axis == None:  # commonly used dropout
        seed = None
        mode = 'downgrade_in_infer' if mode == 'downscale_in_infer' else mode  #semantic transfer

        def get_attrs(prog, dropout_prob, is_test, seed):
            if (seed is None or seed == 0) and prog.random_seed != 0:
                seed = prog.random_seed
            attrs = {
                'dropout_prob': dropout_prob,
                'is_test': is_test,
                'fix_seed': seed is not None,
                'seed': seed if seed is not None else 0,
                'dropout_implementation': mode,
            }
            return attrs

        if in_dygraph_mode():
            if default_main_program().random_seed != 0:
                seed = default_main_program().random_seed
            out, mask = core.ops.dropout(
                x, 'dropout_prob', p, 'is_test', not training, 'fix_seed',
                seed is not None, 'seed', seed
                if seed is not None else 0, 'dropout_implementation', mode)
            return out

        helper = LayerHelper('dropout', **locals())
        check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                                 'dropout')

        out = helper.create_variable_for_type_inference(dtype=x.dtype)
        mask = helper.create_variable_for_type_inference(
            dtype=core.VarDesc.VarType.UINT8, stop_gradient=True)

        attrs = get_attrs(helper.main_program, p, not training, seed)

        helper.append_op(
            type='dropout',
            inputs={'X': [x]},
            outputs={'Out': [out],
                     'Mask': [mask]},
            attrs=attrs)
        return out
    else:  #sometimes called dropout_nd #TODO: optimize with c++
        if not in_dygraph_mode():
            check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'dropout')
        dtype = x.dtype
        keep_prob = 1 - p
        if training:
            if p == 1.:
                return layers.scale(x, scale=0.)

            scale_input = layers.scale(
                x, scale=1 / keep_prob) if mode == 'upscale_in_train' else x

            #get mask shape
            input_shape = x.shape
936
            drop_axes = [axis] if isinstance(axis, int) else list(axis)
937 938
            if min(drop_axes) < 0 or max(drop_axes) > len(input_shape) - 1:
                raise ValueError("axis value should be greater than or equal to 0 and less than dimensions of x:{}, but get axis value:{} " \
939 940 941
                                 .format(len(input_shape), max(drop_axes)))
            if len(drop_axes) > len(input_shape):
                raise ValueError(
942
                    "length of axis should not be greater than dimensions of x:{}, but get length of axis: {}".
943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976
                    format(len(input_shape), len(drop_axes)))
            mask_shape = [1] * len(input_shape)
            for i in drop_axes:
                mask_shape[i] = input_shape[i]

            #get mask
            random_tensor = layers.uniform_random(
                mask_shape, dtype='float32', min=0., max=1.0)
            p = layers.fill_constant(shape=[1], dtype='float32', value=p)
            keep_mask = layers.greater_equal(random_tensor, p)

            scale_input = layers.cast(scale_input, dtype)
            keep_mask = layers.cast(keep_mask, dtype)
            ret = paddle.multiply(scale_input, keep_mask, name=name)
            return ret
        else:  # test
            ret = layers.scale(
                x, scale=keep_prob) if mode == 'downscale_in_infer' else x
            return ret


def dropout2d(x, p=0.5, training=True, data_format='NCHW', name=None):
    """
    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.

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

    Args:
        x (Tensor):  The input is 4-D Tensor with shape [N, C, H, W] or [N, H, W, C].
                     The data type is float32 or float64.
        p (float): Probability of setting units to zero. Default 0.5.
        training (bool): A flag indicating whether it is in train phrase or not. Default True.
977
        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].
978 979 980 981 982
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        A Tensor representing the dropout2d, has same shape and data type as `x` .

983

984 985
    Examples:
        .. code-block:: python
986

987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031
            import paddle
            import numpy as np

            x = np.random.random(size=(2, 3, 4, 5)).astype('float32')
            x = paddle.to_tensor(x)
            y_train = paddle.nn.functional.dropout2d(x)  #train
            y_test = paddle.nn.functional.dropout2d(x, training=False) #test
            for i in range(2):
                for j in range(3):
                    print(x.numpy()[i,j,:,:])
                    print(y_train.numpy()[i,j,:,:]) # may all 0
                    print(y_test.numpy()[i,j,:,:])
    """
    input_shape = x.shape
    if len(input_shape) != 4:
        raise ValueError("dimensions of x should be 4, but received {} != 4"\
        .format(len(input_shape)))

    if data_format not in ["NCHW", "NHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCHW' or 'NHWC'. Received "
            "Attr(data_format): %s." % str(data_format))

    return dropout(
        x,
        p=p,
        axis=[0, 1] if data_format == 'NCHW' else [0, 3],
        training=training,
        mode="upscale_in_train",
        name=name)


def dropout3d(x, p=0.5, training=True, data_format='NCDHW', name=None):
    """
    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.

