GitLab

* remove references to old API in docstring - update data processors

* style

* fix tests - better type checking error messages

* better type checking

* include awesome fix by @LysandreJik for #5310

* updated doc and examples

* Kill model archive maps

* Fixup

* Also kill model_archive_map for MaskedBertPreTrainedModel

* Unhook config_archive_map

* Tokenizers: align with model id changes

* make style && make quality

* Fix CI

Change the example code to use encode_plus since the token_type_id
wasn't being correctly set.

* add multiple choice for longformer

* add models to docs

* adapt docstring

* add test to longformer

* add longformer for mc in init and modeling auto

* fix tests

* Test case for #3936

* multigpu tests pass on pytorch 1.4.0

* Fixup

* multigpu tests pass on pytorch 1.5.0

* Update src/transformers/modeling_utils.py

* Update src/transformers/modeling_utils.py

* rename multigpu to require_multigpu

* mode doc

* [file_utils] use_cdn + documentation

* Move to cdn. urls for weights

* [urls] Hotfix for bert-base-japanese

* change encoder decoder style to bart & t5 style

* make encoder decoder generation dummy work for bert

* make style

* clean init config in encoder decoder

* add tests for encoder decoder models

* refactor and add last tests

* refactor and add last tests

* fix attn masks for bert encoder decoder

* make style

* refactor prepare inputs for Bert

* refactor

* finish encoder decoder

* correct typo

* add docstring to config

* finish

* add tests

* better naming

* make style

* fix flake8

* clean docstring

* make style

* rename

token_type_id is converted into the segment embedding. For question answering,
this needs to highlight whether a token belongs to sequence 0 or 1.
encode_plus takes care of correctly setting this parameter automatically.

* Delete some copy pasted code

* Electra wip

* helpers

* Electra wip

* Electra v1

* ELECTRA may be saved/loaded

* Generator & Discriminator

* Embedding size instead of halving the hidden size

* ELECTRA Tokenizer

* Revert BERT helpers

* ELECTRA Conversion script

* Archive maps

* PyTorch tests

* Start fixing tests

* Tests pass

* Same configuration for both models

* Compatible with base + large

* Simplification + weight tying

* Archives

* Auto + Renaming to standard names

* ELECTRA is uncased

* Tests

* Slight API changes

* Update tests

* wip

* ElectraForTokenClassification

* temp

* Simpler arch + tests

Removed ElectraForPreTraining which will be in a script

* Conversion script

* Auto model

* Update links to S3

* Split ElectraForPreTraining and ElectraForTokenClassification

* Actually test PreTraining model

* Remove num_labels from configuration

* wip

* wip

* From discriminator and generator to electra

* Slight API changes

* Better naming

* TensorFlow ELECTRA tests

* Accurate conversion script

* Added to conversion script

* Fast ELECTRA tokenizer

* Style

* Add ELECTRA to README

* Modeling Pytorch Doc + Real style

* TF Docs

* Docs

* Correct links

* Correct model intialized

* random fixes

* style

* Addressing Patrick's and Sam's comments

* Correct links in docs

* All Tokenizers

BertTokenizer + few fixes
RobertaTokenizer
OpenAIGPTTokenizer + Fixes
GPT2Tokenizer + fixes
TransfoXLTokenizer
Correct rst for TransformerXL
XLMTokenizer + fixes
XLNet Tokenizer + Style
DistilBERT + Fix XLNet RST
CTRLTokenizer
CamemBERT Tokenizer
FlaubertTokenizer
XLMRobertaTokenizer
cleanup

* cleanup

* change masking to direct labelings

* fix black

* switch to ignore index

* .

* fix black

* activations.py contains a mapping from string to activation function
* resolves some `gelu` vs `gelu_new` ambiguity

PyTorch < 1.3 requires multiplication operands to be of the same type.
This was violated when using default attention mask (i.e.,
attention_mask=None in arguments) given BERT in the decoder mode.

In particular, this was breaking Model2Model and made tutorial
from the quickstart failing.

Masked indices should have -1 and not -100. Updating documentation + scripts that were forgotten

Created a link between the linear layer bias and the model attribute bias. This does not change anything for the user nor for the conversion scripts, but allows the `resize_token_embeddings` method to resize the bias as well as the weights of the decoder.

Added a test.

Signed-off-by: Romain Keramitas <r.keramitas@gmail.com>