Inc. (United States), Slidebean Inc. (United States), Zoho (India) are some of the key players that are part of study coverage. Additionally, the players which are also part of the research ...

Presentation Software Market to Witness Major Growth by 2028 | Easy WebContent, Inc. Prezi Inc. , Google

Scrum is a popular framework for agile project management that emphasizes collaboration, adaptability, and iterative development. While Scrum includes ceremonies like Sprint Planning, Daily Stand ...

What's beyond SCRUM ceremonies

This feature can be particularly useful for remote presentations, where the personal connection can sometimes be lost. Slidebean is a web-based presentation tool that revolutionizes the way presentations are made. With just a few clicks, users can create ...

10 Best AI Presentation Generators (September 2023)

And there were even a couple of attempts to build businesses using AI like HustleGPT and Slidebean. HustleGPT was the brainchild of Jackson Fall, who asked ChatGPT to turn $100 into a profitable ...

What roller skating says about AI

10 parimat tehisintellekti esitluse generaatorit (august 2023)

This is just me sharing my thoughts with people who like what I share! I was watching a cool Slidebean video about El Salvador and its failed implementation of Bitcoin as a legal tender.

Where Did El Salvador go Wrong?

Capture your audience's attention with a unique visual storytelling experience. Slidebean: Don't have the time to design your slides? Slidebean uses AI algorithms to automatically create beautiful ...

10 Powerful AI Tools to Create Amazing Presentations: Boost Your Impact and Wow Your Audience!

1. Are you really a start-up or are you simply starting a small business? According to NerdWallet, a start-up is distinct from a small business in that the intention behind it is to make an impact ...

6 things you need to know before you start a start-up

Growing up, I was unfamiliar with what defines an entrepreneur. It didn't know what they looked like, what character traits they embodied, what habits they possessed, what they studied, or their ...

Getting Ahead: At Being Entrepreneurial

Have you ever felt a weird feeling like you do not know what to do at work? If you do, it is likely not your fault. Product roadmaps, backlog, refinement, and planning are common words in the ...

