<html><body>1. Understand Time-scales: Before diving deep to Instant, let’s discover some time-scales. UT (Universal Time) is a time standard based on Earth's rotation, essentially measuring time as observed from the planet. It reflects the natural cycle of day and night but can vary slightly due to irregularities in Earth's rotation. TAI (Temps Atomique International or International Atomic Time) is a highly precise time standard based entirely on atomic clocks, specifically using the vibrations of Cesium-133 atoms. UTC (Coordinated Universal Time): is the global time standard and is the combination of UT and TAI. It is designed to be more consistent than UT by relying on precise atomic clocks (same as TAI).  Because of tidal acceleration, the Earth’s rotation is gradually slowing down. As a result, at some point some time, the day in UT (solar day) will get longer than today by one atomic second, which means some day’s length will be 86400 + 1 atomic second. The day that will last 1-atomic-second-longer than today. That day is not predictable and can only be figured out by measurement. To deal with the issue, UTC is synchronized with UT through the use of leap seconds, ensuring it stays within ±0.9 seconds of UT, which helps keep it in line with the natural day. As the Earth’s rotation is slowing down, UTC will only add a leap second to be consistent with UT. As an observation, the earth's rotational acceleration is on average 0.002 atomic seconds per day, which means 500 days or 1.5 years will end up with 1 atomic second. Let's take a example: Today, Tuesday, January 20, 2024, which is exactly 86400 atomic seconds. 1.5 years from now, Friday, January 02, 2026, a day will be 86400 + 1 atomic seconds. But the difference between UT and UTC should be within ±0.9 seconds, then at some point of time between today and Friday, January 02, 2026, a leap second will be added at the end of that day, which introduces the 61st second of a minute: yyyy-MM-ddT23:59:60Z. While UTC adds a leap second to be synchronized with UT, TAI, on the other hand, does not. As a result, time in UTC and TAI will be further and further apart. At this moment, TAI is 37 seconds ahead of UTC: <ul><li style="text-align:justify;">UTC: 2024-08-20 14:47:55Z </li><li style="text-align:justify;">TAI: 2024-08-20 14:48:32Z  </li><li style="text-align:justify;">See real-time gap here: <a href="http://leapsecond.com/java/gpsclock.htm">http://leapsecond.com/java/gpsclock.htm</a>  </li></ul>Java Time-scale Although Instant represents a point in time in UTC, it follows a 99%-like UTC time-scale, called Java Time-scale. In Java Time-scale, one day includes exactly 86,400 seconds without a leap second. If a day in UTC contains a leap second, that leap second will be spread equally over the last 1000 seconds of the day, maintaining the appearance of exactly 86400 seconds per day. 2. How Instant is stored in Java Application Instant is an object that: <ul><li style="text-align:justify;">represents a point in time in UTC since 1970-01-01T00:00:00Z, we call 1970-01-01T00:00:00Z an EPOCH. </li><li style="text-align:justify;">with nanosecond precision. </li></ul>To represent a point in time, Instant introduces: <pre><code class="language-plaintext">/**  
* The number of seconds from the epoch of 1970-01-01T00:00:00Z.  
*/  
private final long seconds; </code></pre><ul><li style="text-align:justify;">seconds is the number of seconds from the EPOCH.  </li></ul>If seconds are negative, the instant is before the EPOCH. If seconds are positive, the instant is after the EPOCH. Despite Long.MIN_VALUE is -9,223,372,036,854,775,808, the minimum supported epoch second is -31,557,014,167,219,200, which is -1000000000-01-01T00:00Z (one billion year in the past) and the maximum supported epoch second is 31,556,889,864,403,199, which is 1000000000-12-31T23:59:59Z (the end of 1 billion years in the future). To represent a point in time with nanosecond precision, Instant introduces: <pre><code class="language-plaintext">/**  
* The number of nanoseconds, later along the time-line, from the seconds field.  
* This is always positive, and never exceeds 999,999,999.  
*/  
private final int nanos; </code></pre><ul><li style="text-align:justify;">nanos is always positive, and never bigger than 999,999,999 </li></ul>Some reasons why min/max supported epoch seconds is not Long.MAX/MIN_VALUE: From a usage perspective: <ul><li style="text-align:justify;">The universe is said to be aged 13.7 billion years old; Long.MIN_VALUE will introduce the time since more than 292 billion years. -&gt; Long.MIN_VALUE is not practical. </li><li style="text-align:justify;">292 billion years in the future (which is Long.MAX_VALUE) is an extreme value (compared to human lifetime), making it not practical either. </li><li style="text-align:justify;">According to research, the first humans emerged in Africa around 2 million years ago, so there should not be much historical data that modern-universal applications should deal with.</li></ul><figure class="image"><img src="https://api.careers.saigontechnology.com/storage/ck-editor/Why-epoch-seconds-limits-are-not-long.max-min-value-CyrT3tzrMHPr1rGc2zUR0pnPNQWdhdbmKKBY8e53.webp"></figure><ul><li style="text-align:justify;">±1 billion of year make it have a good buffer in long type and make the year (1 billion) fit in an int. </li></ul>From a technical perspective: <ul><li style="text-align:justify;">Storing extreme value will easily cause exceptions. I asked ChatGPT to calculate the instant of Long.MIN_VALUE, ChatGPT could not answer due to an Overflow issue.</li></ul><figure class="image"><img src="https://api.careers.saigontechnology.com/storage/ck-editor/calculate-the-instant-of-Long.MIN_VALUE-jWF0AC5aTwa6CQqYCifAZQX00l4fjUhPAzaMX7fM.webp"></figure><ul><li style="text-align:justify;">Limiting the range of value to be far away from the edge will help operations like addition, subtraction, and comparison can be performed safely. </li><li style="text-align:justify;">Storing extremely large values can pose challenges in serialization, database storage, and data transmission. </li></ul>3. Instant.now() Returns the current moment in UTC with precision in: <ul><li style="text-align:justify;">Millisecond: Java 8 </li><li style="text-align:justify;">Microsecond: Java 9, Java 11 </li><li style="text-align:justify;">Nanosecond: Java 17+ </li></ul>Use the below code snippet to run on <a href="https://www.jdoodle.com/online-java-compiler">https://www.jdoodle.com/online-java-compiler</a> with each JDK, you’ll see the difference. <pre><code class="language-plaintext">import java.time.Instant;  
 
