def __init__(self, noise='bernoulli', xavier_init=False):
super(MNIST, self).__init__()
input_shape = (28, 28)
self.conv1 = NSMConv2d_(1,
32,
kernel_size=(5, 5),
prob=0.25,
bias=True,
input_shape=input_shape,
noise=noise)
input_shape = (12, 12)
self.conv2 = NSMConv2d_(32,
64,
kernel_size=(5, 5),
prob=0.25,
bias=True,
noise=noise,
input_shape=input_shape)
self.max_pool = nn.MaxPool2d(kernel_size=2)
self.fc1 = NSMLinear_(64 * 4 * 4,
512,
prob=0.5,
bias=True,
noise=noise)
self.fc2 = nn.Linear(512, 10)
self.dropout = nn.Dropout(p=0.5)
self.sign = Sign_().apply
self.tanh = nn.Tanh()
if xavier_init is True:
for name in self.named_parameters():
if 'weight' in name[0]:
nn.init.xavier_uniform_(name[1].data)
def __init__(self):
super(allconvnet64, self).__init__()
self.sign = Sign_().apply
self.mpool = nn.MaxPool2d(kernel_size=2, stride=2)
self.dropout = nn.Dropout(.5)
self.apool = nn.AvgPool2d(kernel_size=8)
self.relu = nn.LeakyReLU(negative_slope=0.1)
self.tanh = nn.Tanh()
self.noise = GaussianNoise(sigma=0.15)
self.b1 = nn.BatchNorm2d(N1)
self.input = weight_norm(nn.Conv2d(3,
N1,
kernel_size=(3, 3),
padding=(1, 1)),
name='weight',
dim=None)
self.convset1 = nn.ModuleList([
weight_norm(nn.Conv2d(N1, N1, kernel_size=(3, 3), padding=[1, 1]),
name='weight',
dim=None),
weight_norm(nn.Conv2d(N1, N1, kernel_size=(3, 3), padding=[1, 1]),
name='weight',
dim=None),
weight_norm(nn.Conv2d(N1, N1, kernel_size=(3, 3), padding=[1, 1]),
name='weight',
dim=None)
])
self.convset2 = nn.ModuleList([
weight_norm(nn.Conv2d(N1, N2, kernel_size=(3, 3), padding=[1, 1]),
name='weight',
dim=None),
weight_norm(nn.Conv2d(N2, N2, kernel_size=(3, 3), padding=[1, 1]),
name='weight',
dim=None),
weight_norm(nn.Conv2d(N2, N2, kernel_size=(3, 3), padding=[1, 1]),
name='weight',
dim=None)
])
self.convset3 = nn.ModuleList([
weight_norm(nn.Conv2d(N2, N3, kernel_size=(3, 3), padding=[0, 0]),
name='weight',
dim=None)
])
self.nin1 = weight_norm(nn.Conv2d(N3,
N3,
kernel_size=(1, 1),
padding=[0, 0]),
name='weight',
dim=None)
self.nin2 = weight_norm(nn.Conv2d(N3,
N3,
kernel_size=(1, 1),
padding=[0, 0]),
name='weight',
dim=None)
self.fc_out = weight_norm(nn.Linear(in_features=N3, out_features=10),
name='weight',
dim=None)
def __init__(self, noise='bernoulli', batch_size=32):
super(NMNIST_MODEL, self).__init__()
# self.fc_in = Lambda_(lambda x: self.sign(2 * x - 1))
# self.fc_in = nn.Conv2d(1, 1, 1, bias=True)
input_shape = (34, 34)
self.fc_in = nn.Conv2d(1, 1, kernel_size=(1, 1))
self.conv1 = NSMConv2d_(10,
32,
kernel_size=(5, 5),
prob=0.25,
bias=True,
noise=noise,
input_shape=input_shape)
input_shape = (15, 15)
self.conv2 = NSMConv2d_(32,
64,
kernel_size=(5, 5),
prob=0.25,
bias=True,
noise=noise,
input_shape=input_shape)
self.max_pool = nn.MaxPool2d(kernel_size=2)
self.fc1 = NSMLinear_(64 * 5 * 5, 150, prob=0.5, bias=True)
self.fc2 = nn.Linear(150, 10)
self.dropout = nn.Dropout(p=0.5)
self.sign = Sign_().apply
self.tanh = nn.Tanh()
def __init__(self, noise='bernoulli', batch_size=32):
super(eMNIST, self).__init__()
input_shape = (28, 28)
self.fc_in = nn.Conv2d(1, 1, kernel_size=(1, 1))
self.conv1 = NSMConv2d_(1,
32,
kernel_size=(5, 5),
prob=0.25,
bias=True,
noise=noise,
input_shape=input_shape)
input_shape = (12, 12)
self.conv2 = NSMConv2d_(32,
