This PR rewrites the approx tree method to use codebase from hist for better performance and code sharing.
The rewrite has many benefits:
- Support for both `max_leaves` and `max_depth`.
- Support for `grow_policy`.
- Support for mono constraint.
- Support for feature weights.
- Support for easier bin configuration (`max_bin`).
- Support for categorical data.
- Faster performance for most of the datasets. (many times faster)
- Support for prediction cache.
- Significantly better performance for external memory.
- Unites the code base between approx and hist.
On GPU we use rouding factor to truncate the gradient for deterministic results. This PR changes the gradient representation to fixed point number with exponent aligned with rounding factor.
[breaking] Drop non-deterministic histogram.
Use fixed point for shared memory.
This PR is to improve the performance of GPU Hist.
Co-authored-by: Andy Adinets <aadinets@nvidia.com>
- Reduce dependency on dmlc parsers and provide an interface for users to load data by themselves.
- Remove use of threaded iterator and IO queue.
- Remove `page_size`.
- Make sure the number of pages in memory is bounded.
- Make sure the cache can not be violated.
- Provide an interface for internal algorithms to process data asynchronously.
* Use pre-rounding based method to obtain reproducible floating point
summation.
* GPU Hist for regression and classification are bit-by-bit reproducible.
* Add doc.
* Switch to thrust reduce for `node_sum_gradient`.
* Extract interaction constraints from split evaluator.
The reason for doing so is mostly for model IO, where num_feature and interaction_constraints are copied in split evaluator. Also interaction constraint by itself is a feature selector, acting like column sampler and it's inefficient to bury it deep in the evaluator chain. Lastly removing one another copied parameter is a win.
* Enable inc for approx tree method.
As now the implementation is spited up from evaluator class, it's also enabled for approx method.
* Removing obsoleted code in colmaker.
They are never documented nor actually used in real world. Also there isn't a single test for those code blocks.
* Unifying the types used for row and column.
As the size of input dataset is marching to billion, incorrect use of int is subject to overflow, also singed integer overflow is undefined behaviour. This PR starts the procedure for unifying used index type to unsigned integers. There's optimization that can utilize this undefined behaviour, but after some testings I don't see the optimization is beneficial to XGBoost.
* Refactor configuration [Part II].
* General changes:
** Remove `Init` methods to avoid ambiguity.
** Remove `Configure(std::map<>)` to avoid redundant copying and prepare for
parameter validation. (`std::vector` is returned from `InitAllowUnknown`).
** Add name to tree updaters for easier debugging.
* Learner changes:
** Make `LearnerImpl` the only source of configuration.
All configurations are stored and carried out by `LearnerImpl::Configure()`.
** Remove booster in C API.
Originally kept for "compatibility reason", but did not state why. So here
we just remove it.
** Add a `metric_names_` field in `LearnerImpl`.
** Remove `LazyInit`. Configuration will always be lazy.
** Run `Configure` before every iteration.
* Predictor changes:
** Allocate both cpu and gpu predictor.
** Remove cpu_predictor from gpu_predictor.
`GBTree` is now used to dispatch the predictor.
** Remove some GPU Predictor tests.
* IO
No IO changes. The binary model format stability is tested by comparing
hashing value of save models between two commits
* - do not create device vectors for the entire sparse page while computing histograms...
- while creating the compressed histogram indices, the row vector is created for the entire
sparse page batch. this is needless as we only process chunks at a time based on a slice
of the total gpu memory
- this pr will allocate only as much as required to store the ppropriate row indices and the entries
* - do not dereference row_ptrs once the device_vector has been created to elide host copies of those counts
- instead, grab the entry counts directly from the sparsepage
* - training with external memory - part 2 of 2
- when external memory support is enabled, building of histogram indices are
done incrementally for every sparse page
- the entire set of input data is divided across multiple gpu's and the relative
row positions within each device is tracked when building the compressed histogram buffer
- this was tested using a mortgage dataset containing ~ 670m rows before 4xt4's could be
saturated
* Only define `gpu_id` and `n_gpus` in `LearnerTrainParam`
* Pass LearnerTrainParam through XGBoost vid factory method.
* Disable all GPU usage when GPU related parameters are not specified (fixes XGBoost choosing GPU over aggressively).
* Test learner train param io.
* Fix gpu pickling.
* Combine thread launches into single launch per tree for gpu_hist
algorithm.
* Address deprecation warning
* Add manual column sampler constructor
* Turn off omp dynamic to get a guaranteed number of threads
* Enable openmp in cuda code