Prioritizing tasks for product roadmaps

Slidebean

slidebean

To kick off this year’s “12 startups in 12 months” challenge that I’m building in public as an indie hacker, I started working on my first startup, but my startup failed. Some important lessons were learned. What is the percentage of startups that fail? The failure rate of startups is more than 80%. It is extremely difficult to land a winning idea on the first attempt. Someone has to be really lucky. In this article, I will summarize my story and point out five reasons why my first startup of the year failed. The First Startup — Mixio (No Code Landing Page Builder) To find the idea for my first startup of the year, initially, I wanted to build a generic no-code tool that could be used to build any web application. But the scope of this project would be too much for one month, and also it’s always better to focus on a narrower target audience when starting a project. In the end, I decided that indie hackers were my target audience and chose to build a tool that validates startup ideas faster. Eventually, my product became a tool that entrepreneurs or startup founders can use to validate as many ideas as quickly as possible. I wanted to make the experience less time-consuming and to the point where it would be easy for someone new to start building landing pages. I used David’s six criteria for selecting my startup ideas for this challenge. And it did tick all the checkboxes. ✅ ✅ ✅ ✅ ✅ ✅ Has product-market-fit been proven by something similar? Is it B2B? Can the MVP be developed within ≤ 3 Months (max)? Can it be managed without employees? Can it work without making active sales? Can I identify myself with the product? My “Market Research” I pitched my startup idea to some fellow community members of and also created to see which of my ideas are the most popular. Wannabe Entrepreneur Virtual Co-working Space a poll on indiehakers.com IndieHackers.com Poll Result In the indiehackers.com poll, it was the second favorite idea. If it had been logical, I should have started with the most favorite idea for my startup. Since I’ve decided on my target audience already, I chose the second-best idea anyway. Landing Page of my Landing Page Builder I quickly created a landing page for Mixio using the bare minimum of information necessary. I needed user feedback, so I created the page as quickly as possible. The initial landing page of Mixio.net The signup link was just a Google form. The users had to fill out the information about their product in order to generate the landing page. It was a manual process. I only managed to get one user to sign up for my Google form. This is the landing page of my first user that is created by Mixio — . nuxo.org Launched Improved MVP The current version of the landing page is not converting, which means that the MVP needs to be more convincing in order to get people to convert. I started building a minimal working version of my landing page builder so that the users would be able to see the generated landing page right away. I launched my MVP with the bare minimum features needed for my users to get started. You can start using right away, without having to sign up. my landing page builder Building your landing page is a breeze This is an easy-to-use tool that does not require any coding or design skills to build a landing page. The other notable landing pages generated by Mixio are and . CollabClub.co LabSocial.com Traction Photo by Slidebean on Unsplash I have not seen the classic hockey stick growth in my data, unfortunately. I’m using an analytics tool called to track my user behavior. It’s not looking good — the engagement and conversion rate is really low. Even though I didn’t do much marketing, the conversion rate tells, and the number never lies. Heap As of today, only a small number of people have signed up for the service. And around 36 landing pages are being created by both registered and non-registered users. Status — “Startup Failed” Even though I think this startup is a failure, I won’t give up on it. My goal isn’t to give up after one month of a project — I want to see what happens in the long run. Also, I may have a better idea to pivot to something else in the future. You never know. And my goal for taking this challenge was to try out many ideas that could become a profitable SaaS startup. At least, I will be able to quickly build landing pages for my next 11 startups in this challenge. 😉 Lessons Learned So why did my startup fail? I will outline some reasons why my first startup of the year-long 12-startup challenge failed. The lessons are based on the feedback I have received from my target audience so far. These may not be the only factors at play. Photo by Sarah Kilian on Unsplash 1. I didn’t choose the most popular idea This one is obvious. It would’ve been better to go with the most popular idea from the market research poll that I did on indiehackers.com. I see a lot of potential in the idea that the audience chose. 2. I asked the wrong questions I wasn’t able to read my customers’ minds because I didn’t ask the right questions. The key to finding out the truth about your potential customers is to ask the right questions. To find the truth it’s important to ask questions that will get you accurate information. 3. The market is saturated There are many landing page creators available on the market, which makes it difficult for new businesses to compete in this market. I don’t think I could compete with the big players who are already offering free services, and new tools are coming out every day. 4. Entrepreneurs are emotional about their landing pages Some entrepreneurs or indie hackers find building landing pages to be an emotional process. They want to design it themselves and make it perfect. 5. Technical entrepreneurs are comfortable with HTML templates If the founder is a developer or designer, they may be more comfortable hosting their own landing pages with HTML5 templates. I learned this lesson the hard way. 😞 What’s Next to My Journey as an Indie Hacker? I don’t consider this a failure at a personal level, as most startups are supposed to fail! And I may still have plans for Mixio in the future, as the no-code movement is growing rapidly. I just want to try as many ideas as possible before one really sticks. The first startup I attended taught me a few important lessons and helped me improve myself so that I could execute any startup idea better. That’s the whole point of this 12-month long fun experiment. My next idea is obviously the most voted one — a Twitter growth tool. It’s called . Feel free to sign up if you are curious about this tool. 🤗 LabSocial Follow My Journey of Building in Public Photo by Ricky Kharawala on Unsplash If you want to see how my story unfolds and want to support me or just want to come along for the ride with me, please and . It means a lot to me. subscribe to my newsletter follow me on Twitter If you are a maker yourself or want to dive into the journey of , please feel free to . building in public shout out to me on Twitter Also published . here

