.
In layman's terms, the self-attention mechanism allows the inputs to interact with each other (“self”) and find out who they should pay more attention to (“attention”). The outputs are aggregates of these interactions and attention scores.
.
https://towardsdatascience.com/illustrated-self-attention-2d627e33b20a
.
Illustrated: Self-Attention
A step-by-step guide to self-attention with illustrations and code
The illustrations are best viewed on the Desktop. A Colab version can be found here (thanks to Manuel Romero!).
Changelog:
12 Jan 2022 — Improve clarity
5 Jan 2022 — Fix typos and improve clarity
What do BERT, RoBERTa, ALBERT, SpanBERT, DistilBERT, SesameBERT, SemBERT, SciBERT, BioBERT, MobileBERT, TinyBERT and CamemBERT all have in common? And I’m not looking for the answer “BERT” 🤭.
Answer: self-attention 🤗. We are not only talking about architectures bearing the name “BERT’ but, more correctly, Transformer-based architectures. Transformer-based architectures, which are primarily used in modelling language understanding tasks, eschew recurrence in neural networks and instead trust entirely on self-attention mechanisms to draw global dependencies between inputs and outputs. But what’s the math behind this?
That’s what we’re going to find out today. The main content of this post is to walk you through the mathematical operations involved in a self-attention module. By the end of this article, you should be able to write or code a self-attention module from scratch.
This article does not aim to provide the intuitions and explanations behind the different numerical representations and mathematical operations in the self-attention module. It also does not seek to demonstrate the why’s and how-exactly’s of self-attention in Transformers (I believe there’s a lot out there already). Note that the difference between attention and self-attention is also not detailed in this article.
Content
Now let’s get on to it!
0. What is self-attention?
If you think that self-attention is similar, the answer is yes! They fundamentally share the same concept and many common mathematical operations.
A self-attention module takes in n inputs and returns n outputs. What happens in this module? In layman’s terms, the self-attention mechanism allows the inputs to interact with each other (“self”) and find out who they should pay more attention to (“attention”). The outputs are aggregates of these interactions and attention scores.
1. Illustrations
The illustrations are divided into the following steps:
- Prepare inputs
- Initialise weights
- Derive key, query and value
- Calculate attention scores for Input 1
- Calculate softmax
- Multiply scores with values
- Sum weighted values to get Output 1
- Repeat steps 4–7 for Input 2 & Input 3
Note
In practice, the mathematical operations are vectorised, i.e. all the inputs undergo the mathematical operations together. We’ll see this later in the Code section.
Step 1: Prepare inputs
We start with 3 inputs for this tutorial, each with dimension 4.
Input 1: [1, 0, 1, 0]
Input 2: [0, 2, 0, 2]
Input 3: [1, 1, 1, 1]
Step 2: Initialise weights
Every input must have three representations (see diagram below). These representations are called key (orange), query (red), and value (purple). For this example, let’s take that we want these representations to have a dimension of 3. Because every input has a dimension of 4, each set of the weights must have a shape of 4×3.
Note
We’ll see later that the dimension of value is also the output dimension.
To obtain these representations, every input (green) is multiplied with a set of weights for keys, a set of weights for querys (I know that’s not the correct spelling), and a set of weights for values. In our example, we initialise the three sets of weights as follows.
Weights for key:
[[0, 0, 1],
[1, 1, 0],
[0, 1, 0],
[1, 1, 0]]
Weights for query:
[[1, 0, 1],
[1, 0, 0],
[0, 0, 1],
[0, 1, 1]]
Weights for value:
[[0, 2, 0],
[0, 3, 0],
[1, 0, 3],
[1, 1, 0]]
Notes
In a neural network setting, these weights are usually small numbers, initialised randomly using an appropriate random distribution like Gaussian, Xavier and Kaiming distributions. This initialisation is done once before training.
Step 3: Derive key, query and value
Now that we have the three sets of weights, let’s obtain the key, query and value representations for every input.
Key representation for Input 1:
[0, 0, 1]
[1, 0, 1, 0] x [1, 1, 0] = [0, 1, 1]
[0, 1, 0]
[1, 1, 0]
Use the same set of weights to get the key representation for Input 2:
[0, 0, 1]
[0, 2, 0, 2] x [1, 1, 0] = [4, 4, 0]
[0, 1, 0]
[1, 1, 0]
Use the same set of weights to get the key representation for Input 3:
[0, 0, 1]
[1, 1, 1, 1] x [1, 1, 0] = [2, 3, 1]
[0, 1, 0]
[1, 1, 0]
A faster way is to vectorise the above operations:
[0, 0, 1]
[1, 0, 1, 0] [1, 1, 0] [0, 1, 1]
[0, 2, 0, 2] x [0, 1, 0] = [4, 4, 0]
[1, 1, 1, 1] [1, 1, 0] [2, 3, 1]
Let’s do the same to obtain the value representations for every input:
[0, 2, 0]
[1, 0, 1, 0] [0, 3, 0] [1, 2, 3]
[0, 2, 0, 2] x [1, 0, 3] = [2, 8, 0]
[1, 1, 1, 1] [1, 1, 0] [2, 6, 3]
and finally the query representations:
[1, 0, 1]
[1, 0, 1, 0] [1, 0, 0] [1, 0, 2]
[0, 2, 0, 2] x [0, 0, 1] = [2, 2, 2]
[1, 1, 1, 1] [0, 1, 1] [2, 1, 3]
Notes
In practice, a bias vector may be added to the product of matrix multiplication.
