squeezebert.rst 4.75 KB
Newer Older
1
SqueezeBERT
Lysandre Debut's avatar
Lysandre Debut committed
2
-----------------------------------------------------------------------------------------------------------------------
3
4

Overview
Lysandre Debut's avatar
Lysandre Debut committed
5
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
6

Sylvain Gugger's avatar
Sylvain Gugger committed
7
8
9
10
The SqueezeBERT model was proposed in `SqueezeBERT: What can computer vision teach NLP about efficient neural networks?
<https://arxiv.org/abs/2006.11316>`__ by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, Kurt W. Keutzer. It's a
bidirectional transformer similar to the BERT model. The key difference between the BERT architecture and the
SqueezeBERT architecture is that SqueezeBERT uses `grouped convolutions <https://blog.yani.io/filter-group-tutorial>`__
11
12
13
14
instead of fully-connected layers for the Q, K, V and FFN layers.

The abstract from the paper is the following:

Sylvain Gugger's avatar
Sylvain Gugger committed
15
16
17
18
19
20
21
22
23
24
25
26
*Humans read and write hundreds of billions of messages every day. Further, due to the availability of large datasets,
large computing systems, and better neural network models, natural language processing (NLP) technology has made
significant strides in understanding, proofreading, and organizing these messages. Thus, there is a significant
opportunity to deploy NLP in myriad applications to help web users, social networks, and businesses. In particular, we
consider smartphones and other mobile devices as crucial platforms for deploying NLP models at scale. However, today's
highly-accurate NLP neural network models such as BERT and RoBERTa are extremely computationally expensive, with
BERT-base taking 1.7 seconds to classify a text snippet on a Pixel 3 smartphone. In this work, we observe that methods
such as grouped convolutions have yielded significant speedups for computer vision networks, but many of these
techniques have not been adopted by NLP neural network designers. We demonstrate how to replace several operations in
self-attention layers with grouped convolutions, and we use this technique in a novel network architecture called
SqueezeBERT, which runs 4.3x faster than BERT-base on the Pixel 3 while achieving competitive accuracy on the GLUE test
set. The SqueezeBERT code will be released.*
27
28
29

Tips:

Sylvain Gugger's avatar
Sylvain Gugger committed
30
31
32
33
34
- SqueezeBERT is a model with absolute position embeddings so it's usually advised to pad the inputs on the right
  rather than the left.
- SqueezeBERT is similar to BERT and therefore relies on the masked language modeling (MLM) objective. It is therefore
  efficient at predicting masked tokens and at NLU in general, but is not optimal for text generation. Models trained
  with a causal language modeling (CLM) objective are better in that regard.
35
36
37
38
- For best results when finetuning on sequence classification tasks, it is recommended to start with the
  `squeezebert/squeezebert-mnli-headless` checkpoint.

SqueezeBertConfig
Lysandre Debut's avatar
Lysandre Debut committed
39
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
40
41
42
43
44
45

.. autoclass:: transformers.SqueezeBertConfig
    :members:


SqueezeBertTokenizer
Lysandre Debut's avatar
Lysandre Debut committed
46
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
47
48
49
50
51
52
53

.. autoclass:: transformers.SqueezeBertTokenizer
    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
        create_token_type_ids_from_sequences, save_vocabulary


SqueezeBertTokenizerFast
Lysandre Debut's avatar
Lysandre Debut committed
54
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
55
56
57
58
59
60

.. autoclass:: transformers.SqueezeBertTokenizerFast
    :members:


SqueezeBertModel
Lysandre Debut's avatar
Lysandre Debut committed
61
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
62
63
64
65
66
67

.. autoclass:: transformers.SqueezeBertModel
    :members:


SqueezeBertForMaskedLM
Lysandre Debut's avatar
Lysandre Debut committed
68
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
69
70
71
72
73
74

.. autoclass:: transformers.SqueezeBertForMaskedLM
    :members:


SqueezeBertForSequenceClassification
Lysandre Debut's avatar
Lysandre Debut committed
75
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
76
77
78
79
80
81

.. autoclass:: transformers.SqueezeBertForSequenceClassification
    :members:


SqueezeBertForMultipleChoice
Lysandre Debut's avatar
Lysandre Debut committed
82
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
83
84
85
86
87
88

.. autoclass:: transformers.SqueezeBertForMultipleChoice
    :members:


SqueezeBertForTokenClassification
Lysandre Debut's avatar
Lysandre Debut committed
89
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
90
91
92
93
94
95

.. autoclass:: transformers.SqueezeBertForTokenClassification
    :members:


SqueezeBertForQuestionAnswering
Lysandre Debut's avatar
Lysandre Debut committed
96
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
97
98
99

.. autoclass:: transformers.SqueezeBertForQuestionAnswering
    :members: