def __init__(self):
super(FaceBox, self).__init__()
#model
self.conv1 = nn.Conv2d(3, 24, kernel_size=7, stride=4, padding=3)
self.conv1 = init_model(self.conv1)
self.bn1 = nn.BatchNorm2d(24)
self.bn1 = init_model(self.bn1)
self.conv2 = nn.Conv2d(48, 64, kernel_size=5, stride=2, padding=2)
self.conv2 = init_model(self.conv2)
self.bn2 = nn.BatchNorm2d(64)
self.bn2 = init_model(self.bn2)
self.inception1 = Inception()
self.inception2 = Inception()
self.inception3 = Inception()
self.conv3_1 = conv_bn_relu(128, 128, kernel_size=1)
self.conv3_2 = conv_bn_relu(128,
256,
kernel_size=3,
stride=2,
padding=1)
self.conv4_1 = conv_bn_relu(256, 128, kernel_size=1)
self.conv4_2 = conv_bn_relu(128,
256,
kernel_size=3,
stride=2,
padding=1)
self.multilbox = MultiBoxLayer()
def __init__(self):
super(SSD300, self).__init__()
# model
self.base = self.VGG16()
self.norm4 = L2Norm2d(512,20)
self.conv5_1 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
self.conv5_2 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
self.conv5_3 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
self.conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
self.conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
self.conv8_1 = nn.Conv2d(1024, 256, kernel_size=1)
self.conv8_2 = nn.Conv2d(256, 512, kernel_size=3, padding=1, stride=2)
self.conv9_1 = nn.Conv2d(512, 128, kernel_size=1)
self.conv9_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2)
self.conv10_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv10_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv11_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv11_2 = nn.Conv2d(128, 256, kernel_size=3)
# multibox layer
self.multibox = MultiBoxLayer()
def __init__(self):
super(MDSSD300, self).__init__()
# Attention
self.attn1 = AttentionBlock(3)
# model
self.base = self.VGG16()
self.norm4 = L2Norm(512, 20) # 38
self.conv5_1 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.conv5_2 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.conv5_3 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.bn5 = nn.BatchNorm2d(512)
self.conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
self.bn6 = nn.BatchNorm2d(1024)
self.conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
self.bn7 = nn.BatchNorm2d(1024)
# self.attn2 = AttentionBlock(1024)
self.conv8_1 = nn.Conv2d(1024, 256, kernel_size=1)
self.conv8_2 = nn.Conv2d(256, 512, kernel_size=3, padding=1, stride=2)
self.bn8 = nn.BatchNorm2d(512)
self.conv9_1 = nn.Conv2d(512, 128, kernel_size=1)
self.conv9_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2)
self.bn9 = nn.BatchNorm2d(256)
self.conv10_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv10_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv11_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv11_2 = nn.Conv2d(128, 256, kernel_size=3)
self.Fusion1 = FusionBlock(256, 512)
self.Fusion2 = FusionBlock(512, 256)
self.Fusion3 = FusionBlock(1024, 256)
# multibox layer
self.multibox = MultiBoxLayer()
def __init__(self):
super(SSD300, self).__init__()
# model
self.base = self.VGG16()
self.conv4_1 = nn.Conv2d(256, 512, kernel_size=3, padding=1, dilation=1)
self.conv4_2 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
self.conv4_3 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
# In simple terms, dilated convolution is just a convolution applied to input with defined gaps.
# With this definitions, given our input is an 2D image, dilation rate k=1 is normal convolution and k=2 means skipping one pixel per input and k=4 means skipping 3 pixels.
