def __init__(self,
activation,
initType,
numChannels,
imageHeight,
imageWidth,
batchnorm=False,
softmax=False):
super(SkipLSTMEnDe, self).__init__()
self.batchnorm = batchnorm
self.bias = not self.batchnorm
self.initType = initType
self.activation = None
self.batchsize = 1
self.numChannels = numChannels
self.softmax = softmax
# Encoder
self.conv1 = nn.Conv2d(self.numChannels, 16, 3, 1, 1, bias=self.bias)
self.conv2 = nn.Conv2d(16, 32, 3, 1, 1, bias=self.bias)
self.conv3 = nn.Conv2d(32, 64, 3, 1, 1, bias=self.bias)
# Decoder
self.deconv3 = nn.ConvTranspose2d(64, 32, 3, 2, 1, 1)
self.deconv2 = nn.ConvTranspose2d(32, 16, 3, 2, 1, 1)
self.deconv1 = nn.ConvTranspose2d(16, 8, 3, 2, 1, 1)
# LSTM
self.convlstm = ConvLSTMCell(
(int(imageWidth / 8), int(imageHeight / 8)), 64, 64, (3, 3))
self.h, self.c = None, None
# Skip Connections LSTM
self.skip1 = ConvLSTMCell((int(imageWidth / 2), int(imageHeight / 2)),
16, 16, (3, 3))
self.h1, self.c1 = None, None
self.skip2 = ConvLSTMCell((int(imageWidth / 4), int(imageHeight / 4)),
32, 32, (3, 3))
self.h2, self.c2 = None, None
self.pool = nn.MaxPool2d(2, 2)
self.classifier = nn.Conv2d(8, 1, 1)
if activation == 'relu':
self.activation = nn.ReLU(inplace=True)
else:
self.activation = nn.SELU(inplace=True)
if self.batchnorm:
self.bn1 = nn.BatchNorm2d(16)
self.bn2 = nn.BatchNorm2d(32)
self.bn3 = nn.BatchNorm2d(64)
self.bn4 = nn.BatchNorm2d(32)
self.bn5 = nn.BatchNorm2d(16)
self.bn6 = nn.BatchNorm2d(8)
def __init__(self,
activation,
initType,
numChannels,
imageHeight,
imageWidth,
softmax=False):
super(SkipLSTMLayerNorm1D, self).__init__()
self.initType = initType
self.activation = None
self.batchsize = 1
self.numChannels = numChannels
self.softmax = softmax
# Encoder
self.conv1 = nn.Conv2d(self.numChannels, 16, 3, 1, 1, bias=True)
self.conv2 = nn.Conv2d(16, 32, 3, 1, 1, bias=True)
self.conv3 = nn.Conv2d(32, 64, 3, 1, 1, bias=True)
# Decoder
self.deconv3 = nn.ConvTranspose2d(64, 32, 3, 2, 1, 1)
self.deconv2 = nn.ConvTranspose2d(32, 16, 3, 2, 1, 1)
self.deconv1 = nn.ConvTranspose2d(16, 8, 3, 2, 1, 1)
# LSTM
self.convlstm = ConvLSTMCell(
(int(imageWidth / 8), int(imageHeight / 8)), 64, 64, (3, 3))
self.h, self.c = None, None
# Skip Connections LSTM
self.skip1 = ConvLSTMCell((int(imageWidth / 2), int(imageHeight / 2)),
16, 16, (3, 3))
self.h1, self.c1 = None, None
self.skip2 = ConvLSTMCell((int(imageWidth / 4), int(imageHeight / 4)),
32, 32, (3, 3))
self.h2, self.c2 = None, None
self.pool = nn.MaxPool2d(2, 2)
self.classifier = nn.Conv2d(8, 1, 1)
if activation == 'relu':
self.activation = nn.ReLU(inplace=True)
else:
self.activation = nn.SELU(inplace=True)
self.layerNorm1 = nn.LayerNorm([256, 256])
self.layerNorm2 = nn.LayerNorm([128, 128])
self.layerNorm3 = nn.LayerNorm([64, 64])
self.layerNorm4 = nn.LayerNorm([64, 64])
self.layerNorm5 = nn.LayerNorm([128, 128])
self.layerNorm6 = nn.LayerNorm([256, 256])
def __init__(self,
activation,
initType,
numChannels,
imageHeight,
imageWidth,
softmax=False):
super(EnDeLayerNorm_ws, self).__init__()
self.initType = initType
self.activation = None
self.batchsize = 1
self.numChannels = numChannels
self.softmax = softmax
# Encoder
self.conv1 = nn.Conv2d(self.numChannels, 16, 3, 1, 1, bias=True)
self.conv2 = nn.Conv2d(16, 32, 3, 1, 1, bias=True)
self.conv3 = nn.Conv2d(32, 64, 3, 1, 1, bias=True)
# Decoder
self.deconv3 = nn.ConvTranspose2d(64, 32, 3, 2, 1, 1)
self.deconv2 = nn.ConvTranspose2d(32, 16, 3, 2, 1, 1)
self.deconv1 = nn.ConvTranspose2d(16, 8, 3, 2, 1, 1)
# Conv LSTM
self.convlstm = ConvLSTMCell(
(int(imageWidth / 8), int(imageHeight / 8)), 64, 64, (3, 3))
self.h, self.c = None, None
self.pool = nn.MaxPool2d(2, 2)
self.classifier = nn.Conv2d(8, 1, 1)
if activation == 'relu':
self.activation = nn.ReLU(inplace=True)
else:
self.activation = nn.SELU(inplace=True)
self.layerNorm1 = LayerNormConv2d(16)
self.layerNorm2 = LayerNormConv2d(32)
self.layerNorm3 = LayerNormConv2d(64)
self.layerNorm4 = LayerNormConv2d(32)
self.layerNorm5 = LayerNormConv2d(16)
self.layerNorm6 = LayerNormConv2d(8)