[tune] reuse_actors per default for function trainables #24040
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
added 2 commits
April 20, 2022 09:42
krfricke
commented
Apr 20, 2022
Comment on lines
-64
to
-88
| def create_resettable_function(num_resets: defaultdict): | ||
| def trainable(config, checkpoint_dir=None): | ||
| if checkpoint_dir: | ||
| with open(os.path.join(checkpoint_dir, "chkpt"), "rb") as fp: | ||
| step = pickle.load(fp) | ||
| else: | ||
| step = 0 | ||
|
|
||
| while step < 2: | ||
| step += 1 | ||
| with tune.checkpoint_dir(step) as checkpoint_dir: | ||
| with open(os.path.join(checkpoint_dir, "chkpt"), "wb") as fp: | ||
| pickle.dump(step, fp) | ||
| tune.report(**{"done": step >= 2, "iter": step, "id": config["id"]}) | ||
|
|
||
| trainable = wrap_function(trainable) | ||
|
|
||
| class ResetCountTrainable(trainable): | ||
| def reset_config(self, new_config): | ||
| num_resets[self.trial_id] += 1 | ||
| return super().reset_config(new_config) | ||
|
|
||
| return ResetCountTrainable | ||
|
|
||
|
|
There was a problem hiding this comment.
This test never worked, as num_resets was local to each remote actor.
There is no great way to count the number of resets for function trainables so I've removed this for now
krfricke
commented
Apr 20, 2022
| ) | ||
| self.assertEqual([t.last_result["id"] for t in trials], [0, 1, 2, 3]) | ||
| self.assertEqual([t.last_result["iter"] for t in trials], [2, 2, 2, 2]) | ||
| self.assertEqual([num_resets[t.trial_id] for t in trials], [0, 0, 0, 0]) |
There was a problem hiding this comment.
As you can see, this test doesn't work - the number of resets should have been 4, 5, 6, 7
added 2 commits
April 20, 2022 14:56
|
Is there concern if these function trainables use GPU? |
Yard1
approved these changes
Apr 20, 2022
|
@xwjiang2010 thanks, that is indeed a concern, so lets default to False here (note that it doesn't mean it definitely happens, just that we should default to False to be sure) |
|
Updated, PTAL |
Yard1
approved these changes
Apr 21, 2022
xwjiang2010
approved these changes
Apr 22, 2022
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Why are these changes needed?
Function trainables don't carry state, so they should be reused per default for performance optimization.
Related issue number
Would e.g. simplify issues for frequently pausing trials like in #23470
Checks
scripts/format.shto lint the changes in this PR.