self.norm4 = L2Norm2d(20)
# Before Detection Module
self.conv4_3_1 = nn.Conv2d(512, 128, kernel_size=1, padding=1, dilation=1)
self.conv4_3_2 = nn.Conv2d(128, 128, kernel_size=3, padding=1, dilation=1)
# Detection Module
self.conv4_3_3 = nn.Conv2d(128, 128, kernel_size=3, padding=1, dilation=1)
self.conv4_3_3_1 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
# Context Module
self.context1 = nn.Conv2d(128, 64, kernel_size=3, padding=1, dilation=1)
self.context2 = nn.Conv2d(64, 64, kernel_size=3, padding=1, dilation=1)
self.context1_1 = nn.Conv2d(512,256, kernel_size=3, padding=1, dilation=1)
self.context2_1 = nn.Conv2d(256,256, kernel_size=3, padding=1, dilation=1)
self.conv5_1 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
self.conv5_2 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
self.conv5_3 = nn.Conv2d(512, 512, kernel_size=3, padding=1, dilation=1)
self.convd1 = nn.Conv2d(512, 128, kernel_size=1)
self.conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=1, dilation=1)
self.conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
self.conv8_1 = nn.Conv2d(1024, 256, kernel_size=1)
self.conv8_2 = nn.Conv2d(256, 512, kernel_size=3, padding=1, stride=2)
self.conv9_1 = nn.Conv2d(512, 128, kernel_size=1)
self.conv9_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2)
self.conv10_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv10_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv11_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv11_2 = nn.Conv2d(128, 256, kernel_size=3)
# multibox layer
self.multibox = MultiBoxLayer()
def __init__(self, growthRate, reduction): super(DSOD, self).__init__() # stem self.conv1 = StemLayer(3, 64, stride=2) self.conv2 = StemLayer(64, 64, stride=1) self.conv3 = StemLayer(64, 128, stride=1) self.dense1 = self._make_dense(128, growthRate, 6) nChannels = 128+6*growthRate nOutChannels = int(math.floor(nChannels*reduction)) self.trans1 = Transition(nChannels, nOutChannels) nChannels = nOutChannels self.dense2 = self._make_dense(nChannels, growthRate, 8) nChannels += 8*growthRate nOutChannels1 = int(math.floor(nChannels*reduction)) self.trans_wo = Transition_w_o_pooling(nChannels, nOutChannels1) # self.trans2 = Transition(nChannels, nOutChannels) # self.First = self.trans_wo nChannels = nOutChannels1 self.dense3 = self._make_dense(nChannels, growthRate, 8) nChannels += 8*growthRate nOutChannels = int(math.floor(nChannels*reduction)) self.trans_wo1 = Transition_w_o_pooling(nChannels, nOutChannels) nChannels = nOutChannels self.dense4 = self._make_dense(nChannels, growthRate, 8) nChannels += 8*growthRate # nOutChannels = int(math.floor(nChannels*reduction)) nOutChannels = 256 self.trans_wo2 = Transition_w_o_pooling(nChannels, nOutChannels) nChannels = nOutChannels1 nOutChannels = 256 self.conv4 = SingleLayer(nChannels, nOutChannels, kernel_size=1, stride=1, padding=0) # self.second = torch.cat((self.First, self.conv4), 1) # addExtraLyers nChannels = nOutChannels*2 self.third = SingleLayer2(nChannels, 256) nChannels = 256*2 self.forth = SingleLayer2(nChannels, 128) nChannels = 128*2 self.fifith = SingleLayer2(nChannels, 128) nChannels = 128*2 # self.sixth = SingleLayer2(nChannels, 128) self.sixth = LastLayer(nChannels, 128) # multibox layer self.multibox = MultiBoxLayer()
def __init__(self):
super(MDSSD300, self).__init__()
# model
self.base = self.VGG16()
self.norm4 = L2Norm(512, 20) # 64
self.conv5_1 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.conv5_2 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.conv5_3 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
self.conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
self.conv8_1 = nn.Conv2d(1024, 256, kernel_size=1)
self.conv8_2 = nn.Conv2d(256, 512, kernel_size=3, padding=1,
stride=2) #16
self.conv9_1 = nn.Conv2d(512, 128, kernel_size=1)
self.conv9_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1,
stride=2) #8
self.conv10_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv10_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1,
stride=2) #4
self.conv11_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv11_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1,
stride=2) #2
self.conv12_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv12_2 = nn.Conv2d(128, 256, kernel_size=2)
self.Fusion1 = FusionBlock(256, 512)
self.Fusion2 = FusionBlock(512, 256)
self.Fusion3 = FusionBlock(1024, 256)