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

    Args:
        x (Tensor):  The input is 5-D Tensor with shape [N, C, D, H, W] or [N, D, H, W, C].
                     The data type is float32 or float64.
        p (float): Probability of setting units to zero. Default 0.5.
        training (bool): A flag indicating whether it is in train phrase or not. Default True.
1032
        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].
1033 1034 1035 1036 1037
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        A Tensor representing the dropout3d, has same shape and data type with `x` .

1038

1039 1040
    Examples:
        .. code-block:: python
1041

1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072
            import paddle
            import numpy as np

            x = np.random.random(size=(2, 3, 4, 5, 6)).astype('float32')
            x = paddle.to_tensor(x)
            y_train = paddle.nn.functional.dropout3d(x)  #train
            y_test = paddle.nn.functional.dropout3d(x, training=False) #test
            print(x.numpy()[0,0,:,:,:])
            print(y_train.numpy()[0,0,:,:,:]) # may all 0
            print(y_test.numpy()[0,0,:,:,:])
    """

    input_shape = x.shape
    if len(input_shape) != 5:
        raise ValueError("dimensions of x should be 5, but received {} != 5" \
        .format(len(input_shape)))

    if data_format not in ["NCDHW", "NDHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCDHW' or 'NDHWC'. Received "
            "Attr(data_format): %s." % str(data_format))

    return dropout(
        x,
        p=p,
        axis=[0, 1] if data_format == 'NCDHW' else [0, 4],
        training=training,
        mode="upscale_in_train",
        name=name)


1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090
def alpha_dropout(x, p=0.5, training=True, name=None):
    """
    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.

    Args:
        x (Tensor): The input tensor. The data type is float32 or float64.
        p (float | int): Probability of setting units to zero. Default 0.5.
        training (bool): A flag indicating whether it is in train phrase or not. Default True.
        name (str, optional): Name for the operation (optional, default is None). For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        A Tensor representing the dropout, has same shape and data type as `x`.

    Examples:
        .. code-block:: python
1091

1092 1093 1094 1095 1096 1097 1098
            import paddle
            import numpy as np

            x = np.array([[-1, 1], [-1, 1]]).astype('float32')
            x = paddle.to_tensor(x)
            y_train = paddle.nn.functional.alpha_dropout(x, 0.5)
            y_test = paddle.nn.functional.alpha_dropout(x, 0.5, training=False)
1099 1100
            print(x)
            print(y_train)
1101
            # [[-0.10721093, 1.6655989 ], [-0.7791938, -0.7791938]] (randomly)
1102
            print(y_test)
1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113
    """
    if not isinstance(p, (float, int)):
        raise TypeError("p argument should be a float or int")
    if p < 0 or p > 1:
        raise ValueError("p argument should between 0 and 1")

    if not in_dygraph_mode():
        check_variable_and_dtype(x, 'x', ['float32', 'float64'],
                                 'alpha_dropout')

    if training:
1114 1115
        if p == 1:
            return layers.scale(x, scale=0.)
1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148
        #get transformation params
        alpha = 1.6732632423543772848170429916717
        scale = 1.0507009873554804934193349852946
        alpha_p = -alpha * scale
        a = ((1 - p) * (1 + p * alpha_p**2))**-0.5
        b = -a * alpha_p * p

        dtype = x.dtype
        input_shape = x.shape

        #get mask
        random_tensor = layers.uniform_random(
            input_shape, dtype='float32', min=0., max=1.0)
        p = layers.fill_constant(shape=[1], dtype='float32', value=p)
        keep_mask = layers.greater_equal(random_tensor, p)
        keep_mask = layers.cast(keep_mask, dtype)
        drop_mask = layers.elementwise_sub(
            layers.fill_constant(
                shape=input_shape, dtype=dtype, value=1.),
            keep_mask)

        #apply mask
        b = layers.fill_constant(shape=[1], dtype=dtype, value=b)
        y = layers.elementwise_add(
            paddle.multiply(x, keep_mask),
            layers.scale(
                drop_mask, scale=alpha_p))
        res = layers.elementwise_add(layers.scale(y, scale=a), b, name=name)
        return res
    else:  # test
        return x


L
littletomatodonkey 已提交
1149 1150 1151
def pad(x, pad, mode='constant', value=0, data_format="NCHW", name=None):
    """
    Pad tensor according to 'pad' and 'mode'.
L
littletomatodonkey 已提交
1152 1153 1154
    If mode is 'constant' and length of pad is twice as length of x dimension,
    then the padding will be started from the first dimension and moved back onto x
    according to 'pad' and 'value'.
L
littletomatodonkey 已提交
1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222
    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.

