reformer.rst 10.8 KB
Newer Older
Sylvain Gugger's avatar
Sylvain Gugger committed
1
2
3
4
5
6
7
8
9
10
11
12
.. 
    Copyright 2020 The HuggingFace Team. All rights reserved.

    Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
    the License. You may obtain a copy of the License at

        http://www.apache.org/licenses/LICENSE-2.0

    Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
    an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
    specific language governing permissions and limitations under the License.

Patrick von Platen's avatar
Patrick von Platen committed
13
Reformer
Sylvain Gugger's avatar
Sylvain Gugger committed
14
15
16
17
-----------------------------------------------------------------------------------------------------------------------

**DISCLAIMER:** This model is still a work in progress, if you see something strange, file a `Github Issue
<https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__.
Patrick von Platen's avatar
Patrick von Platen committed
18
19

Overview
Sylvain Gugger's avatar
Sylvain Gugger committed
20
21
22
23
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The Reformer model was proposed in the paper `Reformer: The Efficient Transformer
<https://arxiv.org/abs/2001.04451.pdf>`__ by Nikita Kitaev, 艁ukasz Kaiser, Anselm Levskaya.
Patrick von Platen's avatar
Patrick von Platen committed
24

Sylvain Gugger's avatar
Sylvain Gugger committed
25
The abstract from the paper is the following:
Patrick von Platen's avatar
Patrick von Platen committed
26

Sylvain Gugger's avatar
Sylvain Gugger committed
27
28
29
30
31
32
33
34
*Large Transformer models routinely achieve state-of-the-art results on a number of tasks but training these models can
be prohibitively costly, especially on long sequences. We introduce two techniques to improve the efficiency of
Transformers. For one, we replace dot-product attention by one that uses locality-sensitive hashing, changing its
complexity from O(L^2) to O(Llog(L)), where L is the length of the sequence. Furthermore, we use reversible residual
layers instead of the standard residuals, which allows storing activations only once in the training process instead of
N times, where N is the number of layers. The resulting model, the Reformer, performs on par with Transformer models
while being much more memory-efficient and much faster on long sequences.*

35
36
This model was contributed by `patrickvonplaten <https://huggingface.co/patrickvonplaten>`__. The Authors' code can be
found `here <https://github.com/google/trax/tree/master/trax/models/reformer>`__.
Patrick von Platen's avatar
Patrick von Platen committed
37
38

Axial Positional Encodings
Sylvain Gugger's avatar
Sylvain Gugger committed
39
40
41
42
43
44
45
46
47
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Axial Positional Encodings were first implemented in Google's `trax library
<https://github.com/google/trax/blob/4d99ad4965bab1deba227539758d59f0df0fef48/trax/layers/research/position_encodings.py#L29>`__
and developed by the authors of this model's paper. In models that are treating very long input sequences, the
conventional position id encodings store an embedings vector of size :math:`d` being the :obj:`config.hidden_size` for
every position :math:`i, \ldots, n_s`, with :math:`n_s` being :obj:`config.max_embedding_size`. This means that having
a sequence length of :math:`n_s = 2^{19} \approx 0.5M` and a ``config.hidden_size`` of :math:`d = 2^{10} \approx 1000`
would result in a position encoding matrix:
Patrick von Platen's avatar
Patrick von Platen committed
48
49
50
51

.. math::
    X_{i,j}, \text{ with } i \in \left[1,\ldots, d\right] \text{ and } j \in \left[1,\ldots, n_s\right] 

Sylvain Gugger's avatar
Sylvain Gugger committed
52
which alone has over 500M parameters to store. Axial positional encodings factorize :math:`X_{i,j}` into two matrices:
Patrick von Platen's avatar
Patrick von Platen committed
53
54
55
56

.. math::
    X^{1}_{i,j}, \text{ with } i \in \left[1,\ldots, d^1\right] \text{ and } j \in \left[1,\ldots, n_s^1\right] 

Sylvain Gugger's avatar
Sylvain Gugger committed
57
and
Patrick von Platen's avatar
Patrick von Platen committed
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74

.. math::
    X^{2}_{i,j}, \text{ with } i \in \left[1,\ldots, d^2\right] \text{ and } j \in \left[1,\ldots, n_s^2\right] 

with:

.. math::
    d = d^1 + d^2 \text{ and } n_s = n_s^1 \times n_s^2 .

