We Raised $8M Series A to Continue Building Experiment Tracking and Model Registry That “Just Works”

Read more

Blog » ML Model Management » Best Practices When Working With Docker for Machine Learning

Best Practices When Working With Docker for Machine Learning

Application containers may be created, deployed, and executed using the Docker tool. It’s just a packed bundle of application code and the libraries and other dependencies that are needed for it to run. Once executed, a Docker Image turns into a Container and contains all the components required to run an application.

However, what’s the point of doing so? As a data scientist or machine learning engineer, how is this useful? A keyword, especially for data scientists is reproducibility. With Docker, it is possible to develop a basic classifier that can guarantee great reproducibility if your system has a version of libraries and programs that are not updated and cannot be updated since you have certain applications that operate on these versions. This is a decent and easy use case.

But before we dig deeper, let’s start with some basics.

What is Docker?

Docker is a tool for developing and deploying applications using containers. A container is built on the concept that you may bundle your code together with its dependencies into one deployable unit. Containers have been in use for a long time now. Some claim that they were created by Sun Microsystems and published as part of Solaris 10 in 2005 as Zones, while others claim that BSD Jails were the first container technology to be implemented. Consider the shipping containers that are used for intermodal transport as an example of a visual explanation. 

Once you have placed your product (code and dependencies) in the container, you may send it by boat or train (laptop or cloud), and when it gets to its destination, it continues to function (run) just as before the shipment was made.

Cargo ship analogy

Cargo ship analogy
Docker – cargo ship analogy | Source

There is a useful comparison that may help explain how Docker works. As you may be aware, the items are transported across the world in various containers. Furthermore, this segregation is important so that the commodities in one container do not impact the goods in the container next to them.

In order to carry diverse items, containers may be of a single size and material, or they may be of varied sizes and materials depending on the sort of goods that are being transported.

There is a manager who is generally responsible for this boarding, de-boarding, and maintaining the ship’s containers. A Docker engine running on your server serves as the manager in this scenario.

As the ship’s owner, you may instruct the ship’s manager to load and unload containers from and to the ship. Similarly, docker can also execute, deploy, and shut down your application on a server.

For the container, you’ll need a picture that serves as a template for the size and substance of the container. It’s here that you’ll need a blueprint i.e. the image, from which to build this container. Docker hub is a great place to store your image, and the docker engine on your server will take care of the rest.

What role does Docker play in the field of machine learning?

Reproducibility in Docker

It’s important to be able to reproduce your results as accurately as possible whether you’re developing models or doing data analysis. Consider the following scenario: you’ve been requested to do exploratory data analysis on a data collection. You create a local environment that includes the dataset and a few Python libraries, and then you launch your Jupyter notebook. After analyzing your dataset and developing a preliminary model, you decide to share your findings with your colleagues and friends. 

The problem is that for your coworkers to replicate your analysis, they must replicate your whole stack as well. Docker allows you to quickly and easily replicate your working environment. It makes it possible to standardize the version of used libraries in a project, the random seeds, and even the operating system, so that the same results may be generated over and over again by a colleague working on a different machine.

Portability in Docker

As a Data Scientist, it is very beneficial to be able to move smoothly from working on your local workstation to external clusters that provide extra resources such as CPU, RAM, GPUs, etc. In addition, we want to be able to experiment with new frameworks and tools that are being created and shared by the community.

Docker enables you to bundle your code and dependencies into containers that can subsequently be ported to various computers, even if the underlying hardware, operating system, and so on vary from the original machine(s). Another benefit of mobility is the ability to work on projects with a variety of different colleagues with relative simplicity. 

As a data scientist, before you can begin working on an established codebase, you don’t have to spend an entire day (or week) setting up the environment. If the project has been “Dockerized”, you will be able to begin working on it immediately.

Deployment in Docker

The use of Docker simplifies the process of deploying machine learning models. It’s a matter of whether or not you want to share your model with others. It’s as simple as wrapping your model in an API and putting it in a container utilizing Kubernetes technology. A few small steps have been omitted from this section. Recall that the process of transitioning from Docker to the deployment of a model is rather simple. Container management technologies and methods may be used for our model deployment.

Integration in Docker

Cloud providers and physical servers may be provisioned using Docker Cloud to construct Docker nodes. Install the Docker Cloud agent on your physical server or connect your cloud provider credentials securely. You may then “construct node clusters” in a matter of seconds.

