Hi,
Your attention mechanism is quite slow. Since you compute the linear projections (aw and bw) each time although they do not change, the time is almost quadratic.
I have implemented a faster version of attention which does a lot of precomputation and would like to push it as soon as I am done testing.
Regards.
Hi,
Your attention mechanism is quite slow. Since you compute the linear projections (aw and bw) each time although they do not change, the time is almost quadratic.
I have implemented a faster version of attention which does a lot of precomputation and would like to push it as soon as I am done testing.
Regards.