def __init__(self):
super().__init__()
n_channel = 56
self.preBlock = nn.Sequential(
nn.Conv2d(1, n_channel, 9, stride=1, padding=4, groups=1),
nn.PReLU())
# ResBlock 8
self.blocks = nn.Sequential(
SRBlock(n_channel, n_channel),
SRBlock(n_channel, n_channel),
SRBlock(n_channel, n_channel),
SRBlock(n_channel, n_channel),
SRBlock(n_channel, n_channel),
SRBlock(n_channel, n_channel),
SRBlock(n_channel, n_channel),
SRBlock(n_channel, n_channel),
)
self.postBlock = nn.Sequential(
nn.Conv2d(n_channel, n_channel, 3, stride=1, padding=1,
bias=False),
nn.BatchNorm2d(n_channel),
)
self.final = nn.Sequential(
nn.Conv2d(n_channel, 2, 9, stride=1, padding=4, groups=1), )
self.symmetry_amp = Lambda(partial(symmetry, mode="real"))
self.symmetry_imag = Lambda(partial(symmetry, mode="imag"))
self.elu = GeneralELU(add=+(1 + 1e-10))
def __init__(self, img_size):
super().__init__()
self.conv1_phase = nn.Sequential(
*conv_phase(1, 4, (23, 23), 1, 11, 1, add=-2.1415))
self.conv2_phase = nn.Sequential(
*conv_phase(4, 8, (21, 21), 1, 10, 1, add=-2.1415))
self.conv3_phase = nn.Sequential(
*conv_phase(8, 12, (17, 17), 1, 8, 1, add=-2.1415))
self.conv_con1_phase = nn.Sequential(
LocallyConnected2d(12, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(-2.1415),
)
self.conv4_phase = nn.Sequential(
*conv_phase(1, 4, (5, 5), 1, 3, 2, add=-2.1415))
self.conv5_phase = nn.Sequential(
*conv_phase(4, 8, (5, 5), 1, 2, 1, add=-2.1415))
self.conv6_phase = nn.Sequential(
*conv_phase(8, 12, (3, 3), 1, 3, 2, add=-2.1415))
self.conv7_phase = nn.Sequential(
*conv_phase(12, 16, (3, 3), 1, 1, 1, add=-2.1415))
self.conv_con2_phase = nn.Sequential(
LocallyConnected2d(16, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(-2.1415),
)
self.conv8_phase = nn.Sequential(
*conv_phase(1, 4, (3, 3), 1, 1, 1, add=-2.1415))
self.conv9_phase = nn.Sequential(
*conv_phase(4, 8, (3, 3), 1, 1, 1, add=-2.1415))
self.conv10_phase = nn.Sequential(
*conv_phase(8, 12, (3, 3), 1, 2, 2, add=-2.1415))
self.conv_con3_phase = nn.Sequential(
LocallyConnected2d(12, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(-2.1415),
)
self.symmetry_imag = Lambda(partial(symmetry, mode="imag"))
def __init__(self, img_size):
super().__init__()
self.img_size = img_size
self.conv1_amp = nn.Sequential(
nn.Conv2d(1, 4, stride=2, kernel_size=3, padding=3 // 2),
nn.ReLU())
self.conv1_phase = nn.Sequential(
nn.Conv2d(1, 4, stride=2, kernel_size=3, padding=3 // 2),
GeneralELU(1 - pi))
self.conv2_amp = nn.Sequential(
nn.Conv2d(4, 8, stride=2, kernel_size=3, padding=3 // 2),
nn.ReLU())
self.conv2_phase = nn.Sequential(
nn.Conv2d(4, 8, stride=2, kernel_size=3, padding=3 // 2),
GeneralELU(1 - pi))
self.conv3_amp = nn.Sequential(
nn.Conv2d(8, 16, stride=2, kernel_size=3, padding=3 // 2),
nn.ReLU())
self.conv3_phase = nn.Sequential(
nn.Conv2d(8, 16, stride=2, kernel_size=3, padding=3 // 2),
GeneralELU(1 - pi),
)
self.conv4_amp = nn.Sequential(
nn.Conv2d(16, 32, stride=2, kernel_size=3, padding=3 // 2),
nn.ReLU())
self.conv4_phase = nn.Sequential(
nn.Conv2d(16, 32, stride=2, kernel_size=3, padding=3 // 2),
GeneralELU(1 - pi),
)
self.conv5_amp = nn.Sequential(
