def __init__(self, num_classes, num_distr, num_repeat):
super(ManifoldNetComplex, self).__init__()
params = {
'num_classes': num_classes,
'num_distr': num_distr,
'num_repeat': num_repeat
}
self.complex_conv1 = complex.ComplexConv2Deffgroup(
1, 5, (5, 5), (2, 2))
self.complex_conv2 = complex.ComplexConv2Deffgroup(
5, 10, (5, 5), (2, 2))
self.complex_conv3 = complex.ComplexConv2Deffgroup(
10, 20, (5, 5), (2, 2))
self.complex_conv4 = complex.ComplexConv2Deffgroup(
20, 40, (5, 5), (2, 2))
self.SURE = SURE(params, 40, 80, (3, 3), (1, 1), calc_next(3, 3, 1, 5))
self.complex_bn = complex.ComplexBN()
self.proj1 = complex.manifoldReLUv2angle(5) #complex.ReLU4Dsp(20)
self.proj2 = complex.manifoldReLUv2angle(10) #complex.ReLU4Dsp(40)
self.proj3 = complex.manifoldReLUv2angle(20) #complex.ReLU4Dsp(80)
self.proj4 = complex.manifoldReLUv2angle(40) #complex.ReLU4Dsp(80)
self.proj5 = complex.manifoldReLUv2angle(80) #complex.ReLU4Dsp(80)
self.linear1 = nn.Linear(80, 30)
self.linear2 = nn.Linear(30, 11)
self.relu = nn.ReLU()
self.name = 'complex only'
def __init__(self, num_classes, num_distr, num_repeat):
super(ManifoldNetR, self).__init__()
params = {
'num_classes': num_classes,
'num_distr': num_distr,
'num_repeat': num_repeat
}
self.complex_conv1 = complex.ComplexConv2Deffgroup(
1, 40, (5, 1), (2, 1))
self.complex_conv2 = complex.ComplexConv2Deffgroup(
40, 40, (5, 1), (2, 1))
self.SURE = SURE(params, 40, 40, (5, 1), (1, 1),
calc_next(29, (5, 1), 1, 20))
self.proj2 = complex.manifoldReLUv2angle(40) #complex.ReLU4Dsp(40)
self.relu = nn.ReLU()
self.conv_1 = nn.Conv2d(40, 50, (5, 1))
self.mp_1 = nn.MaxPool2d((3, 1))
self.conv_2 = nn.Conv2d(50, 60, (5, 1))
self.conv_3 = nn.Conv2d(60, 80, (3, 1))
self.bn_1 = nn.BatchNorm2d(50)
self.bn_2 = nn.BatchNorm2d(60)
self.bn_3 = nn.BatchNorm2d(80)
self.linear_2 = nn.Linear(80, 40)
self.linear_4 = nn.Linear(40, 11)
self.name = 'Complex Radio'
self.loss_weight = torch.nn.Parameter(torch.rand(1),
requires_grad=True)
def __init__(self, num_classes, num_distr, num_repeat):
super(Test, self).__init__()
params = {
'num_classes': num_classes,
'num_distr': num_distr,
'num_repeat': num_repeat
}
self.complex_conv1 = complex.ComplexConv2Deffgroup(
1, 20, (5, 5), (5, 5)) #20, 20
self.SURE = SURE_pure(params, calc_next(100, 5, 5, 20), 20)
self.name = 'complex+standardCNN'
self.proj2 = complex.manifoldReLUv2angle(20) #complex.ReLU4Dsp(40)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(0.3)
self.conv_1 = nn.Conv2d(20, 40, (5, 5))
self.mp_1 = nn.MaxPool2d((2, 2))
self.conv_2 = nn.Conv2d(40, 60, (3, 3))
self.mp_2 = nn.MaxPool2d((2, 2))
self.conv_3 = nn.Conv2d(60, 80, (3, 3))
self.bn_1 = nn.BatchNorm2d(40)
self.bn_2 = nn.BatchNorm2d(60)
self.bn_3 = nn.BatchNorm2d(80)
self.linear_2 = nn.Linear(80, 40)
self.linear_4 = nn.Linear(40, 11)
self.loss_weight = torch.nn.Parameter(torch.rand(1),
requires_grad=True)
def __init__(self):
super(ManifoldNetAll, self).__init__()
self.complex_conv1 = complex.ComplexConv2Deffgroup(
1, 5, (5, 5), (2, 2))
self.complex_conv2 = complex.ComplexConv2Deffgroup(
5, 10, (5, 5), (2, 2))
self.complex_conv3 = complex.ComplexConv2Deffgroup(
10, 20, (5, 5), (2, 2))