Step 4: Calculate attention scores for Input 1
To obtain attention scores, we start with taking a dot product between Input 1’s query (red) with all keys (orange), including itself. Since there are 3 key representations (because we have 3 inputs), we obtain 3 attention scores (blue).
[0, 4, 2]
[1, 0, 2] x [1, 4, 3] = [2, 4, 4]
[1, 0, 1]
Notice that we only use the query from Input 1. Later we’ll work on repeating this same step for the other querys.
Note
The above operation is known as dot product attention, one of the several score functions. Other score functions include scaled dot product and additive/concat.
Step 5: Calculate softmax
Take the softmax across these attention scores (blue).
softmax([2, 4, 4]) = [0.0, 0.5, 0.5]
Note that we round off to 1 decimal place here for readability.
Step 6: Multiply scores with values
The softmaxed attention scores for each input (blue) is multiplied by its corresponding value (purple). This results in 3 alignment vectors (yellow). In this tutorial, we’ll refer to them as weighted values.
1: 0.0 * [1, 2, 3] = [0.0, 0.0, 0.0]
2: 0.5 * [2, 8, 0] = [1.0, 4.0, 0.0]
3: 0.5 * [2, 6, 3] = [1.0, 3.0, 1.5]
Step 7: Sum weighted values to get Output 1
Take all the weighted values (yellow) and sum them element-wise:
[0.0, 0.0, 0.0]
+ [1.0, 4.0, 0.0]
+ [1.0, 3.0, 1.5]
-----------------
= [2.0, 7.0, 1.5]
The resulting vector [2.0, 7.0, 1.5] (dark green) is Output 1, which is based on the query representation from Input 1 interacting with all other keys, including itself.
Step 8: Repeat for Input 2 & Input 3
Now that we’re done with Output 1, we repeat Steps 4 to 7 for Output 2 and Output 3. I trust that I can leave you to work out the operations yourself 👍🏼.
Notes
The dimension of query and key must always be the same because of the dot product score function. However, the dimension of value may be different from query and key. The resulting output will consequently follow the dimension of value.
2. Code
Here is the code in PyTorch 🤗, a popular deep learning framework in Python. To enjoy the APIs for @ operator, .T and None indexing in the following code snippets, make sure you’re on Python≥3.6 and PyTorch 1.3.1. Just follow along and copy-paste these in a Python/IPython REPL or Jupyter Notebook.
Step 1: Prepare inputs
Step 2: Initialise weights
Step 3: Derive key, query and value
Step 4: Calculate attention scores
Step 5: Calculate softmax
Step 6: Multiply scores with values
Step 7: Sum weighted values
Note
PyTorch has provided an API for this called nn.MultiheadAttention. However, this API requires that you feed in key, query and value PyTorch tensors. Moreover, the outputs of this module undergo a linear transformation.
3. Extending to Transformers
So, where do we go from here? Transformers! Indeed we live in exciting times of deep learning research and high compute resources. The transformer is the incarnation from Attention Is All You Need, originally born to perform neural machine translation. Researchers picked up from here, reassembling, cutting, adding and extending the parts, and extending its usage to more language tasks.
Here I will briefly mention how we can extend self-attention to a Transformer architecture.
Within the self-attention module:
- Dimension
- Bias
Inputs to the self-attention module:
- Embedding module
- Positional encoding
- Truncating
- Masking
Adding more self-attention modules:
- Multihead
- Layer stacking
Modules between self-attention modules:
- Linear transformations
- LayerNorm
That’s all folks! Hope you find the content easy to digest. Is there something that you think I should add or elaborate on further in this article? Do drop a comment! Also, do check out an illustration I created for attention below
References
Attention Is All You Need (arxiv.org)
The Illustrated Transformer (jalammar.github.io)
Related Articles
Attn: Illustrated Attention (towardsdatascience.com)
If you like my content and haven’t already subscribed to Medium, subscribe via my referral link here! NOTE: A portion of your membership fees will be apportioned to me as referral fees.
Special thanks to Xin Jie, Serene, Ren Jie, Kevin and Wei Yih for ideas, suggestions and corrections to this article.
Follow me on Twitter @remykarem for digested articles and other tweets on AI, ML, Deep Learning and Python.
🇸🇬 Software Engineer at GovTech • MComp AI at NUS
.
