def __init__(self, opt, input_ch, input_resolution=65):
super().__init__()
#get options
nf = opt.aspp_nf
#this rate is dilate size based original paper.
x65_rate = [6, 12, 18]
rate = [round(x * input_resolution / 65) for x in x65_rate]
he_w = HeNormal()
with self.init_scope():
self.x1 = define_conv(opt)(input_ch, nf, ksize=1, initialW=he_w)
self.x1_bn = L.BatchNormalization(nf)
self.x3_small = define_atrous_conv(opt)(input_ch, nf, ksize=3, rate=rate[0], initialW=he_w)
self.x3_small_bn = L.BatchNormalization(nf)
self.x3_middle = define_atrous_conv(opt)(input_ch, nf, ksize=3, rate=rate[1], initialW=he_w)
self.x3_middle_bn = L.BatchNormalization(nf)
self.x3_large = define_atrous_conv(opt)(input_ch, nf, ksize=3, rate=rate[2], initialW=he_w)
self.x3_large_bn = L.BatchNormalization(nf)
self.sum_func = define_conv(opt)(nf * 4, input_ch, ksize=3, pad=1, initialW=he_w)
self.activation = F.relu
def __init__(self, opt):
super().__init__()
he_w = HeNormal()
with self.init_scope():
# This ResNet101 use a pre-trained caffemodel that can be downloaded at GitHub
# <https://github.com/KaimingHe/deep-residual-networks>.
self.resnet101 = L.ResNet101Layers()
self.use_layer = ('res3', 512)
nf = self.use_layer[1]
self.c1 = define_atrous_conv(opt)(nf,
nf,
ksize=3,
rate=2,
initialW=he_w)
self.norm1 = L.BatchNormalization(nf)
self.c2 = define_atrous_conv(opt)(nf,
nf,
ksize=3,
rate=4,
initialW=he_w)
self.norm2 = L.BatchNormalization(nf)
self.aspp = ASPP(opt, nf, input_resolution=32)
self.up1 = PixelShuffler(opt, nf, nf // 2, rate=2) #32 -> 64
self.up2 = PixelShuffler(opt, nf // 2, nf // 4, rate=2) #64 -> 128
self.up3 = PixelShuffler(opt, nf // 4, nf // 8,
rate=2) # 128 -> 256
self.to_class = define_conv(opt)(nf // 8,
opt.class_num,
ksize=3,
pad=1,
initialW=he_w)
self.activation = F.leaky_relu
def __init__(self, opt):
super().__init__()
he_w = HeNormal()
down_sampling_num = 3
ngf = opt.ngf
with self.init_scope():
#[input] 3 x 256 x 256
self.c1 = define_conv(opt)(opt.img_shape[0],
ngf,
ksize=4,
stride=2,
pad=1,
initialW=he_w)
self.norm1 = L.BatchNormalization(ngf)
#[input] ngf x 128 x 128
self.c2 = define_conv(opt)(ngf,
ngf * 2,
ksize=4,
stride=2,
pad=1,
initialW=he_w)
self.norm2 = L.BatchNormalization(ngf * 2)
#[input] ngf*2 x 64 x 64
self.c3 = define_conv(opt)(ngf * 2,
ngf * 4,
ksize=4,
stride=2,
pad=1,
initialW=he_w)
self.norm3 = L.BatchNormalization(ngf * 4)
#[input] ngf*4 x 32 x 32
self.a1 = define_atrous_conv(opt)(ngf * 4,
ngf * 4,
ksize=3,
rate=2,
initialW=he_w)
self.norm4 = L.BatchNormalization(ngf * 4)
#[input] ngf*4 x 32 x 32
self.a2 = define_atrous_conv(opt)(ngf * 4,
ngf * 4,
ksize=3,
rate=4,
initialW=he_w)
self.norm5 = L.BatchNormalization(ngf * 4)
#[input] ngf*4 x 32 x 32
resolution = max(opt.img_shape[1],
opt.img_shape[2]) // 2**down_sampling_num
self.aspp = ASPP(opt, ngf * 4, input_resolution=resolution)
#[input] ngf*4 x 32 x 32
self.up1 = PixelShuffler(opt, ngf * 4, ngf * 2, rate=2) #64 -> 128
self.up2 = PixelShuffler(opt, ngf * 2, ngf, rate=2) # 128 -> 256
self.to_class = define_conv(opt)(ngf,
opt.class_num,
ksize=3,
pad=1,
initialW=he_w)
#[output] class_num x 256 x 256
self.activation = F.relu