nn.Conv2d(32, 64, stride=2, kernel_size=3, padding=3 // 2),
nn.ReLU())
self.conv5_phase = nn.Sequential(
nn.Conv2d(32, 64, stride=2, kernel_size=3, padding=3 // 2),
GeneralELU(1 - pi),
)
self.final_amp = nn.Sequential(nn.AdaptiveAvgPool2d(1), nn.Flatten(),
nn.Linear(64, img_size**2))
self.final_phase = nn.Sequential(nn.AdaptiveAvgPool2d(1), nn.Flatten(),
nn.Linear(64, img_size**2))
def __init__(self, img_size):
super().__init__()
torch.cuda.set_device(1)
self.img_size = img_size
self.preBlock_amp = nn.Sequential(
nn.Conv2d(1, 64, 7, stride=2, padding=3), nn.BatchNorm2d(64),
nn.ReLU())
self.preBlock_phase = nn.Sequential(
nn.Conv2d(1, 64, 7, stride=2, padding=3),
nn.BatchNorm2d(64),
GeneralELU(1 - pi),
)
self.maxpool_amp = nn.MaxPool2d(3, 2, 1)
self.maxpool_phase = nn.MaxPool2d(3, 2, 1)
# first block
self.layer1_amp = nn.Sequential(ResBlock_amp(64, 64),
ResBlock_amp(64, 64))
self.layer1_phase = nn.Sequential(ResBlock_phase(64, 64),
ResBlock_phase(64, 64))
self.layer2_amp = nn.Sequential(ResBlock_amp(64, 128, stride=2),
ResBlock_amp(128, 128))
self.layer2_phase = nn.Sequential(ResBlock_phase(64, 128, stride=2),
ResBlock_phase(128, 128))
self.layer3_amp = nn.Sequential(ResBlock_amp(128, 256, stride=2),
ResBlock_amp(256, 256))
self.layer3_phase = nn.Sequential(ResBlock_phase(128, 256, stride=2),
ResBlock_phase(256, 256))
self.layer4_amp = nn.Sequential(ResBlock_amp(256, 512, stride=2),
ResBlock_amp(512, 512))
self.layer4_phase = nn.Sequential(ResBlock_phase(256, 512, stride=2),
ResBlock_phase(512, 512))
self.final_amp = nn.Sequential(nn.AdaptiveAvgPool2d(1), nn.Flatten(),
nn.Linear(512, img_size**2))
self.final_phase = nn.Sequential(nn.AdaptiveAvgPool2d(1), nn.Flatten(),
nn.Linear(512, img_size**2))
def __init__(self, img_size):
super().__init__()
# ########################## Phase 1
self.conv1_amp = nn.Sequential(*conv_amp(
ni=1,
nc=4,
ks=(round_odd(0.365 * img_size), round_odd(0.365 * img_size)),
stride=1,
padding=make_padding(round_odd(0.365 * img_size), 1, 1),
dilation=1,
))
self.conv1_phase = nn.Sequential(*conv_phase(
ni=1,
nc=4,
ks=(round_odd(0.365 * img_size), round_odd(0.365 * img_size)),
stride=1,
padding=make_padding(round_odd(0.365 * img_size), 1, 1),
dilation=1,
add=1 - pi,
))
self.conv2_amp = nn.Sequential(*conv_amp(
ni=4,
nc=8,
ks=(round_odd(0.333 * img_size), round_odd(0.333 * img_size)),
stride=1,
padding=make_padding(round_odd(0.333 * img_size), 1, 1),
dilation=1,
))
self.conv2_phase = nn.Sequential(*conv_phase(
ni=4,
nc=8,
ks=(round_odd(0.333 * img_size), round_odd(0.333 * img_size)),
stride=1,
padding=make_padding(round_odd(0.333 * img_size), 1, 1),
dilation=1,
add=1 - pi,
))
self.conv3_amp = nn.Sequential(*conv_amp(
ni=8,
nc=12,
ks=(round_odd(0.269 * img_size), round_odd(0.269 * img_size)),
stride=1,
padding=make_padding(round_odd(0.269 * img_size), 1, 1),
dilation=1,
))
self.conv3_phase = nn.Sequential(*conv_phase(
ni=8,
nc=12,
ks=(round_odd(0.269 * img_size), round_odd(0.269 * img_size)),
stride=1,
padding=make_padding(round_odd(0.269 * img_size), 1, 1),
dilation=1,
add=1 - pi,
))
self.conv_con1_amp = nn.Sequential(
LocallyConnected2d(12, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
