There were ideas on this in issue #73 which I wanted to start organizing in a separate issue.
Stop when change in reference-set score is less than, say, 0.1% from the previous epoch. (Comment link)
Tweak the MLP tolerance parameter to make training converge faster. (Comment link)
In general, can potentially deploy more conservative values without much testing, while more aggressive values should be tested more carefully.
Note that at this point, train time increases linearly with the number of epochs. Though, if we could cache feature vectors to local filesystem for quicker access in later training epochs, then each epoch after the first could run faster.
There were ideas on this in issue #73 which I wanted to start organizing in a separate issue.
In general, can potentially deploy more conservative values without much testing, while more aggressive values should be tested more carefully.
Note that at this point, train time increases linearly with the number of epochs. Though, if we could cache feature vectors to local filesystem for quicker access in later training epochs, then each epoch after the first could run faster.