def __init__(self, outputs, inputs):
super(BBB3Conv3FC, self).__init__()
self.conv1 = BBBConv2d(inputs, 32, 5, stride=1, padding=2)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv2 = BBBConv2d(32, 64, 5, stride=1, padding=2)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv3 = BBBConv2d(64, 128, 5, stride=1, padding=1)
self.soft3 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=3, stride=2)
self.flatten = FlattenLayer(2 * 2 * 128)
self.fc1 = BBBLinearFactorial(2 * 2 * 128, 1000)
self.soft5 = nn.Softplus()
self.fc2 = BBBLinearFactorial(1000, 1000)
self.soft6 = nn.Softplus()
self.fc3 = BBBLinearFactorial(1000, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.conv2, self.soft2,
self.pool2, self.conv3, self.soft3, self.pool3, self.flatten,
self.fc1, self.soft5, self.fc2, self.soft6, self.fc3
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs,):
super(_BayesianLeNetD, self).__init__()
self.outputs = outputs
#due to the convlayer change have to assume the logvar initial value.
self.q_logvar_init = 0.05
self.p_logvar_init = math.log(0.05)
self.conv1 = BBBConv2d(self.q_logvar_init, self.p_logvar_init,inputs,6, 5, stride=1)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv2 = BBBConv2d(self.q_logvar_init, self.p_logvar_init,6, 16, 5, stride=1)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = FlattenLayer(5 * 5 * 16)
self.fc1 = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init,5 * 5 * 16, 120)
self.soft3 = nn.Softplus()
self.fc2 = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init,120, 84)
self.soft4 = nn.Softplus()
self.fc3 = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init,84, outputs)
layers = [self.conv1, self.soft1, self.pool1, self.conv2, self.soft2, self.pool2,
self.flatten, self.fc1, self.soft3, self.fc2, self.soft4, self.fc3]
#not sure if this is right, test drive.
self.prob = nn.Sigmoid()
if outputs == 1:
layers.append(self.prob)
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(Hybrid_5Conv3FC, self).__init__()
self.features = nn.Sequential(
nn.Conv2d(inputs, 96, kernel_size=11, stride=4),
nn.ReLU(inplace=True),
# nn.Dropout(p=0.5),
nn.MaxPool2d(kernel_size=3, stride=2),
nn.Conv2d(96, 256, kernel_size=5, stride=1, padding=2),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=3, stride=2),
nn.Conv2d(256, 384, kernel_size=3, stride=1, padding=1),
nn.ReLU(inplace=True),
# nn.Dropout(p=0.5),
nn.Conv2d(384, 384, kernel_size=3, stride=1, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(384, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=3, stride=2),
)
self.flatten = FlattenLayer(6 * 6 * 256)
self.fc1 = BBBLinearFactorial(6 * 6 * 256, 4096)
self.dropout1 = nn.Dropout(p=0.5)
self.fc2 = BBBLinearFactorial(4096, 4096)
self.dropout2 = nn.Dropout(p=0.5)
self.fc3 = BBBLinearFactorial(4096, outputs)
layers = [
self.flatten, self.fc1, self.dropout1, self.fc2, self.dropout2,
self.fc3
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBBLeNet, self).__init__()
self.conv1 = BBBConv2d(inputs, 6, 5, stride=1)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv2 = BBBConv2d(6, 16, 5, stride=1)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = FlattenLayer(5 * 5 * 16)
self.fc1 = BBBLinearFactorial(5 * 5 * 16, 120)
self.soft3 = nn.Softplus()
self.fc2 = BBBLinearFactorial(120, 84)
self.soft4 = nn.Softplus()
self.fc3 = BBBLinearFactorial(84, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.conv2, self.soft2,
self.pool2, self.flatten, self.fc1, self.soft3, self.fc2,
self.soft4, self.fc3
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(_BayesianAlexNetD, self).__init__()
self.q_logvar_init = 0.05
self.p_logvar_init = math.log(0.05)
self.classifier = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init, 1* 1 * 128, outputs)
self.conv1 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, inputs, 64, kernel_size=11, stride=4, padding=5)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv2 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, 64, 192, kernel_size=5, padding=2)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv3 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, 192, 384, kernel_size=3, padding=1)
self.soft3 = nn.Softplus()
self.conv4 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, 384, 256, kernel_size=3, padding=1)
self.soft4 = nn.Softplus()
self.conv5 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, 256, 128, kernel_size=3, padding=1)
self.soft5 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
# self.flatten = FlattenLayer(1 * 1 * 128)
# self.fc1 = BBBLinearFactorial(q_logvar_init, N, p_logvar_init, 1* 1 * 128, outputs)
layers = [self.conv1, self.soft1, self.pool1, self.conv2, self.soft2, self.pool2, self.conv3, self.soft3,
self.conv4, self.soft4, self.conv5, self.soft5, self.pool3]
self.layers = nn.ModuleList(layers)
class _ClassifierD(nn.Module):
'''
the ACGAN part, the code is doing the classification here
'''
def __init__(self, outputs, inputs,):
super(_ClassifierD, self).__init__()
self.outputs = outputs
#due to the convlayer change have to assume the logvar initial value.
