def __init__(self,
N,
K,
D,
M=0,
single_subspace=True,
link="log",
bin_size=1.0,
**kwargs):
super(_PoissonEmissionsMixin,
self).__init__(N,
K,
D,
M,
single_subspace=single_subspace,
**kwargs)
self.link_name = link
self.bin_size = bin_size
mean_functions = dict(log=lambda x: np.exp(x) * self.bin_size,
softplus=lambda x: softplus(x) * self.bin_size)
self.mean = mean_functions[link]
link_functions = dict(
log=lambda rate: np.log(rate) - np.log(self.bin_size),
softplus=lambda rate: inv_softplus(rate / self.bin_size))
self.link = link_functions[link]
# Set the bias to be small if using log link
if link == "log":
self.ds = -3 + .5 * npr.randn(
1, N) if single_subspace else npr.randn(K, N)
def __init__(self, K, D, M, link="softplus", bin_size=1.0, **kwargs):
super(PoissonTransitions, self).__init__(K, D, M)
# uniform transitions
Ps = np.ones((K, K))
Ps /= Ps.sum(axis=1, keepdims=True)
self.log_Ps = np.log(Ps)
self.link_name = link
self.bin_size = bin_size
mean_functions = dict(log=lambda x: np.exp(x) * self.bin_size,
softplus=lambda x: softplus(x) * self.bin_size)
self.mean = mean_functions[link]
link_functions = dict(
log=lambda rate: np.log(rate) - np.log(self.bin_size),
softplus=lambda rate: inv_softplus(rate / self.bin_size))
self.link = link_functions[link]
# parameters are a vector of weights
# no other transition parameters
self.Ws = npr.randn(M)
def _softplus_mean(self, x):
return softplus(x) * self.bin_size