Initial support for multioutput regression. #7514
Conversation
| @@ -83,8 +88,10 @@ class RegLossObj : public ObjFunction { | |||
| // for better performance. | |||
| const size_t n_data_blocks = std::max(static_cast<size_t>(1), (on_device ? ndata : nthreads)); | |||
| const size_t block_size = ndata / n_data_blocks + !!(ndata % n_data_blocks); | |||
| auto const n_targets = std::max(info.labels.Shape(1), static_cast<size_t>(1)); | |||
|
|
|||
There was a problem hiding this comment.
We should use more ndarray friendly ways to express the calculation in the future.
|
Can we throw an error when the user attempt to use other objectives with multi-output labels? |
Added a check along with test. Thanks for the suggestion. |
|
The support is primitive. We need to expand it to other regression objectives and add more documents. |
trivialfis
force-pushed
the
multi-output-reg
branch
from
0a1f857
to
01e8a2c
Compare
|
Is this API suitable for wrapping up in XGBoost4J, or do you want to build out a more general one before doing that? |
|
The API is good for regression and binary classification and should be stable unless major issue is found. So it's good enough to start looking into language bindings. I will try to add xgboost4j support later. Will keep you posted on the progress. It's unlikely that we will implement multi class multi target based on existing interface so no need to worry about it at the moment. :-) |
|
@Craigacp Apologies for the slow progress. I looked into the JVM packages and I can't build a complete stack for jvm from spark down to basic java wrapper. The required change is not trivial since we will have 2-dim inputs for both base_margin and label, I'm not sure what's the best way to implement that for jvm packages. |
|
Do you have a branch somewhere with the current state of it? I'd assumed that DMatrix would grow to accept a 2d matrix for the targets and then most of the rest of the changes would be plumbing, but if there's something that needs some design effort I can take a look. |
|
@Craigacp The master branch contains support for Python. If you are interested in a code walk I'm happy to chat offline.
Yes, that's pretty much the only requirement for regression. For multi-label classification some configuration needs to be done to make sure The difficult part for me is just not being familiar with the jvm stack ... |
|
On Python we use a JSON string to represent the input memory buffer, so we have complete information about shape and strides, etc. I'm not sure what's the best way to handle these numeric data types on JVM. |
|
Ok, I should have some time next week to look through how the python code works. I might take you up on that code walk through once I've familiarised myself with it a bit. The length of the header array should tell us the number of examples, and then for multidimensional things we can require that the target array not be sparse and have explicit zeros (though that might be pretty wasteful for multi-label classification). Then the target array is of known shape |
Close #7309 .