Refactor regressor so one can dynamically add "blocks", currently only flat structure is supported and it only implements existing FFM+LR concept.
This directly opens up the following:
easily add linear regression
easily add an algo such as simple attention FFMs
Performance penalty is about 2%, but it gets us a lot, so i think it is worth it. The architecture allows us to keep using on-stack temporary buffers.
Oh, the run_one.sh builds are now static, since this is the only way to be able to build code that directly runs on our servers (due to differences in Glib). This also required small changes to dependancies.
More cleanup still to be done, including figuring out how to handle ImmutableRegressor.
This is kind of a biggie.
Refactor regressor so one can dynamically add "blocks", currently only flat structure is supported and it only implements existing FFM+LR concept. This directly opens up the following:
Performance penalty is about 2%, but it gets us a lot, so i think it is worth it. The architecture allows us to keep using on-stack temporary buffers.
Oh, the run_one.sh builds are now static, since this is the only way to be able to build code that directly runs on our servers (due to differences in Glib). This also required small changes to dependancies.
More cleanup still to be done, including figuring out how to handle ImmutableRegressor.