<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Millennium Bismay on Medium]]></title>
        <description><![CDATA[Stories by Millennium Bismay on Medium]]></description>
        <link>https://medium.com/@millenniumbismay?source=rss-3758b3f99bfc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*wrQ8aDRDie8SGeBXx5UqbA.jpeg</url>
            <title>Stories by Millennium Bismay on Medium</title>
            <link>https://medium.com/@millenniumbismay?source=rss-3758b3f99bfc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 15:58:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@millenniumbismay/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[All You Need to know about RAG — When, Why, and How to use it?]]></title>
            <link>https://medium.com/analytics-vidhya/all-you-need-to-know-about-rag-when-why-and-how-to-use-it-1d70ec2f2ea4?source=rss-3758b3f99bfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d70ec2f2ea4</guid>
            <category><![CDATA[retrieval-augmented]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[genai]]></category>
            <category><![CDATA[information-retrieval]]></category>
            <category><![CDATA[large-language-models]]></category>
            <dc:creator><![CDATA[Millennium Bismay]]></dc:creator>
            <pubDate>Mon, 02 Sep 2024 15:28:16 GMT</pubDate>
            <atom:updated>2024-09-13T08:05:05.955Z</atom:updated>
            <content:encoded><![CDATA[<h3>All You Need to know about RAG — When, Why, and How to use it?</h3><p>With the rapid evolution of Large Language Models (LLMs), Retrieval Augmented Generation (RAG) pipelines have quickly become the de-facto framework for chatbots. These LLMs come with impressive abilities, that allows them to play the role of your ever-helpful digital assistant by following instructions straight from a prompt. The result? Their responses are far more cohesive than those generated by their earlier, less sophisticated counterparts.</p><p>But of course, no superhero is without their kryptonite. They are trained to be <em>relentlessly</em> helpful. They almost never say “No” — and that’s not always a good thing. You see, the real problem arises when they don’t know that they don’t know.</p><p><strong><em>The</em></strong> <strong><em>Hallucination Problem [2]</em></strong>. LLMs are trained on massive volumes of internet data under a Next-Token prediction (or causal) paradigm. Essentially, their job is to predict the next word that’s most likely to follow, based on all the words that have come before it. Now, this might sound clever — and it is! — but it’s also a bit like trying to write a novel one word at a time, without ever planning ahead! The real kicker is that these models don’t come equipped with a built-in fact-checker. So when they stray off the knowledge map, they’re not aware of it — leading to some confident but completely fabricated responses.<em> In short, they lack the ability to stop, reflect, and correct themselves.</em> It’s a bit like having a friend who insists on giving you directions, even though they’re hopelessly lost themselves!</p><blockquote>Hallucination refers to the behavior of how the next tokens (words) predicted by the LLM drifts away from the groundtruth context, resulting in factually irrelevant or wrong outputs.</blockquote><p>RAG aims in keeping the LLMs on track. How? By serving up the most relevant documents that actually answer the user’s query. That’s the key reason RAG has gained so much popularity lately. It takes the impressive In-Context Learning (ICL) abilities of LLMs and anchors them to solid, relevant information, ensuring they deliver responses that are not only cohesive but also accurate and tailored to the user’s needs.</p><p>One crucial aspect of RAG is its ability to query your own data. Sure, we have Google for searching the web, but would you really feel comfortable sharing your personal documents with Google? And what about a company — can it afford to expose its internal documents on the internet? Of course not! That’s where RAG comes to the rescue. Not only can it search through your private documents, but it can also answer queries and generate summaries tailored to your specific data — all while keeping everything private and secure.</p><p>These days, most tutorials barely skim the surface when it comes to the true potential of a fully-fledged RAG framework. Everyone’s in a rush to showcase the power of Generative AI (though let’s be honest — we still have quite a way to go) with as little code as possible. Call four functions from three different libraries and, <em>voilà!</em> Your RAG is up and running! But do you ever wonder what’s actually happening under the hood of those methods from different libraries? And how can we push past these basic RAG systems, which, let’s face it, sometimes perform like a kindergartner trying to search books to explain Newton’s third law of motion? Well, in this blog series — <strong><em>All You Need to Know About RAG</em> </strong>— I’ll take you on a deep dive into the inner workings of the most customizable parts of a RAG framework, with a special focus on building an <strong><em>Agentic RAG</em></strong>. Let’s explore and build together!</p><p>A RAG pipeline has four core concepts —</p><ul><li>Storage</li><li>Retrieval</li><li>Response Generation</li><li>Citations/References</li></ul><p>We all get excited about the magic of ‘<em>Response Generation</em>,’ but let’s not forget — it’s just the final product of all the hard work that came before it. As the saying goes, “You reap what you sow!” This couldn’t be more true for an effective RAG framework.</p><h3><strong>Why is Storage Imporant? Can’t we just dump all at once? What good is using an LLM if we need to focus on storage?</strong></h3><p>Well, well, well… firing off a bunch of questions at once, huh? Let’s break it down. Picture this: you’re writing some code to analyze millions of rows of data using Python and Pandas. Suddenly, you forget how to perform an “Inner Join” on two DataFrames — happens to the best of us! So, naturally, you Google it (don’t worry, you’re in good company!). Now, imagine that instead of taking you directly to the Pandas merge method, Google throws the entire documentation of the Pandas library at you. Sure, you could scroll through and eventually find what you need, but wouldn’t it be far more helpful if Google just pinpointed the exact page you were looking for?