AWS, Azure, DigitalOcean, Packet, and Softlayer are all connected with Docker Cloud for provisioning and administration. To manage nodes on infrastructure that is not supported by Docker Cloud (Google Cloud, Rackspace), you may also manage nodes on physical Linux systems that are connected to the internet.

Best practices to use Docker for Machine Learning (ML)

Best practices to use Docker for ML
Docker explained | Source

1. Building Docker images

What is a Docker file?

A Dockerfile is a text document containing all the commands required to generate a specific image that is read by Docker to build that image automatically. There should be as little setting or updating as possible when creating an image using the docker image build command, so that containers may be paused and/or destroyed at any moment and rebuilt or replaced with little effort.

What is an image?

Using Docker images is familiar to anybody who has used virtual machines other than Docker. Images are more often referred to as “snapshots” in other virtual machine settings. They’re a snapshot of a Docker virtual machine at a certain time and place. While a virtual machine snapshot is similar to a Docker image, there are several differences. 

As a starting point, Docker images are immutable. It’s possible to erase a record after you’ve produced it, but you can’t change it. Create a fresh image from scratch if you require a new snapshot.

In the Docker world, immutability (the ability to not change) is a valuable tool. Once you’ve set up your Docker virtual machine and created an image, you can be certain the image will always function. This makes it simple to experiment with new features in your environment.

It’s possible that you’ll experiment with new applications or restructure your files. As a result of this, you can be certain that your functioning instance will not be impacted. Your Docker virtual machine will always be able to shut down and resume using your current image as if nothing changed at all.

Best practices for building images

  • Use official images: In projects, it’s suggested that you use the Docker Official Images if you’re a new user of Docker. There is comprehensive documentation for these photos, and they encourage best practices while also catering to the broadest range of potential applications.

A typical reason for modifying Docker images is to reduce the size of the image. When installing modules that need optimized code, a full build toolchain is included with many programming language stack images. A more experienced user may create a custom image that only included the pre-compiled libraries that were absolutely essential, thereby saving disk space.

  • For the best results, use a solid base image: When you build containers on top of an untrusted or unmaintained image, you’ll be inheriting all of that image’s flaws and vulnerabilities. So as a general rule of thumb, you should go by the following rules:
    • Trusted repositories and suppliers should be preferred over pictures created by unidentified individuals.
    • Check the image source and the Dockerfile before building your own base image when utilizing custom images. An image published in a public registry is not guaranteed to be produced from a Dockerfile. There is no guarantee that it will be updated in the future.
  • .dockerignore: To prevent undesirable files and directories from being contained in the image when we run the docker build command, we may create a .dockerignore file that contains the files and directories that we don’t want in the image. The application’s processing time could come down if this file isn’t included.
  • Containers must be ephemeral: Ephemeral containers should be generated by the image specified in your Dockerfile. It is referred to as an ephemeral container when it can be halted and destroyed before being rebuilt and replaced with a new container, requiring the least setup and configuration. Disposable containers such as ephemeral ones might be regarded as such. For each new container instance, there is no connection to the prior ones. You should use as many ephemeral patterns as feasible when creating Docker images.
  • Avoid installing unnecessary applications: With the promise of utilizing it later, it might be tempting to install more than you need but to decrease the image’s size and complexity, just install what is actually required. For example, it’s a good idea to remove the curl package once you’ve used it, since many times it’s just needed once during the building of a container.
  • Spread each argument into several lines: There are many advantages to sorting arguments in this manner. You can more easily determine whether there have been modifications, and it helps in the detection of duplication in the package.

Let’s see how it works in the example below!

Example I

RUN apt-get update && apt-get install -y php5-cli php5 php5-gd php5-LDAP php5-mysql apache2 php5-pgsql libapache2-mod-php5

Example II

RUN apt-get update && \
    apt-get install -y \
      apache2 \
      libapache2-mod-php5 \
      php5 \
      php5-cli \
      php5-gd \
      php5-LDAP \
      php5-mysql \
      php5-pgsql

Lines that are not ordered are difficult to read. Due to the amount of mental effort and the focus required to comprehend them, we tend to skip over them during code reviews. So, which way do you think should be preferred?

  • Take advantage of Docker caches: A fascinating and helpful feature of Docker images is the Layer. Layers are used to build Docker images, and each layer has a specific purpose. Dockerfile instructions relate to each layer and describe the image’s file system changes between pre-execution and post-execution. Layers are cached by Docker in order to save build times. Rather than re-creating a similar layer, Docker will simply reuse an existing layer from the cache.