self.q_logvar_init = 0.05
self.p_logvar_init = math.log(0.05)
self.conv1 = BBBConv2d(self.q_logvar_init, self.p_logvar_init,inputs,6, 5, stride=1)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv2 = BBBConv2d(self.q_logvar_init, self.p_logvar_init,6, 16, 5, stride=1)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = FlattenLayer(5 * 5 * 16)
self.fc1 = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init,5 * 5 * 16, 120)
self.soft3 = nn.Softplus()
self.fc2 = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init,120, 84)
self.soft4 = nn.Softplus()
#self.fc3 = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init,84, outputs)
self.fcA = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init,84, 1)
self.fcB = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init,84, 1)
layers = [self.conv1, self.soft1, self.pool1, self.conv2, self.soft2, self.pool2,
self.flatten, self.fc1, self.soft3, self.fc2, self.soft4]
#not sure if this is right, test drive.
self.prob = nn.Sigmoid()
if outputs == 1:
layers.append(self.prob)
self.layers = nn.ModuleList(layers)
def forward(self, x):#used name: probforward
'Forward pass with Bayesian weights'
kl = 0
for layer in self.layers:
if hasattr(layer, 'convprobforward') and callable(layer.convprobforward):
x, _kl, = layer.convprobforward(x)
kl += _kl
elif hasattr(layer, 'fcprobforward') and callable(layer.fcprobforward):
x, _kl, = layer.fcprobforward(x)
kl += _kl
else:
x = layer(x)
#logits = x
#return logits, kl
logitsA, klA = self.fcA.fcprobforward(x)
logitsB, klB = self.fcB.fcprobforward(x)
return logitsA, logitsB, klA+kl, klB+kl
def __init__(self, outputs, inputs):
super(Hybrid_6Conv3FC, self).__init__()
self.features = nn.Sequential(
# Conv Layer block 1
nn.Conv2d(in_channels=inputs,
out_channels=32,
kernel_size=3,
padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=32,
out_channels=64,
kernel_size=3,
padding=1),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2, stride=2),
# Conv Layer block 2
nn.Conv2d(in_channels=64,
out_channels=128,
kernel_size=3,
padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=128,
out_channels=128,
kernel_size=3,
padding=1),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2, stride=2),
# Conv Layer block 3
nn.Conv2d(in_channels=128,
out_channels=256,
kernel_size=3,
padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(in_channels=256,
out_channels=256,
kernel_size=3,
padding=1),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2, stride=2),
)
self.flatten = FlattenLayer(4 * 4 * 256)
self.fc1 = BBBLinearFactorial(4 * 4 * 256, 1024)
self.soft1 = nn.Softplus()
self.fc2 = BBBLinearFactorial(1024, 512)
self.soft2 = nn.Softplus()
self.dropout2 = nn.Dropout(p=0.1)
self.fc3 = BBBLinearFactorial(512, outputs)
layers = [
self.flatten, self.fc1, self.soft1, self.fc2, self.soft2,
self.dropout2, self.fc3
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBBAlexNet, self).__init__()
# self.conv1 = BBBConv2d(inputs, 64, kernel_size=11, stride=4, padding=5)
# self.soft1 = nn.Softplus()
# self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
#
# self.conv2 = BBBConv2d(64, 192, kernel_size=5, padding=2)
# self.soft2 = nn.Softplus()
# self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
#
# self.conv3 = BBBConv2d(192, 384, kernel_size=3, padding=1)
# self.soft3 = nn.Softplus()
#
# self.conv4 = BBBConv2d(384, 256, kernel_size=3, padding=1)
# self.soft4 = nn.Softplus()
#
# self.conv5 = BBBConv2d(256, 128, kernel_size=3, padding=1)
# self.soft5 = nn.Softplus()
# self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
#
# self.flatten = FlattenLayer(1 * 1 * 128)
# self.fc1 = BBBLinearFactorial(1* 1 * 128, outputs)
self.conv1 = BBBConv2d(inputs, 96, kernel_size=11, stride=4)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv2 = BBBConv2d(96, 256, kernel_size=5, stride=1, padding=2)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv3 = BBBConv2d(256, 384, kernel_size=3, stride=1, padding=1)
