contrib.bayesflow.entropy.renyi_alpha

tf.contrib.bayesflow.entropy.renyi_alpha

tf.contrib.bayesflow.entropy.renyi_alpha

renyi_alpha(
    step,
    decay_time,
    alpha_min,
    alpha_max=0.99999,
    name='renyi_alpha'
)

Defined in tensorflow/contrib/bayesflow/python/ops/entropy_impl.py.

See the guide: BayesFlow Entropy (contrib) > Ops

Exponentially decaying Tensor appropriate for Renyi ratios.

When minimizing the Renyi divergence for 0 <= alpha < 1 (or maximizing the Renyi equivalent of elbo) in high dimensions, it is not uncommon to experience NaN and inf values when alpha is far from 1.

For that reason, it is often desirable to start the optimization with alpha very close to 1, and reduce it to a final alpha_min according to some schedule. The user may even want to optimize using elbo_ratio for some fixed time before switching to Renyi based methods.

This Op returns an alpha decaying exponentially with step:

s(step) = (exp{step / decay_time} - 1) / (e - 1)
t(s) = max(0, min(s, 1)),  (smooth growth from 0 to 1)
alpha(t) = (1 - t) alpha_min + t alpha_max

Args:

• step: Non-negative scalar Tensor. Typically the global step or an offset version thereof.
• decay_time: Positive scalar Tensor.
• alpha_min: float or double Tensor. The minimal, final value of alpha, achieved when step >= decay_time
• alpha_max: Tensor of same dtype as alpha_min. The maximal, beginning value of alpha, achieved when step == 0
• name: A name to give this Op.

Returns:

• alpha: A Tensor of same dtype as alpha_min.