conv.py 67.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
#   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.
14
from __future__ import print_function
15

W
whs 已提交
16 17
__all__ = [
    'conv1d',
18
    'conv_transpose1d',
W
whs 已提交
19 20 21 22 23
    'conv2d',
    'conv_transpose2d',
    'conv3d',
    'conv_transpose3d',
]
24

25
import numpy as np
L
LielinJiang 已提交
26
from ...device import get_cudnn_version
27 28 29 30 31 32
from ...fluid.framework import Variable, in_dygraph_mode
from ...fluid import core, dygraph_utils
from ...fluid.layers import nn, utils
from ...fluid.data_feeder import check_variable_and_dtype
from ...fluid.param_attr import ParamAttr
from ...fluid.layer_helper import LayerHelper
33
from .common import pad2d
34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98


def _is_list_or_tuple(input):
    return isinstance(input, (list, tuple))


def _zero_padding_in_batch_and_channel(padding, channel_last):
    if channel_last:
        return list(padding[0]) == [0, 0] and list(padding[-1]) == [0, 0]
    else:
        return list(padding[0]) == [0, 0] and list(padding[1]) == [0, 0]


def _exclude_padding_in_batch_and_channel(padding, channel_last):
    padding_ = padding[1:-1] if channel_last else padding[2:]
    padding_ = [elem for pad_a_dim in padding_ for elem in pad_a_dim]
    return padding_


def _update_padding_nd(padding, channel_last, num_dims):
    if isinstance(padding, str):
        padding = padding.upper()
        if padding not in ["SAME", "VALID"]:
            raise ValueError(
                "Unknown padding: '{}'. It can only be 'SAME' or 'VALID'.".
                format(padding))
        if padding == "VALID":
            padding_algorithm = "VALID"
            padding = [0] * num_dims
        else:
            padding_algorithm = "SAME"
            padding = [0] * num_dims
    elif _is_list_or_tuple(padding):
        # for padding like
        # [(pad_before, pad_after), (pad_before, pad_after), ...]
        # padding for batch_dim and channel_dim included
        if len(padding) == 2 + num_dims and _is_list_or_tuple(padding[0]):
            if not _zero_padding_in_batch_and_channel(padding, channel_last):
                raise ValueError(
                    "Non-zero padding({}) in the batch or channel dimensions "
                    "is not supported.".format(padding))
            padding_algorithm = "EXPLICIT"
            padding = _exclude_padding_in_batch_and_channel(padding,
                                                            channel_last)
            if utils._is_symmetric_padding(padding, num_dims):
                padding = padding[0::2]
        # for padding like [pad_before, pad_after, pad_before, pad_after, ...]
        elif len(padding) == 2 * num_dims and isinstance(padding[0], int):
            padding_algorithm = "EXPLICIT"
            padding = utils.convert_to_list(padding, 2 * num_dims, 'padding')
            if utils._is_symmetric_padding(padding, num_dims):
                padding = padding[0::2]
        # for padding like [pad_d1, pad_d2, ...]
        elif len(padding) == num_dims and isinstance(padding[0], int):
            padding_algorithm = "EXPLICIT"
            padding = utils.convert_to_list(padding, num_dims, 'padding')
        else:
            raise ValueError("In valid padding: {}".format(padding))
    # for integer padding
    else:
        padding_algorithm = "EXPLICIT"
        padding = utils.convert_to_list(padding, num_dims, 'padding')
    return padding, padding_algorithm


W
whs 已提交
99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160
def conv1d(x,
           weight,
           bias=None,
           stride=1,
           padding=0,
           dilation=1,
           groups=1,
           data_format='NCL',
           name=None):
    """
    The convolution1D layer calculates the output based on the input, filter
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCL format, where N is batch size, C is the number of
    channels, L is the length of the feature.
    Filter is in MCK format, where M is the number of output image channels,
    C is the number of input image channels, K is the size of the kernel.
    If the groups is greater than 1, C will equal the number of input image
    channels divided by the groups. If bias attribution and activation type
    are provided, bias is added to the output of the convolution, and the
    corresponding activation function is applied to the final result.

    For each input :math:`X`, the equation is:

    .. math::

        Out = \sigma (W \\ast X + b)

    Where:

    * :math:`X`: Input value, a tensor with NCL format.
    * :math:`W`: Kernel value, a tensor with MCK format.
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.

    Example:

        - Input:

          Input shape: :math:`(N, C_{in}, L_{in})`

          Filter shape: :math:`(C_{out}, C_{in}, L_f)`

        - Output:

          Output shape: :math:`(N, C_{out}, L_{out})`

        Where

        .. math::

            L_{out}&= \\frac{(L_{in} + 2 * padding - (dilation * (L_f - 1) + 1))}{stride} + 1

    Args:
        x (Tensor): The input is 3-D Tensor with shape [N, C, L], the data type 
            of input is float16 or float32 or float64.
        weight (Tensor): The convolution kernel with shape [M, C/g, K], where M is
            the number of output channels, g is the number of groups, K is the kernel's size. 
        bias (Tensor, optional): The bias with shape [M,]. Default: None.
        stride (int or tuple, optional): The stride size. If stride is a tuple, it must
            contain one integers, (stride_size). Default: 1.
161
        padding(int|str|tuple|list, optional): The padding size. Padding could be in one of the following forms.
W
whs 已提交
162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
            1. a string in ['valid', 'same'].
            2. an int, which means the feature map is zero paded by size of `padding` on both sides.
            3. a list[int] or tuple[int] whose length is 1, which means the feature map is zero paded by size of `padding[0]` on both sides.
            4. a list[int] or tuple[int] whose length is 2. It has the form  [pad_before, pad_after].
            5. a list or tuple of pairs of ints. It has the form [[pad_before, pad_after], [pad_before, pad_after], ...]. Note that, the batch dimension and channel dimension are also included. Each pair of integers correspond to the amount of padding for a dimension of the input. Padding in batch dimension and channel dimension should be [0, 0] or (0, 0).
            The default value is 0.
        dilation (int or tuple, optional): The dilation size. If dilation is a tuple, it must
            contain one integer, (dilation_size). Default: 1.
        groups (int, optional): The groups number of the conv1d function. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: 1.
        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: `"NCL"`, `"NLC"`.
            The default is `"NCL"`. When it is `"NCL"`, the data is stored in the order of:
            `[batch_size, input_channels, feature_length]`.
        name(str, optional): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.

    Returns:
        A tensor representing the conv1d, whose data type is the 
        same with input.

    Raises:
188
        ValueError: If the channel dimension of the input is less than or equal to zero.
W
whs 已提交
189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240
        ValueError: If `data_format` is not "NCL" or "NLC".
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
        ShapeError: If the input is not 3-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 1.
        ShapeError: If the number of input channels is not equal to filter's channels * groups.
        ShapeError: If the number of output channels is not be divided by groups.

