def __init__(self, deviations, exponent, weights=None):
super(_WeightedPower, self).__init__()
self.deviations = util.to_tensor(deviations)
if weights is None:
weights = 1.0 / deviations.pow(2)
self.exponent = util.to_tensor(exponent, device=self.deviations.device)
self.weights = util.to_tensor(weights, device=self.deviations.device)
def __init__(self, weights, threshold=0.5):
if threshold < 0:
raise ValueError("Threshold must be nonnegative, received ",
threshold)
super(Huber, self).__init__()
self.weights = util.to_tensor(weights)
self.threshold = threshold
def __init__(self, weights, exponent=1):
if not (weights <= 0).all():
raise ValueError("Weights must be negative.")
super(InvPower, self).__init__()
self.weights = util.to_tensor(weights)
if not isinstance(exponent, torch.Tensor):
exponent = torch.tensor(exponent)
self.exponent = exponent
def test_to_tensor():
args = [0.0, np.array(1.0), torch.tensor(2.0)]
tensors = util.to_tensor(args)
for t, arg in zip(tensors, args):
assert isinstance(t, torch.Tensor)
assert t.item() == float(arg)
assert str(t.device) == "cpu"
assert id(tensors[-1]) == id(args[-1])
def __init__(self, weights, threshold, alpha=1.0):
if threshold < 0:
raise ValueError("Threshold must be nonnegative, received ",
threshold)
super(Sigmoid, self).__init__()
self.weights = util.to_tensor(weights)
self.threshold = threshold
self.alpha = alpha
def __init__(self, weights, threshold, sigma=None):
if threshold < 0:
raise ValueError("Threshold must be nonnegative, received ",
threshold)
if sigma is None:
sigma = threshold / 2
super(Hinge, self).__init__()
self.weights = util.to_tensor(weights)
self.threshold = threshold
self.sigma = sigma
def __init__(self,
weights,
attractive_penalty=Log1p,
repulsive_penalty=LogRatio):
super(PushAndPull, self).__init__()
self.weights = util.to_tensor(weights)
if weights.nelement() == 1:
raise ValueError("`PushAndPull` requires at least two weights.")
self.pos_idx = weights >= 0
self.attractive_penalty = attractive_penalty(weights[self.pos_idx])
self.repulsive_penalty = repulsive_penalty(weights[~self.pos_idx])
def __init__(self, deviations):
super(Quadratic, self).__init__()
self.deviations = util.to_tensor(deviations)
def __init__(self, deviations, exponent):
super(_Log1p, self).__init__()
self.deviations = util.to_tensor(deviations)
self.exponent = util.to_tensor(exponent, device=self.deviations.device)
def __init__(self, deviations, gamma=10.0):
super(SoftFractional, self).__init__()
self.deviations = util.to_tensor(deviations)
self.gamma = util.to_tensor(gamma, device=self.deviations.device)
if gamma <= 0.0:
raise ValueError("gamma must be positive, received ", float(gamma))
def __init__(self, deviations):
super(Fractional, self).__init__()
self.deviations = util.to_tensor(deviations)
def __init__(self, deviations):
super(Absolute, self).__init__()
self.deviations = util.to_tensor(deviations)
def __init__(self, weights, threshold):
super(_ClippedQuadratic, self).__init__()
self.weights = util.to_tensor(weights)
self.threshold = threshold
def __init__(self, deviations, threshold):
super(Huber, self).__init__()
self.deviations = util.to_tensor(deviations)
self.threshold = threshold
def __init__(self, weights):
super(Linear, self).__init__()
self.weights = util.to_tensor(weights)
def __init__(self, deviations, weights=None):
super(WeightedQuadratic, self).__init__()
self.deviations = util.to_tensor(deviations)
if weights is None:
weights = 1.0 / deviations.pow(2)
self.weights = util.to_tensor(weights, device=self.deviations.device)
def __init__(self, weights, exponent=1.0):
super(Log, self).__init__()
self.weights = util.to_tensor(weights)
if not isinstance(exponent, torch.Tensor):
exponent = torch.tensor(exponent, device=weights.device)
self.exponent = exponent
def __init__(self, weights):
super(Cubic, self).__init__()
self.weights = util.to_tensor(weights)
def __init__(self, weights, threshold):
super(_DeadzoneCubic, self).__init__()
self.weights = util.to_tensor(weights)
self.threshold = threshold
def __init__(self, deviations, threshold):
super(_ClippedQuadratic, self).__init__()
self.deviations = util.to_tensor(deviations)
self.threshold = threshold
def __init__(self, weights, exponent=2):
super(LogRatio, self).__init__()
self.weights = util.to_tensor(weights)
if not isinstance(exponent, torch.Tensor):
exponent = torch.tensor(exponent)
self.exponent = exponent