self.soft3 = nn.Softplus()
self.conv4 = BBBConv2d(384, 384, kernel_size=3, stride=1, padding=1)
self.soft4 = nn.Softplus()
self.conv5 = BBBConv2d(384, 256, kernel_size=3, stride=1, padding=1)
self.soft5 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=3, stride=2)
self.flatten = FlattenLayer(6 * 6 * 256)
self.fc1 = BBBLinearFactorial(6 * 6 * 256, 4096)
self.dropout1 = nn.Dropout(p=0.5)
self.fc2 = BBBLinearFactorial(4096, 4096)
self.dropout2 = nn.Dropout(p=0.5)
self.fc3 = BBBLinearFactorial(4096, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.conv2, self.soft2,
self.pool2, self.conv3, self.soft3, self.conv4, self.soft4,
self.conv5, self.soft5, self.pool3, self.flatten, self.fc1,
self.dropout1, self.fc2, self.dropout2, self.fc3
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBBELUN2, self).__init__()
self.conv1 = BBBConv2d(inputs, 96, 6, stride=1)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv2 = BBBConv2d(96, 512, 3, stride=1)
self.soft2 = nn.Softplus()
self.conv3 = BBBConv2d(512, 512, 3, stride=1)
self.soft3 = nn.Softplus()
self.conv4 = BBBConv2d(512, 512, 3, stride=1)
self.soft4 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv5 = BBBConv2d(512, 768, 3, stride=1)
self.soft5 = nn.Softplus()
self.conv6 = BBBConv2d(768, 768, 3, stride=1)
self.soft6 = nn.Softplus()
self.conv7 = BBBConv2d(768, 768, 2, stride=1)
self.soft7 = nn.Softplus()
self.conv8 = BBBConv2d(768, 768, 2, stride=1)
self.soft8 = nn.Softplus()
self.conv9 = BBBConv2d(768, 768, 1, stride=1)
self.soft9 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv10 = BBBConv2d(768, 1024, 3, stride=1)
self.soft10 = nn.Softplus()
self.conv11 = BBBConv2d(1024, 1024, 3, stride=1)
self.soft11 = nn.Softplus()
self.conv12 = BBBConv2d(1024, 1024, 3, stride=1)
self.soft12 = nn.Softplus()
self.pool4 = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = FlattenLayer(8 * 8 * 1024)
self.fc1 = BBBLinearFactorial(8 * 8 * 1024, 4096)
self.soft13 = nn.Softplus()
self.fc2 = BBBLinearFactorial(4096, 4096)
self.soft14 = nn.Softplus()
self.fc3 = BBBLinearFactorial(4096, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.conv2, self.soft2,
self.conv3, self.soft3, self.conv4, self.soft4, self.pool2,
self.conv5, self.soft5, self.conv6, self.soft6, self.conv7,
self.soft7, self.conv8, self.soft8, self.conv9, self.soft9,
self.pool3, self.conv10, self.soft10, self.conv11, self.soft11,
self.conv12, self.soft12, self.pool4, self.flatten, self.fc1,
self.soft13, self.fc2, self.soft14, self.fc3
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(Hybrid_timeseries_1Conv2FC, self).__init__()
self.features = nn.Sequential(
nn.Conv1d(inputs, 64, kernel_size=11, stride=4, padding=5),
nn.ReLU(inplace=True),
)
self.flatten = FlattenLayer(1 * 1 * 64)
self.fc1 = BBBLinearFactorial(1 * 1 * 64, 512)
self.fc2 = BBBLinearFactorial(512, outputs)
layers = [self.flatten, self.fc1, self.fc2]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(Hybrid_5Conv1FC, self).__init__()
self.features = nn.Sequential(
nn.Conv2d(inputs, 64, kernel_size=11, stride=4, padding=5),
nn.ReLU(inplace=True),
nn.Dropout(p=0.5),
nn.MaxPool2d(kernel_size=2, stride=2),
nn.Conv2d(64, 192, kernel_size=5, padding=2),
nn.ReLU(inplace=True),
nn.MaxPool2d(kernel_size=2, stride=2),
nn.Conv2d(192, 384, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
nn.Dropout(p=0.5),
nn.Conv2d(384, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(256, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
nn.Dropout(p=0.5),
nn.MaxPool2d(kernel_size=2, stride=2),
)
self.fc1 = BBBLinearFactorial(256, outputs)
layers = [self.fc1]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBBAlexNet_5Conv1FC, self).__init__()
