def __init__(self, f_in, ratio, nratio, srange=1, nb_classes=10):
super().__init__()
'''
Scale equivariant arch, based on architecture in Kanazawa's paper https://arxiv.org/abs/1412.5104
selecting srange = 1 is equivalent to the paper
'''
self.f_in = f_in
self.ratio = ratio
self.nratio = nratio
self.srange = srange
self.nb_classes = nb_classes
self.conv1 = ScaleConvolution(self.f_in,
69,
3,
self.ratio,
self.nratio,
srange=0,
boundary_condition="dirichlet",
stride=2)
self.conv2 = ScaleConvolution(69,
122,
3,
self.ratio,
self.nratio,
srange=self.srange,
boundary_condition="dirichlet")
self.pool = ScalePool(self.ratio)
self.fc1 = nn.Linear(122, 288, bias=True)
self.fc2 = nn.Linear(288, self.nb_classes, bias=True)
def __init__(self, f_in, ratio, nratio, nb_classes=10):
super().__init__()
'''
Basic scale equivariant architecture, replicating StdNet
'''
self.f_in = f_in
self.ratio = ratio
self.nratio = nratio
self.nb_classes = nb_classes
self.conv1 = ScaleConvolution(self.f_in,
12,
7,
self.ratio,
self.nratio,
srange=0,
boundary_condition="dirichlet",
stride=2)
self.conv2 = ScaleConvolution(12,
21,
5,
self.ratio,
self.nratio,
srange=1,
boundary_condition="dirichlet")
self.pool = ScalePool(self.ratio)
self.fc1 = nn.Linear(21, 150, bias=True)
self.fc2 = nn.Linear(150, self.nb_classes, bias=True)
def __init__(self, ratio, nratio):
super().__init__()
self.ratio = ratio
self.nratio = nratio
self.conv1 = ScaleConvolution(7,
3,
12,
self.ratio,
self.nratio,
srange=0,
boundary_condition="dirichlet",
stride=2)
self.conv2 = ScaleConvolution(5,
12,
21,
self.ratio,
self.nratio,
srange=1,
boundary_condition="dirichlet")
self.pool = ScalePool(self.ratio)
self.fc1 = nn.Linear(21, 150, bias=True)
self.fc2 = nn.Linear(150, 10, bias=True)
def __init__(self,
f_in=1,
size=5,
ratio=2**(2 / 3),
nratio=3,
srange=2,
padding=0,
nb_classes=10):
super().__init__()
'''
Scale equivariant arch with 3 convolutional layers
'''
self.f_in = f_in
self.size = size
self.ratio = ratio
self.nratio = nratio
self.srange = srange
self.padding = padding
self.nb_classes = nb_classes
self.conv1 = ScaleConvolution(self.f_in,
96,
self.size,
self.ratio,
self.nratio,
srange=0,
boundary_condition="dirichlet",
padding=self.padding,
stride=2)
self.conv2 = ScaleConvolution(96,
96,
self.size,
self.ratio,
self.nratio,
srange=self.srange,
boundary_condition="dirichlet",
padding=self.padding)
self.conv3 = ScaleConvolution(96,
192,
self.size,
self.ratio,
self.nratio,
srange=self.srange,
boundary_condition="dirichlet",
padding=self.padding)
self.pool = ScalePool(self.ratio)
self.fc1 = nn.Linear(192, 150, bias=True)
self.fc2 = nn.Linear(150, self.nb_classes, bias=True)
def __init__(self, f_in, ratio, nratio, nb_classes=10):
super().__init__()
'''
Smaller version of the squale equivariant All CNN
'''
self.f_in = f_in
self.ratio = ratio
self.nratio = nratio
self.nb_classes = nb_classes
self.conv1 = ScaleConvolution(self.f_in,
96,
3,
ratio=self.ratio,
nratio=self.nratio,
srange=0,
boundary_condition="dirichlet",
padding=1)
self.conv2 = ScaleConvolution(96,
96,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1)
self.conv3 = ScaleConvolution(96,
192,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1)
self.conv4 = ScaleConvolution(192,
192,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1)
self.weight5 = nn.Parameter(torch.empty(self.nb_classes, 192))
nn.init.orthogonal_(self.weight5)
def __init__(self, f_in, ratio, nratio, nb_classes=10):
super().__init__()
'''
Squale equivariant All convolutional netw
'''
self.f_in = f_in
self.ratio = ratio
self.nratio = nratio
self.nb_classes = nb_classes
self.conv1 = ScaleConvolution(self.f_in,
96,
3,
ratio=self.ratio,
nratio=self.nratio,
srange=0,
boundary_condition="dirichlet",
padding=1)
self.conv2 = ScaleConvolution(96,
96,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1)
self.conv3 = ScaleConvolution(96,
96,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1,
stride=2)
self.conv4 = ScaleConvolution(96,
192,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1)
self.conv5 = ScaleConvolution(192,
192,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1)
self.conv6 = ScaleConvolution(192,
192,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1,
stride=2)
self.conv7 = ScaleConvolution(192,
192,
3,
self.ratio,
self.nratio,
srange=2,
boundary_condition="dirichlet",
padding=1)
self.weight8 = nn.Parameter(torch.empty(192, 192))
nn.init.orthogonal_(self.weight8)
self.weight9 = nn.Parameter(torch.empty(self.nb_classes, 192))
nn.init.orthogonal_(self.weight9)
def __init__(self,
f_in=1,
size=5,
ratio=2**(2 / 3),
nratio=3,
srange=2,
padding=0,
nb_classes=10,
factor=1):
super().__init__()
'''
Scale equivariant arch with 3 convolutional layers
'''
self.f_in = f_in
self.size = size
self.ratio = ratio
self.nratio = nratio
self.srange = srange
self.padding = padding
self.nb_classes = nb_classes
self.conv1 = ScaleConvolution(self.f_in,
int(factor * 12),
self.size,
self.ratio,
self.nratio,
srange=0,
boundary_condition="dirichlet",
padding=self.padding,
stride=2)
self.conv2 = ScaleConvolution(int(factor * 12),
int(factor * 21),
self.size,
self.ratio,
self.nratio,
srange=self.srange,
boundary_condition="dirichlet",
padding=self.padding)
self.conv3 = ScaleConvolution(int(factor * 21),
int(factor * 36),
self.size,
self.ratio,
self.nratio,
srange=self.srange,
boundary_condition="dirichlet",
padding=self.padding)
self.conv4 = ScaleConvolution(int(factor * 36),
int(factor * 36),
self.size,
self.ratio,
self.nratio,
srange=self.srange,
boundary_condition="dirichlet",
padding=self.padding)
self.conv5 = ScaleConvolution(int(factor * 36),
int(factor * 64),
self.size,
self.ratio,
self.nratio,
srange=self.srange,
boundary_condition="dirichlet",
padding=self.padding)
self.conv6 = ScaleConvolution(int(factor * 64),
int(factor * 64),
self.size,
self.ratio,
self.nratio,
srange=self.srange,
boundary_condition="dirichlet",
padding=self.padding)
self.pool = ScalePool(self.ratio)
self.fc = nn.Linear(int(factor * 64), self.nb_classes, bias=True)