def __init__(self, in_ch, out_ch, initialW, sn=True):
super(DiscriminatorBlock, self).__init__()
with self.init_scope():
if sn:
self.c0 = SNConvolution2D(in_ch,
in_ch,
3,
1,
1,
initialW=initialW)
self.c1 = SNConvolution2D(in_ch,
out_ch,
3,
1,
1,
initialW=initialW)
else:
self.c0 = L.Convolution2D(in_ch,
in_ch,
3,
1,
1,
initialW=initialW)
self.c1 = L.Convolution2D(in_ch,
out_ch,
3,
1,
1,
initialW=initialW)
def __init__(self,
in_channels,
out_channels,
ksize=3,
pad=1,
activation=F.relu):
super(OptimizedBlock, self).__init__()
initializer = chainer.initializers.GlorotUniform(math.sqrt(2))
initializer_sc = chainer.initializers.GlorotUniform()
self.activation = activation
with self.init_scope():
self.c1 = SNConvolution2D(in_channels,
out_channels,
ksize=ksize,
pad=pad,
initialW=initializer)
self.c2 = SNConvolution2D(out_channels,
out_channels,
ksize=ksize,
pad=pad,
initialW=initializer)
self.c_sc = SNConvolution2D(in_channels,
out_channels,
ksize=1,
pad=0,
initialW=initializer_sc)
def __init__(self,
in_channels,
out_channels,
hidden_channels=None,
ksize=3,
pad=1,
activation=F.relu,
downsample=False):
super(BlockDisc, self).__init__()
initializer = chainer.initializers.GlorotUniform(math.sqrt(2))
initializer_sc = chainer.initializers.GlorotUniform()
self.activation = activation
self.downsample = downsample
self.learnable_sc = (in_channels != out_channels) or downsample
hidden_channels = in_channels if hidden_channels is None else hidden_channels
with self.init_scope():
self.c1 = SNConvolution2D(in_channels,
hidden_channels,
ksize=ksize,
pad=pad,
initialW=initializer)
self.c2 = SNConvolution2D(hidden_channels,
out_channels,
ksize=ksize,
pad=pad,
initialW=initializer)
if self.learnable_sc:
self.c_sc = SNConvolution2D(in_channels,
out_channels,
ksize=1,
pad=0,
initialW=initializer_sc)
def __init__(self, output_dim=1, wscale=0.02, ch=512, bottom_width=4):
w = chainer.initializers.Normal(wscale)
super(SNMnistDiscriminator, self).__init__()
with self.init_scope():
self.c0 = SNConvolution2D(1,
ch // 4,
ksize=4,
stride=2,
pad=1,
initialW=w)
self.c1 = SNConvolution2D(ch // 4,
ch // 2,
ksize=4,
stride=2,
pad=2,
initialW=w)
self.c2 = SNConvolution2D(ch // 2,
ch,
ksize=4,
stride=2,
pad=1,
initialW=w)
self.lout = SNLinear(bottom_width * bottom_width * ch,
output_dim,
initialW=w)
def __init__(self, ch):
w = chainer.initializers.HeNormal()
super(SNOptimizedResBlock1, self).__init__()
with self.init_scope():
self.c0 = SNConvolution2D(3, ch, 3, 1, 1, initialW=w)
self.c1 = SNConvolution2D(ch, ch, 3, 1, 1, initialW=w)
self.cs = SNConvolution2D(3, ch, 1, initialW=w)
def __init__(self, ch):
w = chainer.initializers.HeNormal()
super(SNDownResBlock2, self).__init__()
with self.init_scope():
self.c0 = SNConvolution2D(ch, ch, 3, 1, 1, initialW=w) # differ from DownResBlock1
self.c1 = SNConvolution2D(ch, ch, 3, 1, 1, initialW=w)
self.cs = SNConvolution2D(ch, ch, 3, 1, 1, initialW=w) # differ from DownResBlock1
def __init__(self, in_ch, out_ch, initialW):
super(SNDiscriminatorBlock, self).__init__()
with self.init_scope():
self.c0 = SNConvolution2D(in_ch, in_ch, 3, 1, 1, initialW=initialW)
self.c1 = SNConvolution2D(in_ch,
out_ch,
3,
1,
1,
initialW=initialW)
def __init__(self, ch=512, wscale=0.02):
w = chainer.initializers.Normal(wscale)
super(SNPGGANDiscriminator, self).__init__()
with self.init_scope():
self.in1 = SNConvolution2D(3, ch // 8, 1, 1, 0, initialW=w)
self.b1 = SNDiscriminatorBlock(ch // 8, ch // 4, initialW=w)
self.b2 = SNDiscriminatorBlock(ch // 4, ch // 2, initialW=w)
self.b3 = SNDiscriminatorBlock(ch // 2, ch, initialW=w)
self.out0 = SNConvolution2D(ch, ch, 3, 1, 1, initialW=w)
self.out1 = SNConvolution2D(ch, ch, 4, 1, 0, initialW=w)
self.out2 = SNLinear(ch, 1, initialW=w)
def __init__(self, ch=512, wscale=0.02, sn=True):
super(Discriminator32, self).__init__()
w = chainer.initializers.Normal(wscale)
with self.init_scope():
if sn:
print('sn')
self.in_ = SNConvolution2D(3, ch // 8, 1, 1, 0, initialW=w)
else:
print('sn_free')
self.in_ = L.Convolution2D(3, ch // 8, 1, 1, 0, initialW=w)
self.b3 = DiscriminatorBlock(ch // 8, ch // 4, initialW=w, sn=sn)
self.b2 = DiscriminatorBlock(ch // 4, ch // 2, initialW=w, sn=sn)
self.b1 = DiscriminatorBlock(ch // 2, ch, initialW=w, sn=sn)
if sn:
self.out0 = SNConvolution2D(ch, ch, 3, 1, 1, initialW=w)
self.out1 = SNConvolution2D(ch, ch, 4, 1, 0, initialW=w)
self.out2 = SNLinear(ch, 1, initialW=w)
else:
self.out0 = L.Convolution2D(ch, ch, 3, 1, 1, initialW=w)
self.out1 = L.Convolution2D(ch, ch, 4, 1, 0, initialW=w)
self.out2 = L.Linear(ch, 1, initialW=w)
def __init__(self, bottom_width=4, ch=512, wscale=0.02, output_dim=1):
w = chainer.initializers.Normal(wscale)
super(SNDCGANDiscriminator, self).__init__()
with self.init_scope():
self.c0_0 = SNConvolution2D(3, ch // 8, 3, 1, 1, initialW=w)
self.c0_1 = SNConvolution2D(ch // 8, ch // 4, 4, 2, 1, initialW=w)
self.c1_0 = SNConvolution2D(ch // 4, ch // 4, 3, 1, 1, initialW=w)
self.c1_1 = SNConvolution2D(ch // 4, ch // 2, 4, 2, 1, initialW=w)
self.c2_0 = SNConvolution2D(ch // 2, ch // 2, 3, 1, 1, initialW=w)
self.c2_1 = SNConvolution2D(ch // 2, ch // 1, 4, 2, 1, initialW=w)
self.c3_0 = SNConvolution2D(ch // 1, ch // 1, 3, 1, 1, initialW=w)
self.l4 = SNLinear(bottom_width * bottom_width * ch,
output_dim,
initialW=w)