</p><p>You know which columns matter for drawing meaningful insights from your data, and Google helps you find the right tool for the job. In the same way, LLMs excel at understanding the context of user queries and answering based on relevant information. But imagine asking them to shift through an entire document instead of just focusing on the most relevant parts! It would be like handing someone a library when all they needed was one book.</p><p>With Large Contexts, LLMs generally suffer from <em>“lost in the middle”</em> problem [3]. It might sound ridiculous, but it has been obseved that LLMs, almost like humans, focus mostly on the start and end of a large document, and might lead to missing key concepts present in the middle of the document.</p><p><strong>Yes, I heard you screaming — What about LLMs supporting very long context length?</strong></p><p>First thing first, The “lost in the middle” problem doesn’t vanish completely — it’s more like it gets a bit of a band-aid. Sure, it’s somewhat mollified, but this can still lead to inconsistencies across different LLMs. And let’s not forget, LLMs that can handle very long contexts tend to be <em>huge</em> (we’re talking 70B+ parameters). Naturally, this comes at the cost of much slower inference speeds. So, it’s a bit of a trade-off — longer context, slower performance. You see, LLMs don’t generate sentences as a whole. They use Transformers Decoder as their backbone. They generate one word/token at a time. Here in the below Fig 1 — Step 1, given the words in orange, ‘<em>Two roads diverged in’</em>, are the input, as the first step, LLM generates — ‘<em>a</em>’, in green, as the output. To generate the next word, in Step 2, it takes the entire text predecessing it as the input — ‘<em>Two roads diverged in a</em>’, and generates — ‘<em>yellow</em>’, and so on. As, every time to generate a new word, it looks at (and processes) all the tokens predecessing it, it takes a Time Complexity of O(n²).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/551/1*1QnWToIZ6CGUnnHsnm3eSg.png" /><figcaption>Fig.1. LLM Inference — One step at a time</figcaption></figure><p>Secondly, we don’t usually search from just one giant encyclopedia, do we? Instead, we gather information from multiple sources and different types of documents. Even if LLMs get <em>really</em> good at handling super long contexts, do you really want to provide a list of documents for them to search through every single time you ask a question? Probably not. What you’d prefer is to simply add a new document, and the next time you query, the LLM should incorporate that fresh info into its response. And let’s not forget the challenge of staying up-to-date. When it comes to integrating RAG with current events, we need the most recent information available. Unfortunately, LLMs have a built-in limitation — they can only know as much as their pre-trained data, which comes with an inevitable knowledge cut-off date. Oops!</p><p>Most of our data is unstructured — raw text, images, htmls, codes, tables in pdfs, etc. Hence we will cover the key concepts of Storage —<em> </em><strong><em>Types of Chunking, Types of Indexing, </em></strong>and <strong><em>Effective Use of Metadata</em></strong> later in this blog series.</p><h3>Isn’t Retrieval just semantic similarity between user query and stored documents? What’s more to it?</h3><p>To answer this question, let’s understand what is semantic similarity with an example:</p><pre>User Query: When should I go to bank before next weekend? I need urgent money.<br><br>Context1: The colorado river is really pretty. There are tons of bars on its<br>          bank! People say the river walk on its bank is awesome. But its<br>          closed on Fridays!<br><br>Context2: Banks are generally open from Monday to Saturday. However, Next week,<br>          Tuesday is a bank holiday.</pre><p>Which one do you think is the most relevant context for the user query? I am sure, most of you agreed on <em>Context2</em>. Well why? Is it because of word/pharse — ‘<em>bank</em>’, ‘<em>before next weekend</em>’, or ‘<em>need urgent money</em>’? Well, I guess we all know. If you look closely, the word ‘<em>bank</em>’ is present two times in Context1 and once in <em>Context2</em>. And the word ‘<em>money</em>’ is not really present in either of the sentences. But its the ‘<em>need for money</em>’ and the structure of sentence in <em>Context2</em>, which makes it clear that the <em>Context2 </em>talks about <em>Bank</em>, a <em>Financial Institute</em>. But how does a LLM know this? — By understanding the contextual similarity between the user query and contexts. A simple keyword search would return <em>Context1 </em>as the better result because of the presence of ‘<em>bank</em>’ two times. However, LLMs use <strong><em>embedding models </em></strong>which are trained on huge corpus of data. And when we pass a text block through these models, we get a vector representation of the text block which captures its context. It uses the position of the words and looks at every word of the text block to identify its context.</p><p><strong>But do we always query with sufficient context? What if our query is open ended and indeed need diverse contexts to be answered? What exactly is the best context to answer our query?</strong></p><p>Believe it or not, most of the time our queries do not have the entire context. Thats when effective retrieval becomes a key. We can use LLMs to augment the query, and this process is called <strong><em>Query Expansion</em></strong>, to extend the context to better retrieve the document. We can also use the <strong><em>Metadata Filtering</em></strong> to narrow down the search space for the retrieval. Now, if our query is too complex, we can break it down into sub-queries and perform retrieval step-by-step and finally aggregate the contexts to respond — this process is called<em> </em><strong><em>Query Transformation</em></strong>.