SplitBatchRunner #1713
Open
SplitBatchRunner #1713
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
| for i in range(0, batch_size, self._max_batch_size): | ||
| batch_args, batch_kwargs = alf.nest.map_structure( | ||
| lambda x: x[i:i + self._max_batch_size], (args, kwargs)) | ||
| outputs.append( |
There was a problem hiding this comment.
are we supposed to have the same results from the full batch and the split batch? what if there are dropouts and batch norm in the model?
There was a problem hiding this comment.
As the doc-string notes, dropout is not supported for training.
BatchNorm will also be affected. But if the batch_size is not too small, its effect on training should be small.
hnyu
reviewed
Dec 9, 2024
alf/utils/lean_function.py
Outdated
|
|
||
| Args: | ||
| model (nn.Module): the model to run | ||
| max_batch_size (int): the maximum batch size to run the model. |
There was a problem hiding this comment.
Need to add a comment that if max_batch_size <=0, what will happen.
|
|
||
| return alf.nest.map_structure(lambda *x: torch.cat(x, dim=0), *outputs) | ||
|
|
||
| def original_forward(self, *args, **kwargs): |
There was a problem hiding this comment.
This is used by user of this class.
|
|
||
|
|
||
| class SplitBatchRunner(torch.nn.Module): | ||
| """Split the input into smaller batches and run the model on each batch. |
There was a problem hiding this comment.
Need to clarify that lean_function will be used.
hnyu
approved these changes
Jan 13, 2025
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Outdated suggestions cannot be applied.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.
Sometime we want to use a large batch for training but the memory is not enough. In this kind of situation, we can split the large batch into smaller batches and use lean_function to run the small batches. This can reduces the memory usage because the intermediate tensors will all be released.
So SplitBatchRunner is implemented to make this process very simple.