On the other side, adding layers that aren’t absolutely essential increases the amount of work. Many unneeded layers have performance concerns since Docker layers are filesystems. It is more efficient to establish a single cache with a single RUN command that applies all dependencies rather than to divide it into many levels since each run command produces a new layer. By finding cacheable layers and using them, you will save a lot of time in the long run.

  • Reduce the number of layers by condensing your assertions into one: Each instruction in the Dockerfile adds a layer to the image. The number of instructions and layers should be maintained to a minimum since this has an impact on performance and build times. So when performing a RUN, COPY, or ADD command attempt to execute this in a single sentence.
FROM alpine:3.4

RUN apk update
RUN apk add curl
RUN apk add vim
RUN apk add git
FROM alpine:3.4

RUN apk update && \
       apk add curl && \
       apk add vim && \
       apk add git

2. Building containers

What is a container?

  • Applications may be packaged in containers that are nothing but isolated environments. It is the goal of the containers to separate and enable the mobility of applications across multiple platforms.
  • In a container, an image offers all the required files to operate a set of processes. Via containers, applications are isolated from other processes that share the same kernel, so no one process can affect the work of others.
  • A cluster is a collection of containers. In a cluster, resources such as storage and processing power may be shared, making it feasible to operate dozens or even hundreds of containers concurrently.
  • Containers are very useful in the context of software development. The DevOps approach is based on their applicability, which assists both the Operations and Development departments. When it comes to developing applications, containers are agnostic to the environment’s features, such as the operating system, so they can be easily shared and accessed, and when it comes to running applications on the same kernel, they’re not as complex as virtual machines.
  • Despite their resemblance, a container is not a virtual machine because containers may share the same kernel as the operating system, but virtual machines can’t.

Best practices for building containers

  1. One container for each process

It is a typical error to perceive containers as virtual computers that can run several distinct aspects at the same time. A container may work in this fashion, but this diminishes most of the benefits of container architecture. For example, it’s tempting to run all three components of the standard Apache/MySQL/PHP stack simultaneously. However, the suggested approach is to utilize two or three distinct containers: one for Apache, one for MySQL, and maybe one for PHP if you are using PHP-FPM.

Since a container is meant to have the same lifetime as the app it contains, each container can only have one app. When a container is launched, so is the app, and when the app is closed, so is the container. The following diagram depicts this recommended practice.

Good practice
.
└── parent_process/
    ├── child process I/
    │   ├── child process I.I
    │   └── child process I.II
    └── child process II/
        ├── child process II.I
        └── child process II.II
Bad practice
.
├── parent_process I/
│   ├── child process I.I
│   └── child process I.II
--------------------------	
└── parent process II/
    ├── child process II.I
    └── child process II.II

There are several reasons why it is advised to run only one process per container:

  • Let’s say that you have two web servers and a database in your application architecture. It is possible to run all three in one container, but it is preferable to run each service in a distinct container so that it is simpler to reuse and grow.
  • A separate container for each service allows you to grow one of your web servers horizontally to accommodate the extra traffic. 
  • You may need a containerized database in the future for another service. In such a case, the same database container may be reused without the need for two additional services. 
  • Logging becomes significantly more difficult when containers are coupled. 
  • With a smaller surface area to deal with, it’s considerably simpler to produce security updates or troubleshoot a problem.

2. Tag your containers

When it comes to maintaining images, Docker tagging is an incredibly useful tool. It assists in maintaining several versions of a docker image. Below is an example of constructing a docker image with tag name v1.0.1

docker build -t geekflare/ubuntu:v1.0.1

There are a couple of tags that you can use:

  • Use stable tags for retaining the basic image of the container. Avoid utilizing these tags for deployment containers since these tags will get changed often, and it might lead to discrepancies in the production environment.
  • Use unique tags for deployments. Using unique tags, you may extend your production cluster to numerous nodes with ease. It prevents inconsistencies, and hosts will not fetch any other docker image version.

It is a recommended practice to disable write-enable on the deployed image tags. This prevents the deployed image from being accidentally deleted from the registry.

3. Running your model in Docker

Why put a machine learning model in Docker?

Challenges

  • The process of building a machine learning model that runs on our PC isn’t that difficult but working with a client who wants to utilize the model at scale in different sorts of servers can be really challenging. There are a variety of concerns that might occur, including performance issues, program crashes, and poor optimization.
  • Our machine learning model can be implemented in a single programming language like Python, but the application will need to communicate with other apps written in other programming languages for data intake, preprocessing, front-end, etc., which is a challenge on its own. 

