def forward_gpu(self, x):
if (cuda.cudnn_enabled and self.use_cudnn
and pooling_nd._check_cudnn_acceptable_type(x[0].dtype)):
# With cuDNN v3 or greater, use cuDNN implementation for inputs
# with spatial dimensions of two or more.
if _cudnn_version >= 3000 and self.ndim >= 2:
return super(MaxPoolingND, self).forward_gpu(x)
# With cuDNN v2, use cuDNN implementation only for inputs with
# spatial dimensions of two.
elif self.ndim == 2:
return super(MaxPoolingND, self).forward_gpu(x)
n, c = x[0].shape[:2]
dims = x[0].shape[2:]
ys = tuple(
conv_nd.get_conv_outsize(d, k, s, p, self.cover_all)
for (d, k, s,
p) in six.moves.zip(dims, self.ksize, self.stride, self.pad))
# (n, c, y_1, y_2, ..., y_N)
y_shape = (n, c) + ys
y = cuda.cupy.empty(y_shape, dtype=x[0].dtype)
self.indexes = cuda.cupy.empty(y_shape, dtype=numpy.int32)
in_params, out_params, operation, name = \
max_pooling_nd_kernel.MaxPoolingNDKernelForward.generate(self.ndim)
cuda.elementwise(in_params, out_params, operation,
name)(x[0].reduced_view(),
*(dims + ys + self.ksize + self.stride +
self.pad + (y, self.indexes)))
return y,
def forward_gpu(self, x):
if (chainer.should_use_cudnn('>=auto') and
pooling_nd._check_cudnn_acceptable_type(x[0].dtype)):
# With cuDNN v3 or greater, use cuDNN implementation for inputs
# with spatial dimensions of two or more.
if _cudnn_version >= 3000 and self.ndim >= 2:
return super(AveragePoolingND, self).forward_gpu(x)
# With cuDNN v2, use cuDNN implementation only for inputs with
# spatial dimensions of two.
elif self.ndim == 2:
return super(AveragePoolingND, self).forward_gpu(x)
self.retain_inputs(())
self._in_shape = x[0].shape
self._in_dtype = x[0].dtype
n, c = x[0].shape[:2]
dims = x[0].shape[2:]
ys = tuple(conv_nd.get_conv_outsize(d, k, s, p,
cover_all=self.cover_all)
for (d, k, s, p) in six.moves.zip(
dims, self.ksize, self.stride, self.pad))
# (n, c, y_1, y_2, ..., y_N)
y_shape = (n, c) + ys
y = cuda.cupy.empty(y_shape, dtype=x[0].dtype)
coeff = 1. / functools.reduce(operator.mul, self.ksize)
in_params, out_params, operation, name = \
average_pooling_nd_kernel.AveragePoolingNDKernelForward.generate(
self.ndim)
cuda.elementwise(in_params, out_params, operation, name)(
x[0].reduced_view(),
*(dims + ys + self.ksize + self.stride + self.pad + (coeff, y)))
return y,
def forward_gpu(self, x):
if (cuda.cudnn_enabled and self.use_cudnn and
pooling_nd._check_cudnn_acceptable_type(x[0].dtype)):
# With cuDNN v3 or greater, use cuDNN implementation for inputs
# with spatial dimensions of two or more.
if _cudnn_version >= 3000 and self.ndim >= 2:
return super(MaxPoolingND, self).forward_gpu(x)
# With cuDNN v2, use cuDNN implementation only for inputs with
# spatial dimensions of two.
elif self.ndim == 2:
return super(MaxPoolingND, self).forward_gpu(x)
n, c = x[0].shape[:2]
dims = x[0].shape[2:]
ys = tuple(conv_nd.get_conv_outsize(d, k, s, p, self.cover_all)
for (d, k, s, p) in six.moves.zip(
dims, self.ksize, self.stride, self.pad))
# (n, c, y_1, y_2, ..., y_N)
y_shape = (n, c) + ys
y = cuda.cupy.empty(y_shape, dtype=x[0].dtype)
self.indexes = cuda.cupy.empty(y_shape, dtype=numpy.int32)
in_params, out_params, operation, name = \
max_pooling_nd_kernel.MaxPoolingNDKernelForward.generate(self.ndim)
cuda.elementwise(in_params, out_params, operation, name)(
x[0].reduced_view(),
*(dims + ys + self.ksize + self.stride + self.pad +
(y, self.indexes)))
return y,
def backward_gpu(self, x, gy):
if (cuda.cudnn_enabled and self.use_cudnn and
pooling_nd._check_cudnn_acceptable_type(x[0].dtype)):
# With cuDNN v3 or greater, use cuDNN implementation for inputs
# with spatial dimensions of two or more.
if _cudnn_version >= 3000 and self.ndim >= 2:
return super(AveragePoolingND, self).backward_gpu(x, gy)
# With cuDNN v2, use cuDNN implementation only for inputs with
# spatial dimensions of two.
elif self.ndim == 2:
return super(AveragePoolingND, self).backward_gpu(x, gy)
n, c = x[0].shape[:2]
dims = x[0].shape[2:]
ys = gy[0].shape[2:]
gx = cuda.cupy.empty_like(x[0])
coeff = 1. / functools.reduce(operator.mul, self.ksize)
in_params, out_params, operation, name = \
average_pooling_nd_kernel.AveragePoolingNDKernelBackward.generate(
self.ndim)
cuda.elementwise(in_params, out_params, operation, name)(
gy[0].reduced_view(),
*(dims + ys + self.ksize + self.stride + self.pad + (coeff, gx)))
return gx,
def backward_gpu(self, x, gy):
if (cuda.cudnn_enabled and self.use_cudnn
and pooling_nd._check_cudnn_acceptable_type(x[0].dtype)):
# With cuDNN v3 or greater, use cuDNN implementation for inputs
# with spatial dimensions of two or more.
if _cudnn_version >= 3000 and self.ndim >= 2:
return super(AveragePoolingND, self).backward_gpu(x, gy)
# With cuDNN v2, use cuDNN implementation only for inputs with
# spatial dimensions of two.
elif self.ndim == 2:
return super(AveragePoolingND, self).backward_gpu(x, gy)
n, c = x[0].shape[:2]
dims = x[0].shape[2:]
ys = gy[0].shape[2:]
gx = cuda.cupy.empty_like(x[0])
coeff = 1. / functools.reduce(operator.mul, self.ksize)
in_params, out_params, operation, name = \
average_pooling_nd_kernel.AveragePoolingNDKernelBackward.generate(
self.ndim)
cuda.elementwise(in_params, out_params, operation, name)(
gy[0].reduced_view(),
*(dims + ys + self.ksize + self.stride + self.pad + (coeff, gx)))
return gx,
