def __init__(self):
super(ResNet50_mod_name, self).__init__()
self.conv1 = nn.Conv2d(3,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3,
stride=2,
padding=1,
return_indices=True)
self.conv111 = nn.Conv2d(64, 64, kernel_size=1, stride=1, bias=False)
self.bn111 = nn.BatchNorm2d(64)
self.conv112 = nn.Conv2d(64,
64,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn112 = nn.BatchNorm2d(64)
self.conv113 = nn.Conv2d(64, 256, kernel_size=1, stride=1, bias=False)
self.bn113 = nn.BatchNorm2d(256)
# Downsample
self.conv114 = nn.Conv2d(64, 256, kernel_size=1, stride=1, bias=False)
self.bn114 = nn.BatchNorm2d(256)
self.conv121 = nn.Conv2d(256, 64, kernel_size=1, stride=1, bias=False)
self.bn121 = nn.BatchNorm2d(64)
self.conv122 = nn.Conv2d(64,
64,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn122 = nn.BatchNorm2d(64)
self.conv123 = nn.Conv2d(64,
256,
kernel_size=1,
stride=1,
groups=1,
dilation=1,
bias=False)
self.bn123 = nn.BatchNorm2d(256)
self.conv131 = nn.Conv2d(256, 64, kernel_size=1, stride=1, bias=False)
self.bn131 = nn.BatchNorm2d(64)
self.conv132 = nn.Conv2d(64,
64,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn132 = nn.BatchNorm2d(64)
self.conv133 = nn.Conv2d(64,
256,
kernel_size=1,
stride=1,
groups=1,
dilation=1,
bias=False)
self.bn133 = nn.BatchNorm2d(256)
self.conv211 = nn.Conv2d(256, 128, kernel_size=1, stride=1, bias=False)
self.bn211 = nn.BatchNorm2d(128)
self.conv212 = nn.Conv2d(128,
128,
kernel_size=3,
stride=2,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn212 = nn.BatchNorm2d(128)
self.conv213 = nn.Conv2d(128, 512, kernel_size=1, stride=1, bias=False)
self.bn213 = nn.BatchNorm2d(512)
# Downsample
self.conv214 = nn.Conv2d(256, 512, kernel_size=1, stride=2, bias=False)
self.bn214 = nn.BatchNorm2d(512)
self.conv221 = nn.Conv2d(512, 128, kernel_size=1, stride=1, bias=False)
self.bn221 = nn.BatchNorm2d(128)
self.conv222 = nn.Conv2d(128,
128,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn222 = nn.BatchNorm2d(128)
self.conv223 = nn.Conv2d(128, 512, kernel_size=1, stride=1, bias=False)
self.bn223 = nn.BatchNorm2d(512)
self.conv231 = nn.Conv2d(512, 128, kernel_size=1, stride=1, bias=False)
self.bn231 = nn.BatchNorm2d(128)
self.conv232 = nn.Conv2d(128,
128,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn232 = nn.BatchNorm2d(128)
self.conv233 = nn.Conv2d(128, 512, kernel_size=1, stride=1, bias=False)
self.bn233 = nn.BatchNorm2d(512)
self.conv241 = nn.Conv2d(512, 128, kernel_size=1, stride=1, bias=False)
self.bn241 = nn.BatchNorm2d(128)
self.conv242 = nn.Conv2d(128,
128,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn242 = nn.BatchNorm2d(128)
self.conv243 = nn.Conv2d(128, 512, kernel_size=1, stride=1, bias=False)
self.bn243 = nn.BatchNorm2d(512)
self.conv311 = nn.Conv2d(512, 256, kernel_size=1, stride=1, bias=False)
self.bn311 = nn.BatchNorm2d(256)
self.conv312 = nn.Conv2d(256,
256,
kernel_size=3,
stride=2,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn312 = nn.BatchNorm2d(256)
self.conv313 = nn.Conv2d(256,
1024,
kernel_size=1,
stride=1,
bias=False)
self.bn313 = nn.BatchNorm2d(1024)
# Downsample
self.conv314 = nn.Conv2d(512,
1024,
kernel_size=1,
stride=2,
bias=False)
self.bn314 = nn.BatchNorm2d(1024)
self.conv321 = nn.Conv2d(1024,
256,
kernel_size=1,
stride=1,
bias=False)
self.bn321 = nn.BatchNorm2d(256)
self.conv322 = nn.Conv2d(256,
256,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn322 = nn.BatchNorm2d(256)
self.conv323 = nn.Conv2d(256,
1024,
kernel_size=1,
stride=1,
bias=False)
self.bn323 = nn.BatchNorm2d(1024)
self.conv331 = nn.Conv2d(1024,
256,
kernel_size=1,
stride=1,
bias=False)
self.bn331 = nn.BatchNorm2d(256)
self.conv332 = nn.Conv2d(256,
256,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn332 = nn.BatchNorm2d(256)
self.conv333 = nn.Conv2d(256,
1024,
kernel_size=1,
stride=1,
bias=False)
self.bn333 = nn.BatchNorm2d(1024)
self.conv341 = nn.Conv2d(1024,
256,
kernel_size=1,
stride=1,
bias=False)
self.bn341 = nn.BatchNorm2d(256)
self.conv342 = nn.Conv2d(256,
256,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn342 = nn.BatchNorm2d(256)
self.conv343 = nn.Conv2d(256,
1024,
kernel_size=1,
stride=1,
bias=False)
self.bn343 = nn.BatchNorm2d(1024)
self.conv351 = nn.Conv2d(1024,
256,
kernel_size=1,
stride=1,
bias=False)
self.bn351 = nn.BatchNorm2d(256)
self.conv352 = nn.Conv2d(256,
256,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn352 = nn.BatchNorm2d(256)
self.conv353 = nn.Conv2d(256,
1024,
kernel_size=1,
stride=1,
bias=False)
self.bn353 = nn.BatchNorm2d(1024)
self.conv361 = nn.Conv2d(1024,
256,
kernel_size=1,
stride=1,
bias=False)
self.bn361 = nn.BatchNorm2d(256)
self.conv362 = nn.Conv2d(256,
256,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn362 = nn.BatchNorm2d(256)
self.conv363 = nn.Conv2d(256,
1024,
kernel_size=1,
stride=1,
bias=False)
self.bn363 = nn.BatchNorm2d(1024)
self.conv411 = nn.Conv2d(1024,
512,
kernel_size=1,
stride=1,
bias=False)
self.bn411 = nn.BatchNorm2d(512)
self.conv412 = nn.Conv2d(512,
512,
kernel_size=3,
stride=2,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn412 = nn.BatchNorm2d(512)
self.conv413 = nn.Conv2d(512,
2048,
kernel_size=1,
stride=1,
bias=False)
self.bn413 = nn.BatchNorm2d(2048)
# Downsample
self.conv414 = nn.Conv2d(1024,
2048,
kernel_size=1,
stride=2,
bias=False)
self.bn414 = nn.BatchNorm2d(2048)
self.conv421 = nn.Conv2d(2048,
512,
kernel_size=1,
stride=1,
bias=False)
self.bn421 = nn.BatchNorm2d(512)
self.conv422 = nn.Conv2d(512,
512,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn422 = nn.BatchNorm2d(512)
self.conv423 = nn.Conv2d(512,
2048,
kernel_size=1,
stride=1,
bias=False)
self.bn423 = nn.BatchNorm2d(2048)
self.conv431 = nn.Conv2d(2048,
512,
kernel_size=1,
stride=1,
bias=False)
self.bn431 = nn.BatchNorm2d(512)
self.conv432 = nn.Conv2d(512,
512,
kernel_size=3,
stride=1,
groups=1,
dilation=1,
bias=False,
padding=1)
self.bn432 = nn.BatchNorm2d(512)
self.conv433 = nn.Conv2d(512,
2048,
kernel_size=1,
stride=1,
bias=False)
self.bn433 = nn.BatchNorm2d(2048)
self.unpool = nn.MaxUnpool2d(kernel_size=3, stride=2, padding=1)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.unsample = nn.Upsample(size=7, mode='nearest')
def __init__(self):
super(SegNet, self).__init__()
# initialise network parameters
filter = [64, 128, 256, 512, 512]
self.class_nb = 13
# define encoder decoder layers
self.encoder_block = nn.ModuleList([self.conv_layer([3, filter[0]])])
self.decoder_block = nn.ModuleList([self.conv_layer([filter[0], filter[0]])])
for i in range(4):
self.encoder_block.append(self.conv_layer([filter[i], filter[i + 1]]))
self.decoder_block.append(self.conv_layer([filter[i + 1], filter[i]]))
# define convolution layer
self.conv_block_enc = nn.ModuleList([self.conv_layer([filter[0], filter[0]])])
self.conv_block_dec = nn.ModuleList([self.conv_layer([filter[0], filter[0]])])
for i in range(4):
if i == 0:
self.conv_block_enc.append(self.conv_layer([filter[i + 1], filter[i + 1]]))
self.conv_block_dec.append(self.conv_layer([filter[i], filter[i]]))
else:
self.conv_block_enc.append(nn.Sequential(self.conv_layer([filter[i + 1], filter[i + 1]]),
self.conv_layer([filter[i + 1], filter[i + 1]])))
self.conv_block_dec.append(nn.Sequential(self.conv_layer([filter[i], filter[i]]),
self.conv_layer([filter[i], filter[i]])))
# define task attention layers
self.encoder_att = nn.ModuleList([nn.ModuleList([self.att_layer([filter[0], filter[0], filter[0]])])])
self.decoder_att = nn.ModuleList([nn.ModuleList([self.att_layer([2 * filter[0], filter[0], filter[0]])])])
self.encoder_block_att = nn.ModuleList([self.conv_layer([filter[0], filter[1]])])
self.decoder_block_att = nn.ModuleList([self.conv_layer([filter[0], filter[0]])])
for j in range(3):
if j < 2:
self.encoder_att.append(nn.ModuleList([self.att_layer([filter[0], filter[0], filter[0]])]))
self.decoder_att.append(nn.ModuleList([self.att_layer([2 * filter[0], filter[0], filter[0]])]))
for i in range(4):
self.encoder_att[j].append(self.att_layer([2 * filter[i + 1], filter[i + 1], filter[i + 1]]))
self.decoder_att[j].append(self.att_layer([filter[i + 1] + filter[i], filter[i], filter[i]]))
for i in range(4):
if i < 3:
self.encoder_block_att.append(self.conv_layer([filter[i + 1], filter[i + 2]]))
self.decoder_block_att.append(self.conv_layer([filter[i + 1], filter[i]]))
else:
self.encoder_block_att.append(self.conv_layer([filter[i + 1], filter[i + 1]]))
self.decoder_block_att.append(self.conv_layer([filter[i + 1], filter[i + 1]]))
self.pred_task1 = self.conv_layer([filter[0], self.class_nb], pred=True)
self.pred_task2 = self.conv_layer([filter[0], 1], pred=True)
self.pred_task3 = self.conv_layer([filter[0], 3], pred=True)
# define pooling and unpooling functions
self.down_sampling = nn.MaxPool2d(kernel_size=2, stride=2, return_indices=True)
self.up_sampling = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.logsigma = nn.Parameter(torch.FloatTensor([-0.5, -0.5, -0.5]))
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.xavier_normal_(m.weight)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.xavier_normal_(m.weight)
nn.init.constant_(m.bias, 0)
def __init__(self, in_size, out_size):
super(segnetUp3, self).__init__()
self.unpool = nn.MaxUnpool2d(2, 2)