</p><p>Sometimes, retrieving lesser number of documents can lead to a problem called <strong>Recall Ceiling, </strong>i.e. the quality of the generation is constrained by the quality and quantity of retrieval. So, why not retrieve a large number of documents? — Well, that would lead to very large contexts if we just provide all of them to the LLM? You see, that again leads us to the<em> lost-in-the-middle</em> problem. They said truely — “The End is the Beginning and the Beginning is the End! — <em>Dark</em>”. Well, We can use a <strong><em>Neural Re-ranker</em></strong> like ColBert to re-rank the large list of retrieved results. And finally, if our query is too open ended, we might also miss key documents if we just find semantic similar documents, so why not also use the traditional keyword retrievers, like BM25? — Well we can and we call that method <strong><em>Ensemble Retrieval</em></strong>. We will discuss all of these key techniques in blogs to follow.</p><p><strong>But can we guaranteed that all relevant documents would be retrieved by doing all these?</strong></p><p>Unfortunately, No! That’s where<strong><em> Agentic RAG </em></strong>swoops in — cue the drumroll, please! Imagine, you’re searching through multiple books, trying to crack a tough assignment. You dive into the first book, find the relevant topic, and read through it eagerly… but nope, no solution there. So, you move on to the second book. You skim through the familiar sections and check out a few new ones, but still no luck. Determined, you reach for the third book, and <em>voilà!</em> You finally find the approach you need!Now, did you notice something? After each book, you made a decision to move on to the next one because the answer wasn’t in the current book.</p><blockquote><em>That’s exactly what Agentic RAG mimics — </em>it recognizes when the current context isn’t sufficient to answer a query and knows when to dig deeper and fetch more context.</blockquote><p>Pretty smart, right? There are frameworks like <strong><em>Self-RAG</em></strong> that enable this kind of agentic behavior. But let’s not get ahead of ourselves — Self-RAG deserves a tutorial all on its own, which I’ll be covering in future blogs.</p><h3>Wait, That’s a lot of theory! Just tell me quickly — Why do we need Citations? Are we not doing all these other stuffs already ?</h3><p>Alright, let’s get straight to the good stuff — if you handle Storage and Retrieval efficiently, citations won’t be a headache! But yes, we absolutely need citations. In fact, citations are probably the single most important element in building a system that’s <strong><em>faithful </em></strong>and <strong><em>trustworthy</em></strong>. (If only we could design humans this way too!)</p><p>Citations not only provide users with the exact sources from which the responses are generated, but they also help tackle the Hallucination Problem in LLMs. By forcing the model to cite its sources, we create effective guardrails that keep it from wandering too far off track. And don’t worry, the proper use of citations ties into all the other components of RAG, which we’ll be diving into more in upcoming blogs.</p><h3>Hey! Are you cheating? You just missed the most important part — “Response Generation”. Did you do that on purpose?</h3><p>Yes, It might be the most important part as a final product, but as I stated earlier — “You reap what you sow!”, the final Response is ought to be as good as all other components of the RAG framework. And truth be told, given a set of contexts, and a query, LLMs are already very good, sometimes too good, to answer the query in the most coherent manner possible. So, we will see how different RAG frameworks can generate different responses and how we can incrementally improve our final response.</p><p>We can broadly classify the RAG frameworks into the following three categories —</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/987/1*vDGtUbMVXFlx798c0sv-pQ.png" /><figcaption>Fig.2. (left) Iterative retrieval involves alternating between retrieval and generation, allowing for richer and more targeted context from the knowledge base at each step. (Middle) Recursive retrieval involves gradually refining the user query and breaking down the problem into sub-problems, then continuously solving complex problems through retrieval and generation. (Right) Adaptive retrieval focuses on enabling the RAG system to autonomously determine whether external knowledge retrieval is necessary and when to stop retrieval and generation [1]</figcaption></figure><p>We’ll be diving into all these RAG frameworks — with code! — in our upcoming blogs. I know this has been a long, theory-heavy post. Trust me, I wanted to jump straight into the code too! But writing all of this turned out to be way harder than just whipping up some code examples. Still, I realized that understanding <em>when</em> and <em>why</em> to build an advanced RAG system is crucial before diving into the <em>how</em>.</p><p>So, if you’ve scrolled all the way down looking for code and are a bit miffed that it’s not here yet, I encourage you to head back and check out the “When” and “Why” first. If you have read the post, Thank you! I hope you have understood why RAG is here to stay and why we should enhance the basic RAG. I promise, the “How” is coming soon — and I’ll be bringing the code with me! Until then, keep learning, keep growing, and maybe give me a follow if you enjoyed the content and want to know the “How” inside and out!</p><p><strong>References</strong></p><p>[1] <a href="https://arxiv.org/abs/2405.06211">A Survey on RAG Meeting LLMs</a></p><p>[2] <a href="https://arxiv.org/abs/2311.05232">A Survey on Hallucinations in LLMs</a></p><p>[3] <a href="https://arxiv.org/abs/2307.03172">Lost in the Middle</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d70ec2f2ea4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/analytics-vidhya/all-you-need-to-know-about-rag-when-why-and-how-to-use-it-1d70ec2f2ea4">All You Need to know about RAG — When, Why, and How to use it?</a> was originally published in <a href="https://medium.com/analytics-vidhya">Analytics Vidhya</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Getting started with Airflow in 10 minutes…]]></title>
            <link>https://medium.com/analytics-vidhya/getting-started-with-airflow-in-10-minutes-f7c5eef8ebfc?source=rss-3758b3f99bfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/f7c5eef8ebfc</guid>
