vision.py 16.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
#   Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15
from ...device import get_cudnn_version
16
from ...static import Variable
R
ruri 已提交
17
from ...fluid.layer_helper import LayerHelper
18 19 20
from ...fluid.data_feeder import check_variable_and_dtype
from ...fluid import dygraph_utils
import numpy as np
W
wanghuancoder 已提交
21
from paddle import _C_ops
Z
zhiboniu 已提交
22 23
from ...device import is_compiled_with_rocm
from paddle import in_dynamic_mode
24
from paddle.framework import _non_static_mode
R
ruri 已提交
25

26 27
__all__ = []

28 29 30 31 32 33 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

def affine_grid(theta, out_shape, align_corners=True, name=None):
    """
    It generates a grid of (x,y) coordinates using the parameters of
    the affine transformation that correspond to a set of points where
    the input feature map should be sampled to produce the transformed
    output feature map.

    Args:
        theta (Tensor) - A tensor with shape [N, 2, 3]. It contains a batch of affine transform parameters.
                           The data type can be float32 or float64.
        out_shape (Tensor | list | tuple): The shape of target output with format [batch_size, channel, height, width].
                                             ``out_shape`` can be a Tensor or a list or tuple. The data
                                             type must be int32.
        align_corners(bool): Whether to align corners of target feature map and source feature map. Default: True.
        name(str|None): 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, A Tensor with shape [batch_size, H, W, 2] while 'H' and 'W' are the height and width of feature map in affine transformation. The data type is the same as `theta`.

    Raises:
        ValueError: If the type of arguments is not supported.

    Examples:

        .. code-block:: python

            import paddle
            import paddle.nn.functional as F
            import numpy as np
            # theta shape = [1, 2, 3]
            theta = np.array([[[-0.7, -0.4, 0.3],
                               [ 0.6,  0.5, 1.5]]]).astype("float32")
            theta_t = paddle.to_tensor(theta)
            y_t = F.affine_grid(
                    theta_t,
                    [1, 2, 3, 3],
                    align_corners=False)
W
whs 已提交
66
            print(y_t)
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81
            
            #[[[[ 1.0333333   0.76666665]
            #   [ 0.76666665  1.0999999 ]
            #   [ 0.5         1.4333333 ]]
            #
            #  [[ 0.5666667   1.1666666 ]
            #   [ 0.3         1.5       ]
            #   [ 0.03333333  1.8333334 ]]
            #
            #  [[ 0.10000002  1.5666667 ]
            #   [-0.16666666  1.9000001 ]
            #   [-0.43333334  2.2333333 ]]]]
    """
    if not isinstance(theta, Variable):
        raise ValueError("The theta should be a Tensor.")
82

83 84 85 86 87
    cudnn_version = get_cudnn_version()
    if cudnn_version is not None and cudnn_version >= 6000 and align_corners:
        use_cudnn = True
    else:
        use_cudnn = False
Z
zhiboniu 已提交
88
    if is_compiled_with_rocm():
89
        use_cudnn = False  # ROCM platform do not have MIOPEN kernel for affine_grid
90 91 92 93 94

    if not (isinstance(out_shape, list) or isinstance(out_shape, tuple) or \
            isinstance(out_shape, Variable)):
        raise ValueError("The out_shape should be a list, tuple or Tensor.")

Z
zhiboniu 已提交
95
    if in_dynamic_mode():
96 97
        _out_shape = out_shape.numpy().tolist() if isinstance(
            out_shape, Variable) else out_shape
W
wanghuancoder 已提交
98 99 100
        return _C_ops.affine_grid(theta, "output_shape", _out_shape,
                                  "align_corners", align_corners, "use_cudnn",
                                  use_cudnn)
101

