I updated the liftcover pack to now perform expectation maximization and gradient descent calculations in Python (if wanted). The advantage is that they can now be performed on a GPU. Moreover, I added the possibility of using parallelism to speed up computation.
With these improvements, I was able to run parameter learning on a 10Mb input file with 190,000 rules in less than 1 hour.
The updated version can be tried online at https://cplint.eu, albeit without GPU and with a limit of 2 threads per job.
3 Likes