    Examples:
        .. code-block:: python

          import paddle
          import paddle.nn.functional as F
          import numpy as np
          x = np.array([[[4, 8, 1, 9],
            [7, 2, 0, 9],
            [6, 9, 2, 6]]]).astype(np.float32)
          w=np.array(
          [[[9, 3, 4],
            [0, 0, 7],
            [2, 5, 6]],
           [[0, 3, 4],
            [2, 9, 7],
            [5, 6, 8]]]).astype(np.float32)
          paddle.disable_static()
          x_var = paddle.to_tensor(x)
          w_var = paddle.to_tensor(w)
          y_var = F.conv1d(x_var, w_var)
          y_np = y_var.numpy()
          print(y_np)
          
          # [[[133. 238.]
          #   [160. 211.]]]
    """
    cudnn_version = get_cudnn_version()
    if cudnn_version is not None:
        use_cudnn = True
    else:
        use_cudnn = False

    if data_format not in ["NCL", "NLC"]:
        raise ValueError("Attr(data_format) should be 'NCL' or 'NLC'. "
                         "Received Attr(data_format): {}.".format(data_format))

    channel_last = (data_format == "NHWC")
    channel_dim = -1 if channel_last else 1
    conv2d_data_format = "NHWC" if channel_last else "NCHW"
    num_channels = x.shape[channel_dim]
    num_filters = weight.shape[0]
    if num_channels < 0:
241
        raise ValueError("The channel dimension of the input({}) "
W
whs 已提交
242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262
                         "should be defined. Received: {}.".format(
                             x.shape, num_channels))
    if num_channels % groups != 0:
        raise ValueError(
            "the channel of input must be divisible by groups,"
            "received: the channel of input is {}, the shape of input is {}"
            ", the groups is {}".format(num_channels, x.shape, groups))
    if num_filters % groups != 0:
        raise ValueError(
            "the number of filters must be divisible by groups,"
            "received: the number of filters is {}, the shape of weight is {}"
            ", the groups is {}".format(num_filters, weight.shape, groups))

    # update attrs
    padding, padding_algorithm = _update_padding_nd(padding, channel_last, 1)
    if len(padding) == 2:
        padding = padding + [0] * 2
    elif len(padding) == 1:
        padding = padding + [0]
    else:
        raise ValueError(
263
            "The size of padding's dimension should be 1 or 2. But got padding={}".
W
whs 已提交
264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324
            format(padding))

    stride = utils.convert_to_list(stride, 1, 'stride') + [1]
    dilation = utils.convert_to_list(dilation, 1, 'dilation') + [1]

    l_type = "conv2d"
    if (num_channels == groups and num_filters % num_channels == 0 and
            not use_cudnn):
        l_type = 'depthwise_conv2d'
        use_cudnn = False

    inputs = {'Input': [x], 'Filter': [weight]}
    attrs = {
        'strides': stride,
        'paddings': padding,
        'dilations': dilation,
        'groups': groups,
        'use_cudnn': use_cudnn,
        'use_mkldnn': False,
        'fuse_relu_before_depthwise_conv': False,
        "padding_algorithm": padding_algorithm,
        "data_format": conv2d_data_format
    }
    squeeze_aixs = -2 if channel_last else -1
    x = nn.unsqueeze(input=x, axes=[squeeze_aixs])
    weight = nn.unsqueeze(input=weight, axes=[-1])
    if in_dygraph_mode():
        attrs = ('strides', stride, 'paddings', padding, 'dilations', dilation,
                 'groups', groups, 'use_cudnn', use_cudnn, 'use_mkldnn', False,
                 'fuse_relu_before_depthwise_conv', False, "padding_algorithm",
                 padding_algorithm, "data_format", conv2d_data_format)
        out = getattr(core.ops, l_type)(x, weight, *attrs)
        if bias is not None:
            out = nn.elementwise_add(out, bias, axis=channel_dim)
    else:
        inputs = {'Input': [x], 'Filter': [weight]}
        attrs = {
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
            'use_mkldnn': False,
            'fuse_relu_before_depthwise_conv': False,
            "padding_algorithm": padding_algorithm,
            "data_format": conv2d_data_format
        }
        check_variable_and_dtype(x, 'input', ['float16', 'float32', 'float64'],
                                 'conv2d')
        helper = LayerHelper(l_type, **locals())
        dtype = helper.input_dtype()
        out = helper.create_variable_for_type_inference(dtype)
        outputs = {"Output": [out]}
        helper.append_op(
            type=l_type, inputs=inputs, outputs=outputs, attrs=attrs)
        if bias is not None:
            out = nn.elementwise_add(out, bias, axis=channel_dim)
    out = nn.squeeze(input=out, axes=[squeeze_aixs])
    return out


325
def conv2d(x,
326 327 328
           weight,
           bias=None,
           stride=1,
329
           padding=0,
330 331 332 333 334
           dilation=1,
           groups=1,
           data_format="NCHW",
           name=None):
    """
S
swtkiwi 已提交
335

336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352
    The convolution2D layer calculates the output based on the input, filter
    and strides, paddings, dilations, groups parameters. Input and
    Output are in NCHW or NHWC format, where N is batch size, C is the number of
    channels, H is the height of the feature, and W is the width of the feature.
    Filter is in MCHW format, where M is the number of output image channels,
    C is the number of input image channels, H is the height of the filter,
    and W is the width of the filter. If the groups is greater than 1,
    C will equal the number of input image channels divided by the groups.
    Please refer to UFLDL's `convolution
    <http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution/>`_
    for more details.
    If bias attribution and activation type are provided, bias is added to the
    output of the convolution, and the corresponding activation function is
    applied to the final result.

    For each input :math:`X`, the equation is:

353
    ..  math::
354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379

        Out = \sigma (W \\ast X + b)

    Where:

    * :math:`X`: Input value, a tensor with NCHW or NHWC format.
    * :math:`W`: Filter value, a tensor with MCHW format.
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.

    Example:

        - Input:

          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`

          Filter shape: :math:`(C_{out}, C_{in}, H_f, W_f)`

        - Output:

          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`

        Where

380
        ..  math::
381 382 383 384 385

            H_{out}&= \\frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]} + 1 \\\\
            W_{out}&= \\frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]} + 1

    Args:
386
        x (Tensor): The input is 4-D Tensor with shape [N, C, H, W], the data type 
387
            of input is float16 or float32 or float64.
388
        weight (Tensor): The convolution kernel with shape [M, C/g, kH, kW], where M is
389 390
            the number of output channels, g is the number of groups, kH is the filter's
            height, kW is the filter's width. 
391 392 393 394
        bias (Tensor, optional): The bias with shape [M,].
        stride (int|tuple): The stride size. It means the stride in convolution. 
            If stride is a tuple, it must contain two integers, (stride_height, stride_width). 
            Otherwise, stride_height = stride_width = stride. Default: stride = 1.
395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422
        padding (string|int|list|tuple): The padding size. It means the number of zero-paddings
            on both sides for each dimension.If `padding` is a string, either 'VALID' or
            'SAME' which is the padding algorithm. If padding size is a tuple or list,
            it could be in three forms: `[pad_height, pad_width]` or
            `[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`, and when 
            `data_format` is `"NCHW"`, `padding` can be in the form `[[0,0], [0,0], 
            [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `"NHWC"`, `pool_padding` can be in the form
            `[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
        dilation (int|tuple): The dilation size. It means the spacing between the kernel
            points. If dilation is a tuple, it must contain two integers, (dilation_height, 
            dilation_width). Otherwise, dilation_height = dilation_width = dilation. 
            Default: dilation = 1.
        groups (int): The groups number of the Conv2d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: groups=1.
        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"`, `"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]`.
        name(str, optional): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.

    Returns:
423
        A Tensor representing the conv2d result, whose data type is the same with input. 
424 425 426

    Raises:
        ValueError: If `data_format` is not "NCHW" or "NHWC".
427
        ValueError: If the channel dimension of the input is less than or equal to zero.
428 429 430 431 432 433 434 435 436 437 438 439
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
        ShapeError: If the input is not 4-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels * groups.
        ShapeError: If the number of output channels is not be divided by groups.

