DistilBERT for token classification

[stefan-it]

Hi,

this PR adds a DistilBertForTokenClassification implementation (mainly inspired by the BERT implementation) that allows to perform sequence labeling tasks like NER or PoS tagging.

Additionally, the run_ner.py example script was modified to fully support DistilBERT for NER tasks.

I did a small comparison between BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) with the same hyperparameters as specified in the example documentation (one run):

Model	F-Score Dev	F-Score Test
bert-large-cased	95.59	91.70
roberta-large	95.96	91.87
distilbert-base-uncased	94.34	90.32

Unit test for the DistilBertForTokenClassification implementation is also added.

[LysandreJik]

This is great, thanks @stefan-it!

[thomwolf]

This is great, thanks a lot @stefan-it.
I've added your quick benchmark in the readme.

[codecov-io]

Codecov Report

Merging #1792 into master will increase coverage by 0.03%.
The diff coverage is 97.29%.

@@            Coverage Diff             @@
##           master    #1792      +/-   ##
==========================================
+ Coverage   84.03%   84.07%   +0.03%     
==========================================
  Files          94       94              
  Lines       14032    14069      +37     
==========================================
+ Hits        11792    11828      +36     
- Misses       2240     2241       +1

Impacted Files	Coverage Δ
transformers/tests/modeling_distilbert_test.py	99.18% <100%> (+0.08%)	⬆️
transformers/modeling_distilbert.py	95.87% <96.15%> (+0.02%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update b5d330d...05db5bc. Read the comment docs.

[tedgoddard]

I just wanted to say thanks for this PR, it was just what I was looking for at the time.