log probability for all classes given an observation about libspn

[jostosh]

Hi, there's this part:

class_roots = spn.ParallelSums(full_scope_prod, num_sums=num_classes)
root = spn.Sum(class_roots)

You can add something along these lines:

# Make another root that doesn't take in latent indicators, thereby always marginalizing over class
marginalizing_root = spn.Sum(class_roots, weights=root.weights)

log_p_x_given_y = class_roots.get_log_value()
log_p_y = marginalizing_root.weights.node.get_log_value()
log_p_x = marginalizing_root.get_log_value()

# Good ol' Bayes rule: p(y|x) == p(x|y)p(y)/p(x)
log_p_y_given_x = (log_p_x_given_y + log_p_y) - log_p_x
p_y_given_x = tf.exp(log_p_y_given_x)

[anicolson]

Thank you! I'll give it a try