"...lm-evaluation-harness.git" did not exist on "a18104a436bc7fa9ba817e3cd970aa742dc5c129"
bart.rst 2.85 KB
Newer Older
Sam Shleifer's avatar
Sam Shleifer committed
1
2
Bart
----------------------------------------------------
3
**DISCLAIMER:** If you see something strange,
Sam Shleifer's avatar
Sam Shleifer committed
4
5
6
file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
@sshleifer

Sylvain Gugger's avatar
Sylvain Gugger committed
7
8
9
Overview
~~~~~~~~~~~~~~~~~~~~~

Sam Shleifer's avatar
Sam Shleifer committed
10
The Bart model was `proposed <https://arxiv.org/abs/1910.13461>`_ by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer on 29 Oct, 2019.
11
According to the abstract,
Sam Shleifer's avatar
Sam Shleifer committed
12

Sam Shleifer's avatar
Sam Shleifer committed
13
14
15
- Bart uses a standard seq2seq/machine translation architecture with a bidirectional encoder (like BERT) and a left-to-right decoder (like GPT).
- The pretraining task involves randomly shuffling the order of the original sentences and a novel in-filling scheme, where spans of text are replaced with a single mask token.
- BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains of up to 6 ROUGE.
Sam Shleifer's avatar
Sam Shleifer committed
16

Sam Shleifer's avatar
Sam Shleifer committed
17
The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_
Sam Shleifer's avatar
Sam Shleifer committed
18
19


Sylvain Gugger's avatar
Sylvain Gugger committed
20
21
Implementation Notes:

22
23
24
25
- Bart doesn't use :obj:`token_type_ids` for sequence classification. Use BartTokenizer.encode to get the proper splitting.
- The forward pass of ``BartModel`` will create decoder inputs (using the helper function ``transformers.modeling_bart._prepare_bart_decoder_inputs``)  if they are not passed. This is different than some other modeling APIs.
- Model predictions are intended to be identical to the original implementation. This only works, however, if the string you pass to ``fairseq.encode`` starts with a space.
- ``BartForConditionalGeneration.generate`` should be used for conditional generation tasks like summarization, see the example in that docstrings
26
- Models that load the ``"facebook/bart-large-cnn"`` weights will not have a ``mask_token_id``, or be able to perform mask filling tasks.
27

Sylvain Gugger's avatar
Sylvain Gugger committed
28
29
30
31
32
33
34
35
36
37
38
39
BartConfig
~~~~~~~~~~~~~~~~~~~~~

.. autoclass:: transformers.BartConfig
    :members:


BartTokenizer
~~~~~~~~~~~~~~~~~~~~~

.. autoclass:: transformers.BartTokenizer
    :members:
Sam Shleifer's avatar
Sam Shleifer committed
40

Sam Shleifer's avatar
Sam Shleifer committed
41
42

BartModel
43
~~~~~~~~~~~~~
Sam Shleifer's avatar
Sam Shleifer committed
44
45
46
47

.. autoclass:: transformers.BartModel
    :members: forward

48
.. autofunction:: transformers.modeling_bart._prepare_bart_decoder_inputs
Sam Shleifer's avatar
Sam Shleifer committed
49
50
51
52
53
54
55
56
57


BartForSequenceClassification
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

.. autoclass:: transformers.BartForSequenceClassification
    :members: forward


Suraj Patil's avatar
Suraj Patil committed
58
59
60
61
62
63
64
BartForQuestionAnswering
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

.. autoclass:: transformers.BartForQuestionAnswering
    :members: forward


Sylvain Gugger's avatar
Sylvain Gugger committed
65
66
67
68
69
70
BartForConditionalGeneration
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

.. autoclass:: transformers.BartForConditionalGeneration
    :members: generate, forward

Sam Shleifer's avatar
Sam Shleifer committed
71