public class MyClass {  
 
    public static void main(String args[]) {  
        Instant instant = Instant.now();  
        System.out.println(instant);  
        //2024-08-19T15:47:28.487729962Z  
    }  
} </code></pre>The precision of  Instant.now() can vary between different JVMs and OS/Hardware. On some platforms, you might observe higher precision, while on others, it may be lower. Behind the Scenes of Instant.now() The Instant.now() method provides a seemingly straightforward way to retrieve the current moment in time. But what happens beneath the surface? Let's delve into the intricate workings of this method:   <pre><code class="language-plaintext">//This is simplified implementation from Clock.java without retry  
static Instant currentInstant() {  
    long localOffset = System.currentTimeMillis() / 1000 - 1024;  
    long adjustment = VM.getNanoTimeAdjustment(localOffset);  
    return Instant.ofEpochSecond(localOffset, adjustment);  
} </code></pre> Setting the Local Offset: The code starts by calculating a preliminary offset from the current system time. This offset, represented by localOffset, is obtained by dividing the system time in milliseconds (System.currentTimeMillis()) by 1000 (to convert milliseconds to seconds) and then subtracting 1024 seconds (approximately 17 minutes and 4 seconds). Fine-Tuning with Nanosecond Adjustment: The VM.getNanoTimeAdjustment(localOffset) method is called. This method interacts with the underlying system to determine the precise nanosecond adjustment needed to convert the localOffset to the current UTC time. The returned value, nanoTimeAdjustment, represents the time difference (in nanoseconds) between localOffset and the actual UTC time. Constructing the Instant: Finally, an Instant object is created using the localOffset (seconds since the epoch) and the nanoTimeAdjustment (nanoseconds within the second). Example Breakdown: Imagine the current UTC time is 2024-08-20T15:59:50.123Z, which translates to epoch milliseconds of 1724144390123. Below is how currentInstant() would run: <pre><code class="language-plaintext">long localOffset = System.currentTimeMillis() / 1000 - 1024;  
 
// localOffset = 1724144390123 / 1000 - 1024 = 1724143366  
long adjustment = VM.getNanoTimeAdjustment(localOffset);  
 
// 1024044595000  
return Instant.ofEpochSecond(localOffset, adjustment);  
 
// Instant.ofEpochSecond(1724143366, 1024044595000);   
// 2024-08-20T09:12:18.044595Z </code></pre>Nanosecond Adjustment: Suppose VM.getNanoTimeAdjustment(1724143366) returns 1024044595000, which equates to 1024.044595 seconds. The Catch: While the returned value appears to provide nanosecond precision (due to the decimal component), closer inspection reveals that the last three digits are zeros. This implies that the underlying system can only guarantee accuracy up to milliseconds. The Resulting Instant: Instant.ofEpochSecond(1724143366, 1024044595000) This translates to an Instant object representing 2024-08-20T09:12:18.044595Z. However, it's crucial to remember that the nanosecond precision might not be entirely reliable due to the system limitations. In essence, Instant.now() provides a convenient way to access the current time with nanosecond precision (in theory). However, the underlying system might restrict the actual accuracy to milliseconds. Remember, this explanation is for educational purposes and might not reflect the exact implementation details in all Java environments. 4. Conclusion  In conclusion, java.time.Instant is a powerful class that represents a specific point on the timeline, measured to the nanosecond. It offers a reliable way to work with timestamps in UTC, making it ideal for applications that require precise time measurements. With its immutable nature and comprehensive API, Instant simplifies handling of time data in Java, whether you're recording events, synchronizing across time zones, or performing time-based calculations. By leveraging Instant, developers can ensure accurate and consistent time management in their applications, paving the way for robust and future-proof solutions. Thank you all for reading all the way down to this. This article is just the beginning of java.time, let’s wait for upcoming posts! Reference: <a href="https://docs.oracle.com/en/java/javase/17/docs/api/java.base/java/time/Instant.html">https://docs.oracle.com/en/java/javase/17/docs/api/java.base/java/time/Instant.html</a> https://www.timeanddate.com/time/aboututc.html </body></html>

The Instant class in Java is immutable and thread-safe, making it ideal for use in time-sensitive applications where absolute time without any time zone considerations is needed. Let’s give Instant a closer look today.

