[hivemind.Optimizer] ProgressTracker #408
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
Codecov Report
@@ Coverage Diff @@
## master #408 +/- ##
==========================================
+ Coverage 84.17% 84.57% +0.40%
==========================================
Files 75 76 +1
Lines 7121 7280 +159
==========================================
+ Hits 5994 6157 +163
+ Misses 1127 1123 -4
|
|
Massive test results: #5a4c56b test accuracy = 29 / 31 (too flappy, trying more lenient thresholds for epoch==2) |
borzunov
requested changes
Nov 16, 2021
| @contextlib.contextmanager | ||
| def pause_updates(self): | ||
| """Temporarily stop progress tracker from updating global training state""" | ||
| with self.lock_global_progress, self.performance_ema.pause(): |
There was a problem hiding this comment.
(may be out of scope of this PR)
I'm surprised that performance_ema.pause() resets the timer instead of actually pausing it.
Co-authored-by: Alexander Borzunov <[email protected]>
borzunov
requested changes
Nov 16, 2021
| try: | ||
| while not self.shutdown_triggered.is_set(): | ||
| wait_timeout = max(0.0, last_report_time + self.metadata_expiration - get_dht_time()) | ||
| logger.debug(f"Will report progress again in {wait_timeout} seconds or on user command.") |
There was a problem hiding this comment.
nit: Here and further:
Suggested change
| logger.debug(f"Will report progress again in {wait_timeout} seconds or on user command.") | |
| logger.debug(f"Will report progress again in {wait_timeout} seconds or on user command") |
Comment on lines
130
to
136
| if self.global_epoch > self.local_progress.epoch: | ||
| return True | ||
| elif self.global_progress.samples_accumulated >= self.target_batch_size: | ||
| return True | ||
| elif get_dht_time() >= self.global_progress.eta_next_epoch: | ||
| return True | ||
| return False |
There was a problem hiding this comment.
Suggested change
| if self.global_epoch > self.local_progress.epoch: | |
| return True | |
| elif self.global_progress.samples_accumulated >= self.target_batch_size: | |
| return True | |
| elif get_dht_time() >= self.global_progress.eta_next_epoch: | |
| return True | |
| return False | |
| return ( | |
| self.global_epoch > self.local_progress.epoch or | |
| self.global_progress.samples_accumulated >= self.target_batch_size or | |
| get_dht_time() >= self.global_progress.eta_next_epoch | |
| ) |
borzunov
approved these changes
Nov 16, 2021
tests/test_optimizer.py
Outdated
| assert not tracker.is_alive() | ||
|
|
||
| mean_step_time = sum(step_time_deltas) / len(step_time_deltas) | ||
| for i in (0, 1, 5): |
There was a problem hiding this comment.
Suggested change
| for i in (0, 1, 5): | |
| for i in (0, 1, 5): # Without the 4th worker (the fastest one) |
tests/test_optimizer.py
Outdated
| mean_step_time = sum(step_time_deltas) / len(step_time_deltas) | ||
| for i in (0, 1, 5): | ||
| assert 1.05 * mean_step_time < step_time_deltas[i] < 2.0 * mean_step_time | ||
| for i in (2, 3, 4): |
There was a problem hiding this comment.
Suggested change
| for i in (2, 3, 4): | |
| for i in (2, 3, 4): # With the 4th worker |
borzunov
reviewed
Nov 16, 2021
| performance_ema_alpha: float = 0.1, | ||
| metadata_expiration: float = 30.0, | ||
| status_loglevel: int = logging.DEBUG, | ||
| private_key: PrivateKey = None, |
There was a problem hiding this comment.
Suggested change
| private_key: PrivateKey = None, | |
| private_key: Optional[RSAPrivateKey] = None, |
borzunov
reviewed
Nov 16, 2021
tests/test_optimizer.py
Outdated
| tracker.shutdown() | ||
| dht.shutdown() | ||
|
|
||
| # note: we use processes here because RSASignatureValidator inside trackers uses process-wide RSA keys |
There was a problem hiding this comment.
Suggested change
| # note: we use processes here because RSASignatureValidator inside trackers uses process-wide RSA keys |
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.
Auxiliary class that keeps track of local & global training progress, measured in epochs.
An epoch can be incremented after collaboration accumulates a said number of gradients (target_batch_size).
Similarly to pytorch LR scheduler, epoch can be incremented on a single optimizer update or many local updates.