Closed rizar closed 8 years ago
Heads-up: adapted reverse_words
in my branch of Blocks-examples is already functional! In this process of making it work I have found that having a Fork
in Feedback
, even when it seems to be not necessary, can significantly speed up training. Namely, if was using SimpleRecurrent
, which has only one input, so forking was technically redundant. But the bias vector, shared between the embedding seems to be helpful.
To prevent confusions, I made Fork
mandatory for now. Making it optional later will not be difficult.
I have learnt that some people are already trying to use it. Given that this is Blocks-extras, I will merge this PR, hoping than we will finish it in the course of next months and move to Blocks.
This first version of the code that works, see an adapted Markov chain example here.
Also see https://github.com/mila-udem/blocks/issues/988
delegate
where possiblegenerate