# multibox layer
self.multibox = MultiBoxLayer()
def __init__(self):
super(CFENet, self).__init__()
# model
self.base = self.VGG16()
self.norm4 = L2Norm(512, 20) # 38
self.conv5_1 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.conv5_2 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.conv5_3 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
self.conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
self.conv8_1 = nn.Conv2d(1024, 256, kernel_size=1)
self.conv8_2 = nn.Conv2d(256, 512, kernel_size=3, padding=1, stride=2)
self.conv9_1 = nn.Conv2d(512, 128, kernel_size=1)
self.conv9_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2)
self.conv10_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv10_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv11_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv11_2 = nn.Conv2d(128, 256, kernel_size=3)
# CFE
self.cfe1 = CFE(512, 3)
self.cfe2 = CFE(512, 3)
self.cfe3 = CFE(512, 3)
self.cfe4 = CFE(512, 3)
# FFB
self.ffb1 = FFB(512, 512)
self.ffb2 = FFB(512, 512)
# multibox layer
self.multibox = MultiBoxLayer()
def __init__(self):
super(SSD300, self).__init__()
self.batch_norm = True
# modelVGG16
self.base = self.make_layers()
#self.base = self.VGG16()
self.norm4 = L2Norm2d(20)
self.conv5_1 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.norm5_1 = nn.BatchNorm2d(512)
self.conv5_2 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.norm5_2 = nn.BatchNorm2d(512)
self.conv5_3 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.norm5_3 = nn.BatchNorm2d(512)
self.conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
self.conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
self.conv8_1 = nn.Conv2d(1024, 256, kernel_size=1)
self.conv8_2 = nn.Conv2d(256, 512, kernel_size=3, padding=1, stride=2)
self.conv9_1 = nn.Conv2d(512, 128, kernel_size=1)
self.conv9_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2)
self.conv10_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv10_1_dp = nn.Dropout2d(p=0.2)
self.conv10_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv10_2_dp = nn.Dropout2d(p=0.2)
self.conv11_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv11_1_dp = nn.Dropout2d(p=0.2)
self.conv11_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv11_2_dp = nn.Dropout2d(p=0.2)
# multibox layer
self.multibox = MultiBoxLayer()
def __init__(self):
super(FaceBox_512, self).__init__()
# model
# self.conv1 = nn.Conv2d(3, 24, kernel_size=5, stride=2, padding=2)
# self.conv2 = nn.Conv2d(48, 64, kernel_size=5, stride=2, padding=2)
self.conv1 = nn.Conv2d(3, 24, kernel_size=3, stride=2, padding=1)
self.conv2 = nn.Conv2d(48, 64, kernel_size=3, stride=2, padding=1)
self.inception1 = Inception()
self.inception2 = Inception()
self.inception3 = Inception()
self.conv3_1 = nn.Conv2d(128, 128, kernel_size=1)
self.conv3_2 = nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1)
self.conv4_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv4_2 = nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1)
self.multilbox = MultiBoxLayer()
def __init__(self):
super(FaceBox_ori, self).__init__()
#model
self.conv1 = nn.Conv2d(3, 24, kernel_size=7, stride=4, padding=3)
self.bn1 = nn.BatchNorm2d(24)
self.conv2 = nn.Conv2d(48, 64, kernel_size=5, stride=2, padding=2)
self.bn2 = nn.BatchNorm2d(64)
self.inception1 = Inception1()
self.inception2 = Inception1()
self.inception3 = Inception1()
# self.conv3_1 = nn.Conv2d(128,128,kernel_size=1)
self.conv3_1 = BasicConv2d(128, 128, 1)
# self.conv3_2 = nn.Conv2d(128,256,kernel_size=3,stride=2,padding=1)
self.conv3_2 = BasicConv2d(128, 256, 3, 2, 1)
# self.conv4_1 = nn.Conv2d(256,128,kernel_size=1)
self.conv4_1 = BasicConv2d(256, 128, 1)
# self.conv4_2 = nn.Conv2d(128,256,kernel_size=3,stride=2,padding=1)
self.conv4_2 = BasicConv2d(128, 256, 3, 2, 1)
self.multilbox = MultiBoxLayer()
def __init__(self):
super(FaceBox, self).__init__()
#model
self.conv1 = nn.Conv2d(3, 24, kernel_size=7, stride=4, padding=3)
self.conv2 = nn.Conv2d(48, 64, kernel_size=5, stride=2, padding=2)
# self.conv1_1 = nn.Conv2d(3,24,kernel_size=3,stride=2,padding=1)
self.conv1_1 = BasicConv2d(3, 24, kernel_size=3, stride=2, padding=1)
# self.conv1_2 = nn.Conv2d(24,24,kernel_size=3,stride=2,padding=1)
self.conv1_2 = BasicConv2d(24, 24, kernel_size=3, stride=2, padding=1)
# self.conv1_3 = nn.Conv2d(24,24,kernel_size=3,stride=1,padding=1)
self.conv1_3 = BasicConv2d(24, 24, kernel_size=3, stride=1, padding=1)