self.conv1 = BBBConv2d(inputs, 64, kernel_size=11, stride=4, padding=5)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv2 = BBBConv2d(64, 192, kernel_size=5, padding=2)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv3 = BBBConv2d(192, 384, kernel_size=3, padding=1)
self.soft3 = nn.Softplus()
self.conv4 = BBBConv2d(384, 256, kernel_size=3, padding=1)
self.soft4 = nn.Softplus()
self.conv5 = BBBConv2d(256, 256, kernel_size=3, padding=1)
self.soft5 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = FlattenLayer(1 * 1 * 256)
self.fc1 = BBBLinearFactorial(1 * 1 * 256, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.conv2, self.soft2,
self.pool2, self.conv3, self.soft3, self.conv4, self.soft4,
self.conv5, self.soft5, self.pool3, self.flatten, self.fc1
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBBAlexNetTimeSeries_1Conv2FC, self).__init__()
self.conv1 = BBBConv1d(inputs, 64, kernel_size=11, stride=4, padding=5)
self.soft1 = nn.Softplus()
#self.conv2 = BBBConv1d(64, 192, kernel_size=5, padding=2)
#self.soft2 = nn.Softplus()
self.flatten = FlattenLayer(1 * 1 * 64)
self.fc1 = BBBLinearFactorial(1 * 1 * 64, 512)
#self.fc2 = BBBLinearFactorial(4096, 4096)
self.fc3 = BBBLinearFactorial(512, outputs)
layers = [self.conv1, self.soft1, self.flatten, self.fc1, self.fc3]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBBAlexNet_6Conv3FC, self).__init__()
self.conv1 = BBBConv2d(inputs, 32, kernel_size=3, stride=1, padding=1)
self.norm1 = nn.BatchNorm2d(32)
self.soft1 = nn.Softplus()
self.conv2 = BBBConv2d(32, 64, kernel_size=3, stride=1, padding=1)
self.soft2 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv3 = BBBConv2d(64, 128, kernel_size=3, stride=1, padding=1)
self.norm2 = nn.BatchNorm2d(128)
self.soft3 = nn.Softplus()
self.conv4 = BBBConv2d(128, 128, kernel_size=3, stride=1, padding=1)
self.soft4 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.dropout1 = nn.Dropout(p=0.05)
self.conv5 = BBBConv2d(128, 256, kernel_size=3, stride=1, padding=1)
self.norm3 = nn.BatchNorm2d(256)
self.soft5 = nn.Softplus()
self.conv6 = BBBConv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.soft6 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
self.dropout2 = nn.Dropout(p=0.1)
self.flatten = FlattenLayer(4 * 4 * 256)
self.fc1 = BBBLinearFactorial(4 * 4 * 256, 1024)
self.soft7 = nn.Softplus()
self.fc2 = BBBLinearFactorial(1024, 512)
self.soft8 = nn.Softplus()
self.dropout3 = nn.Dropout(p=0.1)
self.fc3 = BBBLinearFactorial(512, outputs)
layers = [
self.conv1, self.soft1, self.conv2, self.soft2, self.pool1,
self.conv3, self.soft3, self.conv4, self.soft4, self.pool2,
self.conv5, self.soft5, self.conv6, self.soft6, self.pool3,
self.flatten, self.fc1, self.soft7, self.fc2, self.soft8,
self.dropout3, self.fc3
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBBCNN1, self).__init__()
self.conv1 = BBBConv2d(inputs, 92, 3, stride=1)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv2 = BBBConv2d(92, 384, 1, stride=1)
self.soft2 = nn.Softplus()
self.conv3 = BBBConv2d(384, 384, 2, stride=1)
self.soft3 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv4 = BBBConv2d(384, 640, 2, stride=1)
self.soft4 = nn.Softplus()
self.conv5 = BBBConv2d(640, 640, 2, stride=1)
self.soft5 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv6 = BBBConv2d(640, 640, 1, stride=1)
self.soft6 = nn.Softplus()
self.conv7 = BBBConv2d(640, 768, 2, stride=1)
self.soft7 = nn.Softplus()
self.pool4 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv8 = BBBConv2d(768, 768, 2, stride=1)
self.soft8 = nn.Softplus()
self.conv9 = BBBConv2d(768, 768, 2, stride=1)
self.soft9 = nn.Softplus()
self.pool5 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv10 = BBBConv2d(768, 768, 1, stride=1)
self.soft10 = nn.Softplus()
self.conv11 = BBBConv2d(768, 640, 2, stride=1)
self.soft11 = nn.Softplus()
