class MVN(Chain):
def __init__(self, mu, sigma, x0):
super(MVN, self).__init__()
self.mu = mu
self.sigma = sigma
self.n_particle = x0.shape[0]
with self.init_scope():
self.theta = Parameter(initializer=x0)
def logp(self):
d = self.mu.shape[0]
mean = np.broadcast_to(self.mu, (self.n_particle, d))
ln_var = np.broadcast_to(2 * np.log(self.sigma), (self.n_particle, d))
logp = -F.gaussian_nll(self.theta, mean, ln_var, reduce='no')
logp = F.sum(logp, axis=1).reshape(-1)
logp = F.broadcast_to(logp, (self.n_particle, self.n_particle))
return logp
def __call__(self):
ker = rbf(self.theta.reshape(self.n_particle, -1))
nlogp = -self.logp()
loss = F.mean(F.sum(ker.data * nlogp + ker, axis=1))
chainer.report(
{
'loss': loss,
'nlogp': F.mean(nlogp[0]),
},
observer=self,
)
return loss
class InitialSkipArchitecture(Chain):
def __init__(self, size, in_channels, out_channels):
super().__init__()
with self.init_scope():
self.c1 = Parameter(shape=(in_channels, 4, 4),
initializer=Normal(1.0))
self.s1 = StyleAffineTransformation(size, in_channels)
self.w1 = WeightModulatedConvolution(in_channels, out_channels)
self.n1 = NoiseAdder()
self.a1 = LeakyRelu()
self.s2 = StyleAffineTransformation(size, out_channels)
self.trgb = ToRGB(out_channels)
def __call__(self, w):
batch = w.shape[0]
h1 = self.c1.reshape(1, *self.c1.shape)
h2 = broadcast_to(h1, (batch, *self.c1.shape))
h3 = self.w1(h2, self.s1(w))
h4 = self.n1(h3)
h5 = self.a1(h4)
return h5, self.trgb(h5, self.s2(w))
class WeightModulatedConvolution(Link):
def __init__(self,
in_channels,
out_channels,
pointwise=False,
demod=True,
gain=root(2)):
super().__init__()
self.demod = demod
self.ksize = 1 if pointwise else 3
self.pad = 0 if pointwise else 1
self.c = gain * root(1 / (in_channels * self.ksize**2))
with self.init_scope():
self.w = Parameter(shape=(out_channels, in_channels, self.ksize,
self.ksize),
initializer=Normal(1.0))
self.b = Parameter(shape=out_channels, initializer=Zero())
def __call__(self, x, y):
out_channels = self.b.shape[0]
batch, in_channels, height, width = x.shape
modulated_w = self.w * y.reshape(batch, 1, in_channels, 1, 1)
w = modulated_w / sqrt(
sum(modulated_w**2, axis=(2, 3, 4), keepdims=True) +
1e-08) if self.demod else modulated_w
grouped_w = w.reshape(batch * out_channels, in_channels, self.ksize,
self.ksize)
grouped_x = x.reshape(1, batch * in_channels, height, width)
padded_grouped_x = pad(grouped_x,
((0, 0), (0, 0), (self.pad, self.pad),
(self.pad, self.pad)),
mode="edge")
h = convolution_2d(padded_grouped_x,
grouped_w,
stride=1,
pad=0,
groups=batch)
return h.reshape(batch, out_channels, height, width) + self.b.reshape(
1, out_channels, 1, 1)