    Parameters:
        x (Tensor): The input tensor with data type float32/double/int32/int64_t.
        pad (Tensor | List[int32]): The padding size with data type int32. [len(padding)/2] dimensions
            of input will be padded. 1. If input dimension is 3, then the pad has the form (pad_left,
            pad_right). 2. If the input dimension is 4, then the pad has the form (pad_left, pad_right, 
            pad_top, pad_bottom). 3. If the input dimension is 5, then the pad has the form 
            (pad_left, pad_right, pad_top, pad_bottom, pad_front, pad_back).
            
        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'
        value (float32): The value to fill the padded areas in 'constant' mode . Default is 0.0
        data_format (str): An string from: "NCL", "NLC", NHWC", "NCHW", "NCDHW", "NDHWC". 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`.
                    
    Returns: a Tensor padded according to pad and mode and data type is same as input.
    Return Type: Tensor

    Examples:
        .. code-block:: text

            x = [[[[[1., 2., 3.],
                    [4., 5., 6.]]]]]

            Case 0:
                pad = [2, 2, 1, 1, 0, 0],
                mode = 'constant'
                value = 0
                Out = [[[[[0. 0. 0. 0. 0. 0. 0.]
                          [0. 0. 1. 2. 3. 0. 0.]
                          [0. 0. 4. 5. 6. 0. 0.]
                          [0. 0. 0. 0. 0. 0. 0.]]]]]

            Case 1:
                pad = [2, 2, 1, 1, 0, 0],
                mode = 'reflect'
                Out = [[[[[6. 5. 4. 5. 6. 5. 4.]
                          [3. 2. 1. 2. 3. 2. 1.]
                          [6. 5. 4. 5. 6. 5. 4.]
                          [3. 2. 1. 2. 3. 2. 1.]]]]]

            Case 2:
                pad = [2, 2, 1, 1, 0, 0],
                mode = 'replicate'
                Out = [[[[[1. 1. 1. 2. 3. 3. 3.]
                          [1. 1. 1. 2. 3. 3. 3.]
                          [4. 4. 4. 5. 6. 6. 6.]
                          [4. 4. 4. 5. 6. 6. 6.]]]]]

            Case 3:
                pad = [2, 2, 1, 1, 0, 0],
                mode = 'circular'
                Out = [[[[[5. 6. 4. 5. 6. 4. 5.]
                          [2. 3. 1. 2. 3. 1. 2.]
                          [5. 6. 4. 5. 6. 4. 5.]
                          [2. 3. 1. 2. 3. 1. 2.]]]]]

    Code Examples:
        .. code-block:: python
L
littletomatodonkey 已提交
1223

L
littletomatodonkey 已提交
1224 1225 1226 1227 1228 1229
            import numpy as np
            import paddle
            import paddle.nn.functional as F
            
            # example 1
            x_shape = (1, 1, 3)
L
littletomatodonkey 已提交
1230 1231 1232
            x = paddle.arange(np.prod(x_shape), dtype="float32").reshape(x_shape) + 1
            y = F.pad(x, [2, 3], value=1, mode='constant', data_format="NCL")
            print(y)
L
littletomatodonkey 已提交
1233
            # [[[1. 1. 1. 2. 3. 1. 1. 1.]]]
1234
            
