Fix dpr<>bart config for RAG #8808
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
| @@ -104,7 +104,8 @@ def prepare_config_and_inputs(self): | |||
| token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels) | |||
| choice_labels = ids_tensor([self.batch_size], self.num_choices) | |||
|
|
|||
| config = BertConfig( | |||
There was a problem hiding this comment.
This is dangerous when new params are added to Bert, it'll lead to silent errors
lhoestq
approved these changes
Nov 26, 2020
| @@ -178,7 +178,7 @@ def __init__(self, config): | |||
|
|
|||
| # position_ids (1, len position emb) is contiguous in memory and exported when serialized | |||
| self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1))) | |||
| self.position_embedding_type = config.position_embedding_type | |||
| self.position_embedding_type = getattr(config, "position_embedding_type", "absolute") | |||
There was a problem hiding this comment.
maybe add a comment that mentions backward compatibility ?
LysandreJik
approved these changes
Nov 27, 2020
| @@ -104,7 +104,8 @@ def prepare_config_and_inputs(self): | |||
| token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels) | |||
| choice_labels = ids_tensor([self.batch_size], self.num_choices) | |||
|
|
|||
| config = BertConfig( | |||
LysandreJik
added a commit
that referenced
this pull request
Nov 30, 2020
* correct dpr test and bert pos fault * fix dpr bert config problem * fix layoutlm * add config to dpr as well
stas00
pushed a commit
to stas00/transformers
that referenced
this pull request
Dec 5, 2020
* correct dpr test and bert pos fault * fix dpr bert config problem * fix layoutlm * add config to dpr as well
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.
What does this PR do?
There was a big silent bug between the interaction of DPR and BERT. Bert added a new config parameter that DPR does not have -> so a "normal" dpr config crashes when used with BERT. This is actually a very big bug and was introduces in #8276 as pointed out by @lhoestq - thanks!
Two things went wrong here.
We should be more careful in general when introducing new config parameters and calling them via
config.<new_parameter>especially for models like BERT that can be used with other configs.The DPR tests should have could that, but instead of using a normal DPR config, a BERT-like DPR config was used in the tests, which is dangerous because it exactly doesn't catch errors like those.
This PR fixes 1) and 2) by calling the config in the case of the newly introduces parameter only with
getattr(config, <param_name>, <default_value>)and adds the config functionality to DPR as well (DPR also should have this functionality over BERT's positional embedding). IMOgetattr(config, <param_name>, <default_value>)should be used for models like BERT in general because they could be used and wrapped in many different ways.Also the DPR test is fixed to use a DPR config instead of a BERT config.
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors which may be interested in your PR.