def __init__(self,
mean=None,
logit=None,
use_gumbel_softmax=True,
tau=0.1):
super().__init__()
if mean is not None:
self.mean = asarray(mean)
assert ((self.mean.value >= 0).all()
and np.allclose(self.mean.value.sum(axis=-1), 1))
self.logit = log(self.mean)
self._log_pdf = self._log_pdf_mean
elif logit is not None:
self.mean = softmax(logit)
self.logit = asarray(logit)
self._log_pdf = self._log_pdf_logit
else:
raise ValueError
self.n_category = self.mean.shape[-1]
if use_gumbel_softmax:
self.forward = self._forward_gumbel_softmax
else:
self.forward = self._forward
self.tau = tau
self.eye = np.eye(self.n_category)
def _forward(self):
if self.mean.ndim == 1:
index = np.random.choice(self.n_category, p=self.mean.value)
return asarray(self.eye[index])
else:
mean = self.mean.value.reshape(-1, self.n_category)
indices = [np.random.choice(self.n_category, p=p) for p in mean]
onehot = self.eye[np.array(indices)]
return asarray(onehot)
def __init__(self, mean=None, logit=None):
super().__init__()
if mean is not None:
self.mean = asarray(mean)
assert ((self.mean.value >= 0).all()
and (self.mean.value <= 1).all())
self.logit = logit_func(mean)
self._log_pdf = self._log_pdf_mu
elif logit is not None:
self.mean = sigmoid(logit)
self.logit = asarray(logit)
self._log_pdf = self._log_pdf_logit
else:
raise ValueError
def _convert2array(arg):
if not isinstance(arg, Array):
return asarray(arg)
else:
return arg
def truncnormal(min, max, scale, size):
return asarray(truncnorm(a=min, b=max, scale=scale).rvs(size))
def normal(mean, std, size):
return asarray(np.random.normal(mean, std, size))
def forward(self):
binary = (np.random.uniform(size=self.mean.shape) < self.mean.value)
return asarray(binary)
def __init__(self, std, ndim: int):
super().__init__()
self.std = asarray(std)
self.ndim = ndim
assert ((self.std.value >= 0).all())
def __init__(self, mean, std):
super().__init__()
self.mean = asarray(mean)
self.std = asarray(std)
assert ((self.std.value > 0).all())
def uniform(min, max, size):
return asarray(np.random.uniform(min, max, size))
def truncnormal(min, max, scale, size):
#https://www.zhihu.com/question/49923924 截断正态分布
return asarray(truncnorm(a=min, b=max, scale=scale).rvs(size))