def __init__(self, num_F=32):
super(feature2d, self).__init__()
self.inplanes = 32
self.F = num_F
self.conv1 = conv2d_bn(3, 32, kernel_size=5, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.block1 = conv_res(32, 32, blocks=8, stride=1)
self.conv2 = nn.Conv2d(32, 32, kernel_size=3, stride=1, padding=1)
def __init__(self, maxdisparity=192):
super(dispnet, self).__init__()
self.name = "dispnet"
#self.D = maxdisparity
self.delt = 1e-6
# 卷积层
self.conv1 = conv2d_bn(6, 64, kernel_size=7, stride=2, bn=flag_bn, activefun=activefun_default)
self.conv2 = conv2d_bn(64, 128, kernel_size=5, stride=2, bn=flag_bn, activefun=activefun_default)
self.conv3a = conv2d_bn(128, 256, kernel_size=5, stride=2, bn=flag_bn, activefun=activefun_default)
self.conv3b = conv2d_bn(256, 256, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
self.conv4a = conv2d_bn(256, 512, kernel_size=3, stride=2, bn=flag_bn, activefun=activefun_default)
self.conv4b = conv2d_bn(512, 512, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
self.conv5a = conv2d_bn(512, 512, kernel_size=3, stride=2, bn=flag_bn, activefun=activefun_default)
self.conv5b = conv2d_bn(512, 512, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
self.conv6a = conv2d_bn(512, 1024, kernel_size=3, stride=2, bn=flag_bn, activefun=activefun_default)
self.conv6b = conv2d_bn(1024, 1024, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
# 解卷积层和视差预测层
self.pr6 = nn.Conv2d(1024, 1, kernel_size=3, stride=1, padding=1)
self.deconv5 = deconv4x4_bn(1024, 512, bn=flag_bn, activefun=activefun_default)
self.iconv5 = conv2d_bn(1025, 512, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
self.pr5 = nn.Conv2d(512, 1, kernel_size=3, stride=1, padding=1)
self.deconv4 = deconv4x4_bn(512, 256, bn=flag_bn, activefun=activefun_default)
self.iconv4 = conv2d_bn(769, 256, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
self.pr4 = nn.Conv2d(256, 1, kernel_size=3, stride=1, padding=1)
self.deconv3 = deconv4x4_bn(256, 128, bn=flag_bn, activefun=activefun_default)
self.iconv3 = conv2d_bn(385, 128, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
self.pr3 = nn.Conv2d(128, 1, kernel_size=3, stride=1, padding=1)
self.deconv2 = deconv4x4_bn(128, 64, bn=flag_bn, activefun=activefun_default)
self.iconv2 = conv2d_bn(193, 64, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
self.pr2 = nn.Conv2d(64, 1, kernel_size=3, stride=1, padding=1)
self.deconv1 = deconv4x4_bn(64, 32, bn=flag_bn, activefun=activefun_default)
self.iconv1 = conv2d_bn(97, 32, kernel_size=3, stride=1, bn=flag_bn, activefun=activefun_default)
self.pr1 = nn.Conv2d(32, 1, kernel_size=3, stride=1, padding=1)
self.upsample = nn.Upsample(scale_factor=2, mode='bilinear')
# 权重初始化
net_init(self)
def __init__(self, maxdisparity=192):
super(dispnet, self).__init__()
self.name = "dispnet"
self.D = maxdisparity
self.delt = 1e-6
self.count_levels = 7 # 分辨率层数
# 上采样（2倍）
self.upsample = nn.Upsample(scale_factor=2, mode='bilinear')
# 卷积层
self.conv1 = conv2d_bn(6, 64, kernel_size=7, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv2 = conv2d_bn(64, 128, kernel_size=5, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv3a = conv2d_bn(128, 256, kernel_size=5, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv3b = conv2d_bn(256, 256, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv4a = conv2d_bn(256, 512, kernel_size=3, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv4b = conv2d_bn(512, 512, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv5a = conv2d_bn(512, 512, kernel_size=3, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv5b = conv2d_bn(512, 512, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv6a = conv2d_bn(512, 1024, kernel_size=3, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.conv6b = conv2d_bn(1024, 1024, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
# 解卷积层和视差预测层
self.pr6 = nn.Conv2d(1024, 1, kernel_size=3, stride=1, padding=1)
self.deconv5 = deconv2d_bn(1024, 512, kernel_size=4, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.iconv5 = conv2d_bn(1025, 512, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.pr5 = nn.Conv2d(512, 1, kernel_size=3, stride=1, padding=1)
self.deconv4 = deconv2d_bn(512, 256, kernel_size=4, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.iconv4 = conv2d_bn(769, 256, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.pr4 = nn.Conv2d(256, 1, kernel_size=3, stride=1, padding=1)
self.deconv3 = deconv2d_bn(256, 128, kernel_size=4, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.iconv3 = conv2d_bn(385, 128, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.pr3 = nn.Conv2d(128, 1, kernel_size=3, stride=1, padding=1)
self.deconv2 = deconv2d_bn(128, 64, kernel_size=4, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.iconv2 = conv2d_bn(193, 64, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.pr2 = nn.Conv2d(64, 1, kernel_size=3, stride=1, padding=1)
