This repository was archived by the owner on Mar 21, 2024. It is now read-only.
Fix recovery of SSL training, scale SSL training to multiple nodes #565
Conversation
This file contains hidden or 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
ant0nsc
commented
Sep 20, 2021
This reverts commit e03f0a9.
ant0nsc
changed the title
ozan-oktay
reviewed
Oct 6, 2021
ozan-oktay
reviewed
Oct 6, 2021
| @@ -1,8 +1,7 @@ | |||
| # ------------------------------------------------------------------------------------------ | |||
| # Copyright (c) Microsoft Corporation. All rights reserved. | |||
There was a problem hiding this comment.
I think it would be good to attach ssl training run results for future reference - before and after this manual optimisation change. (both for SimCLR and BYOL)
ozan-oktay
reviewed
Oct 6, 2021
ozan-oktay
reviewed
Oct 6, 2021
| super().__init__(ssl_training_dataset_name=SSLDatasetName.CIFAR10, | ||
| linear_head_dataset_name=SSLDatasetName.CIFAR10, | ||
| ssl_training_batch_size=512, | ||
| ssl_training_batch_size=64, |
There was a problem hiding this comment.
Weren't we training these models in machines with 4 gpus? Should this be reduced to 128?
ozan-oktay
reviewed
Oct 6, 2021
| recovery_checkpoint_save_interval=200, | ||
| num_epochs=1000, | ||
| ssl_training_batch_size=1200, | ||
| ssl_training_batch_size=75, |
ozan-oktay
reviewed
Oct 6, 2021
| param.Boolean(default=False, | ||
| doc="Controls the PyTorch Lightning flag 'find_unused_parameters' for the DDP plugin. " | ||
| "Setting it to True comes with a performance hit.") | ||
| pl_limit_train_batches: Optional[int] = \ |
There was a problem hiding this comment.
what happens if the user specifies zero? Would it skip train/val automatically, is it tested?
There was a problem hiding this comment.
0 is valid. We are passing this value straight through to PL, and it does whatever it would do then - yes, it would skip training/validation.
|
superseded by #584 |
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
2 participants
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.
When SSL training gets interrupted on low-priority nodes, there are presently weird glitches on the metrics for the linear head. We suspect that those come from the fact that the optimizer for the linear head is not saved to the checkpoint, and hence has to re-learn all of its statistics.
This PR adds the linear head optimizer such that it is accessible to PL, and will be included in the checkpoint.
In addition, the semantics of
batch_sizein SSL training is changed: Previously it was the effective batch size, taking multiple nodes into account. This meant that the code was effectively hardcoding 16 GPUS. New behaviour: Batch size is now the batch size on a single GPU. As a consequence, we can scale to any number of GPUs without code changes.There are also new flags
pl_limit_train_batchesandpl_limit_val_batchesto speed up training, by reducing the number of batches processed.