    Examples:
        .. code-block:: python

440
          import paddle
441 442 443 444 445 446
          import paddle.nn.functional as F
          import numpy as np

          x = np.random.randn(2, 3, 8, 8).astype(np.float32)
          w = np.random.randn(6, 3, 3, 3).astype(np.float32)

447 448 449 450 451 452 453
          paddle.disable_static()

          x_var = paddle.to_tensor(x)
          w_var = paddle.to_tensor(w)
          y_var = F.conv2d(x_var, w_var)
          y_np = y_var.numpy()

454 455 456 457 458 459 460 461 462 463 464
          print(y_np.shape)

          # (2, 6, 6, 6)
    """
    # entry checks
    if data_format not in ["NCHW", "NHWC"]:
        raise ValueError("Attr(data_format) should be 'NCHW' or 'NHWC'. "
                         "Received Attr(data_format): {}.".format(data_format))

    channel_last = (data_format == "NHWC")
    channel_dim = -1 if channel_last else 1
465
    num_channels = x.shape[channel_dim]
466 467
    num_filters = weight.shape[0]
    if num_channels < 0:
468
        raise ValueError("The channel dimension of the input({}) "
469
                         "should be defined. Received: {}.".format(
470
                             x.shape, num_channels))
471 472 473 474
    if num_channels % groups != 0:
        raise ValueError(
            "the channel of input must be divisible by groups,"
            "received: the channel of input is {}, the shape of input is {}"
475
            ", the groups is {}".format(num_channels, x.shape, groups))
476 477 478 479 480 481
    if num_filters % groups != 0:
        raise ValueError(
            "the number of filters must be divisible by groups,"
            "received: the number of filters is {}, the shape of weight is {}"
            ", the groups is {}".format(num_filters, weight.shape, groups))

482 483 484 485 486 487
    # use_cudnn = True if core.is_compiled_with_cuda() else False
    cudnn_version = get_cudnn_version()

    use_cudnn = True if (core.is_compiled_with_cuda() and
                         cudnn_version is not None) else False

488 489 490 491 492 493
    # update attrs
    padding, padding_algorithm = _update_padding_nd(padding, channel_last, 2)
    stride = utils.convert_to_list(stride, 2, 'stride')
    dilation = utils.convert_to_list(dilation, 2, 'dilation')

    l_type = "conv2d"
494
    if (num_channels == groups and num_filters % num_channels == 0):
495
        l_type = 'depthwise_conv2d'
496
        use_cudnn = False
497

498
    inputs = {'Input': [x], 'Filter': [weight]}
499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515
    attrs = {
        'strides': stride,
        'paddings': padding,
        'dilations': dilation,
        'groups': groups,
        'use_cudnn': use_cudnn,
        'use_mkldnn': False,
        'fuse_relu_before_depthwise_conv': False,
        "padding_algorithm": padding_algorithm,
        "data_format": data_format
    }

    if in_dygraph_mode():
        attrs = ('strides', stride, 'paddings', padding, 'dilations', dilation,
                 'groups', groups, 'use_cudnn', use_cudnn, 'use_mkldnn', False,
                 'fuse_relu_before_depthwise_conv', False, "padding_algorithm",
                 padding_algorithm, "data_format", data_format)
516
        pre_bias = getattr(core.ops, l_type)(x, weight, *attrs)
517
        if bias is not None:
518
            out = nn.elementwise_add(pre_bias, bias, axis=channel_dim)
519
        else:
520
            out = pre_bias
521
    else:
522
        inputs = {'Input': [x], 'Filter': [weight]}
523 524 525 526 527 528 529 530 531 532 533
        attrs = {
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
            'use_mkldnn': False,
            'fuse_relu_before_depthwise_conv': False,
            "padding_algorithm": padding_algorithm,
            "data_format": data_format
        }
534 535
        check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                                 'conv2d')
536 537 538 539 540 541 542
        helper = LayerHelper(l_type, **locals())
        dtype = helper.input_dtype()
        pre_bias = helper.create_variable_for_type_inference(dtype)
        outputs = {"Output": [pre_bias]}
        helper.append_op(
            type=l_type, inputs=inputs, outputs=outputs, attrs=attrs)
        if bias is not None:
543
            out = nn.elementwise_add(pre_bias, bias, axis=channel_dim)
544
        else:
545 546
            out = pre_bias

547 548 549
    return out


550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712
def conv_transpose1d(x,
                     weight,
                     bias=None,
                     stride=1,
                     padding=0,
                     output_padding=0,
                     groups=1,
                     dilation=1,
                     output_size=None,
                     data_format="NCL",
                     name=None):
    """
    The 1-D convolution transpose layer calculates the output based on the input,
    filter, and dilation, stride, padding. Input(Input) and output(Output)
    are in 'NCL' format or 'NLC' where N is batch size, C is the number of channels,
    L is the length of the feature. The details of convolution transpose
    layer, please refer to the following explanation and references
    `therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
    If bias attribution and activation type are provided, bias is added to
    the output of the convolution, and the corresponding activation function
    is applied to the final result.

    For each input :math:`X`, the equation is:

    .. math::

        Out = \sigma (W \\ast X + b)

    Where:

    * :math:`X`: Input value, a 3-D Tensor with 'NCL' format or 'NLC' format.
    * :math:`W`: Filter value, a 3-D Tensor with 'MCK' format.
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D Tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, a 3-D Tensor with data format 'NCL' or 'NLC', the shape of :math:`Out` and :math:`X` may be different.

    Example:

        - Input:

          Input shape: :math:`(N, C_{in}, L_{in})`

          Filter shape: :math:`(C_{in}, C_{out}, L_f)`

        - Output:

          Output shape: :math:`(N, C_{out}, L_{out})`

        Where

        .. math::

           L^\prime_{out} &= (L_{in} - 1) * stride - pad_top - pad_bottom + dilation * (L_f - 1) + 1 + output_padding \\\\
           L_{out} &\in [ L^\prime_{out}, L^\prime_{out} + stride ]

    Note:
          The conv1d_transpose can be seen as the backward of the conv1d. For conv1d,
          when stride > 1, conv1d maps multiple input shape to the same output shape,
          so for conv1d_transpose, when stride > 1, input shape maps multiple output shape.
          If output_size is None, :math:`L_{out} = L^\prime_{out}`;
          else, the :math:`L_{out}` of the output size must between :math:`L^\prime_{out}`
          and :math:`L^\prime_{out} + stride`. conv1d_transpose can compute the kernel size automatically.

    Args:
        x(Tensor): 3-D tensor with [N, C, L] or [N, L, C] format,
                         its data type is float32 or float64.
        weight(Tensor): The convolution kernel, a Tensor with shape [C, M/g, K],
            where M is the number of output channels(filters), g is the number of groups,
            K is the size of the kernel.
        bias(Tensor, optional): The bias, a Tensor with shape [M, ].
        stride(int|tuple|list, optional): The stride size. It means the stride in transposed convolution.
            If stride is a tuple, it must contain one integer, `(stride_size)`.
            Default: stride = 1.
        padding(int|list|str|tuple, optional): The padding size. The padding argument effectively adds
             `dilation * (kernel - 1)` amount of zero-padding on both sides of input. If `padding` is a
             string, either 'VALID' or 'SAME' supported, which is the padding algorithm.
             If `padding` is a tuple or list, it could be in two forms:
             `[pad]` or `[pad_left, pad_right]`. Default: padding = 0.
        output_padding(int|list|tuple, optional): The count of zeros to be added to tail of each dimension.
             If it is a tuple, it must contain one integer. Default: 0.
        groups(int, optional): The groups number of the conv1d transpose function. Inspired by
            grouped convolution in Alex Krizhevsky's Deep CNN paper, in which
            when group=2, the first half of the filters is only connected to the
            first half of the input channels, while the second half of the
            filters is only connected to the second half of the input channels.
            Default: groups = 1.
        dilation(int|tuple|list, optional): The dilation size. It means the spacing between the kernel points.
            If dilation is a tuple, it must contain one integer, `(dilation_size)`.
            Default: dilation = 1.
        output_size(int|tuple|list, optional): The output image size. If output size is a
            tuple, it must contain one integer, `(feature_length)`. None if use
            filter_size, padding, and stride to calculate output_size.
            If output_size and filter_size are specified at the same time, They
            should follow the formula above. Default: None. output_size and filter_size
            should not be None at the same time.
        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: `"NCL"`, `"NLC"`.
            The default is `"NCL"`. When it is `"NCL"`, the data is stored in the order of:
            `[batch_size, input_channels, input_length]`.
        name(str, optional): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.