# self.bconv1 = nn.Conv2d(72,48,kernel_size=1)
self.bconv1 = BasicConv2d(72, 48, kernel_size=1)
self.bn1_1 = nn.BatchNorm2d(24, eps=0.001, momentum=0.1, affine=True)
self.bn1_2 = nn.BatchNorm2d(24, eps=0.001, momentum=0.1, affine=True)
self.bn2_1 = nn.BatchNorm2d(64, eps=0.001, momentum=0.1, affine=True)
self.bn2_2 = nn.BatchNorm2d(64, eps=0.001, momentum=0.1, affine=True)
# self.conv2_1 = nn.Conv2d(48,64,kernel_size=3,stride=2,padding=1)
self.conv2_1 = BasicConv2d(48, 64, kernel_size=3, stride=2, padding=1)
# self.conv2_2 = nn.Conv2d(64,64,kernel_size=3,stride=1,padding=1)
self.conv2_2 = BasicConv2d(64, 64, kernel_size=3, stride=1, padding=1)
# self.bconv2 = nn.Conv2d(192,128,kernel_size=1)
self.bconv2 = BasicConv2d(192, 128, kernel_size=1)
# self.inception1 = Inception1()
# self.inception2 = Inception1()
# self.inception3 = Inception1()
self.rfblock1 = BasicRFB_a(128, 128)
self.rfblock2 = BasicRFB_a(128, 128)
self.rfblock3 = BasicRFB_a(128, 128)
self.bconv3 = nn.Conv2d(256, 128, kernel_size=1)
self.bconv4 = nn.Conv2d(384, 128, kernel_size=1)
self.bn3 = nn.BatchNorm2d(256, eps=0.001, momentum=0.1, affine=True)
self.bn4 = nn.BatchNorm2d(384, eps=0.001, momentum=0.1, affine=True)
# self.conv3_1 = nn.Conv2d(128,128,kernel_size=1)
self.conv3_1 = BasicConv2d(128, 128, kernel_size=1, stride=1)
# self.conv3_2 = nn.Conv2d(128,256,kernel_size=3,stride=2,padding=1)
self.conv3_2 = BasicConv2d(128,
256,
kernel_size=3,
stride=2,
padding=1)
# self.conv4_1 = nn.Conv2d(256,128,kernel_size=1)
self.conv4_1 = BasicConv2d(256, 128, kernel_size=1, stride=1)
# self.conv4_2 = nn.Conv2d(128,256,kernel_size=3,stride=2,padding=1)
self.conv4_2 = BasicConv2d(128,
256,
kernel_size=3,
stride=2,
padding=1)
self.multilbox = MultiBoxLayer()
# Official init from torch repo.
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.xavier_uniform(m.weight.data)
# nn.init.kaiming_normal(m.weight.data)
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
m.bias.data.zero_()
def __init__(self):
super(SSD300, self).__init__()
self.batch_norm = True
# modelVGG16
self.base = self.make_layers()
#self.base = self.VGG16()
self.norm4 = L2Norm2d(20)
self.conv5_1 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.norm5_1 = nn.BatchNorm2d(512)
self.conv5_2 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.norm5_2 = nn.BatchNorm2d(512)
self.conv5_3 = nn.Conv2d(512,
512,
kernel_size=3,
padding=1,
dilation=1)
self.norm5_3 = nn.BatchNorm2d(512)
self.conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
self.conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
self.conv8_1 = nn.Conv2d(1024, 256, kernel_size=1)
self.conv8_2 = nn.Conv2d(256, 512, kernel_size=3, padding=1, stride=2)
self.conv9_1 = nn.Conv2d(512, 128, kernel_size=1)
self.conv9_2 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2)
self.conv10_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv10_1_dp = nn.Dropout2d(p=0.2)
if abs(NetworkConfig.input_image_size - 500.) < 1:
self.conv10_2 = nn.Conv2d(128,
256,
kernel_size=3,
padding=1,
stride=2)
self.conv10_2_dp = nn.Dropout2d(p=0.2)
else:
self.conv10_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv10_2_dp = nn.Dropout2d(p=0.2)
self.conv11_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv11_1_dp = nn.Dropout2d(p=0.2)
self.conv11_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv11_2_dp = nn.Dropout2d(p=0.2)
if NetworkConfig.input_image_size - 300. < 2:
pass
elif NetworkConfig.input_image_size - 400. < 2:
self.conv12_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv12_1_dp = nn.Dropout2d(p=0.2)
self.conv12_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv12_2_dp = nn.Dropout2d(p=0.2)
elif NetworkConfig.input_image_size - 500. < 2:
self.conv12_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv12_1_dp = nn.Dropout2d(p=0.2)
self.conv12_2 = nn.Conv2d(128, 256, kernel_size=2)
self.conv12_2_dp = nn.Dropout2d(p=0.2)
elif NetworkConfig.input_image_size - 600. < 2:
self.conv12_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv12_1_dp = nn.Dropout2d(p=0.2)
self.conv12_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv12_2_dp = nn.Dropout2d(p=0.2)
self.conv13_1 = nn.Conv2d(256, 128, kernel_size=1)
self.conv13_1_dp = nn.Dropout2d(p=0.2)
self.conv13_2 = nn.Conv2d(128, 256, kernel_size=3)
self.conv13_2_dp = nn.Dropout2d(p=0.2)
# multibox layer
self.multibox = MultiBoxLayer()