self.conv1 = conv2DBatchNormRelu(in_size, in_size, 3, 1, 1)
self.conv2 = conv2DBatchNormRelu(in_size, in_size, 3, 1, 1)
self.conv3 = conv2DBatchNormRelu(in_size, out_size, 3, 1, 1)
def __init__(self):
super(FCN, self).__init__()
self.conv1_1 = nn.Conv2d(3,
64,
kernel_size=(3, 3),
stride=1,
padding=1)
self.conv1_2 = nn.Conv2d(64,
64,
kernel_size=(3, 3),
stride=1,
padding=1)
self.max1 = nn.MaxPool2d(kernel_size=(2, 2),
stride=2,
return_indices=True)
self.conv2_1 = nn.Conv2d(64,
128,
kernel_size=(3, 3),
stride=1,
padding=1)
self.conv2_2 = nn.Conv2d(128,
128,
kernel_size=(3, 3),
stride=1,
padding=1)
self.max2 = nn.MaxPool2d(kernel_size=(2, 2),
stride=2,
return_indices=True)
self.conv3_1 = nn.Conv2d(128,
256,
kernel_size=(3, 3),
stride=1,
padding=1)
self.conv3_2 = nn.Conv2d(256,
256,
kernel_size=(3, 3),
stride=1,
padding=1)
self.conv3_3 = nn.Conv2d(256,
256,
kernel_size=(3, 3),
stride=1,
padding=1)
self.max3 = nn.MaxPool2d(kernel_size=(2, 2),
stride=2,
return_indices=True)
self.conv4_1 = nn.Conv2d(256,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.conv4_2 = nn.Conv2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.conv4_3 = nn.Conv2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.max4 = nn.MaxPool2d(kernel_size=(2, 2),
stride=2,
return_indices=True)
self.conv5_1 = nn.Conv2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.conv5_2 = nn.Conv2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.conv5_3 = nn.Conv2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.max5 = nn.MaxPool2d(kernel_size=(2, 2),
stride=2,
return_indices=True)
self.deconv1_1 = nn.ConvTranspose2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.deconv1_2 = nn.ConvTranspose2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.deconv1_3 = nn.ConvTranspose2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.unpool1 = nn.MaxUnpool2d((2, 2), stride=2)
self.deconv2_1 = nn.ConvTranspose2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.deconv2_2 = nn.ConvTranspose2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.deconv2_3 = nn.ConvTranspose2d(512,
512,
kernel_size=(3, 3),
stride=1,
padding=1)
self.unpool2 = nn.MaxUnpool2d(2, stride=2)
self.deconv3_1 = nn.ConvTranspose2d(512,
256,
kernel_size=(3, 3),
stride=1,
padding=1)
self.deconv3_2 = nn.ConvTranspose2d(256,
256,
kernel_size=(3, 3),
stride=1,
padding=1)
self.deconv3_3 = nn.ConvTranspose2d(256,
256,
kernel_size=(3, 3),
stride=1,
padding=1)
self.unpool3 = nn.MaxUnpool2d(2, stride=2)
self.deconv4_1 = nn.ConvTranspose2d(256,
128,
kernel_size=(3, 3),
stride=1,
padding=1)
self.deconv4_2 = nn.ConvTranspose2d(128,
128,
kernel_size=(3, 3),
stride=1,
padding=1)
self.unpool4 = nn.MaxUnpool2d(2, stride=2)
self.deconv5_1 = nn.ConvTranspose2d(128,
64,
kernel_size=(3, 3),
stride=1,
padding=1)
self.deconv5_2 = nn.ConvTranspose2d(64,
64,
kernel_size=(3, 3),
stride=1,
padding=1)
self.unpool5 = nn.MaxUnpool2d(2, stride=2)
self.deconv6_1 = nn.ConvTranspose2d(64,
2,
kernel_size=(3, 3),
stride=1,
padding=1)
def __init__(self, in_size, out_size):
super(segnetUp1, self).__init__()
self.unpool = nn.MaxUnpool2d(2, 2)
self.conv = conv2DBatchNormRelu(in_size, out_size, k_size=5, stride=1, padding=2, with_relu=False)
def __init__(self, params, se_block_type=None):
super(DecoderBlock, self).__init__(params, se_block_type=se_block_type)
self.unpool = nn.MaxUnpool2d(kernel_size=params['pool'],
stride=params['stride_pool'])
img_tensor = torch.ones((1, 1, 4, 4))
avgpool_layer = nn.AvgPool2d((2, 2), stride=(2, 2), divisor_override=3)
img_pool = avgpool_layer(img_tensor)
print("raw_img:\n{0}\npooling_img:\n{1}".format(img_tensor, img_pool))
# maxunpool 最大上采样
# maxpool中的return_indices以tensor形式返回所有最大值的下标
flag = 0
if flag:
img_tensor = torch.randint(high=5, size=(1, 1, 4, 4), dtype=torch.float)
maxpool_layer = nn.MaxPool2d((2, 2), stride=(2, 2), return_indices=True)
img_pool, indices = maxpool_layer(img_tensor)
img_reconstruct = torch.randn_like(img_pool, dtype=torch.float)
maxunpool_layer = nn.MaxUnpool2d((2, 2), stride=(2, 2))
img_unpool = maxunpool_layer(img_reconstruct, indices)
print("raw_img:\n{0}\nimg_pool:\n{1}".format(img_tensor, img_pool))
print("img_reconstruct:\n{0}\nimg_unpool:\n{1}".format(
img_reconstruct, img_unpool))
# 线性变换 linear
flag = 0
if flag:
inputs = torch.tensor([[1., 2, 3]])
linear_layer = nn.Linear(3, 4)
linear_layer.weight.data = torch.tensor([[1., 1., 1.], [2., 2., 2.],
[3., 3., 3.], [4., 4., 4.]])
linear_layer.bias.data.fill_(0.5)
def __init__(self, num_classes):
super(DeconvNet,self).__init__()
layers = make_layers()
self.features_map1 = layers[0]
self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True, return_indices=True) #
self.features_map2 = layers[1]
self.pool2 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True, return_indices=True) #
self.features_map3 = layers[2]
self.pool3 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True, return_indices=True) #
self.features_map4 = layers[3]
self.pool4 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True, return_indices=True) #
self.features_map5 = layers[4]
self.pool5 = nn.MaxPool2d(kernel_size=2, stride=2, ceil_mode=True, return_indices=True) #
# fc6 ~ fc7
self.fc = nn.Sequential(nn.Conv2d(512, 4096, kernel_size = 7),
nn.ReLU(inplace=True),
nn.Dropout(),
nn.Conv2d(4096, 4096, kernel_size = 1),
nn.ReLU(inplace=True),
nn.Dropout()
)
# Deconv
self.fc_deconv = DCB(4096, 512, 7, 1,0)
self.unpool5 = nn.MaxUnpool2d(2, stride=2)
self.deconv5 = nn.Sequential(
DCB(512, 512, 3, 1, 1),
DCB(512, 512, 3, 1, 1),
DCB(512, 512, 3, 1, 1)
)
self.unpool4 = nn.MaxUnpool2d(2, stride=2)
self.deconv4 = nn.Sequential(
DCB(512, 512, 3, 1, 1),
DCB(512, 512, 3, 1, 1),
DCB(512, 256, 3, 1, 1)
)
self.unpool3 = nn.MaxUnpool2d(2, stride=2)
self.deconv3 = nn.Sequential(
DCB(256, 256, 3, 1, 1),
DCB(256, 256, 3, 1, 1),
DCB(256, 128, 3, 1, 1)
)
self.unpool2 = nn.MaxUnpool2d(2, stride=2)
self.deconv2 = nn.Sequential(
DCB(128, 128, 3, 1, 1),
DCB(128, 64, 3, 1, 1)
)
self.unpool1 = nn.MaxUnpool2d(2, stride=2)
self.deconv1 = nn.Sequential(
DCB(64, 64, 3, 1, 1),
DCB(64, 64, 3, 1, 1)
)
self.score = nn.Conv2d(64,
num_classes,
kernel_size=1,
stride=1,
padding=0,
dilation=1)
def __init__(self):
super(encoder_decoder, self).__init__()
#Defining the Encoder Network
self.conv11 = nn.Conv2d(in_channels=3,
out_channels=64,
kernel_size=3,
stride=1,
padding=1)
self.conv12 = nn.Conv2d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=1)
self.bn1 = nn.BatchNorm2d(64)
self.conv21 = nn.Conv2d(in_channels=64,
out_channels=128,
kernel_size=3,
stride=1,
padding=1)
self.conv22 = nn.Conv2d(in_channels=128,
out_channels=128,
kernel_size=3,
stride=1,
padding=1)
self.bn2 = nn.BatchNorm2d(128)
self.conv31 = nn.Conv2d(in_channels=128,
out_channels=256,
kernel_size=3,
stride=1,
padding=1)
self.conv32 = nn.Conv2d(in_channels=256,
out_channels=256,
kernel_size=3,
stride=1,
padding=1)
self.bn3 = nn.BatchNorm2d(256)
self.conv41 = nn.Conv2d(in_channels=256,
out_channels=512,
kernel_size=3,
stride=1,
padding=1)
self.conv42 = nn.Conv2d(in_channels=512,
out_channels=512,
kernel_size=3,
stride=1,
padding=1)
self.bn4 = nn.BatchNorm2d(512)
#Defining Relu and Maxpool Layer
self.prelu = nn.PReLU()
self.maxpool = nn.MaxPool2d(kernel_size=2,
stride=2,
return_indices=True)
#Defining Decoder Network
self.deconv41 = nn.ConvTranspose2d(in_channels=512,
out_channels=512,
kernel_size=3,
stride=1,
padding=1)
self.deconv31 = nn.ConvTranspose2d(in_channels=512,
out_channels=256,
kernel_size=3,
stride=1,
padding=1)
self.deconv32 = nn.ConvTranspose2d(in_channels=256,
out_channels=256,
kernel_size=3,
stride=1,
padding=1)
self.deconv21 = nn.ConvTranspose2d(in_channels=256,
out_channels=128,
kernel_size=3,
stride=1,
padding=1)
self.deconv22 = nn.ConvTranspose2d(in_channels=128,
out_channels=128,
kernel_size=3,
stride=1,
padding=1)
self.deconv11 = nn.ConvTranspose2d(in_channels=128,
out_channels=64,
kernel_size=3,
stride=1,
padding=1)
self.deconv12 = nn.ConvTranspose2d(in_channels=64,
out_channels=64,
kernel_size=3,
stride=1,
padding=1)
self.deconv13 = nn.ConvTranspose2d(in_channels=64,
out_channels=11,
kernel_size=3,
stride=1,
padding=1)
self.bn5 = nn.BatchNorm2d(11)
#Defining UnMaxpool Layer
self.maxunpool = nn.MaxUnpool2d(kernel_size=2, stride=2)
#Defining Last Pooling Layers
self.maxpool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.avgpool = nn.AvgPool2d(kernel_size=2, stride=2)
#Inittializing the module
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal(m.weight, mode='fan_out')
if m.bias is not None:
nn.init.constant(m.bias, 0)
elif isinstance(m, nn.ConvTranspose2d):
nn.init.kaiming_normal(m.weight, mode='fan_out')
if m.bias is not None:
nn.init.constant(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant(m.weight, 1)
nn.init.constant(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.normal(m.weight, 0, 0.01)
nn.init.constant(m.bias, 0)
def __init__(self,
state_size=((imgh, imgw), img_in_channels),
latent_size=latent_size,
conv_channels=conv_channels,
kernel_size=kernel_size):
super(VAE, self).__init__()
self.state_size, self.channels_in = state_size
self.conv1 = nn.Conv2d(self.channels_in,
conv_channels,
kernel_size=kernel_size,
stride=1,
padding=0)
