I am using RelaxedOneHotCategorical in Edward model, similar to
logits = nn(..)
tau = Gamma(0.1, 0.5)
y_prob = RelaxedOneHotCategorical(tau, logits)
and use KLqp for inference, but my Loss is negative after few iterations. My question is
a) Is RelaxedOneHotCategorical supported in Edward
b) Is my usage correct?
c) Are there any issues with KLqp inference when using such a model?
a) Is RelaxedOneHotCategorical supported in Edward?
Yes. Edward – ed.models.RelaxedOneHotCategorical
b) Is my usage correct?
I think that modeling of temperature parameters may not be appropriate.
What about trying to set the temperature parameter as a constant?
For example,
tau = tf.constant(0.5)
Please see the paper ([1611.01144] Categorical Reparameterization with Gumbel-Softmax).
In the first experiment, they used a fixed \tau=1.
In the second experiment, they anneal the temperature using the schedule
\tau = max(0.5, exp(−rt)) of the global training step t.
It seems that they never modeled the temperature parameter as a random variable.
c) Are there any issues with KLqp inference when using such a model?
Attempting to approximate the Categorial distribution with OneHot Categorical distribution, nan occurred. I am looking for a good way.
I’m happy if I can keep good discussions with you.