L
littletomatodonkey 已提交
1235 1236
            # example 2
            x_shape = (1, 1, 2, 3)
L
littletomatodonkey 已提交
1237 1238 1239
            x = paddle.arange(np.prod(x_shape), dtype="float32").reshape(x_shape) + 1
            y = F.pad(x, [1, 2, 1, 1], value=1, mode='circular')
            print(y)
L
littletomatodonkey 已提交
1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254
            # [[[[6. 4. 5. 6. 4. 5.]
            #    [3. 1. 2. 3. 1. 2.]
            #    [6. 4. 5. 6. 4. 5.]
            #    [3. 1. 2. 3. 1. 2.]]]]
    """
    assert mode in ['reflect', 'replicate', 'constant', 'circular'], \
            "mode should be one of constant, reflect, replicate, circular, but got {}.".format(mode)

    data_format = data_format.upper()
    assert data_format in ["NCL", "NCHW", "NCDHW", "NLC", "NHWC", "NDHWC"], \
        "data_format should be in one of [NCL, NCHW, NCDHW, NLC, NHWC, NDHWC], " \
        "but got {}".format(data_format)

    x_dim = len(x.shape)

L
littletomatodonkey 已提交
1255 1256 1257
    if mode == "constant" and isinstance(pad, list) and len(pad) == x_dim * 2:
        return layers.pad(x, pad, pad_value=value)

1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270
    assert x_dim in [
        3, 4, 5
    ], "input tesor dimension must be in [3, 4, 5] but got {}".format(x_dim)

    supported_format_map = {
        3: ["NCL", "NLC"],
        4: ["NCHW", "NHWC"],
        5: ["NCDHW", "NDHWC"],
    }
    assert data_format in supported_format_map[x_dim], \
    "input tensor dimension is {}, it's data format should be in {} but got {}".format(
        x_dim, supported_format_map[x_dim], data_format)

L
littletomatodonkey 已提交
1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342
    unsqueezed_dim = []

    if isinstance(pad, Variable):
        if data_format in ["NCL", "NCHW", "NCDHW"]:
            data_format = "NCDHW"
            if x_dim == 3:
                pad = concat([zeros((4, ), dtype="int32"), pad], axis=0)
                unsqueezed_dim = [3, 4]
                x = unsqueeze(x, axes=unsqueezed_dim)
            elif x_dim == 4:
                pad = concat([pad, zeros((2, ), dtype="int32")], axis=0)
                unsqueezed_dim = [2]
                x = unsqueeze(x, axes=unsqueezed_dim)
        elif data_format in ["NLC", "NHWC", "NDHWC"]:
            data_format = "NDHWC"
            if x_dim == 3:
                pad = concat([zeros((4, ), dtype="int32"), pad], axis=0)
                unsqueezed_dim = [2, 3]
                x = unsqueeze(x, axes=unsqueezed_dim)
            elif x_dim == 4:
                pad = concat([pad, zeros((2, ), dtype="int32")], axis=0)
                unsqueezed_dim = [1]
                x = unsqueeze(x, axes=unsqueezed_dim)
    else:
        if data_format in ["NCL", "NCHW", "NCDHW"]:
            data_format = "NCDHW"
            if x_dim == 3:
                pad = [0, 0, 0, 0] + pad
                unsqueezed_dim = [3, 4]
                x = unsqueeze(x, axes=unsqueezed_dim)
            elif x_dim == 4:
                pad = pad + [0, 0]
                unsqueezed_dim = [2]
                x = unsqueeze(x, axes=unsqueezed_dim)
        elif data_format in ["NLC", "NHWC", "NDHWC"]:
            data_format = "NDHWC"
            if x_dim == 3:
                pad = [0, 0, 0, 0] + pad
                unsqueezed_dim = [2, 3]
                x = unsqueeze(x, axes=unsqueezed_dim)
            elif x_dim == 4:
                pad = pad + [0, 0]
                unsqueezed_dim = [1]
                x = unsqueeze(x, axes=unsqueezed_dim)

    if in_dygraph_mode():
        if isinstance(pad, Variable):
            pad = pad.numpy()
        out = core.ops.pad3d(x, "paddings", pad, "mode", mode, "value", value,
                             "data_format", data_format, "name", name)
    else:
        attrs = {'mode': mode, 'value': value, 'data_format': data_format}
        inputs = {'X': [x]}
        if isinstance(pad, Variable):
            inputs['Paddings'] = [pad]
            attrs['paddings'] = []
        else:
            attrs['paddings'] = pad

        helper = LayerHelper('pad3d', **locals())

        dtype = helper.input_dtype(input_param_name='input')
        out = helper.create_variable_for_type_inference(dtype)
        helper.append_op(
            type='pad3d', inputs=inputs, outputs={"Out": out}, attrs=attrs)

    if len(unsqueezed_dim) != 0:
        out = squeeze(out, axes=unsqueezed_dim)

    return out


Y
Yang Zhang 已提交
1343
def cosine_similarity(x1, x2, axis=1, eps=1e-8):
L
littletomatodonkey 已提交
1344
    """
Y
Yang Zhang 已提交
1345
    Compute cosine similarity between x1 and x2 along axis.
L
littletomatodonkey 已提交
1346 1347 1348 1349

