[Seq2Seq Generation] Call encoder before expanding input_ids #3370
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
sshleifer
changed the title
sshleifer
changed the title
Contributor
|
Like the change a lot! I'd just propose some renaming. |
src/transformers/modeling_bart.py
Outdated
| config_class = BartConfig | ||
| base_model_prefix = "model" | ||
| pretrained_model_archive_map = BART_PRETRAINED_MODEL_ARCHIVE_MAP | ||
| encoder_outputs_batch_idx = 1 # outputs shaped (bs, ...) |
Contributor
There was a problem hiding this comment.
would change the name to encoder_outputs_batch_dim_idx
src/transformers/modeling_t5.py
Outdated
| pretrained_model_archive_map = T5_PRETRAINED_MODEL_ARCHIVE_MAP | ||
| load_tf_weights = load_tf_weights_in_t5 | ||
| base_model_prefix = "transformer" | ||
| encoder_outputs_batch_idx = 0 # outputs shaped (bs, ...) |
Contributor
There was a problem hiding this comment.
would change the name to encoder_outputs_batch_dim_idx
src/transformers/modeling_utils.py
Outdated
| device=next(self.parameters()).device, | ||
| ) | ||
| cur_len = 1 | ||
| batch_idx = self.encoder_outputs_batch_idx |
Contributor
There was a problem hiding this comment.
also here batch_dim_idx
src/transformers/modeling_utils.py
Outdated
| assert ( | ||
| batch_size == encoder_outputs[0].shape[batch_idx] | ||
| ), f"expected encoder_outputs[0] to have 1st dimension bs={batch_size}, got {encoder_outputs[0].shape[1]} " | ||
| expanded_index = ( |
Contributor
There was a problem hiding this comment.
maybe also expanded_idx because we always use idx in the function?
thomwolf
approved these changes
Mar 26, 2020
Member
thomwolf
left a comment
thomwolf
left a comment
There was a problem hiding this comment.
LGTM.
I agree with the name changes proposed by @patrickvonplaten
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
Proposing to call model.encoder before expanding
input_idstoeffective_batch_size*num_beams.For Bart, this saves 1.5 GB of GPU mem on batch_size=6. Savings probably similar for T5 (untested).
Requires knowing which index of the encoder_outputs is associated with the batch dim (we need to expand this dimension), which is different between
BartandT5. This difference is encoded in theself.encoder_outputs_batch_idxvariable.This PR is WIP because
encoder_outputs_batch_idxcould be avoided if we transposed Bart's encoder_outputs, which I haven't tried.