self.conv12 = BBBConv2d(640, 384, 2, stride=1)
self.soft12 = nn.Softplus()
self.pool6 = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = FlattenLayer(8 * 8 * 384)
self.fc1 = BBBLinearFactorial(8 * 8 * 384, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.conv2, self.soft2,
self.conv3, self.soft3, self.pool2, self.conv4, self.soft4,
self.conv5, self.soft5, self.pool3, self.conv6, self.soft6,
self.conv7, self.soft7, self.pool4, self.conv8, self.soft8,
self.conv9, self.soft9, self.pool5, self.conv10, self.soft10,
self.conv11, self.soft11, self.conv12, self.soft12, self.pool6,
self.flatten, self.fc1
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBB4Conv3FC, self).__init__()
self.conv1 = BBBConv2d(inputs, 32, kernel_size=5, stride=1, padding=0)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv2 = BBBConv2d(32, 64, kernel_size=5, stride=1, padding=0)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv3 = BBBConv2d(64, 128, kernel_size=5, stride=1, padding=0)
self.soft3 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=3, stride=2)
self.conv4 = BBBConv2d(128, 256, kernel_size=5, stride=1, padding=0)
self.soft4 = nn.Softplus()
self.pool4 = nn.MaxPool2d(kernel_size=3, stride=2)
# self.conv5 = BBBConv2d(256, 512, kernel_size=5, stride=1, padding=0)
# self.soft5 = nn.Softplus()
# self.pool5 = nn.MaxPool2d(kernel_size=3, stride=2)
self.flatten = FlattenLayer(3 * 3 * 256)
self.fc1 = BBBLinearFactorial(3 * 3 * 256, 1000)
self.soft6 = nn.Softplus()
self.fc2 = BBBLinearFactorial(1000, 1000)
self.soft7 = nn.Softplus()
self.fc3 = BBBLinearFactorial(1000, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.conv2, self.soft2,
self.pool2, self.conv3, self.soft3, self.pool3, self.conv4,
self.soft4, self.pool4, self.flatten, self.fc1, self.soft6,
self.fc2, self.soft7, self.fc3
]
self.layers = nn.ModuleList(layers)
class _BayesianAlexNetD(nn.Module):
'''The architecture of AlexNet with Bayesian Layers'''
def __init__(self, outputs, inputs):
super(_BayesianAlexNetD, self).__init__()
self.q_logvar_init = 0.05
self.p_logvar_init = math.log(0.05)
self.classifier = BBBLinearFactorial(self.q_logvar_init, self.p_logvar_init, 1* 1 * 128, outputs)
self.conv1 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, inputs, 64, kernel_size=11, stride=4, padding=5)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv2 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, 64, 192, kernel_size=5, padding=2)
self.soft2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv3 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, 192, 384, kernel_size=3, padding=1)
self.soft3 = nn.Softplus()
self.conv4 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, 384, 256, kernel_size=3, padding=1)
self.soft4 = nn.Softplus()
self.conv5 = BBBConv2d(self.q_logvar_init, self.p_logvar_init, 256, 128, kernel_size=3, padding=1)
self.soft5 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2)
# self.flatten = FlattenLayer(1 * 1 * 128)
# self.fc1 = BBBLinearFactorial(q_logvar_init, N, p_logvar_init, 1* 1 * 128, outputs)
layers = [self.conv1, self.soft1, self.pool1, self.conv2, self.soft2, self.pool2, self.conv3, self.soft3,
self.conv4, self.soft4, self.conv5, self.soft5, self.pool3]
self.layers = nn.ModuleList(layers)
def forward(self, x):
kl = 0
for layer in self.layers:
if hasattr(layer, 'convprobforward') and callable(layer.convprobforward):
x, _kl, = layer.convprobforward(x)
else:
x = layer.forward(x)
x = x.view(x.size(0), -1)
x, _kl = self.classifier.fcprobforward(x)
kl += _kl
logits = x
return logits, kl
def __init__(self, outputs, inputs):
super(FlowTestNet, self).__init__()
flow = False
self.q_logvar_init = 0.05
self.p_logvar_init = math.log(0.05)
n = 3072
self.classifier1 = BBBLinearFactorial(self.q_logvar_init,
self.p_logvar_init,
n,
n,
flow=flow)
self.classifier2 = BBBLinearFactorial(self.q_logvar_init,