Therefore the following holds:

.. math::
    X_{i,j} = \begin{cases}
                X^{1}_{i, k}, & \text{if }\ i < d^1 \text{ with } k = j \mod n_s^1 \\
                X^{2}_{i - d^1, l}, & \text{if } i \ge d^1 \text{ with } l = \lfloor\frac{j}{n_s^1}\rfloor
              \end{cases}

Sylvain Gugger's avatar
Sylvain Gugger committed
75
76
77
78
Intuitively, this means that a position embedding vector :math:`x_j \in \mathbb{R}^{d}` is now the composition of two
factorized embedding vectors: :math:`x^1_{k, l} + x^2_{l, k}`, where as the :obj:`config.max_embedding_size` dimension
:math:`j` is factorized into :math:`k \text{ and } l`. This design ensures that each position embedding vector
:math:`x_j` is unique.
Patrick von Platen's avatar
Patrick von Platen committed
79

Sylvain Gugger's avatar
Sylvain Gugger committed
80
81
Using the above example again, axial position encoding with :math:`d^1 = 2^5, d^2 = 2^5, n_s^1 = 2^9, n_s^2 = 2^{10}`
can drastically reduced the number of parameters to :math:`2^{14} + 2^{15} \approx 49000` parameters.
Patrick von Platen's avatar
Patrick von Platen committed
82

Sylvain Gugger's avatar
Sylvain Gugger committed
83
84
85
86
In practice, the parameter :obj:`config.axial_pos_embds_dim` is set to a tuple :math:`(d^1, d^2)` which sum has to be
equal to :obj:`config.hidden_size` and :obj:`config.axial_pos_shape` is set to a tuple :math:`(n_s^1, n_s^2)` which
product has to be equal to :obj:`config.max_embedding_size`, which during training has to be equal to the `sequence
length` of the :obj:`input_ids`.
Patrick von Platen's avatar
Patrick von Platen committed
87
88
89


LSH Self Attention
Sylvain Gugger's avatar
Sylvain Gugger committed
90
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Sylvain Gugger's avatar
Sylvain Gugger committed
91

Sylvain Gugger's avatar
Sylvain Gugger committed
92
93
94
95
96
In Locality sensitive hashing (LSH) self attention the key and query projection weights are tied. Therefore, the key
query embedding vectors are also tied. LSH self attention uses the locality sensitive hashing mechanism proposed in
`Practical and Optimal LSH for Angular Distance <https://arxiv.org/abs/1509.02897>`__ to assign each of the tied key
query embedding vectors to one of :obj:`config.num_buckets` possible buckets. The premise is that the more "similar"
key query embedding vectors (in terms of *cosine similarity*) are to each other, the more likely they are assigned to
Sylvain Gugger's avatar
Sylvain Gugger committed
97
the same bucket.
Sylvain Gugger's avatar
Sylvain Gugger committed
98

Sylvain Gugger's avatar
Sylvain Gugger committed
99
The accuracy of the LSH mechanism can be improved by increasing :obj:`config.num_hashes` or directly the argument
Sylvain Gugger's avatar
Sylvain Gugger committed
100
101
102
103
104
105
106
107
108
:obj:`num_hashes` of the forward function so that the output of the LSH self attention better approximates the output
of the "normal" full self attention. The buckets are then sorted and chunked into query key embedding vector chunks
each of length :obj:`config.lsh_chunk_length`. For each chunk, the query embedding vectors attend to its key vectors
(which are tied to themselves) and to the key embedding vectors of :obj:`config.lsh_num_chunks_before` previous
neighboring chunks and :obj:`config.lsh_num_chunks_after` following neighboring chunks.

For more information, see the `original Paper <https://arxiv.org/abs/2001.04451>`__ or this great `blog post
<https://www.pragmatic.ml/reformer-deep-dive/>`__.

Sylvain Gugger's avatar
Sylvain Gugger committed
109
110
111
112
113
Note that :obj:`config.num_buckets` can also be factorized into a list :math:`(n_{\text{buckets}}^1,
n_{\text{buckets}}^2)`. This way instead of assigning the query key embedding vectors to one of :math:`(1,\ldots,
n_{\text{buckets}})` they are assigned to one of :math:`(1-1,\ldots, n_{\text{buckets}}^1-1, \ldots,
1-n_{\text{buckets}}^2, \ldots, n_{\text{buckets}}^1-n_{\text{buckets}}^2)`. This is crucial for very long sequences to
save memory.
Sylvain Gugger's avatar
Sylvain Gugger committed
114
115
116
117
118
119
120
121

When training a model from scratch, it is recommended to leave :obj:`config.num_buckets=None`, so that depending on the
sequence length a good value for :obj:`num_buckets` is calculated on the fly. This value will then automatically be
saved in the config and should be reused for inference.