nn.ReLU(),
)
self.conv_con1_phase = nn.Sequential(
LocallyConnected2d(12, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(1 - pi),
)
# #################################### Phase 2
self.conv4_amp = nn.Sequential(*conv_amp(
ni=1,
nc=4,
ks=(round_odd(0.0793 * img_size), round_odd(0.0793 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0793 * img_size), 1, 2),
dilation=2,
))
self.conv4_phase = nn.Sequential(*conv_phase(
ni=1,
nc=4,
ks=(round_odd(0.0793 * img_size), round_odd(0.0793 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0793 * img_size), 1, 2),
dilation=2,
add=1 - pi,
))
self.conv5_amp = nn.Sequential(*conv_amp(
ni=4,
nc=8,
ks=(round_odd(0.0793 * img_size), round_odd(0.0793 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0793 * img_size), 1, 1),
dilation=1,
))
self.conv5_phase = nn.Sequential(*conv_phase(
ni=4,
nc=8,
ks=(round_odd(0.0793 * img_size), round_odd(0.0793 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0793 * img_size), 1, 1),
dilation=1,
add=1 - pi,
))
self.conv6_amp = nn.Sequential(*conv_amp(
ni=8,
nc=12,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 2),
dilation=2,
))
self.conv6_phase = nn.Sequential(*conv_phase(
ni=8,
nc=12,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 2),
dilation=2,
add=1 - pi,
))
self.conv7_amp = nn.Sequential(*conv_amp(
ni=12,
nc=16,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 1),
dilation=1,
))
self.conv7_phase = nn.Sequential(*conv_phase(
ni=12,
nc=16,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 1),
dilation=1,
add=1 - pi,
))
self.conv_con2_amp = nn.Sequential(
LocallyConnected2d(16, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
nn.ReLU(),
)
self.conv_con2_phase = nn.Sequential(
LocallyConnected2d(16, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(1 - pi),
)
# ################################## Phase 3
self.conv8_amp = nn.Sequential(*conv_amp(
ni=1,
nc=4,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 1),
dilation=1,
))
self.conv8_phase = nn.Sequential(*conv_phase(
ni=1,
nc=4,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 1),
dilation=1,
add=1 - pi,
))
self.conv9_amp = nn.Sequential(*conv_amp(
ni=4,
nc=8,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 1),
dilation=1,
))
self.conv9_phase = nn.Sequential(*conv_phase(
ni=4,
nc=8,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 1),
dilation=1,
add=1 - pi,
))
self.conv10_amp = nn.Sequential(*conv_amp(
ni=8,
nc=12,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 2),
dilation=2,
))
self.conv10_phase = nn.Sequential(*conv_phase(
ni=8,
nc=12,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 2),
dilation=2,
add=1 - pi,
))
self.conv11_amp = nn.Sequential(*conv_amp(
ni=12,
nc=20,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 1),
dilation=1,
))
self.conv11_phase = nn.Sequential(*conv_phase(
ni=12,
nc=20,
ks=(round_odd(0.0476 * img_size), round_odd(0.0476 * img_size)),
stride=1,
padding=make_padding(round_odd(0.0476 * img_size), 1, 1),
dilation=1,
add=1 - pi,
))
self.conv_con3_amp = nn.Sequential(
LocallyConnected2d(20, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
nn.ReLU(),
