def kl_div(input, target, reduction='none', log_target=False):
if log_target:
kl_div = ops.exp(target) * (target - input)
else:
output = target * (ops.log(target) - input)
zeros = ops.zeros_like(input)
kl_div = ops.select(target > 0, output, zeros)
if reduction == 'sum':
return kl_div.sum()
if reduction == 'mean':
return kl_div.mean()
return kl_div
def binary_cross_entropy_with_logits(input, target, weight=None, reduction='mean', pos_weight=None):
max_val = ops.maximum(-input, 0)
if pos_weight is not None:
log_weight = ((pos_weight - 1) * target) + 1
loss = (1 - target) * input
loss_1 = ops.log(ops.exp(-max_val) + ops.exp(-input - max_val)) + max_val
loss += log_weight * loss_1
else:
loss = (1 - target) * input
loss += max_val
loss += ops.log(ops.exp(-max_val) + ops.exp(-input - max_val))
if weight is not None:
output = loss * weight
else:
output = loss
if reduction == "mean":
return ops.reduce_mean(output)
elif reduction == "sum":
return ops.reduce_sum(output)
else:
return output
def get_distribution(self, x_d, dp_mean, dp_std):
return self.normal.log_prob(x_d, dp_mean, ops.exp(dp_std))
def logsigmoid(input):
return ops.log(1 / (1 + ops.exp(-input)))
def celu(input, alpha=1.0):
return ops.maximum(0, input) + ops.minimum(0, alpha * (ops.exp(input / alpha) - 1))
def selu(input):
return 1.0507009873554804934193349852946 * \
(ops.maximum(0, input) + ops.minimum(0, 1.6732632423543772848170429916717 * (ops.exp(input) - 1)))