-
Notifications
You must be signed in to change notification settings - Fork 1.9k
Open
Labels
BugBug in learning semantics, critical by defaultBug in learning semantics, critical by default
Description
Describe the bug
When using LDA with --minibatch > 1, the final partial batch is processed for learning but not counted in the reported statistics, causing an incorrect "number of examples" count.
Root Cause:
The update_stats_lda() function only updates statistics when batch_buffer.size() == minibatch, excluding the final partial batch
How to reproduce
./vw --lda 20 -d dataset.vw --minibatch 256 --passes 1
Version
9.10.0
OS
linux
Language
C++
Additional context
I have a fix ready should I PR?
Metadata
Metadata
Assignees
Labels
BugBug in learning semantics, critical by defaultBug in learning semantics, critical by default