            <category><![CDATA[deployment]]></category>
            <category><![CDATA[ci-cd-pipeline]]></category>
            <category><![CDATA[airflow]]></category>
            <category><![CDATA[pipeline]]></category>
            <category><![CDATA[docker]]></category>
            <dc:creator><![CDATA[Millennium Bismay]]></dc:creator>
            <pubDate>Sun, 18 Apr 2021 00:36:12 GMT</pubDate>
            <atom:updated>2021-04-23T20:39:15.142Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*vzU8-67mzlp34Hk3HWDR7g@2x.jpeg" /></figure><h4>Apache Airflow 2.0.1 with Docker</h4><p>The surge in technological advancements has called for faster developments and even faster deployments. That’s where enters the CI/CD (Continous Integration/ Continous Deployment) pipelines. But what makes the backbone of a CI/CD pipeline? The workflow or orchestration pipeline! And that is where cometh the Airflow.</p><blockquote>With great data comes greater responsibilities…</blockquote><p><a href="https://airflow.apache.org/"><strong><em>Airflow</em></strong></a> is a platform to programmatically author, schedule and monitor workflows. It is dynamic, extensible, elegant and scalable. Airflow works on the central idea of <strong>Directed Acyclic Graphs (DAGs)</strong> executing <em>tasks</em> with the help of <strong>Executers</strong><em>,</em> scheduled by the <strong>Schedulers</strong>. Now, this blog is not about describing how Airflow works but if you are new to Airflow, I will be putting a link here soon to explain its concepts. To stay true to the title of this article, I will try to keep it as crisp as possible to our goal — Running an Airflow 2.0 with Docker. This article assumes that you are familiar with Docker and Docker-compose. Well if you are not familiar with Docker and you have to set an Airflow server up and running, just follow this article step by step and I promise you will be done in no time.</p><h3>Setting up Docker</h3><ol><li>Make sure that there are no previous docker installation exists in your system.</li></ol><pre>$ sudo apt-get remove docker docker-engine docker.io containerd runc</pre><p>2. Update apt packages.</p><pre>$ sudo apt-get update</pre><p>3. Enable HTTS for apt.</p><pre>$ sudo apt-get install apt-transport-https ca-certificates curl gnupg-agent software-properties-common</pre><p>4. Add docker identification key.</p><pre>$ curl -fsSL https://download.docker.com/linux/ubuntu/gpg \<br>  | sudo apt-key add - \<br>  &amp;&amp; sudo apt-key fingerprint 0EBFCD88</pre><pre>OK<br>pub rsa4096 2017-02-22 [SCEA]<br>9DC8 5822 9FC7 DD38 854A E2D8 8D81 803C 0EBF CD88<br>uid [ unknown] Docker Release (CE deb) &lt;docker@docker.com&gt;<br>sub rsa4096 2017-02-22 [S]</pre><p>5. Add the stable release repository to your package manager</p><pre>$ sudo add-apt-repository \<br> &quot;deb [arch=amd64] https://download.docker.com/linux/ubuntu \<br> $(lsb_release -cs) \ stable&quot;</pre><p>6. Install the latest stable docker version</p><pre>$ sudo apt-get update &amp;&amp; sudo apt-get install docker-ce docker-ce-cli containerd.io</pre><p>7. Verify Installation</p><pre>$ sudo docker run hello-world</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YDeKbG4DpW8FZGHe6cdK9w.png" /><figcaption>To run the hello-world image, it fetches from the remote repository. This happens when the image is not present in your system. The next time you run this, it will just fetch from the system</figcaption></figure><pre>$ sudo docker images</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5dQFdqXrcH9CakxWXic1mQ.png" /><figcaption>List of all docker images present in your system</figcaption></figure><pre>$ sudo docker ps -a</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wvk31luNfA8752dMspXnzg.png" /><figcaption>List of all docker container in your system</figcaption></figure><p>10. Install docker-compose</p><pre>$ sudo curl -L &quot;https://github.com/docker/compose/releases/download/1.29.1/docker-compose-$(uname -s)-$(uname -m)&quot; -o /usr/local/bin/docker-compose</pre><pre>$ sudo chmod +x /usr/local/bin/docker-compose</pre><p>If all the steps worked till this point.. Kudos! We are all set for the airflow setup!</p><h3>Setting up Airflow</h3><p>We are going to use the official Apache Airflow docker image. This uses the docker-compose in order to build the image. Airflow is not a stand-alone service. It needs a lot of other services to be able to run. All these services are dockered (let&#39;s say, I just created this verb) within and you need not worry about any of them explicitly!</p><p><strong><em>Lets Start!</em></strong></p><ol><li>Create a folder caller <strong>airflow-docker </strong>and move inside the folder.</li></ol><pre>$ mkdir airflow-docker<br>$ cd airflow-docker</pre><p>2. To deploy Airflow on Docker Compose, you should fetch docker-compose.yaml file.</p><pre>$ curl -LfO &#39;https://airflow.apache.org/docs/apache-airflow/2.0.1/docker-compose.yaml&#39;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fGpjkytd1FqRfvEN3Qcc-Q.png" /></figure><p>This will download the <em>docker-compose.yaml</em> file into your directory. Lets understand the important sections of this docker-compose file.</p><h4>Services</h4><p>This section is the most important section of any docker-compose file which describes all the services required to run the image.</p><p>As we can see here, There are multiple services like <strong>postgres</strong>, <strong>redis</strong>, <strong>airflow-webserver, airflow-scheduler, airflow-worker </strong>and <strong>airflow-init</strong>.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/4312fe5895990d83d69036b5cc45603c/href">https://medium.com/media/4312fe5895990d83d69036b5cc45603c/href</a></iframe><p>For every service, we can see their properties like their respective image names and ports described under them. This way of composing the docker-compose refrains you from personally doing all these tasks and the airflow-init initializes all the other services for you. Here for starters, you can see the airflow-webserver is bind with the port 8000.