self.bn1 = nn.BatchNorm2d(num_features=conv_channels)
self.pool1 = nn.MaxPool2d((2, 2), return_indices=True)
self.conv2 = nn.Conv2d(conv_channels,
conv_channels * 2,
kernel_size=kernel_size,
stride=1,
padding=1)
self.bn2 = nn.BatchNorm2d(num_features=conv_channels * 2)
self.pool2 = nn.MaxPool2d((2, 2), return_indices=True)
self.conv3 = nn.Conv2d(conv_channels * 2,
conv_channels * 2,
kernel_size=kernel_size,
stride=1,
padding=1)
self.bn3 = nn.BatchNorm2d(num_features=conv_channels * 2)
self.pool3 = nn.MaxPool2d((2, 2), return_indices=True)
self.flat_size = (-1,
conv_channels * 2 * ceil(self.state_size[0] / 8) *
ceil(self.state_size[1] / 8))
self.fc21 = nn.Linear(
conv_channels * 2 * ceil(self.state_size[0] / 8) *
ceil(self.state_size[1] / 8), latent_size)
self.fc22 = nn.Linear(
conv_channels * 2 * ceil(self.state_size[0] / 8) *
ceil(self.state_size[1] / 8), latent_size)
self.fc3 = nn.Linear(
latent_size, conv_channels * 2 * ceil(self.state_size[0] / 8) *
ceil(self.state_size[1] / 8))
self.unflatt_size = (-1, conv_channels * 2,
ceil(self.state_size[0] / 8),
ceil(self.state_size[1] / 8))
self.unpool1 = nn.MaxUnpool2d(kernel_size)
self.uncov1 = nn.Conv2d(conv_channels * 2,
conv_channels * 2,
kernel_size=kernel_size,
padding=0)
self.upsample1 = nn.Upsample(scale_factor=2, mode='nearest')
self.unpool2 = nn.MaxUnpool2d(kernel_size)
self.uncov2 = nn.Conv2d(conv_channels * 2,
conv_channels,
kernel_size=kernel_size,
padding=1)
self.upsample2 = nn.Upsample(scale_factor=2, mode='nearest')
self.unpool3 = nn.MaxUnpool2d(kernel_size)
self.uncov3 = nn.Conv2d(conv_channels,
self.channels_in,
kernel_size=kernel_size,
padding=1)
self.upsample3 = nn.Upsample(scale_factor=2, mode='nearest')
def __init__(self, in_channels, outsize):
super(ResblocksDeconv, self).__init__()
# conv 1_1:
self.conv1_1 = nn.Conv2d(in_channels, 64, kernel_size=3, stride=1, padding=1)
self.bn1_1 = nn.BatchNorm2d(64)
# conv 1_2
self.conv1_2 = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1)
self.bn1_2 = nn.BatchNorm2d(64)
# pool 1
self.maxp1 = nn.MaxPool2d(kernel_size=2, stride=2, return_indices=True)
# 120 x 120
# conv2_1
self.conv2_1 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
self.bn2_1 = nn.BatchNorm2d(128)
# conv2_2
self.conv2_2 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1)
self.bn2_2 = nn.BatchNorm2d(128)
# pool2
self.maxp2 = nn.MaxPool2d(kernel_size=2, stride=2, return_indices=True)
# 60 x 60
# conv3_1
self.conv3_1 = nn.Conv2d(128, 256, kernel_size=3, stride=1, padding=1)
self.bn3_1 = nn.BatchNorm2d(256)
# conv3_2
self.conv3_2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.bn3_2 = nn.BatchNorm2d(256)
# conv3_3
self.conv3_3 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.bn3_3 = nn.BatchNorm2d(256)
# pool3
self.maxp3 = nn.MaxPool2d(kernel_size=2, stride=2, return_indices=True)
# 30 x 30
# conv4_1
self.conv4_1 = nn.Conv2d(256, 512, kernel_size=3, stride=1, padding=1)
self.bn4_1 = nn.BatchNorm2d(512)
# conv4_2
self.conv4_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.bn4_2 = nn.BatchNorm2d(512)
# conv4_3
self.conv4_3 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.bn4_3 = nn.BatchNorm2d(512)
# pool4
self.maxp4 = nn.MaxPool2d(kernel_size=2, stride=2, return_indices=True)
# 15 x 15
# conv5_1
self.conv5_1 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.bn5_1 = nn.BatchNorm2d(512)
# conv5_2
self.conv5_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.bn5_2 = nn.BatchNorm2d(512)
# conv5_3
self.conv5_3 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.bn5_3 = nn.BatchNorm2d(512)
# pool5
self.maxp5 = nn.MaxPool2d(kernel_size=2, stride=2, return_indices=True)
# 7 x 7
self.residualBlock_1 = ResidualBlock(512, 512)
self.residualBlock_2 = ResidualBlock(512, 512)
self.residualBlock_3 = ResidualBlock(512, 512)
self.residualBlock_4 = ResidualBlock(512, 512)
self.residualBlock_5 = ResidualBlock(512, 512)
# 7 x 7
# unpool5
self.unpool5 = nn.MaxUnpool2d(2,2)
# _ x _
# deconv5_1
self.deconv5_1 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.debn5_1 = nn.BatchNorm2d(512)
# conv3_2
self.deconv5_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.debn5_2 = nn.BatchNorm2d(512)
# conv3_3
self.deconv5_3 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.debn5_3 = nn.BatchNorm2d(512)
# _ x _
# unpool4
self.unpool4 = nn.MaxUnpool2d(2,2) #paper did unpoolsize 28 (we want 30 x 30)
# _ x _
# deconv4_1
self.deconv4_1 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.debn4_1 = nn.BatchNorm2d(512)
# deconv4_2
self.deconv4_2 = nn.Conv2d(512, 512, kernel_size=3, stride=1, padding=1)
self.debn4_2 = nn.BatchNorm2d(512)
# conv4_3
self.deconv4_3 = nn.Conv2d(512, 256, kernel_size=3, stride=1, padding=1)
self.debn4_3 = nn.BatchNorm2d(256)
# _ x _
# unpool3
self.unpool3 = nn.MaxUnpool2d(2,2) #paper did unpoolsize 56 (we want 60 x 60)
# _ x _
# deconv3_1
self.deconv3_1 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.debn3_1 = nn.BatchNorm2d(256)
# deconv3_2
self.deconv3_2 = nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1)
self.debn3_2 = nn.BatchNorm2d(256)
# deconv3_3
self.deconv3_3 = nn.Conv2d(256, 128, kernel_size=3, stride=1, padding=1)
self.debn3_3 = nn.BatchNorm2d(128)
# _ x _
# unpool2
self.unpool2 = nn.MaxUnpool2d(2,2) #paper did unpoolsize 112 (we want 120 x 120)
# _ x _
# deconv3_1
self.deconv2_1 = nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1)
self.debn2_1 = nn.BatchNorm2d(128)
# deconv3_2
self.deconv2_2 = nn.Conv2d(128, 64, kernel_size=3, stride=1, padding=1)
self.debn2_2 = nn.BatchNorm2d(64)
# _ x _
# unpool1
self.unpool1 = nn.MaxUnpool2d(2,2) #paper did unpoolsize 224 (we want 240 x 240)
# _ x _
# deconv3_1
self.deconv1_1 = nn.Conv2d(64, 32, kernel_size=3, stride=1, padding=1)
self.debn1_1 = nn.BatchNorm2d(32)
# deconv3_2
self.deconv1_2 = nn.Conv2d(32, 3, kernel_size=3, stride=1, padding=1)
# self.debn1_2 = nn.BatchNorm2d(3)
self.in_channels = in_channels
self.outsize = outsize
def __append_layer(self, net_style, args_dict):
args_values_list = list(args_dict.values())
if net_style == "Conv2d":
self.layers.append(
nn.Conv2d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5],
args_values_list[6], args_values_list[7]))
elif net_style == "MaxPool2d":
self.layers.append(
nn.MaxPool2d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5]))
elif net_style == "Linear":
self.layers.append(
nn.Linear(args_values_list[0], args_values_list[1],
args_values_list[2]))
elif net_style == "reshape":
# 如果是特殊情况 reshape，就直接将目标向量尺寸传入
# print(type(args_values_list[0]))
self.layers.append(args_values_list[0])
elif net_style == "Conv1d":
self.layers.append(
nn.Conv1d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5],
args_values_list[6], args_values_list[7]))
elif net_style == "Conv3d":
self.layers.append(
nn.Conv3d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5],
args_values_list[6], args_values_list[7]))
elif net_style == "ConvTranspose1d":
self.layers.append(
nn.ConvTranspose1d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5],
args_values_list[6], args_values_list[7],
args_values_list[8]))
elif net_style == "ConvTranspose2d":
self.layers.append(
nn.ConvTranspose2d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5],
args_values_list[6], args_values_list[7],
args_values_list[8]))
elif net_style == "ConvTranspose3d":
self.layers.append(
nn.ConvTranspose3d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5],
args_values_list[6], args_values_list[7],
args_values_list[8]))
elif net_style == "Unfold":
self.layers.append(
nn.Unfold(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3]))
elif net_style == "Fold":
self.layers.append(
nn.Unfold(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "MaxPool1d":
self.layers.append(
nn.MaxPool1d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5]))
elif net_style == "MaxPool3d":
self.layers.append(
nn.MaxPool3d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4], args_values_list[5]))
elif net_style == "MaxUnpool1d":
self.layers.append(
nn.MaxUnpool1d(args_values_list[0], args_values_list[1],
args_values_list[2]))
elif net_style == "MaxUnpool2d":
self.layers.append(
nn.MaxUnpool2d(args_values_list[0], args_values_list[1],
args_values_list[2]))
elif net_style == "MaxUnpool3d":
self.layers.append(
nn.MaxUnpool3d(args_values_list[0], args_values_list[1],
args_values_list[2]))
elif net_style == "AvgPool1d":
self.layers.append(
nn.AvgPool1d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "AvgPool2d":
self.layers.append(
nn.AvgPool2d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "AvgPool3d":
self.layers.append(
nn.AvgPool3d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "FractionalMaxPool2d":
self.layers.append(
nn.FractionalMaxPool2d(args_values_list[0],
args_values_list[1],
args_values_list[2],
args_values_list[3],
args_values_list[4]))
elif net_style == "LPPool1d":
self.layers.append(
nn.LPPool1d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3]))
elif net_style == "LPPool2d":
self.layers.append(
nn.LPPool2d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3]))
elif net_style == "AdaptiveMaxPool1d":
self.layers.append(
nn.AdaptiveMaxPool1d(args_values_list[0], args_values_list[1]))