    Returns:
        A  tensor representing the result of 1-D transpose convolution, whose
        data type is the same with input. And its shape is (num_batches, channels, length)
        when data_format is `"NCL"` and (num_batches, length, channels) when data_format is
        `"NLC"`.

    Raises:
        ValueError: If `data_format` is a string, but not "NCL" or "NLC".
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
        ValueError: If `output_size` and filter_size are None at the same time.
        ValueError: If `output_padding` is greater than `stride`.
        ShapeError: If the input is not 3-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 1.
        ShapeError: If the number of input channels is not equal to filter's channels.
        ShapeError: If the size of `output_size` is not equal to that of `stride`.

    Examples:
        .. code-block:: python



          import paddle
          import paddle.nn.functional as F
          import numpy as np
          
          paddle.disable_static()
          # shape: (1, 2, 4)
          x=np.array([[[4, 0, 9, 7],
                       [8, 0, 9, 2,]]]).astype(np.float32)
          # shape: (2, 1, 2)
          y=np.array([[[7, 0]],
                      [[4, 2]]]).astype(np.float32)
          x_var = paddle.to_tensor(x)
          w_var = paddle.to_tensor(w)
          y_var = F.conv_transpose1d(x_var, w_var)
          y_np = y_var.numpy()
          print y_np
          
          # [[[60. 16. 99. 75.  4.]]]
    """
    cudnn_version = get_cudnn_version()
    if cudnn_version is not None:
        use_cudnn = True
    else:
        use_cudnn = False

    if data_format not in ['NCL', 'NLC']:
        raise ValueError(
            "Attr(data_format) of conv2d_transpose got wrong value: "
            "received {}, but only 'NCL' or 'NLC' are supported.".format(
                data_format))
    channel_last = (data_format == "NLC")
    channel_dim = -1 if channel_last else 1

    num_channels = x.shape[channel_dim]
    if num_channels < 0:
713
        raise ValueError("The channel dimension of the input({}) "
714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730
                         "should be defined. Received: {}.".format(
                             x.shape, num_channels))
    if num_channels % groups != 0:
        raise ValueError(
            "the channel of input must be divisible by groups,"
            "received: the channel of input is {}, the shape of input is {}"
            ", the groups is {}".format(num_channels, x.shape, groups))

    # update attrs
    padding, padding_algorithm = _update_padding_nd(padding, channel_last, 1)

    if len(padding) == 2:
        padding = padding + [0] * 2
    elif len(padding) == 1:
        padding = padding + [0]
    else:
        raise ValueError(
731
            "The size of padding's dimension should 1 or 2. But got padding={}".
732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803
            format(padding))

    stride = utils.convert_to_list(stride, 1, 'stride') + [1]
    dilation = utils.convert_to_list(dilation, 1, 'dilation') + [1]
    output_padding = utils.convert_to_list(output_padding, 1,
                                           'output_padding') + [0]
    if output_padding[0] > stride[0]:
        raise ValueError(
            "The size of output_padding should not be greater than stride."
            "But got output_padding={} and stride={}".format(output_padding[0],
                                                             stride[0]))

    if output_size is None:
        output_size = []
    elif isinstance(output_size, (list, tuple, int)):
        output_size = utils.convert_to_list(output_size, 1, 'output_size') + [1]
    else:
        raise ValueError("output_size should be int, or list, tuple of ints")

    op_type = 'conv2d_transpose'
    num_filters = weight.shape[1]
    if (num_channels == groups and num_filters == 1 and not use_cudnn):
        op_type = 'depthwise_conv2d_transpose'
        use_cudnn = False

    squeeze_axis = -2 if channel_last else -1
    conv2d_data_format = "NHWC" if channel_last else "NCHW"

    x = nn.unsqueeze(input=x, axes=[squeeze_axis])
    weight = nn.unsqueeze(input=weight, axes=[-1])

    if in_dygraph_mode():
        attrs = ('output_size', output_size, 'strides', stride, 'paddings',
                 padding, 'padding_algorithm', padding_algorithm, 'dilations',
                 dilation, 'groups', groups, 'use_cudnn', use_cudnn,
                 'data_format', conv2d_data_format)
        out = getattr(core.ops, op_type)(x, weight, *attrs)
        if bias is not None:
            out = nn.elementwise_add(out, bias, axis=channel_dim)
    else:
        inputs = {'Input': [x], 'Filter': [weight]}
        attrs = {
            'output_size': output_size,
            'strides': stride,
            'paddings': padding,
            'padding_algorithm': padding_algorithm,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
            'data_format': conv2d_data_format
        }
        check_variable_and_dtype(x, 'input', ['float16', 'float32', 'float64'],
                                 'conv2d_transpose')
        helper = LayerHelper(op_type, **locals())
        dtype = helper.input_dtype()
        out = helper.create_variable_for_type_inference(dtype)
        outputs = {"Output": [out]}
        helper.append_op(
            type=op_type, inputs=inputs, outputs=outputs, attrs=attrs)
        if bias is not None:
            out = nn.elementwise_add(out, bias, axis=channel_dim)

    if output_size is None:
        out = pad2d(
            out,
            padding=[0, output_padding, 0, 0],
            data_format=conv2d_data_format,
            name=name)
    out = nn.squeeze(input=out, axes=[squeeze_axis])
    return out


L
LielinJiang 已提交
804
def conv_transpose2d(x,
805 806 807
                     weight,
                     bias=None,
                     stride=1,
L
LielinJiang 已提交
808 809 810
                     padding=0,
                     output_padding=0,
                     dilation=1,
811
                     groups=1,
L
LielinJiang 已提交
812
                     output_size=None,
813
                     data_format='NCHW',
814 815
                     name=None):
    """
S
swtkiwi 已提交
816

817 818 819 820 821 822 823 824 825 826 827
    The convolution2D transpose layer calculates the output based on the input,
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
    are in NCHW or NHWC format. Where N is batch size, C is the number of channels,
    H is the height of the feature, and W is the width of the feature.
    Parameters(dilations, strides, paddings) are two elements. These two elements
    represent height and width, respectively. The details of convolution transpose
    layer, please refer to the following explanation and references
    `therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
    If bias attribution and activation type are provided, bias is added to
    the output of the convolution, and the corresponding activation function
    is applied to the final result.
L
LielinJiang 已提交
828
    See more detail in :ref:`api_nn_conv_ConvTranspose2d` .
829 830 831

    For each input :math:`X`, the equation is:

832
    ..  math::
833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858

        Out = \sigma (W \\ast X + b)

    Where:

    * :math:`X`: Input value, a 4-D Tensor with NCHW or NHWC format.
    * :math:`W`: Filter value, a 4-D Tensor with MCHW format.
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D Tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, a 4-D Tensor with data format 'NCHW' or 'NHWC', the shape of :math:`Out` and :math:`X` may be different.