Deep dive into java.time: Instant

Tech Lead

Phuong Bui

.NET Developer

Vu Truong

Senior Developer

Trung Tran

Tech lead

Wind

Long Nguyen

Business Analyst

Duong Dat Thanh

Are you facing problems with data consistency in a highly concurrent environment? Are you looking for a way to limit and synchronize access to a shared resource across services in a distributed system? Let's see how a distributed lock can help in this situation.

Implement Distributed Lock For a Microservices Software System

<html><body><h2 style="text-align:justify;">1. AWS Là Gì</h2>Trước tiên anh em mình cùng nhau tìm hiểu để biết xem AWS là gì nhé. Amazon Web Service hay còn được gọi với cái tên thân mật là AWS.  Thật ra có hàng trăm thậm chí là hàng ngàn câu hỏi liên quan đến AWS khi mình tự tập tành tìm hiểu về nó và sau một thời gian tìm hiểu cũng như làm thực tế trên các dự án liên quan đến AWS mình xin mạo muội làm một series về AWS để chia sẻ với anh em đang có mong muốn tìm hiểu về nền tảng đám mây này. Ông bà ta có câu: Răng với tóc là gốc con người, cái tên nói lên tất cả. Amazon Web Service là dịch vụ điện toán đám mây (Cloud Services) nó cung cấp rất nhiều dịch vụ từ việc lưu trữ như: Storage, rồi thì Networking, Virtualization đến cả Machine Learning, AR &amp; VR…Mình sẽ liệt kê ra cho mấy bạn những service thông dụng mà anh em lập trình viên cũng như DevOps thường dùng. Mình có thể thao tác với AWS thông qua hai công cụ chính là AWS Web Console hoặc là AWS CLI. Sẽ bật mí với mấy bạn về hai công cụ hữu ích này sau nha, giờ mình sẽ list ra những dịch vụ cơ bản của AWS.Common Services Của AWS <ul><li style="text-align:justify;">Storage: AWS S3, Databases, Elastic Container Registry (ERC)</li><li style="text-align:justify;">Virtualization:  Elastic Compute Cloud (EC2)</li><li style="text-align:justify;">Programming: Elastic Container Service (ECS), Code Pipeline, Code Build</li><li style="text-align:justify;">Networking: Virtual Private Cloud (VPC), Route53, Load Balancing</li><li style="text-align:justify;">Identity: IAM</li><li style="text-align:justify;">Monitoring: CloudWatch</li></ul> Anh em có thể đăng ký một tài khoản AWS từ <a href="https://portal.aws.amazon.com/billing/signup#/start/email">website chính hãng</a> để khám phá tất tần tật các dịch vụ của em nó. Ok giờ mình thảo luận tí về lợi ích và thế mạnh mà AWS mang lại.<h2 style="text-align:justify;">2. Điểm Mạnh Của AWS</h2> Rõ ràng lợi ích đầu tiên mình thấy được đó là số lượng dịch vụ mà AWS cung cấp. Có khoản 200 dịch vụ và hầu hết các dịch vụ này đủ sức đáp ứng được nhu cầu phát triển cơ sở hạ tầng cũng như nền tảng công nghệ cho các doanh nghiệp trong nhiều lĩnh vực trong đó các doanh nghiệp về lĩnh vực công nghệ thông tin chiếm đa số. Nhưng số lượng dịch vụ nhiều chỉ là một trong những điểm mạnh của AWS bên cạnh đó chất lượng của từng dịch vụ có thể mô tả trong hai từ “tuyệt vời”. Do đó có rất nhiều công ty kể cả chính phủ tin tưởng vào AWS. Xin cả nhà một phút quảng cáo, công ty <a href="https://saigontechnology.com/">SaiGon Technology</a> đã và đang phát triển rất nhiều sản phẩm dựa trên nền tảng AWS.  Để nói về các điểm mạnh cũng như tiện ích của AWS, mình sẽ lấy một ví dụ cụ thể để từ đó anh em DEV có cái nhìn thực tiễn và khách quan. Đặc biệt đối với mấy anh em bắt đầu tìm hiểu về AWS cũng có thể hình dung ra được những tính năng mà AWS mang lại. Lấy ví dụ như thế này. Công ty <a href="https://saigontechnology.com/">SaiGon Technology</a> muốn xây dựng một trang web về e-commerce giống như Shopee or Lazada…Mình sẽ list ra những công việc ở mức high level cần phải được hoàn thành như sau: Phát triển những tính năng của trang web: Authorization, Authentication, Purchasing