self.p_logvar_init,
n,
n,
flow=flow)
self.classifier3 = BBBLinearFactorial(self.q_logvar_init,
self.p_logvar_init,
n,
n,
flow=flow)
self.classifier4 = BBBLinearFactorial(self.q_logvar_init,
self.p_logvar_init,
n,
n,
flow=flow)
self.classifier5 = BBBLinearFactorial(self.q_logvar_init,
self.p_logvar_init,
n,
n,
flow=flow)
self.classifier6 = BBBLinearFactorial(self.q_logvar_init,
self.p_logvar_init,
n,
outputs,
flow=flow)
self.flatten_layer = Flatten()
# self.flatten = FlattenLayer(1 * 1 * 128)
# self.fc1 = BBBLinearFactorial(q_logvar_init, N, p_logvar_init, 1* 1 * 128, outputs)
layers = [
self.flatten_layer, self.classifier1, self.classifier2,
self.classifier3, self.classifier4, self.classifier5,
self.classifier6
]
self.layers = nn.ModuleList(layers)
def __init__(self, outputs, inputs):
super(BBBELUN1, self).__init__()
self.conv1 = BBBConv2d(inputs, 384, 3, stride=1)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv2 = BBBConv2d(384, 384, 1, stride=1)
self.soft2 = nn.Softplus()
self.conv3 = BBBConv2d(384, 384, 2, stride=1)
self.soft3 = nn.Softplus()
self.conv4 = BBBConv2d(384, 640, 2, stride=1)
self.soft4 = nn.Softplus()
self.conv5 = BBBConv2d(640, 640, 2, stride=1)
self.soft5 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv6 = BBBConv2d(640, 640, 1, stride=1)
self.soft6 = nn.Softplus()
self.conv7 = BBBConv2d(640, 768, 2, stride=1)
self.soft7 = nn.Softplus()
self.conv8 = BBBConv2d(768, 768, 2, stride=1)
self.soft8 = nn.Softplus()
self.conv9 = BBBConv2d(768, 768, 2, stride=1)
self.soft9 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv10 = BBBConv2d(768, 768, 1, stride=1)
self.soft10 = nn.Softplus()
self.conv11 = BBBConv2d(768, 896, 2, stride=1)
self.soft11 = nn.Softplus()
self.conv12 = BBBConv2d(896, 896, 2, stride=1)
self.soft12 = nn.Softplus()
self.pool4 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv13 = BBBConv2d(896, 896, 3, stride=1)
self.soft13 = nn.Softplus()
self.conv14 = BBBConv2d(896, 1024, 2, stride=1)
self.soft14 = nn.Softplus()
self.conv15 = BBBConv2d(1024, 1024, 2, stride=1)
self.soft15 = nn.Softplus()
self.pool5 = nn.MaxPool2d(kernel_size=2, stride=1)
self.conv16 = BBBConv2d(1024, 1024, 1, stride=1)
self.soft16 = nn.Softplus()
self.conv17 = BBBConv2d(1024, 1152, 2, stride=1)
self.soft17 = nn.Softplus()
self.pool6 = nn.MaxPool2d(kernel_size=2, stride=2)
self.conv18 = BBBConv2d(1152, 1152, 2, stride=1)
self.soft18 = nn.Softplus()
self.pool7 = nn.MaxPool2d(kernel_size=2, stride=2)
self.flatten = FlattenLayer(2 * 2 * 1152)
self.fc1 = BBBLinearFactorial(2 * 2 * 1152, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.conv2, self.soft2,
self.conv3, self.soft3, self.conv4, self.soft4, self.conv5,
self.soft5, self.pool2, self.conv6, self.soft6, self.conv7,
self.soft7, self.conv8, self.soft8, self.conv9, self.soft9,
self.pool3, self.conv10, self.soft10, self.conv11, self.soft11,
self.conv12, self.soft12, self.pool4, self.conv13, self.soft13,
self.conv14, self.soft14, self.conv15, self.soft15, self.pool5,
self.conv16, self.soft16, self.conv17, self.soft17, self.pool6,
self.conv18, self.soft18, self.pool7, self.flatten, self.fc1
]
self.layers = nn.ModuleList(layers)
def __init__(self,
block,
layers,
input_channels,
num_classes=10,
zero_init_residual=False,
groups=1,
width_per_group=64,
replace_stride_with_dilation=None,
norm_layer=None):
super(ResNet, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.inplanes = 64
self.dilation = 1
self.q_logvar_init = 0.05
self.p_logvar_init = math.log(0.05)
if replace_stride_with_dilation is None:
# each element in the tuple indicates if we should replace
# the 2x2 stride with a dilated convolution instead
replace_stride_with_dilation = [False, False, False]
if len(replace_stride_with_dilation) != 3:
raise ValueError("replace_stride_with_dilation should be None "
"or a 3-element tuple, got {}".format(
