Try running Word2vec model on AWS Lambda
Introduction
This is the article on the 23rd day of mediba Advent Calendar 2020.
This is Nozaki from mediba. Currently, as a back-end engineer, I am in charge of products such as au Web Portal, and operate back-end applications and infrastructure range. He also belongs to the Tech Lead team at the Technology Center.
What's new in AWS Lambda – Container Image Support | Amazon Web Services Blog (https://aws.amazon.com/jp/blogs/news/new-for-aws-lambda-container-image-support/)
AWS Lambda container image support announced this month. You can now package and deploy `Lambda functions as container images up to 10 GB. As you can see, it is now possible to create packages of larger sizes than before. It seems that the container can contain machine learning models and dictionary data used in natural language processing (NLP).
Therefore, in this article, we will include a trained Word2vec model in a container and run it on AWS Lambda. The function to create is to return a word similar to the input word and its similarity (value in the range of 0 to 1), and this can be called via API. The trained Word2vec model is used in the part that returns similar words.
The overall picture is as shown in the figure.
background
The background is why this kind of verification is done this time. au Web Portal has a product that distributes news articles. Some of these features are provided by applying natural language processing (NLP) to news articles.
Currently this system is built on EC2, In the future, we are considering changing to containerization or serverless. As part of this study, I came up with something like the subject.
Procedures performed
AWS Serverless Application Model-Amazon Web Services (https://aws.amazon.com/jp/serverless/sam/) This time, we will use SAM as the framework of AWS Lambda.
- sam init
We will proceed with the following version.
$ sam --version
SAM CLI, version 1.13.2
Initialize the sam project.
$ sam init
Which template source would you like to use?
1 - AWS Quick Start Templates
2 - Custom Template Location
Choice: 1
What package type would you like to use?
1 - Zip (artifact is a zip uploaded to S3)
2 - Image (artifact is an image uploaded to an ECR image repository)
Package type: 2
Which base image would you like to use?
1 - amazon/nodejs12.x-base
2 - amazon/nodejs10.x-base
3 - amazon/python3.8-base
4 - amazon/python3.7-base
5 - amazon/python3.6-base
6 - amazon/python2.7-base
7 - amazon/ruby2.7-base
8 - amazon/ruby2.5-base
9 - amazon/go1.x-base
10 - amazon/java11-base
11 - amazon/java8.al2-base
12 - amazon/java8-base
13 - amazon/dotnetcore3.1-base
14 - amazon/dotnetcore2.1-base
Base image: 3
Project name [sam-app]: sam-wiki-entity-vectors
The following projects can be created.
- The directories and files to be included in the subsequent work are also listed.
.
├── README.md
├── __init__.py
├── events
│ └── event.json
├── hello_world
│ ├── Dockerfile
│ ├── __init__.py
│ ├── app.py
│ ├── requirements.txt
│ └── tohoku_entity_vector // 2.Directory created by getting the model
│ ├── entity_vector.model.bin
│ └── entity_vector.model.txt
├── samconfig.toml // 7.Created with sam deploy
├── template.yaml
└── tests
├── __init__.py
└── unit
├── __init__.py
└── test_handler.py
2. Get the model
This time, we will use the trained model published at Tohoku University Inui-Okazaki Laboratory.
Japanese Wikipedia Entity Vector
Obtain the following two files from the above site, create a tohoku_entity_vector directory in the sam project, and store them.
--Binary file (entity_vector.model.bin)
--Text file (entity_vector.model.txt)
$ du -h hello_world/tohoku_entity_vector
2.6G
These file sizes were 2.6GB.
3. Code changes
Modify the following files in the project.
requirements.txt
Add the following,
gensim
Gensim is a Python library for topic modeling, document indexing, and similarity search using a large corpus.
Dockerfile
Modify the COPY location as follows.
In addition to tohoku_entity_vector / to be copied, include the model in the Docker container.
COPY app.py requirements.txt tohoku_entity_vector/ ./
app.py
Modify as follows.
import os
import json
from gensim.models import KeyedVectors
def lambda_handler(event, context):
word = event.get('queryStringParameters').get('word')
model = KeyedVectors.load_word2vec_format(
'./entity_vector.model.bin', binary=True)
result = model.most_similar('[' + word + ']')
return {
"statusCode": 200,
"body": json.dumps(result, indent=2, ensure_ascii=False),
}
most_similar returns words that are similar to the input word and their similarity (0 to 1) in order of similarity.
gensim.models.Word2Vec.most_similar
template.yaml
Change the Globals section, memory and timeout values.
Globals:
Function:
Timeout: 900
MemorySize: 10240
In both cases, try raising the above quota to the upper limit.
- sam build
Execute the following command with docker running.
$ sam build
5. Confirmation of local operation
You can test the api with the following command in the local environment.
$ sam local start-api
Make the following request in your browser.
<img width="500" " src="https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/889583/865510a8-5426-a876-8960-5532b3a76133.png ">
For the input Kawasaki Frontale entered in the query parameter
As similar words, words such as Jubilo Iwata`` Oita Trinita and their similarities were returned as expected. It looks okay.
6. Create ECR repository
Create an ECR repository for pushing Docker images with the following command.
aws ecr create-repository --repository-name repository-name
7.sam deploy
#First time
$ sam deploy --guided
#After the second time
$ sam deploy
For the first deployment, select the deployment settings by adding the --guided option. samconfig.toml is created.
In the second and subsequent deployments, it will be deployed based on this file.
The deploy command pushes the image to ECR and deploys the function to AWS Lambda.
8. Confirmation of operation
First, check the image of ECR. It is PUSHed and you can see that it has about 1.8GB.
<img width="600" " src="https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/889583/ff04e304-9307-13ce-0fa4-4dda7aa0ebcf.png ">
Next, check the API.
This time, enter FC Barcelona entered in the query parameter and enter
As similar words, words such as Real Madrid`` AC Milan and their similarities were returned as expected. As expected.
The API latency was around 10 seconds.
<img width="700" " src="https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/889583/0eee8fa8-ba85-dd7d-9509-32a871b76122.png ">
Machine learning system design pattern
As I learned from writing this article, The design pattern of the machine learning system was published by Mercari. Publish the design pattern of the machine learning system. | Mercari Engineering Use cases and advantages and disadvantages are organized by specific pattern.
The configuration this time is Web single pattern | ml-system-design-pattern It looks like this pattern.
Summary
--I tried running the Word2vec model on AWS Lambda
--The Word2vec model can now be included in a container, simplifying application development.
--Until now, when handling large files with AWS Lambda, service linkage such as EFS was indispensable.
--It is necessary to consider how to optimize the function memory allocation of AWS Lambda.
Reference article
Recommended Posts