Weighted random tensor select in tensorflow -

i have list of tensors , list representing probability mass function. how can each session run tell tensorflow randomly pick 1 tensor according probability mass function.

i see few possible ways that:

one packing list of tensors in rank 1 higher, , select 1 slice & squeeze based on tensorflow variable i'm going assign correct index. performance penalty approach? tensorflow evaluate other, non-needed tensors?

another using tf.case in similar fashion before me picking 1 tensor out of many. same question -> what's performance penalty since plan on having quite few(~100s) conditional statements per 1 graph run.

is there better way of doing this?

i think should use tf.multinomial(logits, num_samples).

say have:

• a batch of tensors of shape [batch_size, num_features]
• a probability distribution of shape [batch_size]

you want output:

• 1 example batch of tensors, of shape [1, num_features]

---

batch_tensors = tf.constant([[0., 1., 2.], [3., 4., 5.]])  # shape [batch_size, num_features] probabilities = tf.constant([0.7, 0.3])  # shape [batch_size]  # need convert probabilities log_probabilities , reshape [1, batch_size] rescaled_probas = tf.expand_dims(tf.log(probabilities), 0)  # shape [1, batch_size]  # can draw 1 example distribution (we draw more) indice = tf.multinomial(rescaled_probas, num_samples=1)  output = tf.gather(batch_tensors, tf.squeeze(indice, [0])) 

---

what's performance penalty since plan on having quite few(~100s) conditional statements per 1 graph run?

if want multiple draws, should in 1 run increasing parameter num_samples. can gather these num_samples examples in 1 run tf.gather.