self.complex_conv4 = complex.ComplexConv2Deffgroup(
20, 40, (5, 5), (2, 2))
self.complex_conv5 = complex.ComplexConv2Deffgroup(
40, 80, (3, 3), (1, 1))
self.complex_bn = complex.ComplexBN()
self.proj1 = complex.manifoldReLUv2angle(5) #complex.ReLU4Dsp(20)
self.proj2 = complex.manifoldReLUv2angle(10) #complex.ReLU4Dsp(40)
self.proj3 = complex.manifoldReLUv2angle(20) #complex.ReLU4Dsp(80)
self.proj4 = complex.manifoldReLUv2angle(40) #complex.ReLU4Dsp(80)
self.proj5 = complex.manifoldReLUv2angle(80) #complex.ReLU4Dsp(80)
self.complex_fc = complex.ComplexLinearangle(80, 1, 30, 11) #10
def make_complex_layer(complex_c, complex_kern, complex_stride):
complex_parts = []
prev_c = 1
prev_relu = None
for (c, kern, stride) in zip(complex_c, complex_kern, complex_stride):
complex_parts.append(
complex.ComplexConv2Deffgroup(prev_c, c, (kern, kern),
(stride, stride)))
if prev_c == c and prev_relu is not None:
complex_parts.append(prev_relu)
else:
prev_relu = complex.manifoldReLUv2angle(c)
complex_parts.append(prev_relu)
prev_c = c
return complex_parts, prev_c
def __init__(self):
super(ManifoldNetComplex_11, self).__init__()
self.complex_conv1 = complex.ComplexConv2Deffgroup(
1, 20, (5, 5), (2, 2))
self.complex_conv2 = complex.ComplexConv2Deffgroup(
20, 20, (5, 5), (2, 2))
self.proj2 = complex.manifoldReLUv2angle(20) #complex.ReLU4Dsp(40)
# self.proj3 = complex.manifoldReLUv2angle(30)
self.relu = nn.ReLU()
self.linear_1 = complex.ComplexLinearangle2Dmw_outfield(20 * 22 * 22)
self.conv_1 = nn.Conv2d(20, 30, (5, 5))
self.mp_1 = nn.MaxPool2d((2, 2))
self.conv_2 = nn.Conv2d(30, 40, (5, 5), (3, 3))
self.bn_1 = nn.BatchNorm2d(30)
self.bn_2 = nn.BatchNorm2d(40)
self.mp_2 = nn.MaxPool2d((3, 3))
self.conv_3 = nn.Conv2d(40, 50, (2, 2))
self.bn_3 = nn.BatchNorm2d(50)
self.linear_2 = nn.Linear(50, 30)
self.linear_4 = nn.Linear(30, 11)
def __init__(self, num_classes, num_distr, num_repeat):
super(ManifoldNetRes1, self).__init__()
params = {
'num_classes': num_classes,
'num_distr': num_distr,
'num_repeat': num_repeat
}
self.complex_conv1 = complex.ComplexConv2Deffgroup(
1, 20, (5, 5), (2, 2))
self.complex_conv2 = complex.ComplexConv2Deffgroup(
20, 20, (5, 5), (2, 2))
self.complex_conv3 = complex.ComplexConv2Deffgroup(
20, 20, (5, 5), (1, 1))
self.SURE = SURE_pure(params, calc_next(22, 5, 1, 20), 20)
self.complex_res1 = complex.ResidualLayer(20, 20, 20, (5, 5), (2, 2))
self.complex_res2 = complex.ResidualLayer(20, 20, 20, (5, 5), (1, 1))
self.proj2 = complex.manifoldReLUv2angle(20) #complex.ReLU4Dsp(40)
self.relu = nn.ReLU()
self.conv_1 = nn.Conv2d(20, 30, (5, 5))
self.mp_1 = nn.MaxPool2d((3, 3))
self.conv_2 = nn.Conv2d(40, 50, (3, 3))
self.conv_3 = nn.Conv2d(60, 70, (2, 2))
self.bn_1 = nn.BatchNorm2d(30)
self.bn_2 = nn.BatchNorm2d(50)
self.bn_3 = nn.BatchNorm2d(70)
self.linear_2 = nn.Linear(70, 40)
self.linear_4 = nn.Linear(40, 11)
self.name = 'Residual complex for mstar'
self.loss_weight = torch.nn.Parameter(torch.rand(1),
requires_grad=True)
self.res1 = nn.Sequential(*self.make_res_block(30, 40))
self.id1 = nn.Conv2d(30, 40, (1, 1))
self.res2 = nn.Sequential(*self.make_res_block(50, 60))
self.id2 = nn.Conv2d(50, 60, (1, 1))
