I would like to adapt the code of PTB Tensorflow code [https://github.com/tensorflow/models/blob/master/tutorials/rnn/ptb/ptb_word_lm.py], in order to calculate top k predicted word samples, on the basis of given input word sequence? I am confused in the answer mentioned here [How Tensorflow text prediction predicts without softmax activation.
Following can be the options:
Option 1: We can use tf.multinomial, which is similar to tf.random.categorical
k=10 self.predictions = tf.multinomial(logits, k)
Option 2: The other option can be to call a specific function, which will output the value coming from the following line, in k number of times.
Which option will give me fair experimental results?