def __init__(self,
delta_v=0.5,
delta_d=1.5,
max_embedding_dim=10,
norm=1,
alpha=1.0,
beta=1.0,
gamma=0.001):
self.delta_v = delta_v
self.delta_d = delta_d
self.alpha = alpha
self.beta = beta
self.gamma = gamma
self.max_embedding_dim = max_embedding_dim
if self.max_embedding_dim <= 0:
raise ValueError("Max number of embeddings has to be positive!")
# L1 or L2 norm is allowed only
if norm == 1:
self.norm = lambda x, axis=None: c_sum(absolute(x), axis=axis)
elif norm == 2:
self.norm = lambda x, axis=None: sqrt(c_sum(x**2, axis=axis))
else:
raise ValueError("For discriminative loss, "
"norm can only be 1 or 2. "
"Obtained the value : {}".format(norm))
def _normalize(self, x):
size = x.shape[1]
mean = broadcast.broadcast_to(
(sum.sum(x, axis=1) / size)[:, None],
x.shape)
std = broadcast.broadcast_to(sqrt.sqrt(
sum.sum(square.square(x - mean), axis=1) / size)[:, None],
x.shape) + self.eps
return (x - mean) / std
def __init__(self, delta_v=0.5, delta_d=1.5,
max_embedding_dim=10, norm=1,
alpha=1.0, beta=1.0, gamma=0.001):
self.delta_v = delta_v
self.delta_d = delta_d
self.alpha = alpha
self.beta = beta
self.gamma = gamma
self.max_embedding_dim = max_embedding_dim
if self.max_embedding_dim <= 0:
raise ValueError("Max number of embeddings has to be positive!")
# L1 or L2 norm is allowed only
if norm == 1:
self.norm = lambda x, axis=None: c_sum(absolute(x), axis=axis)
elif norm == 2:
self.norm = lambda x, axis=None: sqrt(c_sum(x ** 2, axis=axis))
else:
raise ValueError("For discriminative loss, "
"norm can only be 1 or 2. "
"Obtained the value : {}".format(norm))
def stddev(self):
return sqrt.sqrt(self.variance)