def __init__(self, reduction: str = 'mean', **kwargs):
r""""""
super().__init__()
self.reduce = build_reduce(reduction)
self.kwargs = {
k: v for k, v in kwargs.items()
if k not in ['dim', 'keepdim']
}
def __init__(
self,
window_size: int = 11,
n_channels: int = 3,
reduction: str = 'mean',
**kwargs,
):
r""""""
super().__init__()
self.register_buffer('window', create_window(window_size, n_channels))
self.reduce = build_reduce(reduction)
self.kwargs = kwargs
def __init__(
self,
network: str = 'alex',
scaling: bool = True,
dropout: bool = True,
pretrained: bool = True,
reduction: str = 'mean',
):
r""""""
super().__init__()
# ImageNet scaling
self.scaling = scaling
self.register_buffer('shift', _SHIFT.view(1, -1, 1, 1))
self.register_buffer('scale', _SCALE.view(1, -1, 1, 1))
# Perception layers
if network == 'alex': # AlexNet
layers = models.alexnet(pretrained=True).features
targets = [1, 4, 7, 9, 11]
channels = [64, 192, 384, 256, 256]
elif network == 'squeeze': # SqueezeNet
layers = models.squeezenet1_1(pretrained=True).features
targets = [1, 4, 7, 9, 10, 11, 12]
channels = [64, 128, 256, 384, 384, 512, 512]
elif network == 'vgg': # VGG16
layers = models.vgg16(pretrained=True).features
targets = [3, 8, 15, 22, 29]
channels = [64, 128, 256, 512, 512]
else:
raise ValueError('Unknown network architecture ' + network)
self.net = Intermediary(layers, targets)
for p in self.net.parameters():
p.requires_grad = False
# Linear comparators
self.lin = nn.ModuleList([
nn.Sequential(
nn.Dropout(inplace=True) if dropout else nn.Identity(),
nn.Conv2d(c, 1, kernel_size=1, stride=1, padding=0,
bias=False),
) for c in channels
])
if pretrained:
self.lin.load_state_dict(get_weights(network=network))
self.reduce = build_reduce(reduction)
def __init__(self, reduction: str = 'mean', **kwargs):
r""""""
super().__init__()
self.reduce = build_reduce(reduction)
self.kwargs = kwargs