replace_stride_with_dilation))
self.groups = groups
self.base_width = width_per_group
self.conv1 = BBBConv2d(self.q_logvar_init,
self.p_logvar_init,
input_channels,
self.inplanes,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = norm_layer(self.inplanes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block,
128,
layers[1],
stride=2,
dilate=replace_stride_with_dilation[0])
self.layer3 = self._make_layer(block,
256,
layers[2],
stride=2,
dilate=replace_stride_with_dilation[1])
self.layer4 = self._make_layer(block,
512,
layers[3],
stride=2,
dilate=replace_stride_with_dilation[2])
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.drop = nn.Dropout(p=.5)
self.classifier = BBBLinearFactorial(self.q_logvar_init,
self.p_logvar_init,
512 * block.expansion,
num_classes,
flow=False)
print(block.expansion)
layers2 = [
self.conv1, self.bn1, self.relu, self.maxpool, self.layer1,
self.layer2, self.layer3, self.layer4, self.avgpool
]
self.layers2 = nn.ModuleList(layers2)
for m in self.modules():
if isinstance(m, BBBConv2d):
m.reset_parameters()
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
if zero_init_residual:
for m in self.modules():
if isinstance(m, Bottleneck):
nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
class ResNet(nn.Module):
def __init__(self,
block,
layers,
input_channels,
num_classes=10,
zero_init_residual=False,
groups=1,
width_per_group=64,
replace_stride_with_dilation=None,
norm_layer=None):
super(ResNet, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.inplanes = 64
self.dilation = 1
self.q_logvar_init = 0.05
self.p_logvar_init = math.log(0.05)
if replace_stride_with_dilation is None:
# each element in the tuple indicates if we should replace
# the 2x2 stride with a dilated convolution instead
replace_stride_with_dilation = [False, False, False]
if len(replace_stride_with_dilation) != 3:
raise ValueError("replace_stride_with_dilation should be None "
"or a 3-element tuple, got {}".format(
replace_stride_with_dilation))
self.groups = groups
self.base_width = width_per_group
self.conv1 = BBBConv2d(self.q_logvar_init,
self.p_logvar_init,
input_channels,
self.inplanes,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = norm_layer(self.inplanes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block,
128,
layers[1],
stride=2,
dilate=replace_stride_with_dilation[0])
self.layer3 = self._make_layer(block,
256,
layers[2],
stride=2,
dilate=replace_stride_with_dilation[1])
self.layer4 = self._make_layer(block,
512,
layers[3],
stride=2,
dilate=replace_stride_with_dilation[2])
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.drop = nn.Dropout(p=.5)
self.classifier = BBBLinearFactorial(self.q_logvar_init,
self.p_logvar_init,
512 * block.expansion,
num_classes,
flow=False)
print(block.expansion)
layers2 = [
self.conv1, self.bn1, self.relu, self.maxpool, self.layer1,
self.layer2, self.layer3, self.layer4, self.avgpool
]
self.layers2 = nn.ModuleList(layers2)
for m in self.modules():
if isinstance(m, BBBConv2d):
m.reset_parameters()
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
if zero_init_residual:
for m in self.modules():
if isinstance(m, Bottleneck):
nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
norm_layer = self._norm_layer
downsample = None
previous_dilation = self.dilation
if dilate:
self.dilation *= stride
stride = 1
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
conv1x1(self.inplanes, planes * block.expansion, stride),
norm_layer(planes * block.expansion),
)
layers = []
layers.append(
block(self.inplanes, planes, stride, downsample, self.groups,
self.base_width, previous_dilation, norm_layer))
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(
block(self.inplanes,
planes,
groups=self.groups,
base_width=self.base_width,
dilation=self.dilation,
norm_layer=norm_layer))
return nn.Sequential(*layers)
