patrikeh / go-deep / 39

95%

push

travis-ci

Training >40% performance improvement. (#18)

* Add batch training benchmark.

What takes the time varies depending on setup, so assuming here we have
lots of examples and layer sizes that are sufficiently high.

* Some value reuse to avoid perf loss due to reindexing/dereference.

I would think these values would be cached anyway, but the benchmark does
not lie.

* Buffering workCh helps contention.

Seen by training benchmark.

* Updating wait group per batch reduces contention.

As seen by training benchmark. Small but it's there.

8 of 8 new or added lines in 2 files covered. (100.0%)

374 of 392 relevant lines covered (95.41%)

5341527.95 hits per line