Github user avulanov commented on the pull request:
https://github.com/apache/spark/pull/1290#issuecomment-69849072
Added batch processing to forward and back-propagation, i.e. data points
are stacked into matrix and then processed in matrix form which can be
hardware-accelerated with Blas. This feature should speed-up back-propagation
if batch size is chosen properly even if native-Blas is not plugged. I did few
experiments and it turns out that for better performance batch should result in
- Native-system-Blas and Native-reference-blas: matrices of tens
thousands elements. E.g. if each data point is 780 features and batch size is
100 then batch matrix will contain 78000 elements
- No natives, just f2jblas: thousands of elements, e.g.
780*10(batchSize)=7800 elements
These suggestions correlate with graphs from netlib-java:
https://github.com/fommil/netlib-java. I will post graphs and performance
comparisons on larger scale soon.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
