Arg max vs. sample_from_probabilities about tensorflow-rnn-shakespeare HOT 2 CLOSED

Hi

Yes, you are correct, argmax+accuracy is a very bad metric. Fortunately, it is not used anywhere during training other than for display, to give you an indication the something good is happening because some metric is going up. A better metric to use here would be "perplexity".

During inference, "sampling from probabilities" is a critical step. If you just use argmax (top probability) you get a sequence of the most probable words in english. Something like "the a for the is the...". That is not what you want. Sampling from probabilities indeed introduces randomness, which opens up the vocabulary of words and generates much more english-sounding sentences.

An even better approach would be to use an algorithm like "beam search" to generate the most probable sequence of n characters instead of just the most probable next character. For example, after "for" the most probable next char is " " (space) because "for" is such a frequent word. However, based on prior context, the most probable sequence could be "form" or "fortran" or indeed "for_". Beam search takes multiple paths through the sequence of probabilities and selects the most probable sequences. In a way, sampling from probabilities implements a weak form of beam search, one that will consider the most probable sequences by chance: there is a chance it picks "form" rather than "for " even though the most probable char after "for" is " " (space). If " " space had 50% probability and "m" was at 35%, sampling from probabilities can still take the "m".

from tensorflow-rnn-shakespeare.

Thanks for the prompt answer - very helpful!

from tensorflow-rnn-shakespeare.