Benefits of using Docker

  • As each microservice may be in a separate language, docker enables scalability and the simple addition or removal of independent services. Replicability, portability, ease of deployment, incremental updates, lightweight, and simplicity are some of the benefits of Docker.
  • It is a data scientist’s greatest fear that their model will not accurately reflect the outcomes of real-life or when it is shared with other researchers. In certain cases, the model isn’t the issue; rather, it’s the need of reproducing the whole stack. Using Docker, it is possible to replicate the working environment that is needed to train and operate the machine learning model on any system. 
  • With Docker, code and dependencies can be packaged together into portable containers that can be run on a variety of different hosts, regardless of the hardware or operating system. 
  • A training model may be constructed on a local system and simply transferred to remote clusters that have more resources like GPUs, more RAM, or powerful CPUs. OpenShift, a Kubernetes distribution, makes it simple to deploy and makes your model accessible to the world by wrapping it into an API in a container and deploying the container using technologies like this. 
  • Container image versions can be tracked using Docker, as well as who developed them and with what. Versions may be rolled back as well. In addition, your machine learning application may function even if one of the services it relies on is upgrading, fixing, or unavailable. For example, if an output message contained in the whole solution has to be updated, there is no need to interact with other services and update the entire program.

4. Tracking the metrics of your model

Even after model development, the machine learning (ML) lifecycle is far from over; in fact, that is merely the beginning of the process. To put your model into production, the next step is to deploy and monitor your model. To get the most out of your machine learning model, you need to keep an eye on it for as long as you’re using it.

Consider ML model monitoring the same as frequent health checkups. An operational duty, model modeling helps you to make sure your model is working at its best.

The importance of model monitoring

Monitoring your ML models for errors, crashes, and latency is part of the operational stage that follows model deployment in the machine learning life cycle. 

Model drift, or the degradation of models as they are used, is the most straightforward reason why model monitoring is necessary. There are several reasons why a model’s prediction ability might decline-

  1. Previously unseen information: If there is not enough labeled data or there are computational restrictions on training, it’s impossible to avoid data samples where your machine learning model produces results that are either erroneous or wrong.
  2. Variable connections and changes in the surrounding environment: ML models are optimized depending on the variables and parameters at the time they were built. Consider how effective a spam-detection algorithm developed twenty years ago is now. Twenty years ago, the internet didn’t even exist! This approach may not be able to keep up with the ever-changing nature of spam emails, therefore it is doubtful that it would be effective at all. Changes in the environment are exemplified by this instance.
  3. Changes in the upstream data: This is a term used to describe modifications to the operational data pipeline. The data science team often does not have complete control over all of the systems from which input data originates. While the data is recorded in Celsius, a fully isolated software engineering team may alter the measurement of a variable to Fahrenheit. If one was attempting to determine the average temperature, for example, this may have serious repercussions. 

ML experiment tracking tools

For the reasons mentioned in the last section, it is quite obvious that the staging environment for a machine learning model is never equivalent to the production environment and thus you must keep tabs on any ML model you put in production.

There are a number of technologies out there that can help us in doing just that. For starters, we have well-known tools, such as MLFlow, TensorBoard, Neptune, DVC, Amazon SageMaker, and so on.

To get a better understanding, let’s take a look at Neptune and see how we can keep our model in check through this tool, once it has been deployed through docker.

Short overview of Neptune

Neptune is a metadata store for MLOps, built for teams that run a lot of experiments.‌

It gives you a single place to log, store, display, organize, compare, and query all your model-building metadata.

‌Neptune is used for:‌

  • Experiment tracking: Log, display, organize, and compare ML experiments in a single place.
  • Model registry: Version, store, manage, and query trained models, and model building metadata.
  • Monitoring ML runs live: Record and monitor model training, evaluation, or production runs live

If you want to learn more check the documentation or explore an example project in the app (no registration required).

Docker with Neptune

Let’s follow a step-by-step process to see how we can log experimentation metadata in containerized Python scripts or applications. 

STEP 1: Add Neptune to your code

Let’s train a simple linear regression model and put it in a Docker container. 