def probforward(self, x, dropout=False):
loss = 0
i = 0
out, kl = self.conv1.probforward(x)
out = self.relu(self.bn1(out))
loss += kl
out, kl = self.pf(out, self.layer1)
loss += kl
out, kl = self.pf(out, self.layer2)
loss += kl
out, kl = self.pf(out, self.layer3)
loss += kl
out, kl = self.pf(out, self.layer4)
loss += kl
out = F.avg_pool2d(out, 4)
out = out.view(out.size(0), -1)
if (dropout):
x = self.drop(x)
x, _kl = self.classifier.probforward(out)
kl += loss
logits = x
return logits, kl
def pf(self, x, layer):
kl = 0
for l in layer:
#print(l)
if hasattr(l, 'probforward') and callable(l.probforward):
x, _kl, = l.probforward(x)
kl += _kl
else:
print(l)
x = l.forward(x)
return x, kl
def __init__(self, outputs, inputs):
super(BBBSqueezeNet, self).__init__()
self.conv1 = BBBConv2d(inputs, 64, kernel_size=3, stride=2)
self.soft1 = nn.Softplus()
self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
# Fire module 1
self.squeeze1 = BBBConv2d(64, 16, kernel_size=1)
self.squeeze_activation1 = nn.Softplus()
self.expand3x3_1 = BBBConv2d(16, 128, kernel_size=3, padding=1)
self.expand3x3_activation1 = nn.Softplus()
# Fire module 2
self.squeeze2 = BBBConv2d(128, 16, kernel_size=1)
self.squeeze_activation2 = nn.Softplus()
self.expand3x3_2 = BBBConv2d(16, 128, kernel_size=3, padding=1)
self.expand3x3_activation2 = nn.Softplus()
self.pool2 = nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
# Fire module 3
self.squeeze3 = BBBConv2d(128, 32, kernel_size=1)
self.squeeze_activation3 = nn.Softplus()
self.expand3x3_3 = BBBConv2d(32, 256, kernel_size=3, padding=1)
self.expand3x3_activation3 = nn.Softplus()
# Fire module 4
self.squeeze4 = BBBConv2d(256, 32, kernel_size=1)
self.squeeze_activation4 = nn.Softplus()
self.expand3x3_4 = BBBConv2d(32, 256, kernel_size=3, padding=1)
self.expand3x3_activation4 = nn.Softplus()
self.pool3 = nn.MaxPool2d(kernel_size=3, stride=2, ceil_mode=True)
# Fire module 5
self.squeeze5 = BBBConv2d(256, 48, kernel_size=1)
self.squeeze_activation5 = nn.Softplus()
self.expand3x3_5 = BBBConv2d(48, 384, kernel_size=3, padding=1)
self.expand3x3_activation5 = nn.Softplus()
# Fire module 6
self.squeeze6 = BBBConv2d(384, 48, kernel_size=1)
self.squeeze_activation6 = nn.Softplus()
self.expand3x3_6 = BBBConv2d(48, 384, kernel_size=3, padding=1)
self.expand3x3_activation6 = nn.Softplus()
# Fire module 7
self.squeeze7 = BBBConv2d(384, 64, kernel_size=1)
self.squeeze_activation7 = nn.Softplus()
self.expand3x3_7 = BBBConv2d(64, 512, kernel_size=3, padding=1)
self.expand3x3_activation7 = nn.Softplus()
# Fire module 8
self.squeeze8 = BBBConv2d(512, 64, kernel_size=1)
self.squeeze_activation8 = nn.Softplus()
self.expand3x3_8 = BBBConv2d(64, 512, kernel_size=3, padding=1)
self.expand3x3_activation8 = nn.Softplus()
self.drop1 = nn.Dropout(p=0.5)
self.conv2 = BBBConv2d(512, outputs, kernel_size=1)
self.soft2 = nn.Softplus()
self.flatten = FlattenLayer(13 * 13 * 100)
self.fc1 = BBBLinearFactorial(13 * 13 * 100, outputs)
layers = [
self.conv1, self.soft1, self.pool1, self.squeeze1,
self.squeeze_activation1, self.expand3x3_1,
self.expand3x3_activation1, self.squeeze2,
self.squeeze_activation2, self.expand3x3_2,
self.expand3x3_activation2, self.pool2, self.squeeze3,
self.squeeze_activation3, self.expand3x3_3,
self.expand3x3_activation3, self.squeeze4,
self.squeeze_activation4, self.expand3x3_4,
self.expand3x3_activation4, self.pool3, self.squeeze5,
self.squeeze_activation5, self.expand3x3_5,
self.expand3x3_activation5, self.squeeze6,
self.squeeze_activation6, self.expand3x3_6,
self.expand3x3_activation6, self.squeeze7,
self.squeeze_activation7, self.expand3x3_7,
self.expand3x3_activation7, self.squeeze8,
self.squeeze_activation8, self.expand3x3_8,
self.expand3x3_activation8, self.drop1, self.conv2, self.soft2,
self.flatten, self.fc1
]
self.layers = nn.ModuleList(layers)