Product, Account Management, Product Management… Xây dựng cơ sở dữ liệu:<ul><li style="text-align:justify;">Lựa chọn hệ quản trị cơ sở dữ liệu: SQL-Server, MySQL, PostgreSQL, No-SQL…</li><li style="text-align:justify;">Cài đặt DBMS vào máy chủ</li><li style="text-align:justify;">Thiết lập các phương án backup dữ liệu</li></ul> Build Server tạo môi trường cho việc kiểm thử và vận hành:<ul><li style="text-align:justify;">Xây dựng các máy chủ như: UAT, Staging, Production</li><li style="text-align:justify;">Build các docker image</li><li style="text-align:justify;">Publish các docker image lên repository</li><li style="text-align:justify;">Deploy docker image</li><li style="text-align:justify;">Xây dựng hệ thống log để monitor trong quá trình vận hành</li><li style="text-align:justify;">Mua tên miền và mapping tên miền vào server</li></ul> Ngoài ra, vào những dịp lễ tết, trang web của chúng ta thường chạy các event giảm giá sập sàn, siêu sale, flash sale…dẫn đến việc quá tải. Lúc đó mình lại phải cấm đầu vào cái việc load balancing để chia tải cho hệ thống.Chốt: rất nhiều công việc mà anh em Dev phải chạy bằng cơm.Vậy thông qua ví dụ trên, mình sẽ liệt kê ra các tính năng hữu ích của AWS cho các bạn.<h3 style="text-align:justify;">Tính Tiện Dụng</h3>AWS là một web service, là một dạng của SaaS do đó chỉ cần trong điều kiện có kết nối đến internet là chúng ta có thể truy cập và sử dụng mọi lúc, mọi nơi.Ngoài ra, cá nhân mình cảm thấy giao diện của AWS được thiết kế gọn gàng, trực quan, tạo cho người dùng cảm giác thân thiện, dễ sử dụng<figure class="image"><img src="https://api.careers.saigontechnology.com/storage/ck-editor/CKEDITOR-o6bIfFXJ5P6nJBnndHOmFOr6IQ4sHMK1fC4AeMsZ.png"></figure>Quay lại với ví dụ trên, việc lập trình các tính năm của trang web thì chủ yếu liên quan đến nhánh lập trình bạn có thể sử dụng các ngôn ngữ như: <a href="https://www.facebook.com/SaigonTechnology.HCM.Careers">Java, .Net, Python</a>…Nên mình sẽ không bàn về nó tại thời điểm này.  Mình chuyển sang bước tiếp theo là xây dựng CSDL. Giả sử rằng mình chọn PostgreSQL làm CSDL cho trang web. Vậy theo cách thông thường mình phải setup PostgreSQL vào máy rồi sau đó tạo một connection từ application đến CSDL. Giờ đây với tính năng Relational Database Service (RDS) của AWS mình có thể skip qua việc setting up PostgreSQL. Chỉ một vài cái click chuột mình đã có thể tạo một hệ quản trị CSDL PostgreSQL riêng cho mình. Ngoài ra RDS còn support cả: SQL-Server, MySQL, MariaDB, Oracle <figure class="image"><img src="https://api.careers.saigontechnology.com/storage/ck-editor/CKEDITOR-ciEOv3BzNeATVh4G0rIpQk6mqPFvBcRpp2fnzWf5.png"></figure>Free source: <a href="https://scandiplan.dk/en/aws-rds-snowflake">https://scandiplan.dk/en/aws-rds-snowflake</a> Nhưng nếu mình muốn sử dụng Nosql thì sao? Yên tâm bạn có thể sử dụng DynamoDB từ AWS nhé. Tiếp theo là việc build server tạo môi trường cho việc kiểm thử và vận hành trang web của mình. AWS hỗ trợ rất tốt các hệ điều hành Linux, do đó việc tạo và maintain một máy chủ trên AWS có thể được thực hiện thông qua Elastic Compute Cloud (EC2). Với EC2 bạn có thể xây dựng một máy chủ với hệ điều hành mà bạn mong muốn: Fedora, Ubuntu, Window Server, AWS Linux…<figure class="image"><img src="https://api.careers.saigontechnology.com/storage/ck-editor/CKEDITOR-uOMEYXqBcGLWobdTRJJ1RH8vhaxuOEExzZ4McM0C.png"></figure>Free source: <a href="https://docs.aws.amazon.com/images/AWSEC2/latest/UserGuide/images/overview_getting_started.png">https://docs.aws.amazon.com</a>Ngoài ra những năm gần đây với sự phát triển và phổ biến của Docker thì AWS cũng ra đời Elastic Container Registry (ERC) hỗ trợ cho việc host các docker image.AWS Code Build, Code Pipeline hỗ trợ rất tốt trong việc build và deploy ứng dụng.