elif net_style == "AdaptiveMaxPool2d":
self.layers.append(
nn.AdaptiveMaxPool2d(args_values_list[0], args_values_list[1]))
elif net_style == "AdaptiveMaxPool3d":
self.layers.append(
nn.AdaptiveMaxPool3d(args_values_list[0], args_values_list[1]))
elif net_style == "AdaptiveAvgPool1d":
self.layers.append(nn.AdaptiveAvgPool1d(args_values_list[0]))
elif net_style == "AdaptiveAvgPool2d":
self.layers.append(nn.AdaptiveAvgPool2d(args_values_list[0]))
elif net_style == "AdaptiveAvgPool3d":
self.layers.append(nn.AdaptiveAvgPool3d(args_values_list[0]))
elif net_style == "ReflectionPad1d":
self.layers.append(nn.ReflectionPad1d(args_values_list[0]))
elif net_style == "ReflectionPad2d":
self.layers.append(nn.ReflectionPad2d(args_values_list[0]))
elif net_style == "ReplicationPad1d":
self.layers.append(nn.ReplicationPad1d(args_values_list[0]))
elif net_style == "ReplicationPad2d":
self.layers.append(nn.ReplicationPad2d(args_values_list[0]))
elif net_style == "ReplicationPad3d":
self.layers.append(nn.ReplicationPad3d(args_values_list[0]))
elif net_style == "ZeroPad2d":
self.layers.append(nn.ZeroPad2d(args_values_list[0]))
elif net_style == "ConstantPad1d":
self.layers.append(
nn.ConstantPad1d(args_values_list[0], args_values_list[1]))
elif net_style == "ConstantPad2d":
self.layers.append(
nn.ConstantPad2d(args_values_list[0], args_values_list[1]))
elif net_style == "ConstantPad3d":
self.layers.append(
nn.ConstantPad3d(args_values_list[0], args_values_list[1]))
elif net_style == "ELU":
self.layers.append(nn.ELU(args_values_list[0],
args_values_list[1]))
elif net_style == "Hardshrink":
self.layers.append(nn.Hardshrink(args_values_list[0]))
elif net_style == "Hardtanh":
self.layers.append(
nn.Hardtanh(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "LeakyReLU":
self.layers.append(
nn.LeakyReLU(args_values_list[0], args_values_list[1]))
elif net_style == "LogSigmoid":
self.layers.append(nn.LogSigmoid())
elif net_style == "PReLU":
self.layers.append(
nn.PReLU(args_values_list[0], args_values_list[1]))
elif net_style == "ReLU":
self.layers.append(nn.ReLU(args_values_list[0]))
elif net_style == "ReLU6":
self.layers.append(nn.ReLU6(args_values_list[0]))
elif net_style == "RReLU":
self.layers.append(
nn.RReLU(args_values_list[0], args_values_list[1],
args_values_list[2]))
elif net_style == "SELU":
self.layers.append(nn.SELU(args_values_list[0]))
elif net_style == "CELU":
self.layers.append(
nn.CELU(args_values_list[0], args_values_list[1]))
elif net_style == "Sigmoid":
self.layers.append(nn.Sigmoid())
elif net_style == "Softplus":
self.layers.append(
nn.Softplus(args_values_list[0], args_values_list[1]))
elif net_style == "Softshrink":
self.layers.append(nn.Softshrink(args_values_list[0]))
elif net_style == "Softsign":
self.layers.append(nn.Softsign())
elif net_style == "Tanh":
self.layers.append(nn.Tanh())
elif net_style == "Tanhshrink":
self.layers.append(nn.Tanhshrink())
elif net_style == "Threshold":
self.layers.append(
nn.Threshold(args_values_list[0], args_values_list[1],
args_values_list[2]))
elif net_style == "Softmin":
self.layers.append(nn.Softmin(args_values_list[0]))
elif net_style == "Softmax":
self.layers.append(nn.Softmax(args_values_list[0]))
elif net_style == "Softmax2d":
self.layers.append(nn.Softmax2d())
elif net_style == "LogSoftmax":
self.layers.append(nn.LogSoftmax(args_values_list[0]))
elif net_style == "AdaptiveLogSoftmaxWithLoss":
self.layers.append(
nn.AdaptiveLogSoftmaxWithLoss(args_values_list[0],
args_values_list[1],
args_values_list[2],
args_values_list[3],
args_values_list[4]))
elif net_style == "BatchNorm1d":
self.layers.append(
nn.BatchNorm1d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "BatchNorm2d":
self.layers.append(
nn.BatchNorm2d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "BatchNorm3d":
self.layers.append(
nn.BatchNorm3d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "GroupNorm":
self.layers.append(
nn.GroupNorm(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3]))
elif net_style == "InstanceNorm1d":
self.layers.append(
nn.InstanceNorm1d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "InstanceNorm2d":
self.layers.append(
nn.InstanceNorm2d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "InstanceNorm3d":
self.layers.append(
nn.InstanceNorm3d(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3],
args_values_list[4]))
elif net_style == "LayerNorm":
self.layers.append(
nn.LayerNorm(args_values_list[0], args_values_list[1],
args_values_list[2]))
elif net_style == "LocalResponseNorm":
self.layers.append(
nn.LocalResponseNorm(args_values_list[0], args_values_list[1],
args_values_list[2], args_values_list[3]))
elif net_style == "Linear":
self.layers.append(
nn.Linear(args_values_list[0], args_values_list[1],
args_values_list[2]))
elif net_style == "Dropout":
self.layers.append(
nn.Dropout(args_values_list[0], args_values_list[1]))
elif net_style == "Dropout2d":
self.layers.append(
nn.Dropout2d(args_values_list[0], args_values_list[1]))
elif net_style == "Dropout3d":
self.layers.append(
nn.Dropout3d(args_values_list[0], args_values_list[1]))
elif net_style == "AlphaDropout":
self.layers.append(
nn.AlphaDropout(args_values_list[0], args_values_list[1]))
def __init__(self, n_classes=6, learned_billinear=False):
super(patch_deconvnet, self).__init__()
self.learned_billinear = learned_billinear
self.n_classes = n_classes
self.unpool = nn.MaxUnpool2d(2, stride=2)
self.conv_block1 = nn.Sequential(
# conv1_1
nn.Conv2d(1, 64, 3, padding=1),
nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# conv1_2
nn.Conv2d(64, 64, 3, padding=1),
nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# pool1
nn.MaxPool2d(2, stride=2, return_indices=True, ceil_mode=True),
)
# it returns outputs and pool_indices_1
# 48*48
self.conv_block2 = nn.Sequential(
# conv2_1
nn.Conv2d(64, 128, 3, padding=1),
nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# conv2_2
nn.Conv2d(128, 128, 3, padding=1),
nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# pool2
nn.MaxPool2d(2, stride=2, return_indices=True, ceil_mode=True),
)
# it returns outputs and pool_indices_2
# 24*24
self.conv_block3 = nn.Sequential(
# conv3_1
nn.Conv2d(128, 256, 3, padding=1),
nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# conv3_2
nn.Conv2d(256, 256, 3, padding=1),
nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# conv3_3
nn.Conv2d(256, 256, 3, padding=1),
nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# pool3
nn.MaxPool2d(2, stride=2, return_indices=True, ceil_mode=True),
)
# it returns outputs and pool_indices_3
# 12*12
self.conv_block4 = nn.Sequential(
# conv4_1
nn.Conv2d(256, 512, 3, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# conv4_2
nn.Conv2d(512, 512, 3, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# conv4_3
nn.Conv2d(512, 512, 3, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# pool4
nn.MaxPool2d(2, stride=2, return_indices=True, ceil_mode=True),
)
# it returns outputs and pool_indices_4
# 6*6
self.conv_block5 = nn.Sequential(
# conv5_1
nn.Conv2d(512, 512, 3, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# conv5_2
nn.Conv2d(512, 512, 3, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# conv5_3
nn.Conv2d(512, 512, 3, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# pool5
nn.MaxPool2d(2, stride=2, return_indices=True, ceil_mode=True),
)
# it returns outputs and pool_indices_5
# 3*3
self.conv_block6 = nn.Sequential(
# fc6
nn.Conv2d(512, 4096, 3),
# set the filter size and nor padding to make output into 1*1
nn.BatchNorm2d(4096, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
)
# 1*1
self.conv_block7 = nn.Sequential(
# fc7
nn.Conv2d(4096, 4096, 1),
# set the filter size to make output into 1*1
nn.BatchNorm2d(4096, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
)
self.deconv_block8 = nn.Sequential(
# fc6-deconv
nn.ConvTranspose2d(4096, 512, 3, stride=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
)
# 3*3
self.unpool_block9 = nn.Sequential(
# unpool5
nn.MaxUnpool2d(2, stride=2), )
# usage unpool(output, indices)
# 6*6
self.deconv_block10 = nn.Sequential(
# deconv5_1
nn.ConvTranspose2d(512, 512, 3, stride=1, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# deconv5_2
nn.ConvTranspose2d(512, 512, 3, stride=1, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# deconv5_3
nn.ConvTranspose2d(512, 512, 3, stride=1, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
)
self.unpool_block11 = nn.Sequential(
# unpool4
nn.MaxUnpool2d(2, stride=2), )
# 12*12
self.deconv_block12 = nn.Sequential(
# deconv4_1
nn.ConvTranspose2d(512, 512, 3, stride=1, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# deconv4_2
nn.ConvTranspose2d(512, 512, 3, stride=1, padding=1),
nn.BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# deconv4_3
nn.ConvTranspose2d(512, 256, 3, stride=1, padding=1),
nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
)
self.unpool_block13 = nn.Sequential(
# unpool3
nn.MaxUnpool2d(2, stride=2), )
# 24*24
self.deconv_block14 = nn.Sequential(
# deconv3_1
nn.ConvTranspose2d(256, 256, 3, stride=1, padding=1),
nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# deconv3_2
nn.ConvTranspose2d(256, 256, 3, stride=1, padding=1),