self.deconv1 = deconv2d_bn(64, 32, kernel_size=4, stride=2, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.iconv1 = conv2d_bn(97, 32, kernel_size=3, stride=1, flag_bias=flag_bias_t, bn=flag_bn, activefun=activefun_t)
self.pr1 = nn.Conv2d(32, 1, kernel_size=3, stride=1, padding=1)
# 权重初始化
net_init(self)
for m in [self.pr6, self.pr5, self.pr4, self.pr3, self.pr2, self.pr1]:
m.weight.data = m.weight.data*0.1
def __init__(self, maxdisparity=192):
super(iresnet, self).__init__()
self.name = "iresnet"
self.D = maxdisparity
self.delt = 1e-6
self.count_levels = 7 # 分辨率层数
self.upsample = nn.Upsample(scale_factor=2, mode='bilinear')
# Stem Block for Multi-scale Shared Features Extraction
self.conv1 = conv2d_bn(3,
64,
kernel_size=7,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv2 = conv2d_bn(64,
128,
kernel_size=5,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.deconv1_s = deconv2d_bn(64,
32,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.deconv2_s = deconv2d_bn(128,
32,
kernel_size=8,
stride=4,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=4
self.conv_de1_de2 = conv2d_bn(64,
32,
kernel_size=1,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
# Initial Disparity Estimation Sub-network
self.corr = Corr1d(kernel_size=1, stride=1, D=81, simfun=None)
self.redir = conv2d_bn(128,
64,
kernel_size=1,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv3 = conv2d_bn(81 + 64,
256,
kernel_size=3,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv3_1 = conv2d_bn(256,
256,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv4 = conv2d_bn(256,
512,
kernel_size=3,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv4_1 = conv2d_bn(512,
512,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv5 = conv2d_bn(512,
512,
kernel_size=3,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv5_1 = conv2d_bn(512,
512,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv6 = conv2d_bn(512,
1024,
kernel_size=3,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.conv6_1 = conv2d_bn(1024,
1024,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.pr6 = nn.Conv2d(1024, 1, kernel_size=3, stride=1, padding=1)
self.deconv5 = deconv2d_bn(1024,
512,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.iconv5 = conv2d_bn(1025,
512,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.pr5 = nn.Conv2d(512, 1, kernel_size=3, stride=1, padding=1)
self.deconv4 = deconv2d_bn(512,
256,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.iconv4 = conv2d_bn(769,
256,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.pr4 = nn.Conv2d(256, 1, kernel_size=3, stride=1, padding=1)
self.deconv3 = deconv2d_bn(256,
128,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.iconv3 = conv2d_bn(385,
128,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.pr3 = nn.Conv2d(128, 1, kernel_size=3, stride=1, padding=1)
self.deconv2 = deconv2d_bn(128,
64,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.iconv2 = conv2d_bn(193,
64,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.pr2 = nn.Conv2d(64, 1, kernel_size=3, stride=1, padding=1)
self.deconv1 = deconv2d_bn(64,
32,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.iconv1 = conv2d_bn(97,
32,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.pr1 = nn.Conv2d(32, 1, kernel_size=3, stride=1, padding=1)
self.deconv0 = deconv2d_bn(32,
32,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.iconv0 = conv2d_bn(65,
32,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.pr0 = nn.Conv2d(32, 1, kernel_size=3, stride=1, padding=1)
# Disparity Refinement Sub-network
#imwrap up_conv1b2b
self.r_conv0 = conv2d_bn(65,
32,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.r_conv1 = conv2d_bn(32,
64,
kernel_size=3,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.c_conv1 = conv2d_bn(64,
64,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.r_corr = Corr1d(kernel_size=3, stride=2, D=41, simfun=None)
self.r_conv1_1 = conv2d_bn(105,
64,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.r_conv2 = conv2d_bn(64,
128,
kernel_size=3,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.r_conv2_1 = conv2d_bn(128,
128,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.r_res2 = nn.Conv2d(128, 1, kernel_size=3, stride=1, padding=1)
self.r_deconv1 = deconv2d_bn(128,
64,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.r_iconv1 = conv2d_bn(129,
64,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.r_res1 = nn.Conv2d(64, 1, kernel_size=3, stride=1, padding=1)
self.r_deconv0 = deconv2d_bn(64,
32,
kernel_size=4,
stride=2,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t) # stride=2
self.r_iconv0 = conv2d_bn(65,
32,
kernel_size=3,
stride=1,
flag_bias=flag_bias_t,
bn=flag_bn,
activefun=activefun_t)
self.r_res0 = nn.Conv2d(32, 1, kernel_size=3, stride=1, padding=1)
# init weight
net_init(self)
for m in [
self.pr6, self.pr5, self.pr4, self.pr3, self.pr2, self.pr1,
self.r_res2, self.r_res1, self.r_res0
]:
m.weight.data = m.weight.data * 0.1