<figure class="image"><img src="https://api.careers.saigontechnology.com/storage/ck-editor/CKEDITOR-cvdfpnfDPGlXTnCBiGjEXqRsmRMqKigBruY39Krc.png"></figure>Free source: <a href="https://skdevops.files.wordpress.com/2022/03/62.image-1.png?w=656&amp;h=300&amp;crop=1">https://skdevops.files.wordpress.com</a> Trên trang web của chúng ta không thể nào thiếu các hình ảnh liên quan đến sản phẩm, hình ảnh đại diện của user…thay vì mình lưu những file hình đó vào CSDL hoặc lưu trực tiếp vào ổ đĩa thì mình có thể tận dụng Simple Storage Service (S3). Một service của AWS hỗ trợ việc lưu trữ và transfer file khá tốt.Như mình có đề cập trong ví dụ là cần một domain name gắn vào server. Ví dụ mình cần một tên miền: <a href="https://www.facebook.com/SaigonTechnology.HCM.Careers">saigontechnology-career.com</a> Vậy thì mình có thể đăng ký tên miền này thông qua AWS Route53. Chi phí thì tùy vào tên miền mình chọn, trung bình khoản 15$/năm. Việc cấu hình tên miền cũng được thực hiện dễ dàng thông qua AWS Web Console.<h3 style="text-align:justify;">Khả Năng Mở Rộng</h3>Khả năng mở rộng cũng chính là một điểm mạnh khi sử dụng Cloud nói chung và AWS nói riêng. Thử hình dung xem, một ngày nào đó lượng dữ liệu mà mình phải chứa trong csdl lên đến hàng TB/ngày thì anh em Dev mình phải cuốn cuồng xử lý. Dĩ nhiên, điều này dẫn đến rất nhiều vấn đề tuy nhiên mình chỉ đề cập ở khía cạnh dung lượng lưu trữ. Hoặc trong các event flash sale, lượng người truy cập quá lớn, lúc đó có thể gây quá tải trên hệ thống. Hoặc đơn giản là mình cần nâng cấp RAM cho máy chủ.Tất cả những vấn đề trên, nếu mình tự build một server vật lý thì anh em Dev phải rất vất vả đi xử lý tuy nhiên mình hoàn toàn có thể được giải quyết nhanh gọn bằng các service của AWS. <h3 style="text-align:justify;">Tính Bảo Mật Cao</h3>Không chỉ riêng về AWS, bảo mật là tính năng cần thiết hàng đầu đối với tất cả các ứng dụng, đặc biệt là các ứng dụng về Cloud. Và AWS đã đạt được các chứng nhận về việc tuân thủ khung đảm bảo quyền riêng tư.<ul><li style="text-align:justify;">ISO 27017 về bảo mật đám mây.</li><li style="text-align:justify;">ISO 27701 về quản lý thông tin quyền riêng tư.</li><li style="text-align:justify;">ISO 27018 về bảo vệ thông tin cá nhân trên đám mây.</li></ul>Nãy giờ mình cũng chém gió về điểm mạnh, giờ mình xin phép chia sẻ một vài cái mình  cảm thấy hạn chế khi sử dụng AWS. <h2 style="text-align:justify;">3. Hạn Chế Của AWS</h2><h3 style="text-align:justify;">Chi Phí</h3>Đầu tiên mình muốn đề cập là vấn đề chi phí. AWS là một dạng Cloud Services do đó nó không chỉ tính chi phí trên nguồn tài nguyên mình sử dụng mà nó còn tính cả chi phí về thời gian lưu trữ. Do đó, có những lúc chi phí phát sinh rất cao mà mình không thể kiểm soát được. AWS có cung cấp tính năng gửi notification khi chi phí vượt quá mức giới hạn cho phép. Tuy nhiên nó chỉ là notification bạn nhé. Vì vậy, trong quá trình sử dụng, các bạn nên chú ý và tắt đi những service thật sự không cần thiết.<h3 style="text-align:justify;">Internet</h3>Là một cloud services nên mình chỉ có thể sử dụng AWS ở những địa điểm có kết nối internet. AWS không hỗ trợ offline nên anh em chú ý nhé.<h3 style="text-align:justify;">Thiếu Chuyên Gia</h3>AWS có cơ sở hạ tầng vô cùng phức tạp, nó cung cấp hàng trăm service trãi dài hầu hết nhiều lĩnh vực. Do đó, còn khá ít các chuyên gia có kỹ năng về AWS. Vì vậy, một tổ chức hay cá nhân muốn phát triển AWS thì đây là một điểm bất lợi khá lớn.Mình xin phép kết thúc bài viết tại đây. Hy vọng nhận được nhiều đóng góp của mọi người để cộng đồng lập trình ngày càng phát triển và đặc biệt để những bài viết sau được tốt hơn. </body></html>