64,
kernel_size=(5, 5),
prob=0.25,
bias=True,
noise=noise,
input_shape=input_shape)
self.max_pool = nn.MaxPool2d(kernel_size=2)
self.fc1 = NSMLinear_(64 * 4 * 4, 150, prob=0.5, bias=True)
self.fc2 = nn.Linear(150, 26)
self.dropout = nn.Dropout(p=0.5)
self.sign = Sign_().apply
self.tanh = nn.Tanh()
def __init__(self):
super(allconvnet64, self).__init__()
self.sign = Sign_().apply
self.mpool = nn.MaxPool2d(kernel_size=2)
self.apool = nn.AvgPool2d(kernel_size=8)
self.relu = nn.ReLU()
# self.relu = nn.LeakyReLU(negative_slope=0.1)
self.convset1 = nn.ModuleList([
weight_norm(nn.Conv2d(6, N1, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N1),
weight_norm(nn.Conv2d(N1, N1, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N1),
weight_norm(nn.Conv2d(N1, N1, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N1),
nn.MaxPool2d(kernel_size=2),
nn.Dropout(.5)
])
self.convset2 = nn.ModuleList([
weight_norm(nn.Conv2d(N1, N2, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N2),
weight_norm(nn.Conv2d(N2, N2, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N2),
weight_norm(nn.Conv2d(N2, N2, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N2),
nn.MaxPool2d(kernel_size=2),
nn.Dropout(.5)
])
self.convset3 = nn.ModuleList([
weight_norm(nn.Conv2d(N2, N3, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N3),
weight_norm(nn.Conv2d(N3, N3, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N3),
weight_norm(nn.Conv2d(N3, N3, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N3),
nn.MaxPool2d(kernel_size=2),
nn.Dropout(.5)
])
self.convset4 = nn.ModuleList([
weight_norm(nn.Conv2d(N3, N4, kernel_size=(3, 3), padding=[1, 1]),
name='weight'),
# nn.BatchNorm2d(N4),
weight_norm(nn.Conv2d(N4, N4, kernel_size=(1, 1), padding=[0, 0]),
name='weight'),
# nn.BatchNorm2d(N4),
weight_norm(nn.Conv2d(N4, N4, kernel_size=(1, 1), padding=[0, 0]),
name='weight')
])
self.fc_out = nn.Linear(in_features=N4, out_features=11)
# self.bn_in = nn.BatchNorm2d(6)
self.bn = nn.BatchNorm2d(256)
def __init__(self, batch_size=32, device=0):
super(cMNIST, self).__init__()
self.conv1 = nn.Conv2d(1, 32, kernel_size=(5, 5))
self.conv2 = nn.Conv2d(32, 64, kernel_size=(5, 5))
self.max_pool = nn.MaxPool2d(kernel_size=2)
self.fc1 = nn.Linear(64 * 4 * 4, 150)
self.fc2 = nn.Linear(150, 10)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(p=0.5)
self.softmax = nn.Softmax(dim=1)
self.sign = Sign_().apply
def __init__(self, noise='bernoulli', xavier_init=False):
super(mnist_, self).__init__()
self.fc1 = NSMLinear_(784, 300, prob=0.5, bias=True, noise=noise)
self.fc2 = NSMLinear_(300, 300, prob=0.5, bias=True, noise=noise)
self.fc3 = NSMLinear_(300, 300, prob=0.5, bias=True, noise=noise)
self.fc4 = nn.Linear(300, 10)
self.sign = Sign_().apply
self.tanh = nn.Tanh()
if xavier_init is True:
for name in self.named_parameters():
if 'weight' in name[0]:
nn.init.xavier_uniform_(name[1].data)
def __init__(self, xavier_init=False):
super(mnist_, self).__init__()
self.fc1 = nn.Linear(784, 300)
self.fc2 = nn.Linear(300, 300)
self.fc3 = nn.Linear(300, 300)
self.fc4 = nn.Linear(300, 10)
self.sign = Sign_().apply
self.relu = nn.ReLU()
self.dropout = nn.Dropout(.5)
if xavier_init is True:
for name in self.named_parameters():
if 'weight' in name[0]:
