Unblock ProtoMPRS to control determinism of DataPipe in single/multi-processing and dist/non-dist env #827
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
facebook-github-bot
added
the
CLA Signed
|
@ejguan has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
ejguan
commented
Oct 12, 2022
| ): | ||
| pass | ||
|
|
||
|
|
||
| def SpawnProcessForDataPipeline(multiprocessing_ctx, datapipe, call_locally_fn=None, call_on_reset_epoch=None): | ||
| def SpawnProcessForDataPipeline(multiprocessing_ctx, datapipe, call_on_process_init=None, call_on_epoch_reset=None): |
There was a problem hiding this comment.
I change these argument names to clarify the functionalities.
ejguan
commented
Oct 12, 2022
| @@ -174,38 +188,72 @@ def __init__( | |||
| self.multiprocessing_context = multiprocessing_context | |||
| self.processes = [] | |||
| self.datapipes = [] | |||
| self.combined_datapipes = None | |||
| self.end_datapipe = None | |||
There was a problem hiding this comment.
I change it to end_datapipe because we need to store the last DataPipe for both In-process and multiprocessing cases.
| ) | ||
|
|
||
| # Multiprocessing (num_workers > 0) | ||
| if isinstance(self.end_datapipe, _IterateQueueDataPipes): |
There was a problem hiding this comment.
Will merge conflict with my prefetcher fix
There was a problem hiding this comment.
I will rebase when your PR is landed. I still need to add a test for process-local RNGs
NivekT
approved these changes
Oct 13, 2022
|
@ejguan has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
ejguan
force-pushed
the
extend_proto_rs
branch
from
7cc917a
to
ad8da32
Compare
|
@ejguan has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
VitalyFedyunin
approved these changes
Oct 18, 2022
NivekT
approved these changes
Oct 18, 2022
This was referenced Oct 19, 2022
ejguan
added a commit
to ejguan/data
that referenced
this pull request
Oct 21, 2022
…processing and dist/non-dist env (pytorch#827) Summary: This PR temporarily extend `PrototypingMultiProcessingReadingService` to fully control the determinism of the pipeline in the combinations of: - Single/Multi-processing - Distributed/Non-distributed When we have `SequentialReadingService` ready to combine `DistributedReadingService` and `PrototypingMultiProcessingReadingService`, a few code should be removed. And, for in-process reading service, we still need a method to isolate global RNGs to prevent data-pipeline interferes randomness against model. For multiprocessing case, it will set the same random seed for `Shuffler` and set different deterministic seeds for global RNGs including `python.random`, `torch` and `numpy` within each subprocess. For distributed case, it will share the same random seed for `Shuffler` across all distributed subprocesses to guarantee the shuffle order before sharding. Test Plan: All tests are executed in the combinations of the above environments - [x] Validate the same seed will generate the same order of data - [x] Validate different seeds will generate different order of data - [x] Validate the data after shuffle and sharding in each worker are mutually exclusive and collectively exhaustive with/without manual seed There is one missing test I will add tmrw - [x] Validate subprocess-local RNGs like `random`, `torch` and `numpy` are properly set with different seeds. Pull Request resolved: pytorch#827 Reviewed By: VitalyFedyunin, NivekT Differential Revision: D40323946 Pulled By: ejguan fbshipit-source-id: 2997d6d5dce87a6c38d5ebdf64a00f9769bb18fa
ejguan
added a commit
that referenced
this pull request
Oct 21, 2022
…processing and dist/non-dist env (#827) Summary: This PR temporarily extend `PrototypingMultiProcessingReadingService` to fully control the determinism of the pipeline in the combinations of: - Single/Multi-processing - Distributed/Non-distributed When we have `SequentialReadingService` ready to combine `DistributedReadingService` and `PrototypingMultiProcessingReadingService`, a few code should be removed. And, for in-process reading service, we still need a method to isolate global RNGs to prevent data-pipeline interferes randomness against model. For multiprocessing case, it will set the same random seed for `Shuffler` and set different deterministic seeds for global RNGs including `python.random`, `torch` and `numpy` within each subprocess. For distributed case, it will share the same random seed for `Shuffler` across all distributed subprocesses to guarantee the shuffle order before sharding. Test Plan: All tests are executed in the combinations of the above environments - [x] Validate the same seed will generate the same order of data - [x] Validate different seeds will generate different order of data - [x] Validate the data after shuffle and sharding in each worker are mutually exclusive and collectively exhaustive with/without manual seed There is one missing test I will add tmrw - [x] Validate subprocess-local RNGs like `random`, `torch` and `numpy` are properly set with different seeds. Pull Request resolved: #827 Reviewed By: VitalyFedyunin, NivekT Differential Revision: D40323946 Pulled By: ejguan fbshipit-source-id: 2997d6d5dce87a6c38d5ebdf64a00f9769bb18fa
ejguan
added a commit
to ejguan/data
that referenced
this pull request
Oct 23, 2022
…processing and dist/non-dist env (pytorch#827) Summary: This PR temporarily extend `PrototypingMultiProcessingReadingService` to fully control the determinism of the pipeline in the combinations of: - Single/Multi-processing - Distributed/Non-distributed When we have `SequentialReadingService` ready to combine `DistributedReadingService` and `PrototypingMultiProcessingReadingService`, a few code should be removed. And, for in-process reading service, we still need a method to isolate global RNGs to prevent data-pipeline interferes randomness against model. For multiprocessing case, it will set the same random seed for `Shuffler` and set different deterministic seeds for global RNGs including `python.random`, `torch` and `numpy` within each subprocess. For distributed case, it will share the same random seed for `Shuffler` across all distributed subprocesses to guarantee the shuffle order before sharding. Test Plan: All tests are executed in the combinations of the above environments - [x] Validate the same seed will generate the same order of data - [x] Validate different seeds will generate different order of data - [x] Validate the data after shuffle and sharding in each worker are mutually exclusive and collectively exhaustive with/without manual seed There is one missing test I will add tmrw - [x] Validate subprocess-local RNGs like `random`, `torch` and `numpy` are properly set with different seeds. Pull Request resolved: pytorch#827 Reviewed By: VitalyFedyunin, NivekT Differential Revision: D40323946 Pulled By: ejguan fbshipit-source-id: 2997d6d5dce87a6c38d5ebdf64a00f9769bb18fa
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.
This PR temporarily extend
PrototypingMultiProcessingReadingServiceto fully control the determinism of the pipeline in the combinations of:When we have
SequentialReadingServiceready to combineDistributedReadingServiceandPrototypingMultiProcessingReadingService, a few code should be removed. And, for in-process reading service, we still need a method to isolate global RNGs to prevent data-pipeline interferes randomness against model.For multiprocessing case, it will set the same random seed for
Shufflerand set different deterministic seeds for global RNGs includingpython.random,torchandnumpywithin each subprocess.For distributed case, it will share the same random seed for
Shuffleracross all distributed subprocesses to guarantee the shuffle order before sharding.Tests:
All tests are executed in the combinations of the above environments
There is one missing test I will add tmrw
random,torchandnumpyare properly set with different seeds.