-
-
Notifications
You must be signed in to change notification settings - Fork 8.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Remove single_precision_histogram #7828
Conversation
Close #7728 . |
@@ -532,6 +529,13 @@ struct GPUHistMakerDevice { | |||
|
|||
void ApplySplit(const GPUExpandEntry& candidate, RegTree* p_tree) { | |||
RegTree& tree = *p_tree; | |||
|
|||
// Sanity check - have we created a leaf with no training instances? | |||
if (!rabit::IsDistributed() && row_partitioner) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Any potential solution if this check occurs?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I don't think it should happen with double precision. If it does it is a serious bug that will cause accuracy problems and undefined behaviour.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
My mistake. I ran into issues with empty nodes in partitioner UpdatePosition
, where left_count
can be 0 due to floating-point errors. In this ApplySplit
function, if we check on the left/right child instead of the parent (candidate), then we might run into trouble.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Apologies for the conflict. Do you need help with rebasing?
3145e57
to
67a664d
Compare
Removing the
single_precision_histogram
due to major observed accuracy issues. In particular many tests can fail due to numerical error accumulating in gain calculation, resulting in leaf nodes being opened that contain no training data at all.In this PR I have removed a now redundant class in gpu_hist and slightly refactored the TreeUpdater base class to always require a global context for construction.