    Parameters:
        x1 (Tensor): First input. float32/double.
        x2 (Tensor): Second input. float32/double.
Y
Yang Zhang 已提交
1350
        axis (int): Dimension of vectors to compute cosine similarity. Default is 1.
L
littletomatodonkey 已提交
1351 1352
        eps(float): Small value to avoid division by zero. Default is 1e-8.
                    
Y
Yang Zhang 已提交
1353
    Returns: a Tensor representing cosine similarity between x1 and x2 along axis.
L
littletomatodonkey 已提交
1354 1355 1356 1357
    Return Type: Tensor

    Examples:
        .. code-block:: text
1358

L
littletomatodonkey 已提交
1359 1360 1361 1362 1363 1364 1365 1366 1367
            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 ]]
Y
Yang Zhang 已提交
1368
                axis = 1
L
littletomatodonkey 已提交
1369 1370 1371 1372 1373
                eps = 1e-8
                Out: [0.5275037  0.8368967  0.75037485 0.9245899]

    Code Examples:
        .. code-block:: python
1374

L
littletomatodonkey 已提交
1375 1376 1377 1378 1379 1380 1381 1382 1383
            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)
Y
Yang Zhang 已提交
1384
            result = paddle.nn.functional.cosine_similarity(x1, x2, axis=0)
L
littletomatodonkey 已提交
1385
            print(result)
L
littletomatodonkey 已提交
1386 1387 1388
            # [0.99806249 0.9817672  0.94987036]
            
    """
1389 1390 1391
    w12 = sum(paddle.multiply(x1, x2), axis=axis)
    w1 = sum(paddle.multiply(x1, x1), axis=axis)
    w2 = sum(paddle.multiply(x2, x2), axis=axis)
Y
Yang Zhang 已提交
1392
    n12 = sqrt(clip(w1 * w2, min=eps * eps))
L
littletomatodonkey 已提交
1393 1394
    cos_sim = w12 / n12
    return cos_sim
1395 1396 1397


def linear(x, weight, bias=None, name=None):
1398
    r"""
1399

1400 1401
    Fully-connected linear transformation operator. For each input :math:`X` ,
    the equation is:
1402 1403 1404

    .. math::

1405
        Out = XW + b
1406

1407
    where :math:`W` is the weight and :math:`b` is the bias.
1408

1409 1410 1411 1412 1413 1414 1415
    If the weight is a 2-D tensor of shape :math:`[in\_features, out\_features]` ,
    input should be a multi-dimensional tensor of shape
    :math:`[batch\_size, *, in\_features]` , where :math:`*` means any number of
    additional dimensions. The linear operator multiplies input tensor with
    weight and produces an output tensor of shape :math:`[batch\_size, *, out\_features]` , 
    If :math:`bias` is not None, the bias should be a 1-D tensor of shape
    :math:`[out\_features]` and will be added to the output.
1416

1417 1418 1419 1420 1421 1422 1423
    Parameters:
        x (Tensor): Input tensor. The data type should be float16, float32 or float64.
        weight (Tensor): Weight tensor. The data type should be float16, float32 or float64.
        bias (Tensor, optional): Bias tensor. The data type should be float16, float32 or float64.
                                 If it is set to None, no bias will be added to the output units.
        name (str, optional): Normally there is no need for user to set this parameter.
                              For detailed information, please refer to :ref:`api_guide_Name` .
1424 1425

    Returns:
1426 1427
        Tensor, the shape is :math:`[batch\_size, *, out\_features]` and the
        data type is the same with input :math:`x` .
1428 1429 1430 1431 1432 1433