</p><h4>Volumes</h4><p>The volumes are supposed to be <em>mounted</em> to the docker container airflow path so that any changes to your local files can be directly replicated in the actual container airflow path folders. Mounting is a very important concept in docker and helps to keep the CI/CD pipeline effective.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/8a2860d36eb84c92337d32dd9858a2f1/href">https://medium.com/media/8a2860d36eb84c92337d32dd9858a2f1/href</a></iframe><p>3. This brings us to our next step. As we can see, by default the name of the folders are <strong>dags, logs </strong>and <strong>plugins</strong>, we need to create these folders in our current directory.</p><pre>$ mkdir dags logs plugins</pre><p>The current folder structure should look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R-uZz_DKlbsrcxNljOk_Qg.png" /><figcaption>Current folder structure</figcaption></figure><p>4. To give the root permission in Mac OS or Linux, you need to run the following:</p><pre>$ echo -e &quot;AIRFLOW_UID=$(id -u)\nAIRFLOW_GID=0&quot; &gt; .env</pre><p>This will create a new file .env and write the above-mentioned things into it.</p><p>5. Initialize the Airflow</p><pre>$ sudo docker-compose up airflow-init</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XgH6sUVJugUeT7BFzIlxTA.png" /><figcaption>Initializing airflow for first time</figcaption></figure><p>This will take some time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vLyw5LpFYBK-0BbXDSXmGA.png" /></figure><p>Once all the layers are pulled you will be able to see the <em>airflow-docker_redis_1, airflow-docker_postgres_1, airflow-docker_airflow-init_1</em> marked as <em>done</em>.</p><p>Now with the following command, you can see all the images present in your system. Mark that we had the hello-world image already present while setting up docker.</p><pre>$ sudo docker images</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RYblnmBf2-Bewe8iAMaDYA.png" /></figure><p>We have three new images: <strong>redis</strong>, <strong>postgres</strong> and <strong>apache/airflow.</strong></p><p>6. Starting the Airflow</p><p>And finally we are going to start the Airflow. That means we are going to start the Airflow and all its dependency services like redis, postgres, webserver, worker and scheduler.</p><pre>$ sudo docker-compose up</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MqUQbPrYZcwysLSsS57fTQ.png" /><figcaption>Airflow Up and Running</figcaption></figure><p>Now to check if all services are running or not, type the following command to see all the running containers.</p><pre>$ sudo docker ps</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BljcFQjAl8PZIvVL9oVHbA.png" /><figcaption>All running containers</figcaption></figure><p>As we can see, we have the following containers:</p><ul><li>apache/airflow: running at <a href="http://localhost:8080/">http://localhost:8080</a></li><li>postgres:13</li><li>redis:latest</li></ul><p>7. Go to the UI</p><p>As we have the web server up and running at the <a href="http://localhost:8080/">http://localhost:8080</a>, just go to the address and you will be presented with the below screen</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sDJMfo2hFwnhVBSkgUl3pw.png" /><figcaption>Airflow Login Page</figcaption></figure><p>In the username and password enter <em>airflow</em></p><p>And click on <em>Sign In</em></p><blockquote>Congratulations! You have set your first Airflow server up and running.</blockquote><p>You will be able to see a lot of DAGs already present. These are the example dags that are present in the Apache Airflow image by default.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fod7eAiqTm3GGcQztzuzMw.png" /><figcaption>Airflow Web UI</figcaption></figure><p>This set-up procedure is possible because of the beauty of docker, integrated with the simplicity of the orchestration pipeline. Apache Airflow paves the way to a larger ecosystem of developing, integrating and deploying in a continuous fashion by reducing the overhead of maintaining everything on our own. I will be discussing the concept behind Airflow and how to create your first DAG in detail in my future articles.</p><p>I hope you liked it. Please reach out to me in case you face any difficulty while going through this article or setting up the Airflow. Also, please do clap, comment and follow me if you find this useful and I will be creating more contents for everyone!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f7c5eef8ebfc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/analytics-vidhya/getting-started-with-airflow-in-10-minutes-f7c5eef8ebfc">Getting started with Airflow in 10 minutes…</a> was originally published in <a href="https://medium.com/analytics-vidhya">Analytics Vidhya</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Do you want to be 'The White Tiger' ?]]></title>
            <link>https://millenniumbismay.medium.com/do-you-want-to-be-the-white-tiger-b50ffe78760d?source=rss-3758b3f99bfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/b50ffe78760d</guid>
            <category><![CDATA[tragic-and-dark]]></category>
            <category><![CDATA[movies]]></category>
            <category><![CDATA[the-white-tiger]]></category>
            <category><![CDATA[priyanka-chopra]]></category>
            <category><![CDATA[rajkummar-rao]]></category>
            <dc:creator><![CDATA[Millennium Bismay]]></dc:creator>
            <pubDate>Sat, 23 Jan 2021 18:29:55 GMT</pubDate>
            <atom:updated>2021-01-23T18:29:55.047Z</atom:updated>
