def test_log_logit_sigmoid():
utils.log(torch.tensor(0.))
assert utils.log(torch.tensor(math.e)) - 1. < 1e-2
assert (utils.logit(torch.tensor(0.9)) - 2.197224577 < 1e-2)
assert (utils.logit(torch.tensor(0.2)) + 1.38629436112 < 1e-2)
layer = utils.Logit()
assert (layer(torch.tensor(0.9)) - 2.197224577 < 1e-2)
assert (layer(torch.tensor(0.2)) + 1.38629436112 < 1e-2)
layer = utils.Sigmoid()
assert (layer(torch.tensor(0.5)) - 0.6224593312018545646389 < 1e-2)
assert (layer(torch.tensor(3.)) - 0.9525741268224332191212 < 1e-2)
assert (layer(torch.tensor(-3.)) - 0.04742587317756678087885 < 1e-2)
def __init__(self, probs=[0.5, 0.5], learnable=True):
super().__init__()
self.n_dims = len(probs)
if not isinstance(probs, torch.Tensor):
probs = torch.tensor(probs)
self.logits = log(probs.float())
if learnable:
self.logits = Parameter(self.logits)
def __init__(self, probs=[0.5, 0.5], temperature=1.0,
hard=True, learnable=True):
super().__init__()
self.temperature = temperature
self.hard = hard
if not isinstance(probs, torch.Tensor):
probs = torch.tensor(probs)
self.n_dims = probs.shape
self.logits = log(probs.float())
if learnable:
self.logits = Parameter(self.logits)
def __init__(self, total_count=10, probs=[0.5], learnable=True):
super().__init__()
if not isinstance(probs, torch.Tensor):
total_count = torch.tensor(total_count)
if not isinstance(probs, torch.Tensor):
probs = torch.tensor(probs).view(-1)
self.n_dims = len(probs)
self.total_count = total_count.float()
self.logits = log(probs.float())
if learnable:
self.total_count = Parameter(self.total_count)
self.logits = Parameter(self.logits)
def log_abs_det_jacobian(self, x, y):
return (-(self.alpha.log() + self.beta.log() \
+ (self.alpha - 1) * utils.log(y) + (self.beta - 1) * (-y.pow(self.alpha)).log1p())).sum(-1)
def log_abs_det_jacobian(self, x, y):
return (-(x / self.scale).pow(self.concentration) \
+ (self.concentration * utils.log(x)) + self.concentration.log() \
- self.concentration * self.scale.log()).sum(-1)
def log_prob(self, value):
return (value * (-self.probs).log1p() + log(self.probs)).sum(-1)