GitLab

Attention output was in bnij ordering instead of ijbn which everything
else will expect. This was an oversight on my part, and keeps the
attention inputs/outputs identical to the original code.

Also moved back from tensor slicing to index_select in rel_shift_bnij to
make the tracer happy.

Significant performance boost over the original orderings
on an already somewhat optimised branch this gave me > 2x end-to-end
throughput on a squad xlnet fine-tuning task (batch 8, seq-length 612,
fp16)

Typo with LM Fine tuning script

Fix fp16 masking in PoolerEndLogits

Merge branch 'pooler_end_logits_fp16_fix' of https://github.com/slayton58/pytorch-transformers into pr/1284

Lines 183 - 200, fixed indentation. Line 198, replaced `tokenizer_class` with `BertTokenizer`, since `tokenizer_class` is not defined in the loop it belongs to.

fix link

Fix some typos

Fix tensorflow_dataset glue support

Update run_lm_finetuning.py

The previous method, just as phrased, did not exist in the class.

`glue_convert_examples_to_features` assumed that tensorflow_dataset
examples contains the features `'sentence1'` and `'sentence2'`. This
commit encapsulates the choice of features in the glue processor and
uses that to parse examples.

Just some typos

faster dataset building

Add small  note about the output of hidden states (closes #1332)

Co-Authored-By: Lysandre Debut <lysandre.debut@reseau.eseo.fr>

Co-Authored-By: Lysandre Debut <lysandre.debut@reseau.eseo.fr>