)
self.conv_con3_phase = nn.Sequential(
LocallyConnected2d(20, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(1 - pi),
)
self.symmetry_real = Lambda(symmetry)
self.symmetry_imag = Lambda(partial(symmetry, mode="imag"))
def __init__(self, img_size):
super().__init__()
self.conv1_amp = nn.Sequential(*conv_amp(1, 4, (23, 23), 1, 11, 1))
self.conv1_phase = nn.Sequential(
*conv_phase(1, 4, (23, 23), 1, 11, 1, add=1 - pi))
self.conv2_amp = nn.Sequential(*conv_amp(4, 8, (21, 21), 1, 10, 1))
self.conv2_phase = nn.Sequential(
*conv_phase(4, 8, (21, 21), 1, 10, 1, add=1 - pi))
self.conv3_amp = nn.Sequential(*conv_amp(8, 12, (17, 17), 1, 8, 1))
self.conv3_phase = nn.Sequential(
*conv_phase(8, 12, (17, 17), 1, 8, 1, add=1 - pi))
self.conv_con1_amp = nn.Sequential(
LocallyConnected2d(12, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
nn.ReLU(),
)
self.conv_con1_phase = nn.Sequential(
LocallyConnected2d(12, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(1 - pi),
)
self.conv4_amp = nn.Sequential(*conv_amp(1, 4, (5, 5), 1, 3, 2))
self.conv4_phase = nn.Sequential(
*conv_phase(1, 4, (5, 5), 1, 3, 2, add=1 - pi))
self.conv5_amp = nn.Sequential(*conv_amp(4, 8, (5, 5), 1, 2, 1))
self.conv5_phase = nn.Sequential(
*conv_phase(4, 8, (5, 5), 1, 2, 1, add=1 - pi))
self.conv6_amp = nn.Sequential(*conv_amp(8, 12, (3, 3), 1, 3, 2))
self.conv6_phase = nn.Sequential(
*conv_phase(8, 12, (3, 3), 1, 3, 2, add=1 - pi))
self.conv7_amp = nn.Sequential(*conv_amp(12, 16, (3, 3), 1, 1, 1))
self.conv7_phase = nn.Sequential(
*conv_phase(12, 16, (3, 3), 1, 1, 1, add=1 - pi))
self.conv_con2_amp = nn.Sequential(
LocallyConnected2d(16, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
nn.ReLU(),
)
self.conv_con2_phase = nn.Sequential(
LocallyConnected2d(16, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(1 - pi),
)
self.conv8_amp = nn.Sequential(*conv_amp(1, 4, (3, 3), 1, 1, 1))
self.conv8_phase = nn.Sequential(
*conv_phase(1, 4, (3, 3), 1, 1, 1, add=1 - pi))
self.conv9_amp = nn.Sequential(*conv_amp(4, 8, (3, 3), 1, 1, 1))
self.conv9_phase = nn.Sequential(
*conv_phase(4, 8, (3, 3), 1, 1, 1, add=1 - pi))
self.conv10_amp = nn.Sequential(*conv_amp(8, 12, (3, 3), 1, 2, 2))
self.conv10_phase = nn.Sequential(
*conv_phase(8, 12, (3, 3), 1, 2, 2, add=1 - pi))
self.conv11_amp = nn.Sequential(*conv_amp(12, 20, (3, 3), 1, 1, 1))
self.conv11_phase = nn.Sequential(
*conv_phase(12, 20, (3, 3), 1, 1, 1, add=1 - pi))
self.conv_con3_amp = nn.Sequential(
LocallyConnected2d(20, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
nn.ReLU(),
)
self.conv_con3_phase = nn.Sequential(
LocallyConnected2d(20, 1, img_size, 1, stride=1, bias=False),
nn.BatchNorm2d(1),
GeneralELU(1 - pi),
)
self.symmetry_real = Lambda(symmetry)
self.symmetry_imag = Lambda(partial(symmetry, mode="imag"))
self.flatten = Lambda(flatten)
# self.fully_connected = nn.Linear(3969, 54)
# self.fully_connected = nn.Linear(7938, 5)
# self.fully_connected = nn.Linear(3969, 1)
self.fully_connected = nn.Linear(7938, 3)
self.vaild_gauss_bs = Lambda(vaild_gauss_bs)
self.Relu = nn.ReLU()
self.fft = Lambda(fft)
self.euler = Lambda(euler)
self.shape = Lambda(shape)