102 103 104
    helper = LayerHelper('affine_grid')
    check_variable_and_dtype(theta, 'theta', ['float32', 'float64'],
                             'affine_grid')
105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120
    out = helper.create_variable_for_type_inference(theta.dtype)
    ipts = {'Theta': theta}
    attrs = {"align_corners": align_corners, "use_cudnn": use_cudnn}
    if isinstance(out_shape, Variable):
        ipts['OutputShape'] = out_shape
        check_variable_and_dtype(out_shape, 'out_shape', ['int32'],
                                 'affine_grid')
    else:
        attrs['output_shape'] = out_shape

    helper.append_op(
        type='affine_grid',
        inputs=ipts,
        outputs={'Output': out},
        attrs=None if len(attrs) == 0 else attrs)
    return out
121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137


def grid_sample(x,
                grid,
                mode='bilinear',
                padding_mode='zeros',
                align_corners=True,
                name=None):
    """
    This operation samples input X by using bilinear interpolation or
    nearest interpolation based on flow field grid, which is usually
    generated by :code:`affine_grid` . The grid of shape [N, H, W, 2]
    is the concatenation of (x, y) coordinates with shape [N, H, W] each,
    where x is indexing the 4th dimension (in width dimension) of input
    data x and y is indexing the 3rd dimension (in height dimension),
    finally results is the bilinear interpolation or nearest value of 4 nearest corner
    points. The output tensor shape will be [N, C, H, W].
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154


    Step 1:

    Get (x, y) grid coordinates and scale to [0, H-1/W-1].

    .. code-block:: text

        grid_x = 0.5 * (grid[:, :, :, 0] + 1) * (W - 1)
        grid_y = 0.5 * (grid[:, :, :, 1] + 1) * (H - 1)

    Step 2:
    
    Indices input data X with grid (x, y) in each [H, W] area, and bilinear
    interpolate point value by 4 nearest points or nearest interpolate point value
    by nearest point.

155
    .. code-block:: text
156 157 158 159 160 161 162 163 164 165 166

        wn ------- y_n ------- en
        |           |           |
        |          d_n          |
        |           |           |
        x_w --d_w-- grid--d_e-- x_e
        |           |           |
        |          d_s          |
        |           |           |
        ws ------- y_s ------- wn

167 168 169 170 171 172 173 174 175 176 177 178 179
        For bilinear interpolation:
        x_w = floor(x)              // west side x coord
        x_e = x_w + 1               // east side x coord
        y_n = floor(y)              // north side y coord
        y_s = y_s + 1               // south side y coord
        d_w = grid_x - x_w          // distance to west side
        d_e = x_e - grid_x          // distance to east side
        d_n = grid_y - y_n          // distance to north side
        d_s = y_s - grid_y          // distance to south side
        wn = X[:, :, y_n, x_w]      // north-west point value
        en = X[:, :, y_n, x_e]      // north-east point value
        ws = X[:, :, y_s, x_w]      // south-east point value
        es = X[:, :, y_s, x_w]      // north-east point value
180

181
        output = wn * d_e * d_s + en * d_w * d_s
182 183
                + ws * d_e * d_n + es * d_w * d_n

184 185 186 187 188 189 190 191 192 193
    Args:
        x(Tensor): The input tensor, which is a 4-d tensor with shape
                     [N, C, H, W], N is the batch size, C is the channel
                     number, H and W is the feature height and width.
                     The data type is float32 or float64.
        grid(Tensor): Input grid tensor of shape [N, grid_H, grid_W, 2]. The
                        data type is float32 or float64.
        mode(str, optional): The interpolation method which can be 'bilinear' or 'nearest'.
                         Default: 'bilinear'.
        padding_mode(str, optional) The padding method used when source index
194
                   is out of input images. It can be 'zeros', 'reflection' and 'border'.
195 196 197 198 199 200 201
                   Default: zeros.
        align_corners(bool, optional): If `align_corners` is true, it will projects
                   -1 and 1 to the centers of the corner pixels. Otherwise, it will
                   projects -1 and 1 to the image edges.
        name(str, optional): For detailed information, please refer
                             to :ref:`api_guide_Name`. Usually name is no need to set and
                             None by default.
202 203 204 205

    Returns:
        Tensor, The shape of output is [N, C, grid_H, grid_W] in which `grid_H` is the height of grid and `grid_W` is the width of grid. The data type is same as input tensor.