    Examples:
        .. code-block:: python
          
          import paddle
          
1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446
          x = paddle.randn((3, 2), dtype="float32")
          # x: [[-0.32342386 -1.200079  ]
          #     [ 0.7979031  -0.90978354]
          #     [ 0.40597573  1.8095392 ]]
          weight = paddle.full(shape=[2, 4], fill_value="0.5", dtype="float32", name="weight")
          # weight: [[0.5 0.5 0.5 0.5]
          #          [0.5 0.5 0.5 0.5]]
          bias = paddle.ones(shape=[4], dtype="float32", name="bias")
          # bias: [1. 1. 1. 1.]
          y = paddle.nn.functional.linear(x, weight, bias)
          # 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 ]]
1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481
    """
    if in_dygraph_mode():
        pre_bias = _varbase_creator(dtype=x.dtype)
        core.ops.matmul(x, weight, pre_bias, 'transpose_X', False,
                        'transpose_Y', False, "alpha", 1)
        return dygraph_utils._append_bias_in_dygraph(
            pre_bias, bias, axis=len(x.shape) - 1)
    else:
        helper = LayerHelper('linear', **locals())
        dtype = x.dtype

        check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                                 'linear')
        check_dtype(dtype, 'dtype', ['float16', 'float32', 'float64'], 'linear')

        inputs = {'X': [x], 'Y': [weight]}
        attrs = {
            'transpose_X': False,
            'transpose_Y': False,
            'alpha': 1,
        }
        tmp = helper.create_variable_for_type_inference(dtype)
        helper.append_op(
            type='matmul', inputs=inputs, outputs={'Out': tmp}, attrs=attrs)
        if bias is not None:
            res = helper.create_variable_for_type_inference(dtype)
            helper.append_op(
                type='elementwise_add',
                inputs={'X': [tmp],
                        'Y': [bias]},
                outputs={'Out': [res]},
                attrs={'axis': len(x.shape) - 1})
        else:
            res = tmp
        return res
1482 1483 1484


def label_smooth(label, prior_dist=None, epsilon=0.1, name=None):
1485
    r"""
1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536
    Label smoothing is a mechanism to regularize the classifier layer and is called
    label-smoothing regularization (LSR).

    Label smoothing is proposed to encourage the model to be less confident,
    since optimizing the log-likelihood of the correct label directly may
    cause overfitting and reduce the ability of the model to adapt. Label
    smoothing replaces the ground-truth label :math:`y` with the weighted sum
    of itself and some fixed distribution :math:`\mu`. For class :math:`k`,
    i.e.

    .. math::

        \\tilde{y_k} = (1 - \epsilon) * y_k + \epsilon * \mu_k,

    where :math:`1 - \epsilon` and :math:`\epsilon` are the weights
    respectively, and :math:`\\tilde{y}_k` is the smoothed label. Usually
    uniform distribution is used for :math:`\mu`.

    See more details about label smoothing in https://arxiv.org/abs/1512.00567.

    Parameters:
        label(Tensor): The input variable containing the label data. The
                        label data should use one-hot representation. It's
                        a multidimensional tensor with a shape of
                        :math:`[N_1, ..., Depth]`, where Depth is class number. The dtype can be "float32" and "float64".
        prior_dist(Tensor, optional): The prior distribution to be used to smooth
                        labels. If not provided, an uniform distribution
                        is used. It's a multidimensional tensor with a shape of
                        :math:`[1, class\_num]` . The default value is None.
        epsilon(float, optional): The weight used to mix up the original ground-truth
                        distribution and the fixed distribution. The default value is
                        0.1.
        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`.

    Returns:
        Tensor: The tensor containing the smoothed labels.

    Examples:
        .. code-block:: python

            import paddle
            import numpy as np
            
            x_data = np.array([[[0, 1, 0],
                                [ 1,  0, 1]]]).astype("float32")
            print(x_data.shape)
            paddle.disable_static()
            x = paddle.to_tensor(x_data, stop_gradient=False)
            output = paddle.nn.functional.label_smooth(x)
1537
            print(output)
1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561
            
            #[[[0.03333334 0.93333334 0.03333334]
            #  [0.93333334 0.03333334 0.93333334]]]
    """
    if epsilon > 1. or epsilon < 0.:
        raise ValueError("The value of epsilon must be between 0 and 1.")

    if in_dygraph_mode():
        return core.ops.label_smooth(label, prior_dist, 'epsilon',
                                     float(epsilon))

    check_variable_and_dtype(label, 'label', ['float32', 'float64'],
                             'label_smooth')

    helper = LayerHelper("label_smooth", **locals())
    label.stop_gradient = True
    smooth_label = helper.create_variable_for_type_inference(label.dtype)
    helper.append_op(
        type="label_smooth",
        inputs={"X": label,
                "PriorDist": prior_dist} if prior_dist else {"X": label},
        outputs={"Out": smooth_label},
        attrs={"epsilon": float(epsilon)})
    return smooth_label