            <content:encoded><![CDATA[<p>This week, Netflix brought to us the movie - &#39;The White Tiger’, based on the novel by Aravind Adiga, with the same name, which depicts the rags-to-rich story of an Indian boy, born in a poverty ridden, small village of India. Now this is not a movie review, but not speaking of the brilliant performances by Adarsh Gourav (Balram Halwai), Priyanka Chopra Jonas (Pinky) and Rajkummar Rao (Ashok) would not let me sleep in the night peacefully. The movie which Netflix characterizes as - Offbeat, Provocative, Gritty, Social Issue drama and Underdog is also Dark and Tragic. And the three main actors have dignified their characters by keeping themselves grounded and accepting the reality of the story. So from here on, we shall not discuss the movie plot, but the question asked in the title of this piece. Also there might be a few spoilers ahead, for references, so I would request you to watch the movie before reading from this point onwards or make peace with the spoilers!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/312/1*I-GLOW4ive0Kq2LKXepVwA.jpeg" /></figure><p>Its not the shackles that hold us but the lack of will power to break the shackles. We have heard a lot of stories to break apart the invisible sackles and the social stigma and grow more into oneself. We prepare for it. We decide we shall break the shackles but we wait for that one lucky and pretty day. And alas!, most of us never made to that day. And we blame the situational irony to hide lack of will power. The story revolves around the dreams of its characters and value of their true nature. All the main characters have different upbringings at different places hugely different from each other. The societal differences made their shackles different. It need not be one fit-for-all solution to delve into the ocean and learn swimming, but every one has to find their ocean. The movie shows two dreams, changing into other duties and way of living. It shows the shackles can&#39;t just be broken by money and power. And it shows the will power might come from something tragic. The dream might not be an inherent one, but how one grows into it.</p><p>And then comes - &quot;The way for a poor person to reach the top is either by Crime or by Politics&quot;! Now this line might not be a fit-for-all epilogue, but it tells one thing, you need to get your hands dirty. The movie shows Balram killing his master, Ashok, to free himself from slavery, he had thought himself to be destined for quite long, and just before being replaced from his job by another person. We, most of the times, respond to occurence of an even rather than to a possibility of occurence of the event. This sometimes put us at some unprecedented results to be followed by. The connection between the upbringing of a person and his/her response to a situation might be highly correlated but its also the survival instinct that brings the best or worst in us out. And as they said, &#39;The White Tiger&#39; is born once in many generations, to break the notions set by the society, was actually lesser born than transformed into one.</p><p>Well, so coming back to the question... &quot;Do you need to be &#39;The White Tiger&#39;?&quot; Well the White Tiger is not really a person or a conscience. It&#39;s the moment, the moment which makes you realise your worth. And that moment, that very moment is when the shackles are the weakest. Ofcourse I don&#39;t agree completely with all the dialogues of the movie, but I do feel strongly, every person has his/her own &quot;White Tiger&quot; moments! And yes they are scarce, but they are not once in a life time moments. So you defintely need to have your &#39;White Tiger&#39; moment and cut loose of your shackles to achieve what you want for yourself, cause destiny was never written, its a word choosen by cowards for coincidence!</p><p>CIAO!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b50ffe78760d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Training a Tensorflow model on Browser: Columnar Data (Iris Dataset)]]></title>
            <link>https://medium.com/analytics-vidhya/training-a-tensorflow-model-on-browser-columnar-data-iris-dataset-ff451c5468e4?source=rss-3758b3f99bfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff451c5468e4</guid>
            <category><![CDATA[tensorflowjs]]></category>
            <category><![CDATA[teachable-machine]]></category>
            <category><![CDATA[training-on-the-browser]]></category>
            <category><![CDATA[tensorflow]]></category>
            <dc:creator><![CDATA[Millennium Bismay]]></dc:creator>
            <pubDate>Fri, 14 Feb 2020 13:51:42 GMT</pubDate>
            <atom:updated>2020-02-18T05:55:18.714Z</atom:updated>
            <content:encoded><![CDATA[<h3>Training a Tensorflow JS model on Browser: Columnar Data (Iris Dataset)</h3><p>With the inception of the <a href="https://teachablemachine.withgoogle.com/train">Teachable Machine</a>, there is a new excitement around the world for training ML models in the browser itself. Everyone wants to see the pipeline on the go. Gone are those days when you train a model on a separate machine, create the backend on another machine and deploy the whole ecosystem on another. And hence, comes the Tensorflow JS (TFjs) in action. There are many advantages to this approach as well:</p><ul><li>You could distributively train an ML system using the <strong>user’s resources</strong>.</li><li>It could give <strong>higher privacy</strong> to users who don’t want their resources to leave their system.</li><li><strong>Faster Training</strong> than keeping the model running on a server.</li><li>Using a trained big network is not very slow (the training is slow), so you could <strong>deploy the network on the client-side</strong> instead of on the server to reduce lag when the model has to be used in real-time.</li><li>The models generated are <strong>smaller (in KBs)</strong> compared to their contemporary Tensorflow developed models (in MBs).</li><li>The TFjs Framework is very much similar to the TF framework and hence easy to get your groove on.</li></ul><p>So now, get ready to have your hands dirty with some JavaScript codes to train your model on the browser.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*_goV6BV1jk2gdkPf9MlAXg.png" /></figure><p>For this exercise, we will go with a very famous dataset: <a href="https://archive.ics.uci.edu/ml/datasets/Iris">The Iris Dataset</a>. Ohh! But wait… Machine Learning is a funny business. You need to set your environment damn right to work everything out. So here is a list of things you need to install before you get going (can be installed using pip):</p><ol><li><a href="https://pypi.org/project/tensorflowjs/">Tensorflow JS 1.5.</a>2</li><li><a href="https://linuxize.com/post/how-to-install-sublime-text-3-on-ubuntu-18-04/">Sublime</a> or <a href="https://linuxize.com/post/how-to-install-visual-studio-code-on-ubuntu-18-04/">VSCode</a></li><li><a href="https://chrome.google.com/webstore/detail/web-server-for-chrome/ofhbbkphhbklhfoeikjpcbhemlocgigb/related?hl=en">Web Server for Chrome</a></li></ol><p>Okay! So now as you are ready, let’s create a folder named <strong>iris_classifier</strong> and dive in.