transformers

transformers copied to clipboard

[WIP] Extend VisualQuestionAnsweringPipeline to support QuestionAnwering models (e.g. LayoutLM)

21

# What does this PR do? This PR extends VisualQuestionAnsweringPipeline to accept `words` and `boxes` as input, passes them into the tokenizer/model (along with the question), and post-processes their `QuestionAnsweringModelOutput`...

ankrgyl

Flax models should allow `inputs_embeds`

4

### Feature request Currently, non-Flax models allow `inputs_embeds` instead of `input_ids` (e.g., GPT2) ```python def forward( self, input_ids: Optional[torch.LongTensor] = None, ... inputs_embeds: Optional[torch.FloatTensor] = None, ... ) -> Union[Tuple,...

mattf1n

Allow user-managed Pool in Wav2Vec2ProcessorWithLM.batch_decode

6

# What does this PR do? There are two issues being attacked: 1. Always creating a fresh pool within `Wav2Vec2ProcessorWithLM.batch_decode` generates a big overhead if it's called multiple times (this...

falcaopetri

Update the original mapping in _LazyConfigMapping to fix AutoTokenizer registration

3

# What does this PR do? Currently when we want to register a new config+tokenizer+model, per [the instructions](https://huggingface.co/docs/transformers/model_doc/auto), it seems we should do the following: ``` from transformers import AutoConfig,...

lolipopshock

Fix ORTTrainer failure on DeBERTa(base/v2/sew_d) fp16 training

1

# What does this PR do? __Context__ It was reported in optimum https://github.com/huggingface/optimum/issues/305 that the training on DeBERTa with optimum.onnxruntime.ORTTrainer is broken. After investigation, the break comes from two causes:...

JingyaHuang

Compute true loss Flax examples

1

# What does this PR do? 'True' losses should be computed in Flax examples, as [discussed](https://github.com/huggingface/transformers/pull/18297#discussion_r931971230) with @sanchit-gandhi. ## Who can review? cc @sanchit-gandhi @patrickvonplaten

duongna21

Update BLOOM parameter counts

11

Update parameter counts of BLOOM models. The original counts were incorrect & have already been updated on the hub. I can't add reviewers, but @younesbelkada @thomasw21 may want to review...

Muennighoff

[TensorFlow] Adding GroupViT

18

# Adding TensorFlow version of GroupViT This PR adds the TensorFlow version of [GroupViT](https://github.com/NVlabs/GroupViT). ## Before submitting - [ ] This PR fixes a typo or improves the docs (you...

ariG23498

Added option for users to modify config parameter when calling pytess…

7

# What does this PR do? This is a feature addition to LayoutLMV3, a follow up to the same feature added to [LayoutLMV2's feature extractor ](https://github.com/huggingface/transformers/pull/17733) Giving user option to...

kelvinAI

Typo in configuration

2

Hey @NielsRogge I found an inconsistency in the documentation and the code for configuration of GroupViT. The default for `num_output_groups` is `[64, 8, 8]` (notice the last element in the...

ariG23498