def __init__(self,
dim: tuple = (4, 4),
wave_name='haar',
init_mode='he_normal'):
super(PrimaryDWT, self).__init__()
self.caps_conv = CapsConv2d(dim)
self.dwt = DWTForward(wave_name)
def _make_layer(self, down_sample_times: int, channel: int):
layer_list = []
for i in range(down_sample_times):
layer_list.append(nn.Conv2d(channel, channel // 4, 1))
layer_list.append(DWTForward())
layer_list.append(nn.BatchNorm2d(channel))
return nn.Sequential(*layer_list)
def __init__(self, in_planes, planes, stride=1):
super(BasicBlockDWT, self).__init__()
self.in_planes = in_planes
self.planes = planes
self.stride = stride
self.conv1 = nn.Conv2d(self.in_planes,
self.planes,
kernel_size=3,
stride=self.stride,
padding=1,
bias=False)
self.bn1 = nn.BatchNorm2d(self.planes)
self.conv2 = nn.Conv2d(self.planes,
self.planes,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn2 = nn.BatchNorm2d(self.planes)
if self.stride == 2:
self.conv1 = DWTForward()
self.bn1 = nn.BatchNorm2d(self.in_planes * 4)
self.conv2 = nn.Conv2d(self.in_planes * 4,
self.planes,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn2 = nn.BatchNorm2d(self.planes)
else:
self.conv1 = nn.Conv2d(self.in_planes,
self.planes,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn1 = nn.BatchNorm2d(self.planes)
self.conv2 = nn.Conv2d(self.planes,
self.planes,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn2 = nn.BatchNorm2d(self.planes)
self.relu = nn.ReLU()
self.shortcut = nn.Sequential()
if self.stride != 1 or self.in_planes != self.expansion * self.planes:
self.shortcut = nn.Sequential(
nn.Conv2d(self.in_planes,
self.expansion * self.planes,
kernel_size=1,
stride=self.stride,
bias=False),
nn.BatchNorm2d(self.expansion * self.planes))