nn.BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# deconv3_3
nn.ConvTranspose2d(256, 128, 3, stride=1, padding=1),
nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
)
self.unpool_block15 = nn.Sequential(
# unpool2
nn.MaxUnpool2d(2, stride=2), )
# 48*48
self.deconv_block16 = nn.Sequential(
# deconv2_1
nn.ConvTranspose2d(128, 128, 3, stride=1, padding=1),
nn.BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# deconv2_2
nn.ConvTranspose2d(128, 64, 3, stride=1, padding=1),
nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
)
self.unpool_block17 = nn.Sequential(
# unpool1
nn.MaxUnpool2d(2, stride=2), )
# 96*96
self.deconv_block18 = nn.Sequential(
# deconv1_1
nn.ConvTranspose2d(64, 64, 3, stride=1, padding=1),
nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
# deconv1_2
nn.ConvTranspose2d(64, 64, 3, stride=1, padding=1),
nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True),
nn.ReLU(inplace=True),
)
self.seg_score19 = nn.Sequential(
# seg-score
nn.Conv2d(64, self.n_classes, 1), )
if self.learned_billinear:
raise NotImplementedError
def __init__(self, num_classes, init_weights=True, pretrained=True):
super(DeconvNet, self).__init__()
layers = make_layers(pretrained)
self.conv1 = layers[0]
self.pool1 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
return_indices=True)
self.conv2 = layers[1]
self.pool2 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
return_indices=True)
self.conv3 = layers[2]
self.pool3 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
return_indices=True)
self.conv4 = layers[3]
self.pool4 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
return_indices=True)
self.conv5 = layers[4]
self.pool5 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
return_indices=True)
self.conv67 = nn.Sequential(layers[5], nn.BatchNorm2d(4096), nn.ReLU(),
layers[6], nn.BatchNorm2d(4096), nn.ReLU())
self.deconv67 = nn.Sequential(
nn.ConvTranspose2d(4096, 512, kernel_size=7, stride=1, padding=0),
nn.BatchNorm2d(512), nn.ReLU())
self.unpool5 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconv5 = nn.Sequential(
nn.ConvTranspose2d(512, 512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(512), nn.ReLU(),
nn.ConvTranspose2d(512, 512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(512), nn.ReLU(),
nn.ConvTranspose2d(512, 512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(512), nn.ReLU())
self.unpool4 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconv4 = nn.Sequential(
nn.ConvTranspose2d(512, 512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(512), nn.ReLU(),
nn.ConvTranspose2d(512, 512, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(512), nn.ReLU(),
nn.ConvTranspose2d(512, 256, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(256), nn.ReLU())
self.unpool3 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconv3 = nn.Sequential(
nn.ConvTranspose2d(256, 256, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(256), nn.ReLU(),
nn.ConvTranspose2d(256, 256, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(256), nn.ReLU(),
nn.ConvTranspose2d(256, 128, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(128), nn.ReLU())
self.unpool2 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconv2 = nn.Sequential(
nn.ConvTranspose2d(128, 128, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(128), nn.ReLU(),
nn.ConvTranspose2d(128, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64), nn.ReLU())
self.unpool1 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconv1 = nn.Sequential(
nn.ConvTranspose2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64), nn.ReLU(),
nn.ConvTranspose2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64), nn.ReLU(),
nn.ConvTranspose2d(64,
num_classes,
kernel_size=1,
stride=1,
padding=0))
if init_weights:
self._initialize_weights()
def __init__(self, encoder_block, decoder_block, layers, num_classes=1000):
self.inplanes = 64
super(ResAE, self).__init__()
self.conv1 = nn.Conv2d(3,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.relu1 = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3,
stride=2,
padding=1,
return_indices=True)
self.layer1 = self._make_encoder_layer(encoder_block, 64, layers[0])
self.fc_drop1 = nn.Dropout(p=0.5)
self.layer2 = self._make_encoder_layer(encoder_block,
128,
layers[1],
stride=2)
self.fc_drop2 = nn.Dropout(p=0.5)
self.layer3 = self._make_encoder_layer(encoder_block,
256,
layers[2],
stride=2)
self.fc_drop3 = nn.Dropout(p=0.5)
self.layer4 = self._make_encoder_layer(encoder_block,
512,
layers[3],
stride=2)
self.avgpool = nn.AvgPool2d(7)
self.fc_drop4 = nn.Dropout(p=0.5)
self.fc = nn.Linear(2048 * 1 * 1, p_transform["n_labels"])
self.fc_drop5 = nn.Dropout(p=0.5)
self.layer5 = self._make_decoder_layer(decoder_block,
512,
layers[3],
outplanes=1024,
stride=2)
self.fc_drop6 = nn.Dropout(p=0.5)
self.layer6 = self._make_decoder_layer(decoder_block,
256,
layers[2],
outplanes=512,
stride=2)
self.fc_drop7 = nn.Dropout(p=0.5)
self.layer7 = self._make_decoder_layer(decoder_block,
128,
layers[1],
outplanes=256,
stride=2)
self.fc_drop8 = nn.Dropout(p=0.5)
self.layer8 = self._make_decoder_layer(decoder_block,
64,
layers[0],
outplanes=64,
stride=1)
self.max_unpool = nn.MaxUnpool2d(kernel_size=3, stride=2, padding=1)
self.deconv1 = nn.ConvTranspose2d(64,
64,
7,
stride=2,
padding=3,
output_padding=1,
bias=False)
self.bn2 = nn.BatchNorm2d(64)
self.relu2 = nn.ReLU(inplace=True)
self.c1_conv = nn.Conv2d(64,
p_transform['channels'],
kernel_size=1,
stride=1,
padding=0,
bias=False)
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def __init__(self, kernel_size, stride):
super(CropMaxUnpool2d, self).__init__()
self.unpool = nn.MaxUnpool2d(kernel_size, stride)
def __init__(self, pooling):
super(StatefulMaxUnpool2d, self).__init__()
self.pooling = pooling
self.unpooling = nn.MaxUnpool2d(pooling.kernel_size, pooling.stride, pooling.padding)
def __init__(self, logsigma=True):
"""
:param logsigma: for uncert weighting
"""
super().__init__()
filter = [64, 128, 256, 512, 1024]
self.class_nb = 13
# define encoder decoder layers
self.encoder_block = nn.ModuleList([self.conv_layer([3, filter[0]])])
self.decoder_block = nn.ModuleList(
[self.conv_layer([filter[0], filter[0]])])
for i in range(4):
self.encoder_block.append(
self.conv_layer([filter[i], filter[i + 1]]))
self.decoder_block.append(
self.conv_layer([filter[i + 1], filter[i]]))
# define convolution layer
self.conv_block_enc = nn.ModuleList(
[self.conv_layer([filter[0], filter[0]])])
self.conv_block_dec = nn.ModuleList(
[self.conv_layer([filter[0], filter[0]])])
for i in range(4):
if i == 0:
self.conv_block_enc.append(
self.conv_layer([filter[i + 1], filter[i + 1]]))
self.conv_block_dec.append(
self.conv_layer([filter[i], filter[i]]))
else:
self.conv_block_enc.append(
nn.Sequential(
self.conv_layer([filter[i + 1], filter[i + 1]]),
self.conv_layer([filter[i + 1], filter[i + 1]])))
self.conv_block_dec.append(
nn.Sequential(self.conv_layer([filter[i], filter[i]]),
self.conv_layer([filter[i], filter[i]])))
# define task specific layers
self.pred_task1 = nn.Sequential(
nn.Conv2d(in_channels=filter[0],
out_channels=filter[0],
kernel_size=3,
padding=1),
nn.Conv2d(in_channels=filter[0],
out_channels=self.class_nb,
kernel_size=1,
padding=0))
self.pred_task2 = nn.Sequential(
nn.Conv2d(in_channels=filter[0],
out_channels=filter[0],
kernel_size=3,
padding=1),
nn.Conv2d(in_channels=filter[0],
out_channels=1,
kernel_size=1,
padding=0))
self.pred_task3 = nn.Sequential(
nn.Conv2d(in_channels=filter[0],
out_channels=filter[0],
kernel_size=3,
padding=1),
nn.Conv2d(in_channels=filter[0],
out_channels=3,
kernel_size=1,
padding=0))
# define pooling and unpooling functions
self.down_sampling = nn.MaxPool2d(kernel_size=2,
stride=2,
return_indices=True)
self.up_sampling = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.logsigma = None
if logsigma:
self.logsigma = nn.Parameter(torch.FloatTensor([-0.5, -0.5, -0.5]))
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.xavier_normal_(m.weight)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.xavier_normal_(m.weight)
nn.init.constant_(m.bias, 0)
def __init__(self,input_c,output_c,P,
Type='downsampling',pool_size=(128,64),
pad=0,ratio=2):
super(Bottleneck,self).__init__()
self.Type=Type
if self.Type=='downsampling':
self.net=nn.ModuleDict({'block1':nn.Sequential(nn.Conv2d(input_c,output_c,2,stride=2),
nn.PReLU(),
nn.BatchNorm2d(output_c)),
'block2':nn.Sequential(nn.Conv2d(output_c,output_c,3,padding=pad),
nn.PReLU(),
nn.BatchNorm2d(output_c)),
'block3':nn.Sequential(nn.Conv2d(output_c,output_c,1),
nn.PReLU(),
nn.BatchNorm2d(output_c)),
'dropout':nn.Dropout2d(p=P),
'Pooling':nn.AdaptiveMaxPool2d(pool_size,return_indices=True)})
elif self.Type.split()[0]=='asymmetric':
self.net=nn.ModuleDict({'block1':nn.Sequential(nn.Conv2d(input_c,input_c//ratio,1),
nn.PReLU(),
nn.BatchNorm2d(input_c//ratio)),
'block2':nn.Sequential(nn.Sequential(nn.Conv2d(input_c//ratio,input_c//ratio,(5,1),padding=(2,0)),
nn.Conv2d(input_c//ratio,input_c//ratio,(1,5),padding=(0,2))),
nn.PReLU(),
nn.BatchNorm2d(input_c//ratio)),
'block3':nn.Sequential(nn.Conv2d(input_c//ratio,output_c,1),
nn.PReLU(),
nn.BatchNorm2d(output_c)),