</p><h3>A Simple Web Page</h3><ul><li>Create a file named <strong>Iris_Classifier.html</strong> inside the folder created.</li></ul><pre>&lt;html&gt;</pre><pre>&lt;head&gt;&lt;/head&gt;</pre><pre>&lt;script lang=”js”&gt;&lt;/script&gt;</pre><pre>&lt;body&gt;</pre><pre>&lt;h1&gt;Iris_Classifier&lt;/h1&gt;</pre><pre>&lt;/body&gt;</pre><pre>&lt;/html&gt;</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VS_763m5MQKg63oNtDIfDA.png" /><figcaption>The web page would look like this</figcaption></figure><ul><li>We will not go into any fancy web page designing stuff in this tutorial.</li></ul><h3>Understand the Data</h3><ul><li>Download the data from this <a href="https://archive.ics.uci.edu/ml/datasets/Iris">link</a>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/428/1*kz0KdlHxL_u-iS4m1LslrA.png" /><figcaption>Sample data from the dataset</figcaption></figure><ul><li>Here the attributes are <strong>sepal_length</strong>, <strong>sepal_width</strong>, <strong>petal_length</strong>, <strong>petal_width</strong> and <strong>species</strong>.</li><li><strong>species</strong> is the label given and hence the objective of the model to be trained is predicting the label of a given set of data.</li><li>In the dataset, there are 3 different species: S<strong>etosa</strong>, <strong>Versicolor</strong> and V<strong>irginica</strong>. As we have to predict one of the three, we should go for <a href="https://machinelearningmastery.com/why-one-hot-encode-data-in-machine-learning/"><strong>one-hot encoding</strong></a>.</li></ul><blockquote>Setosa: 1 0 0</blockquote><blockquote>Versicolor: 0 1 0</blockquote><blockquote>Virginica: 0 0 1</blockquote><ul><li>The test data should not contain the <strong>species</strong> column.</li></ul><h3>Let’s Start Coding</h3><ul><li><strong>The codes should be written within the script tags in an HTML file.</strong></li></ul><h4>Importing Tensorflow JS</h4><p>We know there are some packages to import. It is achieved in JavaScript via the following snippet</p><pre>&lt;script src=”https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@latest&quot;&gt;&lt;/script&gt;</pre><h4>Reading the Dataset</h4><ul><li>We need to read the file from the <strong>iris.csv</strong> file.</li><li>We need to specify the label of the dataset by setting the <strong>isLabel </strong>to<strong> true.</strong></li></ul><pre>const csvUrl = ‘iris.csv’;</pre><pre>const trainingData = tf.data.csv(csvUrl, {</pre><pre>columnConfigs: {</pre><pre>species: {</pre><pre>isLabel: true</pre><pre>}}});</pre><h4>Data Transformation</h4><ul><li>Now we need to convert the data into the format we need to feed to the model.</li><li>The first part is mapping the <strong>xs (attributes)</strong> to their respective <strong>ys (target label)</strong>.</li></ul><pre>const convertedData =</pre><pre>trainingData.map(({xs, ys}) =&gt; {</pre><pre>const labels = [</pre><pre>ys.species == “setosa” ? 1 : 0,</pre><pre>ys.species == “virginica” ? 1 : 0,</pre><pre>ys.species == “versicolor” ? 1 : 0</pre><pre>]</pre><pre>return{ xs: Object.values(xs), ys: Object.values(labels)};</pre><pre>}).batch(10);</pre><ul><li>The second part is the <strong>one-hot encoding</strong>. Hold on guys, it gets a little tricky around here, we do not have our beloved sklearn.preprocessing.OneHotEncoder as in Python to do the same. We are assigning <strong>ys.species = [1 0 0] if its a Setosa</strong>,<strong> ys.species = [0 1 0] if its a Virginica </strong>and<strong> ys.species = [0 0 1] if its Versicolor</strong>.</li><li>We create a <strong>convertedData </strong>by passing the <strong>xs</strong> and the <strong>modified ys</strong> in <strong>batch of 10</strong>.</li></ul><p>Before getting into the model architecture, we need to specify one last thing — the <strong>number of input features/neurons</strong>.</p><pre>const numOfFeatures = (await trainingData.columnNames()).length — 1;</pre><h4>Building and Compiling the Model</h4><ul><li>This starts in a very familiar way, if you have already experience with building Keras models, the<strong> tf.sequential()</strong></li><li>Here I am presenting a very basic neural network with one hidden layer with 5 units. You need to mention the <strong>inputShape (<em>numOfFeatures</em>)</strong>, <strong>activation (<em>sigmoid</em>) </strong>and <strong>units (<em>5</em>) </strong>while adding hidden layers to the model.</li><li>In the final layer, as we have used One-hot encoding for the target output, we should expect 3 output units. The activation we will be using is <strong><em>softmax</em></strong>.</li></ul><pre>const model = tf.sequential();</pre><pre>model.add(tf.layers.dense({inputShape: [numOfFeatures], activation: “sigmoid”, units: 5}))</pre><pre>model.add(tf.layers.dense({activation: “softmax”, units: 3}));</pre><p>Once we have done building the model, we need to compile it.</p><ul><li>We need to provide the <strong>loss (<em>categoricalCrossentropy</em>)</strong> and <strong>optimizer (<em>adam</em>)</strong> with a <strong>learning rate (<em>0.06</em>)</strong>.</li></ul><pre>model.compile({loss: “categoricalCrossentropy”, optimizer: tf.train.adam(0.06)});</pre><pre>console.log(“Model Summary:\n”, model.summary());</pre><ul><li><strong>model.summary() </strong>gives the layer by layer architecture of the model with the number of parameters in each layer.</li></ul><h4>Training the Model</h4><ul><li>The <strong>model.fitDataset()</strong> method does the model training.</li></ul><blockquote>The very interesting part is the <strong>await </strong>expression. <strong>An await expression can only be run in an async function</strong>. The <em>await</em> expression causes <em>async</em> function execution to pause until a <em>Promise</em> is settled (that is, fulfilled or rejected), and to resume execution of the <em>async</em> function after fulfillment. When resumed, the value of the <em>await</em> expression is that of the fulfilled <em>Promise</em>.</blockquote><ul><li>We feed the number of <strong>epochs</strong> we need the model to be trained on.