junrushao commented on code in PR #13195:
URL: https://github.com/apache/tvm/pull/13195#discussion_r1016887588
##########
python/tvm/meta_schedule/search_strategy/evolutionary_search.py:
##########
@@ -58,7 +58,7 @@ class EvolutionarySearch(SearchStrategy):
def __init__(
self,
*,
- population_size: int = 2048,
+ population_size: int = 512,
Review Comment:
@masahi Ask for asking, and it's definitely a valid concern. The reason is
that we found it's somehow set to an unreasonably large number (we were
consistently using 512 on downstream), and now wanted to scale it back. Also
note that the population calculation is slightly different than AutoScheduler,
which includes invalid candidates in population, while ours always prune ahead
of time.
We did make sure we have proper numbers in hand before merging, and please
refer to the table below for details:
| | Mainline (ms) | This PR (ms) | Difference |
|---------------------|---------------|--------------|------------|
| resnet_50 | 1.829128008 | 1.73111331 | 5.66% |
| mobilenet_v2 | 0.4773168361 | 0.4806015715 | -0.68% |
| resnet_18 | 0.678600832 | 0.6301909627 | 7.68% |
| mobilenet_v3 | 0.6649458484 | 0.668666894 | -0.56% |
| wide_resnet_50 | 3.753781691 | 3.077549256 | 21.97% |
| densenet_121 | 2.361973117 | 2.288411393 | 3.21% |
| inception_v3 | 3.53431975 | 3.478368823 | 1.61% |
| resnet3d_18 | 7.791658449 | 7.465777971 | 4.36% |
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
