GitLab

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.

cc @julien-c

* fill_mask helper

* [poc] FillMaskPipeline

* Revert "[poc] FillMaskPipeline"

This reverts commit 67eeea55b0f97b46c2b828de0f4ee97d87338335.

* Revert "fill_mask helper"

This reverts commit cacc17b884e14bb6b07989110ffe884ad9e36eaa.

* README: clarify that Pipelines can also do text-classification

cf. question at the AI&ML meetup last week, @mfuntowicz

* Fix test: test feature-extraction pipeline

* Test tweaks

* Slight refactor of existing pipeline (in preparation of new FillMaskPipeline)

* Extraneous doc

* More robust way of doing this

@mfuntowicz as we don't rely on the model name anymore (see AutoConfig)

* Also add RobertaConfig as a quickfix for wrong token_type_ids

* cs

* [BIG] FillMaskPipeline

cc @julien-c @thomwolf

cc @julien-c @@thomwolf

All tests should now work fine.

The CircleCI test should fail.

* add failing test

* fix call to _run_split_on_punc

* format with black

model.parameters() order is apparently not stable (only for xlnet, for some reason)

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.