'dropout':nn.Dropout2d(p=P)})
elif self.Type.split()[0]=='dilated':
self.net=nn.ModuleDict({'block1':nn.Sequential(nn.Conv2d(input_c,input_c//ratio,1),
nn.PReLU(),
nn.BatchNorm2d(input_c//ratio)),
'block2':nn.Sequential(nn.Conv2d(input_c//ratio,input_c//ratio,3,dilation=int(Type.split()[1]),padding=pad),
nn.PReLU(),
nn.BatchNorm2d(input_c//ratio)),
'block3':nn.Sequential(nn.Conv2d(input_c//ratio,output_c,1),
nn.PReLU(),
nn.BatchNorm2d(output_c)),
'dropout':nn.Dropout2d(p=P)})
elif self.Type=='normal':
self.net=nn.ModuleDict({'block1':nn.Sequential(nn.Conv2d(input_c,input_c//ratio,1),
nn.PReLU(),
nn.BatchNorm2d(input_c//ratio)),
'block2':nn.Sequential(nn.Conv2d(input_c//ratio,input_c//ratio,3,padding=pad),
nn.PReLU(),
nn.BatchNorm2d(input_c//ratio)),
'block3':nn.Sequential(nn.Conv2d(input_c//ratio,output_c,1),
nn.PReLU(),
nn.BatchNorm2d(output_c)),
'dropout':nn.Dropout2d(p=P)})
elif self.Type=='upsampling':
self.net=nn.ModuleDict({'block1':nn.Sequential(nn.Conv2d(input_c,input_c//ratio,1),
nn.PReLU(),
nn.BatchNorm2d(input_c//ratio)),
'block2':nn.Sequential(nn.ConvTranspose2d(input_c//ratio,input_c//ratio,2,stride=2,padding=pad),
nn.PReLU(),
nn.BatchNorm2d(input_c//ratio)),
'block3':nn.Sequential(nn.Conv2d(input_c//ratio,output_c,1),
nn.PReLU(),
nn.BatchNorm2d(output_c)),
'block4':nn.Conv2d(input_c,output_c,1),
'unpooling':nn.MaxUnpool2d(2,stride=2),
'dropout':nn.Dropout2d(p=P)})
def __init__(self,
pool_scales,
in_channels,
channels,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
align_corners=False):
super().__init__()
assert isinstance(pool_scales, (list, tuple))
# Pyramid Pooling Module
self.pool_scales = pool_scales
self.in_channels = in_channels
self.channels = channels
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.align_corners = align_corners
self.batch_norm = False
self.ppm = []
for scale in self.pool_scales:
self.ppm.append(
nn.Sequential(
nn.AdaptiveAvgPool2d(scale),
*(ConvModule(self.in_channels,
self.channels,
kernel_size=1,
bias=True,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg).children())))
self.ppm = nn.ModuleList(self.ppm)
# Followed the author's implementation that
# concatenate conv layers described in the supplementary
# material between up operations
self.conv_up1 = nn.Sequential(*(list(
ConvModule(self.in_channels + len(pool_scales) * 256,
self.channels,
padding=1,
kernel_size=3,
bias=True,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg).children()) + list(
ConvModule(self.channels,
self.channels,
padding=1,
bias=True,
kernel_size=3,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg).children())))
self.conv_up2 = nn.Sequential(*(list(
ConvModule(self.channels * 2,
self.channels,
padding=1,
kernel_size=3,
bias=True,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg).children())))
if (self.norm_cfg['type'] == 'BN'):
d_up3 = 128
else:
d_up3 = 64
self.conv_up3 = nn.Sequential(*(list(
ConvModule(self.channels + d_up3,
64,
padding=1,
kernel_size=3,
bias=True,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg).children())))
self.unpool = nn.MaxUnpool2d(2, stride=2)
self.conv_up4 = nn.Sequential(
*(list(
ConvModule(64 + 3 + 3 + 2,
32,
padding=1,
kernel_size=3,
bias=True,
act_cfg=self.act_cfg).children()) + list(
ConvModule(32,
16,
padding=1,
kernel_size=3,
bias=True,
act_cfg=self.act_cfg).children()) +
list(
ConvModule(
16, 7, padding=0, kernel_size=1, bias=True,
act_cfg=None).children())))
def __init__(self, kernel_size=3, padding=1, stride=2):
super().__init__()
self.pool = nn.MaxPool2d(kernel_size=kernel_size, padding=1, stride=stride, return_indices=True)
self.unpool = nn.MaxUnpool2d(kernel_size=kernel_size, padding=padding, stride=stride)
def __init__(self):
super(AutoEncoder, self).__init__()
"""
self.encoder = nn.Sequential(
nn.Conv2d(1, 16, 3, stride=1, padding=1),
nn.BatchNorm2d(16),
nn.ReLU(),
nn.Conv2d(16, 32, 3, stride=1, padding=1),
nn.BatchNorm2d(32),
nn.ReLU(),
nn.MaxPool2d(2, stride=2, return_indices=True),
nn.Conv2d(32, 64, 3, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(),
# nn.Conv2d(64, 64, 3, stride=1, padding=1),
# nn.BatchNorm2d(64),
# nn.ReLU(),
nn.MaxPool2d(2, stride=2, return_indices=True),
# nn.Linear(256*128*87, 1024),
# nn.BatchNorm2d(1024),
# nn.ReLU()
)
self.decoder = nn.Sequential(
nn.MaxUnpool2d(2, stride=2),
# nn.ConvTranspose2d(64, 64, 3, padding=1),
nn.ConvTranspose2d(64, 32, 3, padding=1),
nn.MaxUnpool2d(2, stride=2),
nn.ConvTranspose2d(32, 16, 3, padding=1),
nn.ConvTranspose2d(16, 1, 3, padding=1),
# nn.Sigmoid()
)
"""
self.conv1 = nn.Conv2d(1, 64, 3, stride=1, padding=1)
self.bn1 = nn.BatchNorm2d(64)
self.r1 = nn.ReLU()
self.conv2 = nn.Conv2d(64, 64, 3, stride=1, padding=1)
self.bn2 = nn.BatchNorm2d(64)
self.r2 = nn.ReLU()
self.maxpool1 = nn.MaxPool2d(2, stride=2, return_indices=True)
self.conv3 = nn.Conv2d(64, 128, 3, stride=1, padding=1)
self.bn3 = nn.BatchNorm2d(128)
self.r3 = nn.ReLU()
self.conv4 = nn.Conv2d(128, 128, 3, stride=1, padding=1)
self.bn4 = nn.BatchNorm2d(128)
self.r4 = nn.ReLU()
self.maxpool2 = nn.MaxPool2d(2, stride=2, return_indices=True)
self.conv5 = nn.Conv2d(128, 256, 3, stride=1, padding=1)
self.bn5 = nn.BatchNorm2d(256)
self.r5 = nn.ReLU()
self.conv6 = nn.Conv2d(256, 64, 1, stride=1, padding=0)
self.bn6 = nn.BatchNorm2d(64)
self.r6 = nn.ReLU()
self.conv7 = nn.Conv2d(64, 64, 3, stride=1, padding=1)
self.bn7 = nn.BatchNorm2d(64)
self.r7 = nn.ReLU()
self.conv8 = nn.Conv2d(64, 256, 1, stride=1, padding=0)
self.bn8 = nn.BatchNorm2d(256)
self.r8 = nn.ReLU()
# self.maxpool2 = nn.MaxPool2d(2, stride=2, return_indices=True)
# self.linear1 = nn.Linear(50*128*87, 80)
# self.bn4 = nn.BatchNorm2d(80)
# self.r4 = nn.ReLU()
# self.linear2 = nn.Linear(80, 50*128*87)
# self.r5 = nn.ReLU()
# self.maxunpool2 = nn.MaxUnpool2d(2, stride=2)
self.deconv8 = nn.ConvTranspose2d(256, 64, 1, padding=0)
self.bn9 = nn.BatchNorm2d(64)
self.r9 = nn.ReLU()
self.deconv7 = nn.ConvTranspose2d(64, 64, 3, padding=1)
self.bn10 = nn.BatchNorm2d(64)
self.r10 = nn.ReLU()
self.deconv6 = nn.ConvTranspose2d(64, 256, 1, padding=0)
self.bn11 = nn.BatchNorm2d(256)
self.r11 = nn.ReLU()
self.deconv5 = nn.ConvTranspose2d(256, 128, 1, padding=0)
self.bn12 = nn.BatchNorm2d(128)
self.r12 = nn.ReLU()
self.maxunpool2 = nn.MaxUnpool2d(2, stride=2)
self.deconv4 = nn.ConvTranspose2d(128, 128, 3, padding=1)
self.bn13 = nn.BatchNorm2d(128)
self.r13 = nn.ReLU()
self.deconv3 = nn.ConvTranspose2d(128, 64, 3, padding=1)
self.bn14 = nn.BatchNorm2d(64)
self.r14 = nn.ReLU()
self.maxunpool1 = nn.MaxUnpool2d(2, stride=2)
self.deconv2 = nn.ConvTranspose2d(64, 64, 3, padding=1)
self.bn15 = nn.BatchNorm2d(64)
self.r15 = nn.ReLU()
self.deconv1 = nn.ConvTranspose2d(64, 1, 3, padding=1)
def __init__(self):
super(conv_deconv, self).__init__()
self.convblock1 = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=3, padding=1),
nn.BatchNorm2d(64), #add batchnorm after each convolution,
nn.ReLU(inplace=True),
nn.Conv2d(64, 64, kernel_size=3, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True))
self.maxpool1 = nn.MaxPool2d(kernel_size=2,
stride=2,
return_indices=True)
self.convblock2 = nn.Sequential(
nn.Conv2d(64, 128, kernel_size=3, padding=1), nn.BatchNorm2d(128),
nn.ReLU(inplace=True), nn.Conv2d(128,
128,
kernel_size=3,
padding=1), nn.BatchNorm2d(128),
nn.ReLU(inplace=True))
self.maxpool2 = nn.MaxPool2d(kernel_size=2,
stride=2,
return_indices=True)
self.convblock3 = nn.Sequential(
nn.Conv2d(128, 256, kernel_size=3, padding=1), nn.BatchNorm2d(256),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.BatchNorm2d(256),
nn.ReLU(inplace=True), nn.Conv2d(256,
256,
kernel_size=3,
padding=1), nn.BatchNorm2d(256),
nn.ReLU(inplace=True))
self.maxpool3 = nn.MaxPool2d(kernel_size=2,
stride=2,
return_indices=True)
self.convblock4 = nn.Sequential(
nn.Conv2d(256, 512, kernel_size=3, padding=1), nn.BatchNorm2d(512),
nn.ReLU(inplace=True), nn.Conv2d(512,
512,
kernel_size=3,
padding=1), nn.BatchNorm2d(512),
nn.ReLU(inplace=True), nn.Conv2d(512,
512,
kernel_size=3,
padding=1), nn.BatchNorm2d(512),
nn.ReLU(inplace=True))
self.maxpool4 = nn.MaxPool2d(kernel_size=2,
stride=2,
return_indices=True)
self.convblock5 = nn.Sequential(
nn.Conv2d(512, 512, kernel_size=3, padding=1), nn.BatchNorm2d(512),
nn.ReLU(inplace=True), nn.Conv2d(512,
512,
kernel_size=3,
padding=1), nn.BatchNorm2d(512),
nn.ReLU(inplace=True), nn.Conv2d(512,
512,
kernel_size=3,
padding=1), nn.ReLU(inplace=True),
nn.BatchNorm2d(512))
self.maxpool5 = nn.MaxPool2d(kernel_size=2,
stride=2,
return_indices=True)
#kernel with the same size as image 7 for each convolution we get one value 4096 times in output with 0 padding
self.fc6 = nn.Sequential(
nn.Conv2d(512, 4096, kernel_size=7, padding=0), nn.ReLU(True))
self.fc7 = nn.Sequential(
nn.Conv2d(4096, out_channels=4096,
kernel_size=1), #kernel size 1 to match each pixel
nn.ReLU(True))
self.fc6_deconv = nn.Sequential(
nn.ConvTranspose2d(4096, out_channels=512, kernel_size=7),
nn.ReLU(True))
self.maxunpool5 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconvblock5 = nn.Sequential(
nn.ConvTranspose2d(512, 512, kernel_size=3, padding=1),
nn.BatchNorm2d(512), nn.ReLU(inplace=True),
nn.ConvTranspose2d(512, 512, kernel_size=3, padding=1),
nn.BatchNorm2d(512), nn.ReLU(inplace=True),