    Example:

        - Input:

          Input shape: :math:`(N, C_{in}, H_{in}, W_{in})`

          Filter shape: :math:`(C_{in}, C_{out}, H_f, W_f)`

        - Output:

          Output shape: :math:`(N, C_{out}, H_{out}, W_{out})`

        Where

859
        ..  math::
860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876

           H^\prime_{out} &= (H_{in} - 1) * strides[0] - pad_height_top - pad_height_bottom + dilations[0] * (H_f - 1) + 1 \\\\
           W^\prime_{out} &= (W_{in} - 1) * strides[1] - pad_width_left - pad_width_right + dilations[1] * (W_f - 1) + 1 \\\\
           H_{out} &\in [ H^\prime_{out}, H^\prime_{out} + strides[0] ] \\\\
           W_{out} &\in [ W^\prime_{out}, W^\prime_{out} + strides[1] ]

    Note:
          The conv2d_transpose can be seen as the backward of the conv2d. For conv2d, 
          when stride > 1, conv2d maps multiple input shape to the same output shape, 
          so for conv2d_transpose, when stride > 1, input shape maps multiple output shape.
          If output_size is None, :math:`H_{out} = H^\prime_{out}, W_{out} = W^\prime_{out}`; 
          else, the :math:`H_{out}` of the output size must between :math:`H^\prime_{out}` 
          and :math:`H^\prime_{out} + strides[0]`, and the :math:`W_{out}` of the output size must 
          between :math:`W^\prime_{out}` and :math:`W^\prime_{out} + strides[1]`, 
          conv2d_transpose can compute the kernel size automatically.

    Args:
L
LielinJiang 已提交
877
        x(Tensor): 4-D Tensor with [N, C, H, W] or [N, H, W, C] format,
878
            whose data type is float32 or float64.
L
LielinJiang 已提交
879
        weight(Tensor): The convolution kernel, a Tensor with shape [C, M/g, kH, kW],
880 881
            where M is the number of output channels(filters), g is the number of groups,
            kH is the height of the kernel, and kW is the width of the kernel.
L
LielinJiang 已提交
882 883 884 885
        bias(Tensor, optional): The bias, a Tensor with shape [M, ].
        stride(int|list|tuple, optional): The stride size. It means the stride in transposed convolution. 
            If stride is a tuple, it must contain two integers, (stride_height, stride_width). 
            Otherwise, stride_height = stride_width = stride. Default: stride = 1.
886 887 888 889 890 891 892 893
        padding(str|int|list|tuple, optional): The padding size. It means the number of zero-paddings 
            on both sides for each dimension. If `padding` is a string, either 'VALID' or 
            'SAME' which is the padding algorithm. If padding size is a tuple or list,
            it could be in three forms: `[pad_height, pad_width]` or 
            `[pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
            and when `data_format` is `"NCHW"`, `pool_padding` can be in the form 
            `[[0,0], [0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `"NHWC"`, `pool_padding` can be in the form 
894 895
            `[[0,0], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
L
LielinJiang 已提交
896 897
        output_padding(int|list|tuple, optional): Additional size added to one side
            of each dimension in the output shape. Default: 0.
898 899 900 901 902 903
        groups(int, optional): The groups number of the Conv2d transpose layer. Inspired by
            grouped convolution in Alex Krizhevsky's Deep CNN paper, in which
            when group=2, the first half of the filters is only connected to the
            first half of the input channels, while the second half of the
            filters is only connected to the second half of the input channels.
            Default: groups = 1.
904 905 906
        dilation(int|list|tuple, optional): The dilation size. It means the spacing between the kernel points. 
            If dilation is a tuple, it must contain two integers, (dilation_height, dilation_width). 
            Otherwise, dilation_height = dilation_width = dilation. Default: dilation = 1.
L
LielinJiang 已提交
907 908 909 910 911 912
        output_size(int|tuple|list, optional): The output image size. If output size is a
            tuple, it must contain two integers, (image_height, image_width). None if use
            filter_size, padding, and stride to calculate output_size.
            If output_size is specified, output_size and filter_size (weight)'s shape 
            should follow the formula above. Default: None. output_size and filter_size 
            should not be None at the same time.
913 914 915 916 917 918 919 920 921
        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"`, `"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]`.
        name(str, optional): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.

    Returns:
L
LielinJiang 已提交
922
        A Tensor representing the conv_transpose2d, whose 
923
        data type is the same with input and shape is (num_batches, channels, out_h, 
L
LielinJiang 已提交
924 925
        out_w) or (num_batches, out_h, out_w, channels). The tensor variable storing 
        transposed convolution result.
926 927 928 929 930 931

    Raises:
        ValueError: If `data_format` is not "NCHW" or "NHWC".
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
L
LielinJiang 已提交
932
        ValueError: If `output_size` and kernel_size are None at the same time.
933 934 935 936 937 938 939 940 941 942
        ShapeError: If the input is not 4-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels.
        ShapeError: If the size of `output_size` is not equal to that of `stride`.

    Examples:
        .. code-block:: python

          import numpy as np
L
LielinJiang 已提交
943 944
          import paddle
          import paddle.nn.functional as F
945 946 947 948

          x = np.random.randn(2, 3, 8, 8).astype(np.float32)
          w = np.random.randn(3, 6, 3, 3).astype(np.float32)

L
LielinJiang 已提交
949 950 951
          paddle.disable_static()
          x_var = paddle.to_tensor(x)
          w_var = paddle.to_tensor(w)
952
          y_var = F.conv_transpose2d(x_var, w_var)
L
LielinJiang 已提交
953
          y_np = y_var.numpy()
954 955 956 957 958 959 960 961 962 963 964 965
          print(y_np.shape)

          # (2, 6, 10, 10)
    """

    if data_format not in ['NCHW', 'NHWC']:
        raise ValueError(
            "Attr(data_format) of conv2d_transpose got wrong value: "
            "received {}, but only 'NCHW' or 'NHWC' are supported.".format(
                data_format))
    channel_last = (data_format == "NHWC")
    channel_dim = -1 if channel_last else 1
L
LielinJiang 已提交
966
    num_channels = x.shape[channel_dim]
967
    if num_channels < 0:
968
        raise ValueError("The channel dimension of the input({}) "
969
                         "should be defined. Received: {}.".format(
L
LielinJiang 已提交
970
                             x.shape, num_channels))
971 972 973 974
    if num_channels % groups != 0:
        raise ValueError(
            "the channel of input must be divisible by groups,"
            "received: the channel of input is {}, the shape of input is {}"
L
LielinJiang 已提交
975 976 977 978 979 980
            ", the groups is {}".format(num_channels, x.shape, groups))

    cudnn_version = get_cudnn_version()

    use_cudnn = True if (core.is_compiled_with_cuda() and
                         cudnn_version is not None) else False
981 982 983 984 985

    # update attrs
    padding, padding_algorithm = _update_padding_nd(padding, channel_last, 2)
    stride = utils.convert_to_list(stride, 2, 'stride')
    dilation = utils.convert_to_list(dilation, 2, 'dilation')
L
LielinJiang 已提交
986

