def to_dist_fn(h):
loc, scale = h
xp = cuda.get_array_module(loc, scale)
return distributions.HyperbolicWrapped(
distributions.Independent(
D.Normal(loc=xp.zeros(shape=scale.shape,
dtype=scale.dtype),
scale=scale)),
functions.pseudo_polar_projection(loc))
def to_dist_fn(h):
xp = cuda.get_array_module(h)
scale = F.softplus(h[..., n_latent:])
shape = scale.shape[:-1] + (n_latent, )
return distributions.HyperbolicWrapped(
distributions.Independent(
D.Normal(loc=xp.zeros(shape=shape, dtype=scale.dtype),
scale=scale)),
functions.pseudo_polar_projection(h[..., :n_latent]))
def _log_det_jacobian(self, x, y):
if not self.parts:
xp = cuda.get_array_module(x, y)
return xp.zeros_like(x.array)
result = 0
for part in self.parts:
y = part.forward(x)
result = result + _sum_rightmost(part.log_det_jacobian(x, y),
self.event_dim - part.event_dim)
x = y
return result
def _log_det_jacobian(self, x, y):
shape = x.shape
scale = self.scale
if isinstance(scale, numbers.Number):
xp = cuda.get_array_module(x, y)
result = exponential.log(basic_math.absolute(scale)) \
* xp.ones(shape, dtype=x.dtype)
else:
result = exponential.log(basic_math.absolute(scale))
if self.event_dim:
result_size = result.shape[:-self.event_dim] + (-1, )
result = sum_mod.sum(result.view(result_size), axis=-1)
shape = shape[:-self.event_dim]
return broadcast.broadcast_to(result, shape)
def __call__(self, x):
h = F.relu(self.bn1(self.conv1(x)))
h = self.bn2(self.conv2(h))
# 如果residual block的输入输出维度不一致，则对增加的维度用0来填充
if x.data.shape != h.data.shape:
xp = cuda.get_array_module(x.data)
n, c, hh, ww = x.data.shape # 获取输入数据的数量n、通道数c、高度hh、宽度ww
pad_c = h.data.shape[1] - c # 获取需要增加的通道数
p = xp.zeros((n, pad_c, hh, ww), dtype=xp.float32) # 对增加的维度用0填充
p = chainer.Variable(p)
with chainer.no_backprop_mode():
x = F.concat((p, x))
if x.data.shape[2:] != h.data.shape[2:]:
x = F.average_pooling_2d(x, 1, 2) # 平均池化减半
return F.relu(h + x) # 返回残差函数 Relu(h+x)