Bài viết nhằm hướng dẫn cho anh em lập trình viên nhà mình có một cái nhìn tổng quan nhất về Amazon Web Service (AWS) chẳng hạn như: AWS là gì? Có gì trong AWS? Lợi ích của AWS? Nên tìm hiểu AWS từ đâu.
Qua bài viết hy vọng có thể support được mọi người có thêm kiến thức về AWS.
OK Let’s Go.

[AWS] Bài 01: Giới Thiệu Về Amazon Web Service (AWS)

Explore the enchanting world of CSS animation, elevating web designs with captivating transitions and dynamic effects. Unleash creativity as we delve into key techniques and practices, empowering you to create immersive experiences. Join our exciting CSS animation journey!

Animation in CSS

Vien Truong

Cau Hai

The AI, Machine Learning, and Deep Learning definition are becoming viral recently. Let's see how we could apply them to address our problems in the real world, especially in the Computer Vision field for the Image Classification task.

Build Deep Learning model for the Image Classification task (Part1): Pytorch model

&lt;html&gt;&lt;body&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;Welcome back to Part 2 of my tutorial on building a deep learning model for image classification. In Part 1, It covered the basics of deep learning and walked through the steps to prepare our data for training and basic deep learning model or environment as well. Now, we will focus on the structure of the deep learning project and how to monitor it during training. In the previous blog, you could see I follow an &ldquo;all-in-one&rdquo; for combining all code into one file. In this blog I will introduce you to the basic deep learning project structure and some monitors to help you to track the model performance during training.&lt;/p&gt;&lt;h2 style=&quot;text-align:justify;&quot;&gt;1. Project Structure&lt;/h2&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;When starting a new deep learning project, it's essential to plan ahead and think about the overall structure of the project. This will make it easier to maintain your code and data as your project grows, and to collaborate with other developers and researchers.&lt;/p&gt;&lt;h3 style=&quot;text-align:justify;&quot;&gt;Key components&lt;/h3&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;Here are the key components of a deep learning project:&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Data&lt;/strong&gt;: The first step in any machine learning project is to gather and preprocess data. In the case of image classification, this means collecting a large dataset of images that are labeled with the corresponding classes. There are several popular datasets for image classification, such as CIFAR-10, MNIST, and ImageNet. You can also create your dataset by collecting images from various sources.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;Once you have your dataset, you will need to preprocess it by resizing the images, normalizing the pixel values, and splitting it into training, validation, and testing sets. It's essential to ensure that your data is well-organized and that the directory structure is consistent across all subsets.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Model&lt;/strong&gt;: The next component of a deep learning project is the model itself. There are many deep learning architectures to choose from, such as Convolutional Neural Networks (CNNs), Residual Networks (ResNets), and Recurrent Neural Networks (RNNs). You can also use transfer learning by fine-tuning pre-trained models such as VGG16 or Inception.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;It's crucial to define the model architecture clearly and to document any hyperparameters and design decisions. This will make it easier to reproduce your results and to make changes to the model in the future.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Training&lt;/strong&gt;: Once you have your data and model, it's time to train the model on the training set. This involves running the images through the model and adjusting the weights and biases to minimize the loss function. You can use various optimization algorithms, such as stochastic gradient descent (SGD), Adam, or RMSprop.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;It's important to monitor the training process by logging the loss and accuracy metrics, and to use techniques such as early stopping and learning rate scheduling to prevent overfitting and speed up convergence.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Evaluation&lt;/strong&gt;: After training, you need to evaluate the performance of the model on the validation set. This involves running the images through the model and computing metrics such as accuracy, precision, recall, and F1 score. You can also use techniques such as confusion matrices and ROC curves to analyze the model's behavior.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;It's important to report the evaluation results clearly and to compare them with the state-of-the-art models in the literature. This will help you determine whether your model is good enough for the intended application.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Deployment&lt;/strong&gt;: Finally, once you have a trained and evaluated model, it's time to deploy it in the real world. This involves converting the model into a format that can be run on different platforms, such as TensorFlow Lite or ONNX. You can also deploy the model on the cloud or on edge devices such as Raspberry Pi or Jetson Nano.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;It's important to document the deployment process clearly and to test the model thoroughly on new data to ensure that it behaves as expected.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;Now that we have covered the key components of a deep learning project, let's discuss how to organize your code and data to ensure efficiency and reproducibility.&lt;/p&gt;&lt;h3 style=&quot;text-align:justify;&quot;&gt;Code structure&lt;/h3&gt;&lt;h4 style=&quot;text-align:justify;&quot;&gt;Data Organization&lt;/h4&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;In this structure, we have separate directories for the training, validation, and testing data, with subdirectories for each class. This makes it easy to load the data using data generators or PyTorch datasets.&lt;/p&gt;&lt;pre&gt;&lt;code class=&quot;language-plaintext&quot;&gt;```
project/
  data/
    train/
      class1/
        image1.jpg
        image2.jpg
        ...
      class2/
        image1.jpg
        image2.jpg
        ...
      classN/
        image1.jpg
        image2.jpg
        ...
    val/
      class1/
        image1.jpg
        image2.jpg
        ...
      class2/
        image1.jpg
        image2.jpg
        ...
      classN/
        image1.jpg
        image2.jpg
        ...
    test/
      class1/
        image1.jpg
        image2.jpg
        ...
      class2/
        image1.jpg
        image2.jpg
        ...