987 988 989
    if output_size is None:
        output_size = []
    else:
L
LielinJiang 已提交
990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003
        if output_padding != 0:
            raise ValueError('output_padding option is mutually exclusive with '
                             'output_size')
        if isinstance(output_size, (list, tuple, int)):
            output_size = utils.convert_to_list(output_size, 2, 'output_size')
        else:
            raise ValueError(
                "output_size should be int, or list, tuple of ints")

    if output_padding == 0:
        output_padding = []
    else:
        output_padding = utils.convert_to_list(output_padding, 2,
                                               'output_padding')
1004 1005 1006

    op_type = 'conv2d_transpose'
    num_filters = weight.shape[1]
L
LielinJiang 已提交
1007
    if (num_channels == groups and num_filters == 1):
1008
        op_type = 'depthwise_conv2d_transpose'
L
LielinJiang 已提交
1009
        use_cudnn = False
1010 1011

    if in_dygraph_mode():
L
LielinJiang 已提交
1012 1013 1014 1015 1016
        attrs = ('output_padding', output_padding, 'output_size', output_size,
                 'strides', stride, 'paddings', padding, 'padding_algorithm',
                 padding_algorithm, 'dilations', dilation, 'groups', groups,
                 'use_cudnn', use_cudnn, 'data_format', data_format)
        pre_bias = getattr(core.ops, op_type)(x, weight, *attrs)
1017
        if bias is not None:
L
LielinJiang 已提交
1018
            out = nn.elementwise_add(pre_bias, bias, axis=channel_dim)
1019
        else:
L
LielinJiang 已提交
1020
            out = pre_bias
1021
    else:
L
LielinJiang 已提交
1022
        inputs = {'Input': [x], 'Filter': [weight]}
1023
        attrs = {
L
LielinJiang 已提交
1024
            'output_padding': output_padding,
1025 1026 1027 1028 1029 1030 1031 1032 1033
            'output_size': output_size,
            'strides': stride,
            'paddings': padding,
            'padding_algorithm': padding_algorithm,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
            'data_format': data_format
        }
L
LielinJiang 已提交
1034
        check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
1035 1036
                                 'conv2d_transpose')
        helper = LayerHelper(op_type, **locals())
L
LielinJiang 已提交
1037
        pre_bias = helper.create_variable_for_type_inference(x.dtype)
1038 1039 1040
        outputs = {"Output": [pre_bias]}
        helper.append_op(
            type=op_type, inputs=inputs, outputs=outputs, attrs=attrs)
L
LielinJiang 已提交
1041

1042
        if bias is not None:
L
LielinJiang 已提交
1043
            out = nn.elementwise_add(pre_bias, bias, axis=channel_dim)
1044
        else:
L
LielinJiang 已提交
1045 1046
            out = pre_bias

1047 1048 1049
    return out


1050
def conv3d(x,
1051 1052 1053
           weight,
           bias=None,
           stride=1,
1054
           padding=0,
1055 1056 1057 1058 1059
           dilation=1,
           groups=1,
           data_format="NCDHW",
           name=None):
    """
S
swtkiwi 已提交
1060

1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
    The convolution3D layer calculates the output based on the input, filter
    and strides, paddings, dilations, groups parameters. Input(Input) and
    Output(Output) are in NCDHW or NDHWC format. Where N is batch size C is the number of
    channels, D is the depth of the feature, H is the height of the feature,
    and W is the width of the feature. Convlution3D is similar with Convlution2D
    but adds one dimension(depth). If bias attribution and activation type are
    provided, bias is added to the output of the convolution, and the
    corresponding activation function is applied to the final result.

    For each input :math:`X`, the equation is:

1072
    ..  math::
1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097

        Out = \sigma (W \\ast X + b)

    In the above equation:

    * :math:`X`: Input value, a tensor with NCDHW or NDHWC format.
    * :math:`W`: Filter value, a tensor with MCDHW format.
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.

    Example:

        - Input:

          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`

          Filter shape: :math:`(C_{out}, C_{in}, D_f, H_f, W_f)`

        - Output:
          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`

        Where

1098
        ..  math::
1099 1100 1101 1102 1103 1104

            D_{out}&= \\frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{strides[0]} + 1 \\\\
            H_{out}&= \\frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{strides[1]} + 1 \\\\
            W_{out}&= \\frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{strides[2]} + 1

    Args:
1105
        x (Tensor): The input is 5-D Tensor with shape [N, C, D, H, W], the data 
1106 1107 1108 1109
            type of input is float16 or float32 or float64.
        weight (Variable): The convolution kernel, a Tensor with shape [M, C/g, kD, kH, kW],
            where M is the number of filters(output channels), g is the number of groups,
            kD, kH, kW are the filter's depth, height and width respectively.
1110 1111 1112 1113
        bias (Tensor, optional): The bias, a Tensor of shape [M, ].
        stride (int|tuple): The stride size. It means the stride in convolution. If stride is a 
            tuple, it must contain three integers, (stride_depth, stride_height, stride_width). 
            Otherwise, stride_depth = stride_height = stride_width = stride. Default: stride = 1.
1114 1115 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
        padding (string|int|list|tuple): The padding size. It means the number of zero-paddings 
            on both sides for each dimension. If `padding` is a string, either 'VALID' or
            'SAME' which is the padding algorithm. If padding size is a tuple or list,
            it could be in three forms: `[pad_depth, pad_height, pad_width]` or
            `[pad_depth_front, pad_depth_back, pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
            and when `data_format` is `"NCDHW"`, `pool_padding` can be in the form
            `[[0,0], [0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
            when `data_format` is `"NDHWC"`, `pool_padding` can be in the form
            `[[0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
        dilation (int|tuple): The dilation size. It means the spacing between the kernel points. 
            If dilation is a tuple, it must contain three integers, (dilation_depth, dilation_height,
            dilation_width). Otherwise, dilation_depth = dilation_height = dilation_width = dilation. 
            Default: dilation = 1.
        groups (int): The groups number of the Conv3d Layer. According to grouped
            convolution in Alex Krizhevsky's Deep CNN paper: when group=2,
            the first half of the filters is only connected to the first half
            of the input channels, while the second half of the filters is only
            connected to the second half of the input channels. Default: groups=1
        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"`, `"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]`.
        name(str|None): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.

    Returns:
1142
        A Tensor representing the conv3d, whose data type is 
1143 1144 1145 1146 1147 1148
        the same with input. If act is None, the tensor variable storing the 
        convolution result, and if act is not None, the tensor variable storing 
        convolution and non-linearity activation result.

    Raises:
        ValueError: If `data_format` is not "NCDHW" or "NDHWC".
1149
        ValueError: If the channel dimension of the input is less than or equal to zero.
1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
        ShapeError: If the input is not 5-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels * groups.
        ShapeError: If the number of output channels is not be divided by groups.

    Examples:
        .. code-block:: python

            import numpy as np
1163 1164
            import paddle
            import paddle.nn.functional as F
1165 1166 1167 1168

            x = np.random.randn(2, 3, 8, 8, 8).astype(np.float32)
            w = np.random.randn(6, 3, 3, 3, 3).astype(np.float32)

1169 1170 1171 1172 1173
            paddle.disable_static()
            x_var = paddle.to_tensor(x)
            w_var = paddle.to_tensor(w)
            y_var = F.conv3d(x_var, w_var)
            y_np = y_var.numpy()
1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185
            print(y_np.shape)

            # (2, 6, 6, 6, 6)
    """
    # entry check
    if data_format not in ["NCDHW", "NDHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCDHW' or 'NDHWC'. Received "
            "Attr(data_format): {}.".format(data_format))

    channel_last = (data_format == "NDHWC")
    channel_dim = -1 if channel_last else 1
1186
    num_channels = x.shape[channel_dim]
1187 1188 1189
    num_filters = weight.shape[0]
    if num_channels < 0:
        raise ValueError(
1190
            "The channel dimension of the input({}) should be defined. "
1191
            "Received: {}.".format(x.shape, num_channels))
1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202
    if num_channels % groups != 0:
        raise ValueError(
            "The number of input channels must be divisible by Attr(groups). "
            "Received: number of channels({}), groups({}).".format(num_channels,
                                                                   groups))
    if num_filters % groups != 0:
        raise ValueError(
            "The number of filters must be divisible by Attr(groups). "
            "Received: number of filters({}), groups({}).".format(num_filters,
                                                                  groups))