nn.ConvTranspose2d(512, 512, kernel_size=3, padding=1),
nn.BatchNorm2d(512), nn.ReLU(inplace=True))
self.maxunpool4 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconvblock4 = nn.Sequential(
nn.ConvTranspose2d(512, 512, kernel_size=3, padding=1),
nn.BatchNorm2d(512), nn.ReLU(inplace=True),
nn.ConvTranspose2d(512, 512, kernel_size=3, padding=1),
nn.BatchNorm2d(512), nn.ReLU(inplace=True),
nn.ConvTranspose2d(512, 256, kernel_size=3, padding=1),
nn.BatchNorm2d(256), nn.ReLU(inplace=True))
self.maxunpool3 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconvblock3 = nn.Sequential(
nn.ConvTranspose2d(256, 256, kernel_size=3, padding=1),
nn.BatchNorm2d(256), nn.ReLU(inplace=True),
nn.ConvTranspose2d(256, 256, kernel_size=3, padding=1),
nn.BatchNorm2d(256), nn.ReLU(inplace=True),
nn.ConvTranspose2d(256, 128, kernel_size=3, padding=1),
nn.BatchNorm2d(128), nn.ReLU(inplace=True))
self.maxunpool2 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconvblock2 = nn.Sequential(
nn.ConvTranspose2d(128, 128, kernel_size=3, padding=1),
nn.BatchNorm2d(128), nn.ReLU(inplace=True),
nn.ConvTranspose2d(128, 64, kernel_size=3, padding=1),
nn.BatchNorm2d(64), nn.ReLU(inplace=True))
self.maxunpool1 = nn.MaxUnpool2d(kernel_size=2, stride=2)
self.deconvblock1 = nn.Sequential(
nn.ConvTranspose2d(64, 64, kernel_size=3, padding=1),
nn.BatchNorm2d(64), nn.ReLU(inplace=True),
nn.ConvTranspose2d(64, 1, kernel_size=3, padding=1),
nn.BatchNorm2d(1), nn.ReLU(inplace=True))
def __init__(self,
input_shape,
block,
layers,
num_classes=1,
dilation=1,
dilation_growth=2,
block_up=None,
layers_up=None,
skip_connection=False,
inplanes=None,
preblock=None):
super(ResNetDeconv, self).__init__()
self.skip_connection = skip_connection
if block_up is None:
block_up = BasicBlockUp if block.__name__ == 'BasicBlock' \
else BottleneckUp
elif (block_up is not BasicBlockUp) and (block_up is not BottleneckUp):
raise ValueError('Invalid upsampling block {}'.format(block_up))
if layers_up is None:
layers_up = layers[::-1]
self.nb_blocks = len(layers_up)
self.base = ResNet(input_shape,
block,
layers,
num_classes=num_classes,
dilation=dilation,
dilation_growth=dilation_growth,
return_indices=True,
return_sizes=True,
skip_connection=skip_connection,
preblock=preblock)
last_layer = [
obj for obj in self.base.modules() if isinstance(obj, nn.Conv2d)
][-1]
self.inplanes = last_layer.out_channels
for i, layer in enumerate(layers_up):
if isinstance(layer, int):
blocks = layer
channels = 2**(6 + (len(layers_up) - 1) - i)
stride = 2 if i < (len(layers_up) - 1) else 1
dilate = dilation / dilation_growth**i
elif isinstance(layer, list) and isinstance(layer[1], dict):
blocks = layer[0]
channels = layer[1].get('depth', 2**(6 + i))
stride = layer[1].get('stride', 2 if i <
(len(layers_up) - 1) else 1)
block_up = layer[1].get('block', block_up)
dilate = layer[1].get('dilation',
dilation / dilation_growth**i)
else:
raise ValueError(
'Elements of {} should either be int '.format(layers) +
'or list of [int, dict]')
outplanes = 64 if i == (len(layers_up) - 1) else None
self.__dict__['_modules']['layer{}'.format(
i + 1)] = self._make_up_layer(block_up,
channels,
blocks,
stride=stride,
outplanes=outplanes).apply(
partial(
self._nostride_dilate,
dilate=dilate))
# this deconv/unpool should be condicioned to conv1 (preblock)
# in ResNet which is now configurable
if preblock is None:
self.unpool = nn.MaxUnpool2d(kernel_size=(3, 3),
stride=(2, 2),
padding=1)
self.deconv = nn.ConvTranspose2d(self.inplanes,
2,
kernel_size=7,
stride=2,
padding=3,
bias=False)
else:
if preblock['pool'] is None:
self.unpool = identity_up
else:
self.unpool = nn.MaxUnpool2d(*preblock['pool'])
self.deconv = nn.ConvTranspose2d(self.inplanes,
2,
*preblock['conv'][1:],
bias=False)
self.bn = nn.BatchNorm2d(2)
self.relu = nn.ReLU(inplace=True)
self.classifier = nn.Conv2d(2, num_classes, kernel_size=1, padding=0)
def __init__(self, nin=1, num_classes=2, dtype=torch.float32):
super().__init__()
self.projecting_factor = 4
self.n_kernels = 16
# Initial
self.conv0 = nn.Conv2d(nin, 15, kernel_size=3, stride=2, padding=1)
self.maxpool0 = nn.MaxPool2d(2, return_indices=True)
# First group
self.bottleNeck1_0 = BottleNeckDownSampling(self.n_kernels,
self.projecting_factor,
self.n_kernels * 4, dtype)
self.bottleNeck1_1 = BottleNeckNormal(self.n_kernels * 4,
self.n_kernels * 4,
self.projecting_factor, 0.01)
self.bottleNeck1_2 = BottleNeckNormal(self.n_kernels * 4,
self.n_kernels * 4,
self.projecting_factor, 0.01)
self.bottleNeck1_3 = BottleNeckNormal(self.n_kernels * 4,
self.n_kernels * 4,
self.projecting_factor, 0.01)
self.bottleNeck1_4 = BottleNeckNormal(self.n_kernels * 4,
self.n_kernels * 4,
self.projecting_factor, 0.01)
# Second group
self.bottleNeck2_0 = BottleNeckDownSampling(self.n_kernels * 4,
self.projecting_factor,
self.n_kernels * 8, dtype)
self.bottleNeck2_1 = BottleNeckNormal(self.n_kernels * 8,
self.n_kernels * 8,
self.projecting_factor, 0.1)
self.bottleNeck2_2 = BottleNeckDownSamplingDilatedConv(
self.n_kernels * 8, self.projecting_factor, self.n_kernels * 8, 2)
self.bottleNeck2_3 = BottleNeckNormal_Asym(self.n_kernels * 8,
self.n_kernels * 8,
self.projecting_factor, 0.1)
self.bottleNeck2_4 = BottleNeckDownSamplingDilatedConv(
self.n_kernels * 8, self.projecting_factor, self.n_kernels * 8, 4)
self.bottleNeck2_5 = BottleNeckNormal(self.n_kernels * 8,
self.n_kernels * 8,
self.projecting_factor, 0.1)
self.bottleNeck2_6 = BottleNeckDownSamplingDilatedConv(
self.n_kernels * 8, self.projecting_factor, self.n_kernels * 8, 8)
self.bottleNeck2_7 = BottleNeckNormal_Asym(self.n_kernels * 8,
self.n_kernels * 8,
self.projecting_factor, 0.1)
self.bottleNeck2_8 = BottleNeckDownSamplingDilatedConv(
self.n_kernels * 8, self.projecting_factor, self.n_kernels * 8, 16)
# Third group
self.bottleNeck3_1 = BottleNeckNormal(self.n_kernels * 8,
self.n_kernels * 8,
self.projecting_factor, 0.1)
self.bottleNeck3_2 = BottleNeckDownSamplingDilatedConv(
self.n_kernels * 8, self.projecting_factor, self.n_kernels * 8, 2)
self.bottleNeck3_3 = BottleNeckNormal_Asym(self.n_kernels * 8,
self.n_kernels * 8,
self.projecting_factor, 0.1)
self.bottleNeck3_4 = BottleNeckDownSamplingDilatedConv(
self.n_kernels * 8, self.projecting_factor, self.n_kernels * 8, 4)
self.bottleNeck3_5 = BottleNeckNormal(self.n_kernels * 8,
self.n_kernels * 8,
self.projecting_factor, 0.1)
self.bottleNeck3_6 = BottleNeckDownSamplingDilatedConv(
self.n_kernels * 8, self.projecting_factor, self.n_kernels * 8, 8)
self.bottleNeck3_7 = BottleNeckNormal_Asym(self.n_kernels * 8,
self.n_kernels * 8,
self.projecting_factor, 0.1)
self.bottleNeck3_8 = BottleNeckDownSamplingDilatedConvLast(
self.n_kernels * 8, self.projecting_factor, self.n_kernels * 4, 16)
# ### Decoding path ####
# Unpooling 1
self.unpool_0 = nn.MaxUnpool2d(2)
self.bottleNeck_Up_1_0 = BottleNeckUpSampling(self.n_kernels * 8,
self.projecting_factor,
self.n_kernels * 4)
self.PReLU_Up_1 = nn.PReLU()
self.bottleNeck_Up_1_1 = BottleNeckNormal(self.n_kernels * 4,
self.n_kernels * 4,
self.projecting_factor, 0.1)
self.bottleNeck_Up_1_2 = BottleNeckNormal(self.n_kernels * 4,
self.n_kernels,
self.projecting_factor, 0.1)
# Unpooling 2
self.unpool_1 = nn.MaxUnpool2d(2)
self.bottleNeck_Up_2_1 = BottleNeckUpSampling(self.n_kernels * 2,
self.projecting_factor,
self.n_kernels)
self.bottleNeck_Up_2_2 = BottleNeckNormal(self.n_kernels,
self.n_kernels,
self.projecting_factor, 0.1)
self.PReLU_Up_2 = nn.PReLU()
# Unpooling Last
self.deconv3 = upSampleConv(self.n_kernels, self.n_kernels)
self.out_025 = nn.Conv2d(self.n_kernels * 8,
num_classes,
kernel_size=3,
stride=1,
padding=1)
self.out_05 = nn.Conv2d(self.n_kernels,
num_classes,
kernel_size=3,
stride=1,
padding=1)
self.final = nn.Conv2d(self.n_kernels, num_classes, kernel_size=1)
def __init__(self, class_nb=13, task='semantic', ignore_index=-1):
super(SegNet, self).__init__()
# initialise network parameters
filter = [64, 128, 256, 512, 512]
self.class_nb = class_nb
self.ignore_index = ignore_index
self.task = task
# define encoder decoder layers
self.encoder_block = nn.ModuleList([self.conv_layer([3, filter[0]])])
self.decoder_block = nn.ModuleList(
[self.conv_layer([filter[0], filter[0]])])
for i in range(4):
self.encoder_block.append(
self.conv_layer([filter[i], filter[i + 1]]))
self.decoder_block.append(
self.conv_layer([filter[i + 1], filter[i]]))
# define convolution layer
self.conv_block_enc = nn.ModuleList(
[self.conv_layer([filter[0], filter[0]])])
self.conv_block_dec = nn.ModuleList(
[self.conv_layer([filter[0], filter[0]])])
for i in range(4):
if i == 0:
self.conv_block_enc.append(
self.conv_layer([filter[i + 1], filter[i + 1]]))
self.conv_block_dec.append(
self.conv_layer([filter[i], filter[i]]))
else:
self.conv_block_enc.append(
nn.Sequential(
self.conv_layer([filter[i + 1], filter[i + 1]]),
self.conv_layer([filter[i + 1], filter[i + 1]])))
self.conv_block_dec.append(
nn.Sequential(self.conv_layer([filter[i], filter[i]]),
self.conv_layer([filter[i], filter[i]])))
self.encoder_att = nn.ModuleList([
nn.ModuleList([self.att_layer([filter[0], filter[0], filter[0]])])
])
self.decoder_att = nn.ModuleList([
nn.ModuleList(
[self.att_layer([2 * filter[0], filter[0], filter[0]])])