</li><li>The callbacks are used to keep a track of how the model has performed from the starting of the training phase.</li></ul><pre>await model.fitDataset(convertedData,</pre><pre>{epochs:100,</pre><pre>callbacks:{</pre><pre>onEpochEnd: async(epoch, logs) =&gt;{</pre><pre>console.log(“Epoch: “ + epoch + “ Loss: “ + logs.loss);</pre><pre>}</pre><pre>}});</pre><h4>Testing the Model</h4><p>Finally, we reach the moment of truth. The result.</p><ul><li>The test set should not contain the <strong>label</strong>.</li><li>There is no <strong>numpy</strong> array in JavaScript, hence we need to send the data in a <strong><em>2d tensor format</em></strong>: <em>tf.tensor2d() which also needs the shape of the 2d tensor</em>.</li></ul><pre>const testVal = tf.tensor2d([5.0, 1.0, 2.0, 8.0], [1, 4]);</pre><pre>const prediction = model.predict(testVal);</pre><pre>const pIndex = tf.argMax(prediction, axis=1).dataSync();</pre><pre>const classNames = [“Setosa”, “Virginica”, “Versicolor”];</pre><pre>alert(classNames[pIndex])</pre><ul><li>We get the model to predict via <strong>model.predict()</strong>.</li></ul><h4>Final Nail in the Coffin</h4><p>We have come to the end of our coding session. Just one last thing. The procedure that we are discussing, need to run without the user waiting to load the webpage but simultaneously the flow must be waiting up for the model to be trained, else all of these will be meaningless.</p><p>Hence, We have to run it in an <strong><em>async</em></strong> function. We have already discussed the <strong><em>await</em></strong> expression which can only be declared inside an async function.</p><blockquote>The purpose of <em>async/await</em> is to simplify using promises synchronously, and to perform some behavior on a group of <em>Promises</em>. As <em>Promises</em> are similar to structured callbacks, <em>async/await</em> is similar to combining <a href="https://developer.mozilla.org/en-US/docs/Web/JavaScript/Guide/Iterators_and_Generators">generators</a> and promises.</blockquote><ul><li>We need to put the whole code inside an async function, i.e. run()</li></ul><pre>async function run(){</pre><pre>const csvUrl = ‘iris.csv’;</pre><pre>const trainingData = tf.data.csv(csvUrl, {</pre><pre>columnConfigs: {</pre><pre>species: {</pre><pre>isLabel: true</pre><pre>}</pre><pre>}</pre><pre>});</pre><pre>const numOfFeatures = (await trainingData.columnNames()).length — 1;</pre><pre>const convertedData =</pre><pre>trainingData.map(({xs, ys}) =&gt; {</pre><pre>const labels = [</pre><pre>ys.species == “setosa” ? 1 : 0,</pre><pre>ys.species == “virginica” ? 1 : 0,</pre><pre>ys.species == “versicolor” ? 1 : 0</pre><pre>]</pre><pre>return{ xs: Object.values(xs), ys: Object.values(labels)};</pre><pre>}).batch(10);</pre><pre>const model = tf.sequential();</pre><pre>model.add(tf.layers.dense({inputShape: [numOfFeatures], activation: “sigmoid”, units: 5}))</pre><pre>model.add(tf.layers.dense({activation: “softmax”, units: 3}));</pre><pre>model.compile({loss: “categoricalCrossentropy”, optimizer: tf.train.adam(0.06)});</pre><pre>await model.fitDataset(convertedData,</pre><pre>{epochs:100,</pre><pre>callbacks:{</pre><pre>onEpochEnd: async(epoch, logs) =&gt;{</pre><pre>console.log(“Epoch: “ + epoch + “ Loss: “ + logs.loss);</pre><pre>}</pre><pre>}});</pre><pre>const testVal = tf.tensor2d([5.0, 1.0, 2.0, 8.0], [1, 4]);</pre><pre>const prediction = model.predict(testVal);</pre><pre>const pIndex = tf.argMax(prediction, axis=1).dataSync();</pre><pre>const classNames = [“Setosa”, “Virginica”, “Versicolor”];</pre><pre>// alert(prediction)</pre><pre>alert(classNames[pIndex])</pre><pre>}</pre><ul><li>Now call the <strong>run()</strong> function to make this whole thing run.</li></ul><pre>run();</pre><h3>Let’s see the Action</h3><ul><li>Go to the <strong>Web Server for Chrome</strong> window and add the folder we created in the starting — Iris_Classifier.</li><li>Select <strong>Run in background</strong>.</li><li>Click on the Web Server URL(s) which will be a localhost address with a port, i.e. <a href="http://127.0.0.1:8887/"><strong>http://127.0.0.1:8887</strong></a></li><li>Once you land on the <strong>Index of Current Directory</strong> page, click on our HTML file — iris-classifier.html</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EBjih9JsQ5ipEK8sPSHXUQ.png" /></figure><ul><li>This will lend you on a page where you will just see a big blank white web page with Iris_Classifier scribbled on the top left. Don’t worry. Press <strong><em>Ctrl + Shift + I</em></strong> <strong>(Developer Tools). </strong>You will get to see the console readings.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aQjGQIq4LRQVaZ7ofM_ZUw.png" /><figcaption>The initial phases of Training</figcaption></figure><ul><li>The Console log shows us the Epoch number and its corresponding loss.</li><li>The model is being trained in the background. Once the model is trained, you see an alert with the name of the species for the test set you provided in the code.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lMqO_l9V3vBc0tNG1wuNPw.png" /><figcaption>Prediction for the test set after the model is run for 100 epochs</figcaption></figure><ul><li>If you see closely, the loss has been decreasing from <em>epoch 0</em> to <em>epoch 99</em>.</li></ul><h3>Wrapping it Up</h3><p>Here we are at the end of this tutorial and I hope you have a very basic idea of how to train a model on the browser. This is my first try at technical blogging in Medium. Please provide me with suggestions to improve any part as well as reach out to me whenever you have any doubts on the coding front. Also, here is the <a href="https://github.com/millenniumbismay/Medium-Blogs/tree/master/Iris_Classifier">link</a> to the code in my <a href="https://github.com/millenniumbismay">GitHub</a> account.</p><p>Thank you. Hope you like it. Please leave your comments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff451c5468e4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/analytics-vidhya/training-a-tensorflow-model-on-browser-columnar-data-iris-dataset-ff451c5468e4">Training a Tensorflow model on Browser: Columnar Data (Iris Dataset)</a> was originally published in <a href="https://medium.com/analytics-vidhya">Analytics Vidhya</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>