1203 1204 1205 1206
    cudnn_version = get_cudnn_version()
    use_cudnn = True if (core.is_compiled_with_cuda() and
                         cudnn_version is not None) else False

1207 1208 1209 1210 1211 1212 1213 1214 1215 1216
    padding, padding_algorithm = _update_padding_nd(padding, channel_last, 3)
    stride = utils.convert_to_list(stride, 3, 'stride')
    dilation = utils.convert_to_list(dilation, 3, 'dilation')
    op_type = "conv3d"

    if in_dygraph_mode():
        attrs = ('strides', stride, 'paddings', padding, 'dilations', dilation,
                 'groups', groups, 'use_cudnn', use_cudnn, 'use_mkldnn', False,
                 "padding_algorithm", padding_algorithm, "data_format",
                 data_format)
1217
        pre_bias = getattr(core.ops, op_type)(x, weight, *attrs)
1218
        if bias is not None:
1219
            out = nn.elementwise_add(pre_bias, bias, axis=channel_dim)
1220
        else:
1221
            out = pre_bias
1222
    else:
1223
        inputs = {'Input': [x], 'Filter': [weight]}
1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235
        attrs = {
            'strides': stride,
            'paddings': padding,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
            'use_mkldnn': False,
            "padding_algorithm": padding_algorithm,
            "data_format": data_format
        }
        helper = LayerHelper(op_type, **locals())
        dtype = helper.input_dtype()
1236 1237
        check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                                 'conv3d')
1238 1239 1240 1241 1242 1243 1244

        pre_bias = helper.create_variable_for_type_inference(dtype)
        outputs = {"Output": [pre_bias]}

        helper.append_op(
            type=op_type, inputs=inputs, outputs=outputs, attrs=attrs)
        if bias is not None:
1245
            out = nn.elementwise_add(pre_bias, bias, axis=channel_dim)
1246
        else:
1247
            out = pre_bias
1248 1249 1250 1251

    return out


L
LielinJiang 已提交
1252
def conv_transpose3d(x,
1253 1254 1255
                     weight,
                     bias=None,
                     stride=1,
L
LielinJiang 已提交
1256 1257
                     padding=0,
                     output_padding=0,
1258
                     groups=1,
L
LielinJiang 已提交
1259 1260
                     dilation=1,
                     output_size=None,
1261
                     data_format='NCDHW',
1262 1263
                     name=None):
    """
L
LielinJiang 已提交
1264
    The convolution3d transpose layer calculates the output based on the input,
1265 1266 1267 1268 1269 1270 1271 1272 1273 1274
    filter, and dilations, strides, paddings. Input(Input) and output(Output)
    are in NCDHW or NDHWC format. Where N is batch size, C is the number of channels,
    D is the depth of the feature, H is the height of the feature, and W
    is the width of the feature. Parameters(dilations, strides, paddings) are
    two elements. These two elements represent height and width, respectively.
    The details of convolution transpose layer, please refer to the following
    explanation and references `therein <https://arxiv.org/pdf/1603.07285.pdf>`_.
    If bias attribution and activation type are provided, bias is added to
    the output of the convolution, and the corresponding activation function
    is applied to the final result.
L
LielinJiang 已提交
1275
    See more detail in :ref:`api_nn_conv_ConvTranspose3d` .
1276 1277 1278

    For each input :math:`X`, the equation is:

1279
    ..  math::
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

        Out = \sigma (W \\ast X + b)

    In the above equation:

    * :math:`X`: Input value, a Tensor with NCDHW or NDHWC format.
    * :math:`W`: Filter value, a Tensor with MCDHW format.
    * :math:`\\ast`: Convolution operation.
    * :math:`b`: Bias value, a 2-D Tensor with shape [M, 1].
    * :math:`\\sigma`: Activation function.
    * :math:`Out`: Output value, the shape of :math:`Out` and :math:`X` may be different.

    Example:

        - Input:

          Input shape: :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`

          Filter shape: :math:`(C_{in}, C_{out}, D_f, H_f, W_f)`

        - Output:

          Output shape: :math:`(N, C_{out}, D_{out}, H_{out}, W_{out})`

        Where

1306
        ..  math::
1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327

           D^\prime_{out} &= (D_{in} - 1) * strides[0] - 2 * paddings[0] + dilations[0] * (D_f - 1) + 1 \\\\
           H^\prime_{out} &= (H_{in} - 1) * strides[1] - 2 * paddings[1] + dilations[1] * (H_f - 1) + 1 \\\\
           W^\prime_{out} &= (W_{in} - 1) * strides[2] - 2 * paddings[2] + dilations[2] * (W_f - 1) + 1 \\\\
           D_{out} &\in [ D^\prime_{out}, D^\prime_{out} + strides[0] ] \\\\
           H_{out} &\in [ H^\prime_{out}, H^\prime_{out} + strides[1] ] \\\\
           W_{out} &\in [ W^\prime_{out}, W^\prime_{out} + strides[2] ]

    Note:
          The conv3d_transpose can be seen as the backward of the conv3d. For conv3d, 
          when stride > 1, conv3d maps multiple input shape to the same output shape, 
          so for conv3d_transpose, when stride > 1, input shape maps multiple output shape.
          If output_size is None, :math:`H_{out} = H^\prime_{out}, :math:`H_{out} = \
          H^\prime_{out}, W_{out} = W^\prime_{out}`; else, the :math:`D_{out}` of the output 
          size must between :math:`D^\prime_{out}` and :math:`D^\prime_{out} + strides[0]`, 
          the :math:`H_{out}` of the output size must between :math:`H^\prime_{out}` 
          and :math:`H^\prime_{out} + strides[1]`, and the :math:`W_{out}` of the output size must 
          between :math:`W^\prime_{out}` and :math:`W^\prime_{out} + strides[2]`, 
          conv3d_transpose can compute the kernel size automatically.

    Args:
L
LielinJiang 已提交
1328
        x(Tensor): The input is 5-D Tensor with shape [N, C, D, H, W] or [N, D, H, W, C], the data type 
1329
            of input is float32 or float64.
L
LielinJiang 已提交
1330
        weight (Tensor): The convolution kernel, a Tensor with shape [C, M/g, kD, kH, kW],
1331 1332
            where M is the number of filters(output channels), g is the number of groups,
            kD, kH, kW are the filter's depth, height and width respectively.
L
LielinJiang 已提交
1333 1334 1335 1336 1337
        bias (Tensor, optional): The bias, a Tensor of shape [M, ].
        stride(int|list|tuple, optional): The stride size. It means the stride in transposed convolution. 
            If stride is a tuple, it must contain three integers, (stride_depth, stride_height, 
            stride_width). Otherwise, stride_depth = stride_height = stride_width = stride. 
            Default: stride = 1.
1338 1339 1340 1341
        padding (string|int|list|tuple, optional): The padding size. It means the number of zero-paddings 
            on both sides for each dimension. If `padding` is a string, either 'VALID' or
            'SAME' which is the padding algorithm. If padding size is a tuple or list,
            it could be in three forms: `[pad_depth, pad_height, pad_width]` or
1342
            `[pad_depth_front, pad_depth_back, pad_height_top, pad_height_bottom, pad_width_left, pad_width_right]`,
1343
            and when `data_format` is `"NCDHW"`, `pool_padding` can be in the form
1344
            `[[0,0], [0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right]]`.
1345
            when `data_format` is `"NDHWC"`, `pool_padding` can be in the form
1346 1347
            `[[0,0], [pad_depth_front, pad_depth_back], [pad_height_top, pad_height_bottom], [pad_width_left, pad_width_right], [0,0]]`.
            Default: padding = 0.
L
LielinJiang 已提交
1348 1349
        output_padding(int|list|tuple, optional): Additional size added to one side
            of each dimension in the output shape. Default: 0.
1350 1351 1352 1353 1354 1355
        groups(int, optional): The groups number of the Conv3d transpose layer. Inspired by
            grouped convolution in Alex Krizhevsky's Deep CNN paper, in which
            when group=2, the first half of the filters is only connected to the
            first half of the input channels, while the second half of the
            filters is only connected to the second half of the input channels.
            Default: groups=1
1356 1357 1358 1359
        dilation(int|list|tuple, optional): The dilation size. It means the spacing between the kernel points. 
            If dilation is a tuple, it must contain three integers, (dilation_depth, dilation_height, 
            dilation_width). Otherwise, dilation_depth = dilation_height = dilation_width = dilation. 
            Default: dilation = 1.
L
LielinJiang 已提交
1360 1361 1362 1363 1364
        output_size(int|list|tuple, optional): The output image size. If output size is a
            tuple, it must contain three integers, (image_depth, image_height, image_width). This
            parameter only works when filter_size is None. If output_size and filter_size are 
            specified at the same time, They should follow the formula above. Default: None. 
            Output_size and filter_size should not be None at the same time.
1365 1366 1367 1368
        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"`, `"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]`.
1369 1370 1371 1372 1373
        name(str, optional): For detailed information, please refer 
           to :ref:`api_guide_Name`. Usually name is no need to set and 
           None by default.

