def __init__(self, num_scales, in_channels, out_channels, kernel_size, num_groups, num_layers, kernel_sizes):
super(KernelDecoder, self).__init__()
self.num_scales = num_scales
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.num_groups = num_groups
self.num_layers = num_layers
self.kernel_sizes = kernel_sizes
# channels
self.num_channels = self.num_scales * self.out_channels * (self.in_channels // self.num_groups)
# modules
self.decoder = nn.Sequential(
ConvPool2D(
channels = [self.num_channels] * 2,
sampling_type = 'UP-DECONV',
sampling_sizes = self.kernel_size
),
ConvPool2D(
channels = [self.num_channels] * self.num_layers,
kernel_sizes = self.kernel_sizes
),
nn.BatchNorm2d(self.num_channels)
)
self.apply(weights_init)
def __init__(self, scales, channels, kernel_sizes):
super(MotionDecoder, self).__init__()
self.scales = scales
self.channels = channels
self.kernel_sizes = kernel_sizes
# modules
self.decoder = ConvPool2D(channels=self.channels,
kernel_sizes=self.kernel_sizes)
self.apply(weights_init)
def __init__(self, channels, kernel_sizes, sampling_sizes):
super(MotionEncoder, self).__init__()
self.channels = channels
self.kernel_sizes = kernel_sizes
self.sampling_sizes = sampling_sizes
# modules
self.encoder = ConvPool2D(channels=self.channels,
kernel_sizes=self.kernel_sizes,
last_nonlinear=False,
sampling_type='SUB-MAXPOOL',
sampling_sizes=self.sampling_sizes)
self.apply(weights_init)
def __init__(self, num_scales, channels, kernel_sizes, sampling_sizes):
super(ImageEncoder, self).__init__()
self.num_scales = num_scales
self.channels = channels
self.kernel_sizes = kernel_sizes
self.sampling_sizes = sampling_sizes
# modules
self.encoders = []
for k in range(num_scales):
self.encoders.append(
ConvPool2D(channels=self.channels,
kernel_sizes=self.kernel_sizes,
last_nonlinear=True,
sampling_type='SUB-MAXPOOL',
sampling_sizes=self.sampling_sizes))
self.add_module('encoder-{0}'.format(k + 1), self.encoders[-1])
self.apply(weights_init)