206
    Examples:
207

208
        .. code-block:: python
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
            import paddle
            import paddle.nn.functional as F
            import numpy as np
            
            # shape=[1, 1, 3, 3]
            x = np.array([[[[-0.6,  0.8, -0.5],
                            [-0.5,  0.2,  1.2],
                            [ 1.4,  0.3, -0.2]]]]).astype("float64")
            
            # grid shape = [1, 3, 4, 2]
            grid = np.array(
                         [[[[ 0.2,  0.3],
                            [-0.4, -0.3],
                            [-0.9,  0.3],
                            [-0.9, -0.6]],
                           [[ 0.4,  0.1],
                            [ 0.9, -0.8],
                            [ 0.4,  0.5],
                            [ 0.5, -0.2]],
                           [[ 0.1, -0.8],
                            [-0.3, -1. ],
                            [ 0.7,  0.4],
                            [ 0.2,  0.8]]]]).astype("float64")
            
234
            
235 236 237 238 239 240 241 242
            x = paddle.to_tensor(x)
            grid = paddle.to_tensor(grid)
            y_t = F.grid_sample(
                x,
                grid,
                mode='bilinear',
                padding_mode='border',
                align_corners=True)
W
whs 已提交
243
            print(y_t)
244 245 246 247 248 249
            
            # output shape = [1, 1, 3, 4]
            # [[[[ 0.34   0.016  0.086 -0.448]
            #    [ 0.55  -0.076  0.35   0.59 ]
            #    [ 0.596  0.38   0.52   0.24 ]]]]
    """
250

251
    _modes = ['bilinear', 'nearest']
252
    _padding_modes = ['zeros', 'reflection', 'border']
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
    if mode not in _modes:
        raise ValueError(
            "The mode of grid sample function should be in {}, but got: {}".
            format(_modes, mode))
    if padding_mode not in _padding_modes:
        raise ValueError(
            "The padding mode of grid sample function should be in {}, but got: {}".
            format(_padding_modes, padding_mode))

    if not isinstance(align_corners, bool):
        raise ValueError("The align corners should be bool, but got: {}".format(
            align_corners))

    cudnn_version = get_cudnn_version()
    use_cudnn = False
Z
zhiboniu 已提交
268
    if not is_compiled_with_rocm() and (
269 270
            cudnn_version is not None
    ) and align_corners and mode == 'bilinear' and padding_mode == 'zeros':
271
        use_cudnn = True
W
whs 已提交
272 273 274
        # CUDNN always computes gradients for all inputs
        x.stop_gradient = False
        grid.stop_gradient = False
275

Z
zhiboniu 已提交
276
    if in_dynamic_mode():
277 278
        attrs = ('mode', mode, 'padding_mode', padding_mode, 'align_corners',
                 align_corners, 'use_cudnn', use_cudnn)
W
wanghuancoder 已提交
279
        out = getattr(_C_ops, 'grid_sampler')(x, grid, *attrs)
280
    else:
281 282 283 284 285 286 287 288 289 290 291
        helper = LayerHelper("grid_sample", **locals())
        check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'grid_sample')
        check_variable_and_dtype(grid, 'grid', ['float32', 'float64'],
                                 'grid_sample')
        ipts = {'X': x, 'Grid': grid}
        attrs = {
            'mode': mode,
            'padding_mode': padding_mode,
            'align_corners': align_corners,
            'use_cudnn': use_cudnn
        }
292 293 294 295 296 297 298
        out = helper.create_variable_for_type_inference(x.dtype)
        helper.append_op(
            type='grid_sampler',
            inputs=ipts,
            attrs=attrs,
            outputs={'Output': out})
    return out
R
ruri 已提交
299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315


def pixel_shuffle(x, upscale_factor, data_format="NCHW", name=None):
    """
    This API implements pixel shuffle operation.
    See more details in :ref:`api_nn_vision_PixelShuffle` .
    Parameters:
        x(Tensor): 4-D tensor, the data type should be float32 or float64.
        upscale_factor(int): factor to increase spatial resolution.
        data_format (str): The data format of the input and output data. 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): The default value is None.  Normally there is no need for user to set this property.
    Returns:
        Out(tensor): Reshaped tensor according to the new dimension.
    Raises:
        ValueError: If the square of upscale_factor cannot divide the channels of input.
    Examples:
        .. code-block:: python
316