])
self.encoder_block_att = nn.ModuleList(
[self.conv_layer([filter[0], filter[1]])])
self.decoder_block_att = nn.ModuleList(
[self.conv_layer([filter[0], filter[0]])])
for j in range(1):
for i in range(4):
self.encoder_att[j].append(
self.att_layer(
[2 * filter[i + 1], filter[i + 1], filter[i + 1]]))
self.decoder_att[j].append(
self.att_layer(
[filter[i + 1] + filter[i], filter[i], filter[i]]))
for i in range(4):
if i < 3:
self.encoder_block_att.append(
self.conv_layer([filter[i + 1], filter[i + 2]]))
self.decoder_block_att.append(
self.conv_layer([filter[i + 1], filter[i]]))
else:
self.encoder_block_att.append(
self.conv_layer([filter[i + 1], filter[i + 1]]))
self.decoder_block_att.append(
self.conv_layer([filter[i + 1], filter[i + 1]]))
if self.task == 'semantic':
self.pred_task = self.conv_layer([filter[0], self.class_nb],
pred=True)
if self.task == 'depth':
self.pred_task = self.conv_layer([filter[0], 1], pred=True)
if self.task == 'normal':
self.pred_task = self.conv_layer([filter[0], 3], pred=True)
# define pooling and unpooling functions
self.down_sampling = nn.MaxPool2d(kernel_size=2,
stride=2,
return_indices=True)
self.up_sampling = nn.MaxUnpool2d(kernel_size=2, stride=2)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.xavier_normal_(m.weight)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.xavier_normal_(m.weight)
nn.init.constant_(m.bias, 0)
def __init__(self, in_ch, out_ch, is_bn=False):
super(SegNetUpx3, self).__init__()
# upsampling and convolution block
self.unpool = nn.MaxUnpool2d(2, 2)
self.block = UNetDownx3(in_ch, out_ch, is_bn)
def __init__(self):
super(EncoderDecoder, self).__init__()
##Encoder
# vgg = models.vgg11_bn(pretrained=True)
# self.features = vgg
# self.features.features[0] = nn.Conv2d(4, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
# weight = torch.zeros([64, 4, 3, 3], dtype=torch.float)
# weight[:, 0:3, :, :] = vgg.features[0].weight
# state_dict = vgg.state_dict()
# state_dict['features.0.weight'] = weight
# self.features.load_state_dict(state_dict)
self.layer0 = nn.Conv2d(4,
64,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1))
self.layer1 = nn.BatchNorm2d(64,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.layer2 = nn.ReLU(inplace=True)
self.layer3 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
dilation=1,
ceil_mode=False,
return_indices=True)
self.layer4 = nn.Conv2d(64,
128,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1))
self.layer5 = nn.BatchNorm2d(128,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.layer6 = nn.ReLU(inplace=True)
self.layer7 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
dilation=1,
ceil_mode=False,
return_indices=True)
self.layer8 = nn.Conv2d(128,
256,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1))
self.layer9 = nn.BatchNorm2d(256,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.layer10 = nn.ReLU(inplace=True)
self.layer11 = nn.Conv2d(256,
256,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1))
self.layer12 = nn.BatchNorm2d(256,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.layer13 = nn.ReLU(inplace=True)
self.layer14 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
dilation=1,
ceil_mode=False,
return_indices=True)
self.layer15 = nn.Conv2d(256,
512,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1))
self.layer16 = nn.BatchNorm2d(512,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.layer17 = nn.ReLU(inplace=True)
self.layer18 = nn.Conv2d(512,
512,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1))
self.layer19 = nn.BatchNorm2d(512,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.layer20 = nn.ReLU(inplace=True)
self.layer21 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
dilation=1,
ceil_mode=False,
return_indices=True)
self.layer22 = nn.Conv2d(512,
512,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1))
self.layer23 = nn.BatchNorm2d(512,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.layer24 = nn.ReLU(inplace=True)
self.layer25 = nn.Conv2d(512,
512,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1))
self.layer26 = nn.BatchNorm2d(512,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.layer27 = nn.ReLU(inplace=True)
self.layer28 = nn.MaxPool2d(kernel_size=2,
stride=2,
padding=0,
dilation=1,
ceil_mode=False,
return_indices=True)
##Decoder
self.deconv6 = nn.Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1))
self.deconvBN6 = nn.BatchNorm2d(512)
self.Unpooling5 = nn.MaxUnpool2d(2, stride=2)
self.deconv5 = nn.Conv2d(512,
512,
kernel_size=(5, 5),
stride=(1, 1),
padding=(2, 2))
self.deconvBN5 = nn.BatchNorm2d(512)
self.Unpooling4 = nn.MaxUnpool2d(2, stride=2)
self.deconv4 = nn.Conv2d(512,
256,
kernel_size=(5, 5),
stride=(1, 1),
padding=(2, 2))
self.deconvBN4 = nn.BatchNorm2d(256)
self.Unpooling3 = nn.MaxUnpool2d(2, stride=2)
self.deconv3 = nn.Conv2d(256,
128,
kernel_size=(5, 5),
stride=(1, 1),
padding=(2, 2))
self.deconvBN3 = nn.BatchNorm2d(128)
self.Unpooling2 = nn.MaxUnpool2d(2, stride=2)
self.deconv2 = nn.Conv2d(128,
64,
kernel_size=(5, 5),
stride=(1, 1),
padding=(2, 2))
self.deconvBN2 = nn.BatchNorm2d(64)
self.Unpooling1 = nn.MaxUnpool2d(2, stride=2)
self.deconv1 = nn.Conv2d(64,
64,
kernel_size=(5, 5),
stride=(1, 1),
padding=(2, 2))
self.deconvBN1 = nn.BatchNorm2d(64)
self.conv_final = nn.Conv2d(64,
1,
kernel_size=(5, 5),
stride=(1, 1),
padding=(2, 2))
self.initialize()
def __init__(self,
in_channels,
out_channels,
internal_ratio=4,
dropout_prob=0,
bias=False,
relu=True):
super().__init__()
# Check in the internal_scale parameter is within the expected range
# [1, channels]
if internal_ratio <= 1 or internal_ratio > in_channels:
raise RuntimeError(
"Value out of range. Expected value in the "
"interval [1, {0}], got internal_scale={1}. ".format(
in_channels, internal_ratio))
internal_channels = in_channels // internal_ratio
if relu:
activation = nn.ReLU
else:
activation = nn.PReLU
# Main branch - max pooling followed by feature map (channels) padding
self.main_conv1 = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=bias),
nn.BatchNorm2d(out_channels))
# Remember that the stride is the same as the kernel_size, just like
# the max pooling layers
self.main_unpool1 = nn.MaxUnpool2d(kernel_size=2)
# Extension branch - 1x1 convolution, followed by a regular, dilated or
# asymmetric convolution, followed by another 1x1 convolution. Number
# of channels is doubled.
# 1x1 projection convolution with stride 1
self.ext_conv1 = nn.Sequential(
nn.Conv2d(in_channels, internal_channels, kernel_size=1,
bias=bias), nn.BatchNorm2d(internal_channels),
activation())
# Transposed convolution
self.ext_tconv1 = nn.ConvTranspose2d(internal_channels,
internal_channels,
kernel_size=2,
stride=2,
bias=bias)
self.ext_tconv1_bnorm = nn.BatchNorm2d(internal_channels)
self.ext_tconv1_activation = activation()
# 1x1 expansion convolution
self.ext_conv2 = nn.Sequential(
nn.Conv2d(internal_channels,
out_channels,
kernel_size=1,
bias=bias), nn.BatchNorm2d(out_channels))
self.ext_regul = nn.Dropout2d(p=dropout_prob)
# PReLU layer to apply after concatenating the branches
self.out_activation = activation()
def __init__(self, out_channels=1, in_channels=3, pretrained=False):
super(SegNet_VGG_GN, self).__init__()
vgg_bn = vgg16_bn(pretrained=pretrained)
encoder = list(vgg_bn.features.children())
# Adjust the input size
if in_channels != 3:
encoder[0] = nn.Conv2d(in_channels,
64,
kernel_size=3,
stride=1,
padding=1)
#
for i in range(len(encoder)):
if isinstance(encoder[i], nn.BatchNorm2d):
encoder[i] = nn.GroupNorm(32, encoder[i].num_features)
# Encoder, VGG without any maxpooling
self.stage1_encoder = nn.Sequential(*encoder[:6])
self.stage2_encoder = nn.Sequential(*encoder[7:13])
self.stage3_encoder = nn.Sequential(*encoder[14:23])
self.stage4_encoder = nn.Sequential(*encoder[24:33])
self.stage5_encoder = nn.Sequential(*encoder[34:-1])
self.pool = nn.MaxPool2d(kernel_size=2, stride=2, return_indices=True)
# Decoder, same as the encoder but reversed, maxpool will not be used
decoder = encoder
decoder = [
i for i in list(reversed(decoder))
if not isinstance(i, nn.MaxPool2d)
]
# Replace the last conv layer
decoder[-1] = nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1)
# When reversing, we also reversed conv->batchN->relu, correct it
decoder = [
item for i in range(0, len(decoder), 3)
for item in decoder[i:i + 3][::-1]
]
# Replace some conv layers & batchN after them
for i, module in enumerate(decoder):
if isinstance(module, nn.Conv2d):
if module.in_channels != module.out_channels:
decoder[i + 1] = nn.GroupNorm(32, module.in_channels)
decoder[i] = nn.Conv2d(module.out_channels,
module.in_channels,
kernel_size=3,
stride=1,
padding=1)
self.stage1_decoder = nn.Sequential(*decoder[0:9])
self.stage2_decoder = nn.Sequential(*decoder[9:18])
self.stage3_decoder = nn.Sequential(*decoder[18:27])
self.stage4_decoder = nn.Sequential(*decoder[27:33])
self.stage5_decoder = nn.Sequential(
*decoder[33:],
nn.Conv2d(64, out_channels, kernel_size=3, stride=1, padding=1))
self.unpool = nn.MaxUnpool2d(kernel_size=2, stride=2)
self._initialize_weights(self.stage1_decoder, self.stage2_decoder,
self.stage3_decoder, self.stage4_decoder,
self.stage5_decoder)