1
$\begingroup$

TensorFlow's Magenta melody generation module has a parameter for branch_factor1. What is it?

$\endgroup$
1
$\begingroup$

Branch factor in beam search refers to average number of branches at each level of a search tree. When predicting a sequence in an LSTM model, such as note selection in melody generation, $b$ branches will be decoded. Increasing the branch factor from 1 to 4 in the call to melody_rnn_generate will maintain the average number of branches to be at $b$.

Specifically for Magenta, it generates 4 branches for each event sequence. It will increase the accuracy of the prediction at the expense of time, and in melody creation at the risk of getting stuck in a local optimum. For example, increasing the branch factor from 1 to 2 decreased the average negative log-likelihood of a melody generation implementation from 36 to 8, but also made the sequence prediction more dependent on the primer sequence (repeating the primer note).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.