def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
dilation=1,
bias=False,
transpose=False):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = make_list(kernel_size, 4)
self.stride = make_list(stride, 4)
self.dilation = make_list(dilation, 4)
if isinstance(self.kernel_size, str):
if self.kernel_size == 'hypercross':
self.kernel = nn.Parameter(torch.zeros(9, in_channels, out_channels))
else:
raise NotImplementedError
elif np.prod(self.kernel_size) > 1:
self.kernel = nn.Parameter(torch.zeros(np.prod(self.kernel_size), in_channels, out_channels))
else:
assert not transpose
self.kernel = nn.Parameter(torch.zeros(in_channels, out_channels))
self.bias = None if not bias else nn.Parameter(torch.zeros(out_channels))
self.t = transpose
self.init_weight()
if kernel_size == 1:
assert not transpose
def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
dilation=1,
bias=False,
transpose=False):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = make_list(kernel_size)
self.stride = make_list(stride)
self.dilation = make_list(dilation)
if np.prod(self.kernel_size) > 1:
self.kernel = nn.Parameter(torch.zeros(np.prod(self.kernel_size), in_channels, out_channels))
else:
assert not transpose
self.kernel = nn.Parameter(torch.zeros(in_channels, out_channels))
self.bias = None if not bias else nn.Parameter(torch.zeros(out_channels))
self.t = transpose
self.init_weight()
def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
dilation=1,
bias=False,
transpose=False):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = make_list(kernel_size)
self.stride = make_list(stride)
self.dilation = make_list(dilation)
assert(np.prod(kernel_size) > 1)
assert(not transpose)
self.kernel = nn.Parameter(torch.zeros(np.prod(self.kernel_size), in_channels, out_channels))
self.bias = None if not bias else nn.Parameter(torch.zeros(out_channels))
self.pkernel = nn.Parameter(torch.zeros(np.prod(self.kernel_size), in_channels, 3 * np.prod(self.kernel_size)))
self.pkernel.register_hook(self._set_lr)
self.init_weight()
def __init__(
self,
kernel_size,
stride,
dilation,
):
if isinstance(kernel_size, str):
if kernel_size == 'hypercross':
self.stride = make_list(stride, 4)
self.dilation = make_list(dilation, 4)
kernel_offset = np.array([[0, 0, 0, 0], [1, 0, 0, 0],
[-1, 0, 0, 0], [0, 1, 0, 0],
[0, -1, 0, 0], [0, 0, 1, 0],
[0, 0, -1, 0], [0, 0, 0, 1],
[0, 0, 0, -1]])
kernel_offset *= self.stride * self.dilation
else:
raise NotImplementedError
else:
self.kernel_size = make_list(kernel_size, 4)
self.stride = make_list(stride, 4)
self.dilation = make_list(dilation, 4)
t_offset = (np.arange(-self.kernel_size[0] // 2 + 1,
self.kernel_size[0] // 2 + 1) *
self.stride[0] * self.dilation[0]).tolist()
x_offset = (np.arange(-self.kernel_size[1] // 2 + 1,
self.kernel_size[1] // 2 + 1) *
self.stride[1] * self.dilation[1]).tolist()
y_offset = (np.arange(-self.kernel_size[2] // 2 + 1,
self.kernel_size[2] // 2 + 1) *
self.stride[2] * self.dilation[2]).tolist()
z_offset = (np.arange(-self.kernel_size[3] // 2 + 1,
self.kernel_size[3] // 2 + 1) *
self.stride[3] * self.dilation[3]).tolist()
kernel_offset = np.array([[t, x, y, z] for t in t_offset
for z in z_offset for y in y_offset
for x in x_offset])
self.kernel_offset = torch.from_numpy(kernel_offset).int()
def __init__(self, kernel_size, stride, dilation):
self.kernel_size = make_list(kernel_size, 3)
self.stride = make_list(stride, 3)
self.dilation = make_list(dilation, 3)
x_offset = (np.arange(-self.kernel_size[0] // 2 + 1,
self.kernel_size[0] // 2 + 1) * self.stride[0] *
self.dilation[0]).tolist()
y_offset = (np.arange(-self.kernel_size[1] // 2 + 1,
self.kernel_size[1] // 2 + 1) * self.stride[1] *
self.dilation[1]).tolist()
z_offset = (np.arange(-self.kernel_size[2] // 2 + 1,
self.kernel_size[2] // 2 + 1) * self.stride[2] *
self.dilation[2]).tolist()
kernel_offset = np.array([[x, y, z] for z in z_offset for y in y_offset
for x in x_offset])
self.kernel_offset = torch.from_numpy(kernel_offset).int()
def __init__(self,
in_channels,
kernel_size=3,
stride=1,
dilation=1,
bias=False,
transpose=False):
super().__init__()
self.in_channels = in_channels
self.kernel_size = make_list(kernel_size, 4)
self.stride = make_list(stride, 4)
self.dilation = make_list(dilation, 4)
assert (np.prod(self.kernel_size) > 1)
self.kernel = nn.Parameter(torch.zeros(np.prod(self.kernel_size), in_channels, 1))
self.bias = None
self.t = transpose
self.init_weight()