sample() with temperature

[RodinIvan]

I've implemented run your network on my data, and seems that it struggles sometimes in generating texts - it just generates the sequence of dots of 'a a a a a a'.
Can you advise, how can I modify sample() function to generate texts with temperature parameter? Seems that it can help.

[alecwangcq]

You can divide outputs (which is the logit) in the sample function of models.py by the desired temperature, the feed it torch.multinomial for sampling.
But if you do greedy decoding, the temperature will not make any differences to the result.