nn.init.xavier_uniform_(name[1].data)
def __init__(self, noise='bernoulli'):
super(allnsmconvnet64, self).__init__()
self.sign = Sign_().apply
self.mpool = nn.MaxPool2d(kernel_size=2, stride=2)
self.apool = nn.AvgPool2d(kernel_size=8)
self.noise = GaussianNoise(sigma=0.15)
self.input = nn.Conv2d(3, N1, kernel_size=(3, 3), padding=(1, 1))
self.b0 = nn.BatchNorm2d(3)
self.b1 = nn.BatchNorm2d(N1)
self.b2 = nn.BatchNorm1d(N3)
self.convset1 = nn.ModuleList([
NSMConv2d_(N1,
N1,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[32, 32],
noise=noise,
bias=False),
NSMConv2d_(N1,
N1,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[32, 32],
noise=noise,
bias=False),
NSMConv2d_(N1,
N1,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[32, 32],
noise=noise,
bias=False)
])
self.convset2 = nn.ModuleList([
NSMConv2d_(N1,
N2,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[16, 16],
noise=noise,
bias=False),
NSMConv2d_(N2,
N2,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[16, 16],
noise=noise,
bias=False),
NSMConv2d_(N2,
N2,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[16, 16],
noise=noise,
bias=False)
])
self.convset3 = nn.ModuleList([
NSMConv2d_(N2,
N3,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[8, 8],
noise=noise,
bias=False),
NSMConv2d_(N3,
N3,
kernel_size=(1, 1),
prob=.5,
padding=[0, 0],
input_shape=[8, 8],
noise=noise,
bias=False),
NSMConv2d_(N3,
N3,
kernel_size=(1, 1),
prob=.5,
padding=[0, 0],
input_shape=[8, 8],
noise=noise,
bias=False)
])
self.fc_out = nn.Linear(in_features=N3, out_features=100)
self.tanh = nn.Tanh()
def __init__(self, noise='bernoulli'):
super(allnsmconvnet64, self).__init__()
self.sign = Sign_().apply
self.mpool = nn.MaxPool2d(kernel_size=2)
self.apool = nn.AvgPool2d(kernel_size=8)
self.convset1 = nn.ModuleList([
NSMConv2d_(6,
N1,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[64, 64],
noise=noise,
bias=True),
NSMConv2d_(N1,
N1,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[64, 64],
noise=noise,
bias=True),
NSMConv2d_(N1,
N1,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[64, 64],
noise=noise,
bias=True)
])
self.convset2 = nn.ModuleList([
NSMConv2d_(N1,
N2,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[32, 32],
noise=noise,
bias=True),
NSMConv2d_(N2,
N2,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[32, 32],
noise=noise,
bias=True),
NSMConv2d_(N2,
N2,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[32, 32],
noise=noise,
bias=True)
])
self.convset3 = nn.ModuleList([
NSMConv2d_(N2,
N3,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[16, 16],
noise=noise,
bias=True),
NSMConv2d_(N3,
N3,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[16, 16],
noise=noise,
bias=True),
NSMConv2d_(N3,
N3,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[16, 16],
noise=noise,
bias=True)
])
self.convset4 = nn.ModuleList([
NSMConv2d_(N3,
N4,
kernel_size=(3, 3),
prob=.5,
padding=[1, 1],
input_shape=[8, 8],
noise=noise,
bias=True),
NSMConv2d_(N4,
N4,
kernel_size=(1, 1),
prob=.5,
padding=[0, 0],
input_shape=[8, 8],
noise=noise,
bias=True),
NSMConv2d_(N4,
N4,
kernel_size=(1, 1),
prob=.5,
padding=[0, 0],
input_shape=[8, 8],
noise=noise,
bias=True)
])
self.fc_out = nn.Linear(in_features=N4, out_features=11)