def __init__(self, in_ch, out_ch):
super(GeneratorBlock, self).__init__()
self.out_ch = out_ch
irelu = chainer.initializers.HeNormal(scale=1.0)
with self.init_scope():
self.c0 = EqualizedConvolution2d(in_ch,
out_ch,
3,
1,
1,
initialW=irelu)
self.c1 = EqualizedConvolution2d(out_ch,
out_ch,
3,
1,
1,
initialW=irelu)
def __init__(self, in_ch, out_ch):
super(DiscriminatorBlock, self).__init__()
#self.pooling_comp = pooling_comp
self.out_ch = out_ch
irelu = chainer.initializers.HeNormal(scale=1.0)
with self.init_scope():
self.c0 = EqualizedConvolution2d(in_ch,
in_ch,
3,
1,
1,
initialW=irelu)
self.c1 = EqualizedConvolution2d(in_ch,
out_ch,
3,
1,
1,
initialW=irelu)
def __init__(self, max_stage=12):
super(Discriminator, self).__init__()
self.max_stage = max_stage
#self.pooling_comp = pooling_comp # compensation of ave_pool is 0.5-Lipshitz
self.R = (16, 32, 64, 128, 128, 128, 128, 128)
irelu = chainer.initializers.HeNormal(scale=1.0)
ilinear = chainer.initializers.HeNormal(scale=1.0 / np.sqrt(2))
with self.init_scope():
self.in7 = EqualizedConvolution2d(3,
self.R[0],
1,
1,
0,
initialW=irelu)
self.b7 = DiscriminatorBlock(self.R[0], self.R[1])
self.in6 = EqualizedConvolution2d(3,
self.R[1],
1,
1,
0,
initialW=irelu)
self.b6 = DiscriminatorBlock(self.R[1], self.R[2])
self.in5 = EqualizedConvolution2d(3,
self.R[2],
1,
1,
0,
initialW=irelu)
self.b5 = DiscriminatorBlock(self.R[2], self.R[3])
self.in4 = EqualizedConvolution2d(3,
self.R[3],
1,
1,
0,
initialW=irelu)
self.b4 = DiscriminatorBlock(self.R[3], self.R[4])
self.in3 = EqualizedConvolution2d(3,
self.R[4],
1,
1,
0,
initialW=irelu)
self.b3 = DiscriminatorBlock(self.R[4], self.R[5])
self.in2 = EqualizedConvolution2d(3,
self.R[5],
1,
1,
0,
initialW=irelu)
self.b2 = DiscriminatorBlock(self.R[5], self.R[6])
self.in1 = EqualizedConvolution2d(3,
self.R[6],
1,
1,
0,
initialW=irelu)
self.b1 = DiscriminatorBlock(self.R[6], self.R[7])
self.in0 = EqualizedConvolution2d(3,
self.R[7],
1,
1,
0,
initialW=irelu)
self.out0 = EqualizedConvolution2d(self.R[7] + 1,
self.R[7],
3,
1,
1,
initialW=irelu)
self.out1 = EqualizedConvolution2d(self.R[7],
self.R[7],
4,
1,
0,
initialW=irelu)
self.out2 = EqualizedLinear(self.R[7], 1, initialW=ilinear)
def __init__(self, n_hidden=512, max_stage=12):
super(Generator, self).__init__()
self.n_hidden = n_hidden
self.R = (128, 128, 128, 128, 128, 64, 32, 16)
self.max_stage = max_stage
irelu = chainer.initializers.HeNormal(scale=1.0)
ilinear = chainer.initializers.HeNormal(scale=1.0 / np.sqrt(2))
with self.init_scope():
self.c0 = EqualizedConvolution2d(n_hidden,
self.R[0],
4,
1,
3,
initialW=irelu)
self.c1 = EqualizedConvolution2d(self.R[0],
self.R[0],
3,
1,
1,
initialW=irelu)
self.out0 = EqualizedConvolution2d(self.R[0],
3,
1,
1,
0,
initialW=ilinear)
self.b1 = GeneratorBlock(self.R[0], self.R[1])
self.out1 = EqualizedConvolution2d(self.R[1],
3,
1,
1,
0,
initialW=ilinear)
self.b2 = GeneratorBlock(self.R[1], self.R[2])
self.out2 = EqualizedConvolution2d(self.R[2],
3,
1,
1,
0,
initialW=ilinear)
self.b3 = GeneratorBlock(self.R[2], self.R[3])
self.out3 = EqualizedConvolution2d(self.R[3],
3,
1,
1,
0,
initialW=ilinear)
self.b4 = GeneratorBlock(self.R[3], self.R[4])
self.out4 = EqualizedConvolution2d(self.R[4],
3,
1,
1,
0,
initialW=ilinear)
self.b5 = GeneratorBlock(self.R[4], self.R[5])
self.out5 = EqualizedConvolution2d(self.R[5],
3,
1,
1,
0,
initialW=ilinear)
self.b6 = GeneratorBlock(self.R[5], self.R[6])
self.out6 = EqualizedConvolution2d(self.R[6],
3,
1,
1,
0,
initialW=ilinear)
self.b7 = GeneratorBlock(self.R[6], self.R[7])
self.out7 = EqualizedConvolution2d(self.R[7],
3,
1,
1,
0,
initialW=ilinear)