Performs deformable region-of-interest pooling on inputs. As described
in `Deformable Convolutional Networks <https://arxiv.org/abs/1703.06211>`_, it will get offset for each bin after
roi pooling so that pooling at correct region. Batch_size will change to the number of region bounding boxes after deformable_roi_pooling.
The operation has three steps:
Args:
1. Dividing each region proposal into equal-sized sections with the pooled_width and pooled_height.
input (Variable):The input of Deformable PSROIPooling.The shape of input tensor is
[N,C,H,W]. Where N is batch size,C is number of input channels,H
2. Add offset to pixel in ROI to get new location and the new value which are computed directly through
is height of the feature, and W is the width of the feature.
bilinear interpolation with four nearest pixel.
rois (Variable): ROIs (Regions of Interest) to pool over.It should be
a 2-D LoDTensor of shape (num_rois, 4), the lod level
3. Sample several points in each bin to get average values as output.
is 1. Given as [[x1, y1, x2, y2], ...], (x1, y1) is
the top left coordinates, and (x2, y2) is the bottom
right coordinates.
Args:
trans (Variable): Offset of features on ROIs while pooling.The format is NCHW, where
input (Variable):The input of deformable roi pooling and it is tensor which value type is float32. The shape of input is
N is number of ROIs, C is number of channels, which indicate the offset distance
[N, C, H, W]. Where N is batch size, C is number of input channels,
in the x and y directions, H is pooled height, and W is pooled width.
H is height of the feature, and W is the width of the feature.
no_trans (bool): Whether to add offset to get new value or not while roi pooling, which
rois (Variable): ROIs (Regions of Interest) with type float32 to pool over. It should be
value is True or False. Default: False.
a 2-D LoDTensor of shape (num_rois, 4), and the lod level
spatial_scale (float): Ratio of input feature map height (or width) to raw image height (or width).
is 1. Given as [[x1, y1, x2, y2], ...], (x1, y1) is
Equals the reciprocal of total stride in convolutional layers, Default: 1.0.
the top left coordinates, and (x2, y2) is the bottom
group_size (list|tuple): The number of groups which input channels are divided.(eg.number of input channels
right coordinates, which value type is float32.
is k1*k2*(C+1), which k1 and k2 are group width and height and C+1 is number of output
trans (Variable): Offset of features on ROIs while pooling which value type is float32. The format is [N, C, H, W], where
chanels. eg.(4, 6), which 4 is height of group and 6 is width of group. Default: [1, 1].
N is number of ROIs, C is number of channels, which indicate the offset distance
pooled_height (integer): The pooled output height. Default: 1.
in the x and y directions, H is pooled height, and W is pooled width.
pooled_width (integer): The pooled output width. Default: 1.
no_trans (bool): Whether to add offset to get new value or not while roi pooling, which value with type bool is True or False.
part_size (list|tuple): The height and width of offset, eg.(4, 6), which height is 4 and width is 6, Default:
If value is True, no offset will be added in operation. Default: False.
if None, default value is [pooled_height, pooled_width].
spatial_scale (float): Ratio of input feature map height (or width) to raw image height (or width), which value type is float32.
sample_per_part (integer): The number of samples in each bin. Default: 1.
Equals the reciprocal of total stride in convolutional layers, Default: 1.0.
trans_std (float): Coefficient of offset. Default: 0.1.
group_size (list|tuple): The number of groups which input channels are divided and the input is list or tuple, which value type is int32. (eg.number of input channels
position_sensitive (bool): Whether to choose deformable psroi pooling mode or not. Default: False.
is k1 * k2 * (C + 1), which k1 and k2 are group width and height and C+1 is number of output
name (str): Name of layer. Default: None.
chanels.) eg.(4, 6), which 4 is height of group and 6 is width of group. Default: [1, 1].
Returns:
pooled_height (int): The pooled output height which value type is int32. Default: 1.
Variable: The tensor variable storing the deformable psroi pooling \
pooled_width (int): The pooled output width which value type is int32. Default: 1.
result.
part_size (list|tuple): The height and width of offset which values in list or tuple is int32, eg.(4, 6), which height is 4 and width is 6, and values always equal to pooled_height \
and pooled_width. Default: if None, default value is [pooled_height, pooled_width].
sample_per_part (int): The number of samples in each bin which value type is int32. If value is bigger, it will consume more performance. Default: 1.
trans_std (float): Coefficient of offset which value type is float32. It controls weight of offset. Default: 0.1.
position_sensitive (bool): Whether to choose deformable psroi pooling mode or not, and value type is bool(True or False). If value is False, input dimension equals to output dimension. \
If value is True, input dimension shoule be output dimension * pooled_height * pooled_width. Default: False.
name (str|None): Name of layer. Default: None.
Returns:
Variable: Output of deformable roi pooling is that, if position sensitive is False, input dimension equals to output dimension. If position sensitive is True,\
input dimension should be the result of output dimension divided by pooled height and pooled width.
Examples:
Examples:
.. code-block:: python
.. code-block:: python
# position_sensitive=True
import paddle.fluid as fluid
input = fluid.layers.data(name="input",
shape=[2, 192, 64, 64],
dtype='float32',
append_batch_size=False)
rois = fluid.layers.data(name="rois",
shape=[4],
dtype='float32',
lod_level=1)
trans = fluid.layers.data(name="trans",
shape=[2, 384, 64, 64],
dtype='float32',
append_batch_size=False)
x = fluid.layers.nn.deformable_roi_pooling(input=input,