      classN/
        image1.jpg
        image2.jpg
        ...        
```&lt;/code&gt;&lt;/pre&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Fig1: Dataset structure&lt;/strong&gt;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;It's essential to organize the data in a consistent and reproducible manner. Here are some tips for organizing the data:&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;- Use a consistent directory structure across all subsets&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;- Name the images using a clear and descriptive convention, such as `{class}-{index}.jpg`&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;- Use a separate CSV file or JSON file to store the class labels and image paths&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;- Use data augmentation techniques such as flipping and rotation to increase the size of the dataset&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;- Normalize the pixel values using a standard technique, such as dividing by 255 or subtracting the mean and dividing by the standard deviation&lt;/p&gt;&lt;h4 style=&quot;text-align:justify;&quot;&gt;Code Organization&lt;/h4&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;For the code organization, you could apply the basic structure like below:&lt;/p&gt;&lt;pre&gt;&lt;code class=&quot;language-plaintext&quot;&gt;project/
  data/
    train/
      ...
    val/
      ...
    test/
      ...
  config/
    config.yml
  src/
    model/
      __init__.py
      cnn1_model.py
      cnn2_model.py
      vgg16_model.py
      layers.py
    config.py
    dataset.py
    loss.py
    optim.py
    util.py
  train.py
  eval.py
  predict.py
  log/
  mlruns/
  model/
  requirements.txt
  README.md&lt;/code&gt;&lt;/pre&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Fig2: Project structure&lt;/strong&gt;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;The `&lt;strong&gt;data/&lt;/strong&gt;` directory contains the dataset&nbsp; for our project, you could follow the previous section to understand how to organize data for the Image classification task, and I will not mention it again.&nbsp;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;The `&lt;strong&gt;src/&lt;/strong&gt;` directory will contain all of our modules for building models like model architecture, loss modules, optimizer modules, or some support modules like utilities, config reader... It's important to modularize the code and to use clear naming conventions to make it easy to understand and modify.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;The `&lt;strong&gt;src/model/&lt;/strong&gt;` directory contains the the code for the model architecture, including any custom layers, e.g: `&lt;strong&gt;vgg16_model.py`&lt;/strong&gt;, `&lt;strong&gt;cnn_model.py`&lt;/strong&gt; or `&lt;strong&gt;layers.py`&lt;/strong&gt;&nbsp;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;The `&lt;strong&gt;src/model/config.py&lt;/strong&gt;` file contains the the code for configuration object class, it will load the configuration for your model from the yaml file in `&lt;strong&gt;config/config.yml&lt;/strong&gt;` file, and then parsing the dict data into the Python Class object for you to use in the model.&nbsp;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;The `&lt;strong&gt;src/model/dataset.py&lt;/strong&gt;` file contains the code for the dataset reader object, for example we will define the Pytorch dataset loader in this module to use during the training process.&nbsp;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;The `&lt;strong&gt;src/model/loss.py&lt;/strong&gt;` file contains the code for the loss function of our models, we could define some custom loss function here.&nbsp;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;The `&lt;strong&gt;src/model/optim.py&lt;/strong&gt;` file contains the code for the optimizer module for the training process.&nbsp;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;The `&lt;strong&gt;train.py&lt;/strong&gt;`, `&lt;strong&gt;eval.py&lt;/strong&gt;`, and `&lt;strong&gt;predict.py&lt;/strong&gt;` scripts are responsible for training, evaluating, and deploying the model, respectively. These scripts should use the modules in the `&lt;strong&gt;src/&lt;/strong&gt;` directory and should take command-line arguments for flexibility.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;Some other folders, like `&lt;strong&gt;log/&lt;/strong&gt;` will be used to store the training result using tensorboard tool, `&lt;strong&gt;mlrun/&lt;/strong&gt;` will be the same but for using mlflow tool, `&lt;strong&gt;model/&lt;/strong&gt;` will be used to store the trained model after finishing.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;Finally, the `&lt;strong&gt;requirements.txt&lt;/strong&gt;` file contains a list of dependencies required to run the code, and the `&lt;strong&gt;README.md&lt;/strong&gt;` file provides documentation and instructions for running the project.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;For more detail, you could check the full demo source code here for the project structure:&nbsp;&lt;a href=&quot;https://github.com/phonglesaigontechnology/Image-Classification-Model-2&quot;&gt;&lt;u&gt;https://github.com/phonglesaigontechnology/Image-Classification-Model-2&lt;/u&gt;&lt;/a&gt;&nbsp;&lt;/p&gt;&lt;h2 style=&quot;text-align:justify;&quot;&gt;2. Monitoring&lt;/h2&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;When training a deep learning model for image classification, it's essential to monitor its performance closely to ensure that it's learning effectively and not overfitting to the training data. Overfitting occurs when the model performs well on the training data but poorly on the test data, which indicates that it has memorized the training data instead of learning the underlying patterns.&lt;/p&gt;&lt;h3 style=&quot;text-align:justify;&quot;&gt;Monitor Values&lt;/h3&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;There are several ways to monitor the performance of your deep learning model during training.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Loss and accuracy&lt;/strong&gt;: The loss function is a measure of how well the model is able to predict the correct output for a given input. During training, the goal is to minimize the loss function, which means that the model is improving its ability to make accurate predictions. The accuracy metric measures the percentage of correctly classified images during training. It's a useful metric for evaluating the performance of the model on the training data, but it doesn't necessarily indicate how well the model will perform on new, unseen data.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;=&amp;gt; It's important to monitor the loss and accuracy metrics during training and to plot them over time using a visualization tool such as TensorBoard. This will allow you to see how the model is improving over time and to identify any potential issues, such as overfitting or underfitting.