# training.py

import numpy as np
from sklearn.linear_model import LinearRegression
from sklearn.metrics import mean_squared_error
from sklearn.model_selection import train_test_split
import joblib
import neptune.new as neptune
from neptune.new.types import File
import os


run = neptune.init(
    project="PROJECT_WORKSPACE/PROJECT_NAME",
    api_token="YOUR_API_TOKEN",
    tags='a_tag_to_help_you'
) 

lr = LinearRegression()
# #############################################################################
# Load and split data
for _ in range(100):
    rng = np.random.RandomState(_)
    x = 10 * rng.rand(1000).reshape(-1,1)
    y = 2 * x - 5 + rng.randn(1000).reshape(-1,1)

    X_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=123)

    # #############################################################################
    # Fitting the model

    lr.fit(X_train, y_train)
    test_mse = mean_squared_error(y_test, lr.predict(X_test))


    run['test/mse'].log(test_mse)

joblib.dump(lr, '/any_name.pkl')
run['model'].upload(File.as_pickle('any_name.pkl'))

To observe how our model behaves with different data, we run it 100 times, each time with different random data using Numpy’s .RandomState(). So, each time, we receive an updated value of our metric, which gets logged with the .log function.

Now, lr is our model, and test_mse is our metric. How do we get this same result to be reproduced on a different machine? We use Docker!

STEP 2: Update the requirements.txt file and create the Dockerfile

Since now we’re using the Neptune package, we need to include this in our requirements file.

# requirements.txt
joblib==1.1.0
neptune-client
numpy==1.21.4
scikit-learn==1.0.1

Now create a Dockerfile that will:

  • Specify the base container image from which we will build ours from
  • Install dependencies in the updated requirements.txt
  • Copy our training script to the container image and define the command to execute it.
# syntax=docker/dockerfile:1
FROM python:3.8-slim-buster

RUN apt-get update
RUN apt-get -y install gcc

COPY requirements.txt requirements.txt
RUN pip3 install -r requirements.txt

# Copy all files in the current dir to the main dir of the container
COPY . .
CMD [ "python3", "-W ignore" ,"training.py" ]

STEP 3: Build and run docker container passing API token

Build a docker image from the Dockerfile created in the last step by running the following command:

# image-name: neptune-docker
docker build --tag <image-name> . 

Now, we run the container with our credentials.

# NEPTUNE_API_TOKEN: YOUR_API_TOKEN
# image-name: neptune-docker 
docker run -e NEPTUNE_API_TOKEN="<YOUR_API_TOKEN>" <image-name>

After you run the docker container you will get a link on the terminal. Click on the link to open the Run in Neptune to watch your model training live.

Monitor model training runs live in Neptune
Monitor model training runs live in Neptune

In our example we are logging MSE, this can be changed to any metric which makes sense for your model and use case.                           

Conclusion

Many tutorials and articles on the internet are focused only on how to model and comprehend metrics. A prominent scientist or an employee at a tiny corporation will have to undergo peer review in the scientific process.

Keep an eye on how your model performs in light of fresh data and changes. Even if you’re not interested in learning about MLOps, you won’t be able to benefit from intensive courses. I hope you found the material helpful in your everyday routines!

Here are some references that can help you better understand Docker best practices and how a Machine Learning practitioner can benefit from this.

  • For more information on how to build and run a container for your machine learning model, see this article.
  • If you are a beginner, this article will help you with the basics of model monitoring.
  • This article has more detail on using Docker for deep learning applications.
  • Finally, it’s a 3-articles series about Docker for machine learning.

READ NEXT

MLOps at GreenSteam: Shipping Machine Learning [Case Study]

7 mins read | Tymoteusz Wołodźko | Posted March 31, 2021

GreenSteam is a company that provides software solutions for the marine industry that help reduce fuel usage. Excess fuel usage is both costly and bad for the environment, and vessel operators are obliged to get more green by the International Marine Organization and reduce the CO2 emissions by 50 percent by 2050.

Greensteam logo

Even though we are not a big company (50 people including business, devs, domain experts, researchers, and data scientists), we have already built several machine learning products over the last 13 years that help some major shipping companies make informed performance optimization decisions.

MLOps shipping

In this blog post, I want to share our journey to building the MLOps stack. Specifically, how we:

  • dealt with code dependencies
  • approached testing ML models  
  • built automated training and evaluation pipelines 
  • deployed and served our models
  • managed to keep human-in-the-loop in MLOps
Continue reading ->
Experiment tracking Experiment management

15 Best Tools for ML Experiment Tracking and Management

Read more
ML model management

Machine Learning Model Management: What It Is, Why You Should Care, and How to Implement It

Read more
A Comprehensive Guide on How to Monitor Your Models in Production

A Comprehensive Guide on How to Monitor Your Models in Production

Read more
How to Serve Machine Learning Models With TensorFlow Serving and Docker

How to Serve Machine Learning Models With TensorFlow Serving and Docker

Read more