    Returns:
L
LielinJiang 已提交
1374
        A Tensor representing the conv_transpose3d, whose data 
1375 1376 1377 1378 1379 1380 1381 1382 1383 1384
        type is the same with input and shape is (num_batches, channels, out_d, out_h, 
        out_w) or (num_batches, out_d, out_h, out_w, channels). If act is None, the tensor 
        variable storing the transposed convolution result, and if act is not None, the tensor 
        variable storing transposed convolution and non-linearity activation result.

    Raises:
        ValueError: If `data_format` is not "NCDHW" or "NDHWC".
        ValueError: If `padding` is a string, but not "SAME" or "VALID".
        ValueError: If `padding` is a tuple, but the element corresponding to the input's batch size is not 0 
            or the element corresponding to the input's channel is not 0.
L
LielinJiang 已提交
1385
        ValueError: If `output_size` and kernel_size are None at the same time.
1386 1387 1388 1389 1390 1391 1392 1393
        ShapeError: If the input is not 5-D Tensor.
        ShapeError: If the input's dimension size and filter's dimension size not equal.
        ShapeError: If the dimension size of input minus the size of `stride` is not 2.
        ShapeError: If the number of input channels is not equal to filter's channels.
        ShapeError: If the size of `output_size` is not equal to that of `stride`.

    Examples:
       .. code-block:: python
L
LielinJiang 已提交
1394 1395
          
          import numpy as np
1396

L
LielinJiang 已提交
1397
          import paddle
1398 1399 1400 1401 1402
          import paddle.nn.functional as F

          x = np.random.randn(2, 3, 8, 8, 8).astype(np.float32)
          w = np.random.randn(3, 6, 3, 3, 3).astype(np.float32)

L
LielinJiang 已提交
1403 1404 1405 1406 1407 1408
          paddle.disable_static()

          x_var = paddle.to_tensor(x)
          w_var = paddle.to_tensor(w)
          y_var = F.conv_transpose3d(x_var, w_var)
          y_np = y_var.numpy()
1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420
          print(y_np.shape)

          # (2, 6, 10, 10, 10)
    """
    # entry checks
    if data_format not in ["NCDHW", "NDHWC"]:
        raise ValueError(
            "Attr(data_format) should be 'NCDHW' or 'NDHWC'. Received "
            "Attr(data_format): {}.".format(data_format))

    channel_last = (data_format == "NDHWC")
    channel_dim = -1 if channel_last else 1
L
LielinJiang 已提交
1421
    num_channels = x.shape[channel_dim]
1422 1423 1424
    num_filters = weight.shape[1]
    if num_channels < 0:
        raise ValueError(
1425
            "The channel dimension of the input({}) should be defined. "
L
LielinJiang 已提交
1426
            "Received: {}.".format(x.shape, num_channels))
1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438
    if num_channels % groups != 0:
        raise ValueError(
            "The number of input channels must be divisible by Attr(groups). "
            "Received: number of channels({}), groups({}).".format(num_channels,
                                                                   groups))

    padding, padding_algorithm = _update_padding_nd(padding, channel_last, 3)
    stride = utils.convert_to_list(stride, 3, 'stride')
    dilation = utils.convert_to_list(dilation, 3, 'dilation')
    if output_size is None:
        output_size = []
    else:
L
LielinJiang 已提交
1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458
        if output_padding != 0:
            raise ValueError('output_padding option is mutually exclusive with '
                             'output_size')
        if isinstance(output_size, (list, tuple, int)):
            output_size = utils.convert_to_list(output_size, 3, 'output_size')
        else:
            raise ValueError(
                "output_size should be int, or list, tuple of ints")

    if output_padding == 0:
        output_padding = []
    else:
        output_padding = utils.convert_to_list(output_padding, 3,
                                               'output_padding')

    cudnn_version = get_cudnn_version()

    #TODO(LielinJiang): whether to use cudnn according to the version of cudnn
    use_cudnn = True if (core.is_compiled_with_cuda() and
                         cudnn_version is not None) else False
1459 1460 1461 1462 1463

    op_type = 'conv3d_transpose'
    data_format_ = "NHWC" if channel_last else "NCHW"

    if in_dygraph_mode():
L
LielinJiang 已提交
1464 1465 1466 1467 1468
        attrs = ('output_padding', output_padding, 'output_size', output_size,
                 'paddings', padding, "padding_algorithm", padding_algorithm,
                 'strides', stride, 'dilations', dilation, 'groups', groups,
                 'use_cudnn', use_cudnn, "data_format", data_format_)
        pre_bias = getattr(core.ops, op_type)(x, weight, *attrs)
1469
        if bias is not None:
L
LielinJiang 已提交
1470
            out = nn.elementwise_add(pre_bias, bias, axis=channel_dim)
1471
        else:
L
LielinJiang 已提交
1472
            out = pre_bias
1473
    else:
L
LielinJiang 已提交
1474
        inputs = {'Input': [x], 'Filter': [weight]}
1475
        attrs = {
L
LielinJiang 已提交
1476
            'output_padding': output_padding,
1477 1478 1479 1480 1481 1482 1483 1484 1485 1486
            'output_size': output_size,
            'paddings': padding,
            "padding_algorithm": padding_algorithm,
            'strides': stride,
            'dilations': dilation,
            'groups': groups,
            'use_cudnn': use_cudnn,
            "data_format": data_format_
        }
        helper = LayerHelper(op_type, **locals())
L
LielinJiang 已提交
1487 1488
        check_variable_and_dtype(x, 'x', ['float16', 'float32', 'float64'],
                                 'conv3d')
1489

L
LielinJiang 已提交
1490
        pre_bias = helper.create_variable_for_type_inference(x.dtype)
1491 1492 1493 1494 1495
        outputs = {"Output": [pre_bias]}

        helper.append_op(
            type=op_type, inputs=inputs, outputs=outputs, attrs=attrs)
        if bias is not None:
L
LielinJiang 已提交
1496
            out = nn.elementwise_add(pre_bias, bias, axis=channel_dim)
1497
        else:
L
LielinJiang 已提交
1498
            out = pre_bias
1499 1500

    return out