R
ruri 已提交
317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333
            import paddle
            import paddle.nn.functional as F
            import numpy as np
            x = np.random.randn(2, 9, 4, 4).astype(np.float32)
            x_var = paddle.to_tensor(x)
            out_var = F.pixel_shuffle(x_var, 3)
            out = out_var.numpy()
            # (2, 1, 12, 12)
    """
    if not isinstance(upscale_factor, int):
        raise TypeError("upscale factor must be int type")

    if data_format not in ["NCHW", "NHWC"]:
        raise ValueError("Attr(data_format) should be 'NCHW' or 'NHWC'."
                         "But recevie Attr(data_format): {} ".format(
                             data_format))

Z
zhiboniu 已提交
334
    if in_dynamic_mode():
W
wanghuancoder 已提交
335 336
        return _C_ops.pixel_shuffle(x, "upscale_factor", upscale_factor,
                                    "data_format", data_format)
R
ruri 已提交
337 338

    helper = LayerHelper("pixel_shuffle", **locals())
339
    check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'pixel_shuffle')
R
ruri 已提交
340 341 342 343 344 345 346 347
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type="pixel_shuffle",
        inputs={"X": x},
        outputs={"Out": out},
        attrs={"upscale_factor": upscale_factor,
               "data_format": data_format})
    return out
348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415


def channel_shuffle(x, groups, data_format="NCHW", name=None):
    """
    This API implements channel shuffle operation.
    See more details in :ref:`api_nn_vision_ChannelShuffle` .

    Parameters:
        x (Tensor): 4-D tensor, the data type should be float32 or float64.
        groups (int): Number of groups to divide channels in.
        data_format (str): The data format of the input and output data. An optional string of 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].
        name (str, optional): Name for the operation (optional, default is None). Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`.

    Returns:
        Out (Tensor): Rearranged tensor keeping the original tensor shape.

    Examples:
        .. code-block:: python
            :name: channel_shuffle-example

            import paddle
            import paddle.nn.functional as F
            x = paddle.arange(0, 0.6, 0.1, 'float32')
            x = paddle.reshape(x, [1, 6, 1, 1])
            # [[[[0.        ]],
            #   [[0.10000000]],
            #   [[0.20000000]],
            #   [[0.30000001]],
            #   [[0.40000001]],
            #   [[0.50000000]]]]
            y = F.channel_shuffle(x, 3)
            # [[[[0.        ]],
            #   [[0.20000000]],
            #   [[0.40000001]],
            #   [[0.10000000]],
            #   [[0.30000001]],
            #   [[0.50000000]]]]
    """
    if len(x.shape) != 4:
        raise ValueError(
            "Input x should be 4D tensor, but received x with the shape of {}".
            format(x.shape))

    if not isinstance(groups, int):
        raise TypeError("groups must be int type")

    if groups <= 0:
        raise ValueError("groups must be positive")

    if data_format not in ["NCHW", "NHWC"]:
        raise ValueError("Attr(data_format) should be 'NCHW' or 'NHWC'."
                         "But recevie Attr(data_format): {} ".format(
                             data_format))

    if _non_static_mode():
        return _C_ops.channel_shuffle(x, "groups", groups, "data_format",
                                      data_format)

    helper = LayerHelper("channel_shuffle", **locals())
    check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'channel_shuffle')
    out = helper.create_variable_for_type_inference(dtype=x.dtype)
    helper.append_op(
        type="channel_shuffle",
        inputs={"X": x},
        outputs={"Out": out},
        attrs={"groups": groups,
               "data_format": data_format})
    return out