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Learning rate&lt;/strong&gt;: The learning rate is a hyperparameter that controls the step size taken during gradient descent optimization. It determines how quickly the model updates its weights during training.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;=&amp;gt; Monitoring the learning rate during training can help you to optimize it for better performance. If the learning rate is too high, the model may overshoot the optimal weights and converge to a suboptimal solution. If the learning rate is too low, the model may take too long to converge or get stuck in a local minimum.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Early stopping&lt;/strong&gt;: Early stopping is a technique used to prevent overfitting during training. It involves monitoring the performance of the model on a validation set and stopping the training process when the performance on the validation set stops improving.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;=&amp;gt; To implement early stopping, you can use a callback function in your training script that monitors the performance on the validation set and stops the training process when the performance stops improving for a certain number of epochs.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Regularization techniques&lt;/strong&gt;: Regularization techniques are used to prevent overfitting by adding additional constraints to the model during training. There are several types of regularization techniques, including L1 and L2 regularization, dropout, and data augmentation.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;=&amp;gt; Monitoring the impact of these techniques on the training performance can help you to optimize them for better performance. For example, you can experiment with different dropout rates or regularization strengths to see how they affect the model's performance.&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;In conclusion, monitoring the performance of your deep learning model during training is essential for ensuring that it's learning effectively and not overfitting to the training data. By monitoring metrics such as loss, accuracy, learning rate, and regularization strength, you can optimize your model for better performance and prevent overfitting.&lt;/p&gt;&lt;h3 style=&quot;text-align:justify;&quot;&gt;Tools&lt;/h3&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;There are several other monitoring tools available for deep learning models. Here are a few examples: MLflow, Tensorboard, Neptune.ai, Weights &amp;amp; Biases, Comet.ml...&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;But I will only introduce you to MLflow and Tensorboard because they are not cloud-based tools. They are open-source platforms, free to use, and can be run locally or on a cluster, and it integrates with many popular machine learning libraries such as TensorFlow, PyTorch, and scikit-learn&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;MLflow&lt;/strong&gt; provides tools for managing experiments, packaging code into reproducible runs, and sharing and deploying models. MLflow allows you to track model training runs, experiment results, and model artifacts. You can use it to compare model performance across different experiments and to reproduce and share your results.&lt;/p&gt;&lt;pre&gt;&lt;code class=&quot;language-plaintext&quot;&gt;import mlflow
# Setup MLflow
mlflow.set_tracking_uri(&quot;./mlruns&quot;)
mlflow.set_experiment(&quot;MyModel CNN&quot;)
# Log params
for key, value in config.__dict__.items():
   mlflow.log_param(key, value)
# Log training loss
mlflow.log_metric('train_loss', loss.item(), step=global_step)
# Log training accuracy
mlflow.log_metric('val_accuracy', accuracy, step=epoch)
&lt;/code&gt;&lt;/pre&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;To view the logs in MLFlow, run the command below in the terminal at the `&lt;strong&gt;project/&lt;/strong&gt;` folder and navigate to&nbsp;&lt;i&gt;http://localhost:5000&lt;/i&gt; in your browser.&nbsp;&lt;/p&gt;&lt;pre&gt;&lt;code class=&quot;language-plaintext&quot;&gt;(venv)$ mlflow ui&lt;/code&gt;&lt;/pre&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;img src=&quot;https://api.careers.saigontechnology.com/storage/ck-editor/CKEDITOR-oNVFUFUjLmuGCuvmAgpczmhDMTG332h78Hu6Qcsi.png&quot;&gt;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Fig3: MLFlow Dashboard UI&lt;/strong&gt;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;TensorBoard&lt;/strong&gt; is a visualization toolkit that comes with TensorFlow. It allows you to visualize your TensorFlow model graph, training curves, and other metrics during training. TensorBoard can also help you debug your model and analyze its performance. It supports features such as profiling, graph visualization, and embedding visualizations.&lt;/p&gt;&lt;pre&gt;&lt;code class=&quot;language-plaintext&quot;&gt;from tensorboardX import SummaryWriter
# Setup TensorBoard
writer = SummaryWriter( &lsquo;log/tensorboard_cnn&rsquo;, time.strftime(&quot;%Y-%m-%d-%H-%M-%S&quot;)))
# Log training loss
writer.add_scalar('train/loss', loss.item(), global_step)
# Log training accuracy
writer.add_scalar('val/accuracy', accuracy, epoch)&lt;/code&gt;&lt;/pre&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;To view the logs in TensorBoard, run the command below in the terminal at `&lt;strong&gt;project/&lt;/strong&gt;` folder and navigate to&nbsp;&lt;i&gt;http://localhost:6006&lt;/i&gt; in your browser.&nbsp;&lt;/p&gt;&lt;pre&gt;&lt;code class=&quot;language-plaintext&quot;&gt;(venv)$ tensorboard --logdir=./logs/tensorboard_cnn&lt;/code&gt;&lt;/pre&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;img src=&quot;https://api.careers.saigontechnology.com/storage/ck-editor/CKEDITOR-oIJlMKbfLMk7v2BWxfNmJU7BWg8WPSzEGeocUbRE.png&quot;&gt;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Fig4: Tensorboard Dashboard UI&lt;/strong&gt;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;In this code, we use the `&lt;strong&gt;SummaryWriter&lt;/strong&gt;` from TensorBoardX and the `&lt;strong&gt;mlflow`&lt;/strong&gt; library to log the training loss and test accuracy to both TensorBoard and MLflow.&nbsp;&lt;/p&gt;&lt;p style=&quot;text-align:justify;&quot;&gt;&lt;strong&gt;Source: &lt;/strong&gt;https://saigontechnology.com/blog/build-deep-learning-model-for-the-image-classification-task-part2-structure-monitoring-for-training&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;

When starting a new deep learning project, it's essential to plan ahead and think about the overall structure of the project. This will make it easier to maintain your code and data as your project grows, and to collaborate with other developers and researchers.

Hung.NguyenThanh