update docs for gpu external memory #5332
Conversation
Codecov Report
@@ Coverage Diff @@
## master #5332 +/- ##
=======================================
Coverage 83.76% 83.76%
=======================================
Files 11 11
Lines 2409 2409
=======================================
Hits 2018 2018
Misses 391 391 Continue to review full report at Codecov.
|
doc/parameter.rst
Outdated
| - ``gradient_based``: the selection probability for each training instance is proportional to the | ||
| *regularized absolute value* of gradients (more specifically, :math:`\sqrt{g^2+\lambda h^2}`). | ||
| ``subsample`` may be set to as low as 0.1 without loss of model accuracy. Note that this | ||
| sampling method is only supported when ``tree_method`` is set to ``gpu_hist``. |
There was a problem hiding this comment.
Maybe leave a note that other tree methods support only the uniform sampling?
| @@ -1,6 +1,6 @@ | |||
| ############################################ | |||
| Using XGBoost External Memory Version (beta) | |||
There was a problem hiding this comment.
Em ... the external memory support is not available for CPU Hist, if we were to declare it stable, we should at least document it in the first section that Hist is not supported.
There was a problem hiding this comment.
Are you sure? I think I've tried hist for comparison in external memory mode. The limitation is right now the CPU code keeps the whole histogram in memory, which may still cause out-of-memory errors if the dataset is too large, but the mechanism of using #cacheprefix and dealing with multiple pages is supported.
There was a problem hiding this comment.
Ok, added hist not well tested as a limitation. If you want, I can add the beta back. :)
|
Otherwise looks good to me. |
Last bit for #4357
@RAMitchell @trivialfis