import numpy
import six
from chainer import cuda
from chainer.functions.array import concat
from chainer.functions.pooling import max_pooling_2d
from chainer.functions.pooling import pooling_2d
class SpatialPyramidPooling2D(pooling_2d.Pooling2D):
"""Spatial pyramid pooling over a set of 2d planes."""
def __init__(self, x_shape, pyramid_height, pooling_class, use_cudnn=True):
bottom_c, bottom_h, bottom_w = x_shape
self.pyramid_height = pyramid_height
# create pooling functions for different pyramid levels
out_dim = 0
self.split_inds = []
self.poolers = []
for pyramid_level in six.moves.range(pyramid_height):
num_bins = int(2 ** pyramid_level)
ksize_h = int(numpy.ceil(bottom_h / (float(num_bins))))
remainder_h = ksize_h * num_bins - bottom_h
pad_h = remainder_h // 2
ksize_w = int(numpy.ceil(bottom_w / (float(num_bins))))
remainder_w = ksize_w * num_bins - bottom_w
pad_w = remainder_w // 2
ksize = (ksize_h, ksize_w)
pad = (pad_h, pad_w)
if pooling_class is max_pooling_2d.MaxPooling2D:
pooler = pooling_class(ksize=ksize, stride=None, pad=pad,
cover_all=True, use_cudnn=use_cudnn)
self.poolers.append(pooler)
else:
raise NotImplementedError()
out_dim += bottom_c * (num_bins ** 2)
if pyramid_level < pyramid_height - 1:
self.split_inds.append(out_dim)
def forward(self, x):
self.ys = []
for pooler in self.poolers:
y = pooler.forward(x)[0]
n, c, h, w = pooler.out_shape = y.shape
self.ys.append(y.reshape((n, c * h * w, 1, 1)))
return concat.Concat(axis=1).forward(self.ys)
def backward(self, x, gy):
xp = cuda.get_array_module(*x)
gx = xp.zeros_like(x[0])
gys = xp.split(gy[0], self.split_inds, axis=1)
for pooler, gy in zip(self.poolers, gys):
gy = gy.reshape(pooler.out_shape)
gx += pooler.backward(x, (gy,))[0]
return gx,
[docs]def spatial_pyramid_pooling_2d(x, pyramid_height, pooling_class,
use_cudnn=True):
"""Spatial pyramid pooling function.
It outputs a fixed-length vector regardless of input feature map size.
It performs pooling operation to the input 4D-array ``x`` with different
kernel sizes and padding sizes, and then flattens all dimensions except
first dimension of all pooling results, and finally concatenates them along
second dimension.
At :math:`i`-th pyramid level, the kernel size
:math:`(k_h^{(i)}, k_w^{(i)})` and padding size
:math:`(p_h^{(i)}, p_w^{(i)})` of pooling operation are calculated as
below:
.. math::
k_h^{(i)} &= \\lceil b_h / 2^i \\rceil, \\\\
k_w^{(i)} &= \\lceil b_w / 2^i \\rceil, \\\\
p_h^{(i)} &= (2^i k_h^{(i)} - b_h) / 2, \\\\
p_w^{(i)} &= (2^i k_w^{(i)} - b_w) / 2,
where :math:`\\lceil \\cdot \\rceil` denotes the ceiling function, and
:math:`b_h, b_w` are height and width of input variable ``x``,
respectively. Note that index of pyramid level :math:`i` is zero-based.
See detail in paper: `Spatial Pyramid Pooling in Deep Convolutional \
Networks for Visual Recognition \
<https://arxiv.org/abs/1406.4729>`_.
Args:
x (~chainer.Variable): Input variable. The shape of ``x`` should be
``(batchsize, # of channels, height, width)``.
pyramid_height (int): Number of pyramid levels
pooling_class (MaxPooling2D or AveragePooling2D):
Only MaxPooling2D class can be available for now.
use_cudnn (bool): If ``True`` and cuDNN is enabled, then this function
uses cuDNN as the core implementation.
Returns:
~chainer.Variable: Output variable. The shape of the output variable
will be :math:`(batchsize, c \\sum_{h=0}^{H-1} 2^{2h}, 1, 1)`,
where :math:`c` is the number of channels of input variable ``x``
and :math:`H` is the number of pyramid levels.
.. note::
This function uses some pooling classes as components to perform
spatial pyramid pooling. Now it supports only
:class:`~functions.MaxPooling2D` as elemental pooling operator so far.
"""
return SpatialPyramidPooling2D(x.shape[1:], pyramid_height,
pooling_class, use_cudnn=use_cudnn)(x)