Using LSH self attention, the memory and time complexity of the query-key matmul operation can be reduced from
:math:`\mathcal{O}(n_s \times n_s)` to :math:`\mathcal{O}(n_s \times \log(n_s))`, which usually represents the memory
and time bottleneck in a transformer model, with :math:`n_s` being the sequence length.
Patrick von Platen's avatar
Patrick von Platen committed
122
123


Sylvain Gugger's avatar
Sylvain Gugger committed
124
125
Local Self Attention
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Patrick von Platen's avatar
Patrick von Platen committed
126

Sylvain Gugger's avatar
Sylvain Gugger committed
127
128
129
130
Local self attention is essentially a "normal" self attention layer with key, query and value projections, but is
chunked so that in each chunk of length :obj:`config.local_chunk_length` the query embedding vectors only attends to
the key embedding vectors in its chunk and to the key embedding vectors of :obj:`config.local_num_chunks_before`
previous neighboring chunks and :obj:`config.local_num_chunks_after` following neighboring chunks.
Patrick von Platen's avatar
Patrick von Platen committed
131

Sylvain Gugger's avatar
Sylvain Gugger committed
132
133
134
Using Local self attention, the memory and time complexity of the query-key matmul operation can be reduced from
:math:`\mathcal{O}(n_s \times n_s)` to :math:`\mathcal{O}(n_s \times \log(n_s))`, which usually represents the memory
and time bottleneck in a transformer model, with :math:`n_s` being the sequence length.
Patrick von Platen's avatar
Patrick von Platen committed
135
136


Sylvain Gugger's avatar
Sylvain Gugger committed
137
138
Training
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Patrick von Platen's avatar
Patrick von Platen committed
139

Sylvain Gugger's avatar
Sylvain Gugger committed
140
141
142
143
During training, we must ensure that the sequence length is set to a value that can be divided by the least common
multiple of :obj:`config.lsh_chunk_length` and :obj:`config.local_chunk_length` and that the parameters of the Axial
Positional Encodings are correctly set as described above. Reformer is very memory efficient so that the model can
easily be trained on sequences as long as 64000 tokens.
Patrick von Platen's avatar
Patrick von Platen committed
144

Sylvain Gugger's avatar
Sylvain Gugger committed
145
For training, the :class:`~transformers.ReformerModelWithLMHead` should be used as follows:
Patrick von Platen's avatar
Patrick von Platen committed
146

Sylvain Gugger's avatar
Sylvain Gugger committed
147
.. code-block::
Patrick von Platen's avatar
Patrick von Platen committed
148

149
150
    input_ids = tokenizer.encode('This is a sentence from the training data', return_tensors='pt')
    loss = model(input_ids, labels=input_ids)[0]
Patrick von Platen's avatar
Patrick von Platen committed
151
152
153


ReformerConfig
Sylvain Gugger's avatar
Sylvain Gugger committed
154
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Patrick von Platen's avatar
Patrick von Platen committed
155
156
157
158
159
160

.. autoclass:: transformers.ReformerConfig
    :members:


ReformerTokenizer
Sylvain Gugger's avatar
Sylvain Gugger committed
161
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Patrick von Platen's avatar
Patrick von Platen committed
162
163

.. autoclass:: transformers.ReformerTokenizer
Sylvain Gugger's avatar
Sylvain Gugger committed
164
    :members: save_vocabulary
Patrick von Platen's avatar
Patrick von Platen committed
165
166


167
168
169
170
171
172
173
ReformerTokenizerFast
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

.. autoclass:: transformers.ReformerTokenizerFast
    :members:


Patrick von Platen's avatar
Patrick von Platen committed
174
ReformerModel
Sylvain Gugger's avatar
Sylvain Gugger committed
175
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Patrick von Platen's avatar
Patrick von Platen committed
176
177

.. autoclass:: transformers.ReformerModel
Sylvain Gugger's avatar
Sylvain Gugger committed
178
    :members: forward
Patrick von Platen's avatar
Patrick von Platen committed
179
180
181


ReformerModelWithLMHead
Sylvain Gugger's avatar
Sylvain Gugger committed
182
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Patrick von Platen's avatar
Patrick von Platen committed
183
184

.. autoclass:: transformers.ReformerModelWithLMHead
Sylvain Gugger's avatar
Sylvain Gugger committed
185
    :members: forward
186
187


188
ReformerForMaskedLM
Sylvain Gugger's avatar
Sylvain Gugger committed
189
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
190
191

.. autoclass:: transformers.ReformerForMaskedLM
Sylvain Gugger's avatar
Sylvain Gugger committed
192
    :members: forward
193
194


195
ReformerForSequenceClassification
Sylvain Gugger's avatar
Sylvain Gugger committed
196
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
197
198

.. autoclass:: transformers.ReformerForSequenceClassification
Sylvain Gugger's avatar
Sylvain Gugger committed
199
    :members: forward
200
201


202
ReformerForQuestionAnswering
Sylvain Gugger's avatar
Sylvain Gugger committed
203
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
204
205

.. autoclass:: transformers.ReformerForQuestionAnswering
Sylvain Gugger's avatar
Sylvain Gugger committed
206
    :members: forward