Why My First Startup Failed: 5 Lessons Learned

Growing up and entering adulthood post-college, I always knew that I wanted to be in a position to help the Black community succeed and break barriers. Prior to becoming an entrepreneur, I did this while working at my old 9-5 where I led teams and was in a position to hire. I made it a habit to seek out Black candidates and made it a point to mentor and prepare my Black colleagues and employees into higher positions of power. Even outside of hiring, I intentionally sought out Black-owned businesses and Black creatives for special projects and needs that my company had. Upon leaving my 9-5, I knew that there was an opportunity for me to do to help advance the Black community. In response to that, I launched , and we're a media platform created to help Black entrepreneurs and professionals succeed through smart, actionable content and resources. more Mogul Millennial Since launching, we've been able to reach hundreds of thousands of young, ambitious Black professionals and business owners across the U.S., and have connected with Black leaders at top organizations and startups such as Harlem Capital, Snap Inc., Calendly, and Squire to teach our community valuable and tangible lessons to help them professionally grow. Currently, Black people are making tremendous strides in corporate, and are among the fastest-growing group of entrepreneurs. However, unfortunately, there is still an alarming wealth and funding gap, and a lack of resources that's given and created specifically for Black entrepreneurs. For this reason and more is why Mogul Millennial exists, and why Mozilla Builders to provide resources for the Black entrepreneurial community. we're working with was created to join forces with people who want to build technology that fixes the internet and helps the world. At Mogul Millennial, we love working with companies and allies who want to help us build the blueprint for Black success, and provide the resources that Black founders need to build profitable businesses and create generational wealth. Mozilla Builders With the support of Mozilla Builders, Mogul Millennial is launching a new educational series for early-stage Black founders, . Launch Like a Mogul Launch Like a Mogul is a free, 6-week that's teaching early-stage Black founders how to launch a winning growth and monetization plan for their current business (or business idea if they're still in the idea phase). series In , the founders will have access to learn everything from how to build an MVP, how to acquire customers, develop a monetization strategy, how to create a launch plan, will get tips on raising venture capital and alternative funding, and so much more! Launch Like a Mogul For Launch Like a Mogul, we've connected with some of the most promising Black startup leaders and investors to help teach our founders. In the 6-week series, the founders will learn from: Keenan Beasley - the CEO of Supply Factory Brands, Venture Noire, and Sunday II Sunday Jasmine Jones, the founder & CEO of Cherry Blossom Intimates, a breast health facility, and intimates boutique Sydney Thomas, a Senior Associate at Precursor Ventures Jose 'Caya' Cayasso, the co-founder and CEO at Slidebean, a venture-backed, multi-million dollar company founded in 2014 Obi Omile Jr., the co-founder & CEO of theCut, a tech platform modernizing the barbershop experience And more! Click here to check out the full list and to sign up for Launch Like a Mogul if you're an aspiring or early-stage Black founder. For sponsorship or media requests regarding Mogul Millennial or Launch Like a Mogul, please send an email to brittani@themogulmillennial.com. Also published at https://www.linkedin.com/pulse/helping-black-founders-succeed-mogul-millennial-brittani-hunter-mba/?published=t

How We're Helping Black Founders Succeed with Launch Like a Mogul

Too often people expect their first launch to be a huge success. If their expectations aren’t met, they’re quick to throw in the towel and move on. While others won’t bother launching until their product is perfect. And a third group decides to give up even before getting their product out there. The problem is people expect too much from their first launch. They expect customers to whip out their credit card and purchase their product right away. If you’re a startup or this is your first time launching a product or service, people will be reluctant to purchase based on you not having a track record. Couple that with . It depends on price point and how much time it will take to install or use your product. buy time Buy time as I define it in this post is how long it takes before a customer is ready to buy your product. Establishing Credibility For Your Product There are a number of ways to establish credibility for your product or service. Here are a few of the most popular: having people who have tested out your product or used it for some time provide a brief quote. The quote should accurately describe their experience using the product and how it benefited them. Customer testimonial: provide in-depth analysis of a customer’s experience using your product. You want to make sure the experience is replicable. So if they experienced something out of the ordinary, others may question their experience. You also want to show why someone chose to use your product versus everything else in the market, and why they would recommend it to others. Customer case study: Here’s an example of a case study from a student who participated in my in 2016. You’ll notice how specific it is in terms of the customer’s persona: being a software engineer who is mid-career. The challenges the customer had on their own, resistance they had to our product, how they got over that resistance, how they benefited from the product, and why they’d recommend it to others. Confident Communicator Course you can leverage someone else’s credibility and audience. The key here is to make sure their audience is one that will want your product. Otherwise, it can call into question their credibility and come off as pushy. The best influencers are ones whose use of your product seems natural. Influencer: everyone wants to make the cover of front page news, or be featured on a post in TechCrunch. However, this can be a very broad strokes approach, which may build awareness for your product but not lead to any sales. So like influencers, you want to find press whose audience would want your product. Press: often one of the above isn’t enough. It can take up to 7 touch points to get someone to actually purchase your product. Here’s a recent example from someone who purchased my book: Multiple touch points: I’ve learned about the book from your guest post on Slidebean’s blog: https://slidebean.com/blog/design/presenters-most-common-mistakes I stumbled upon it while searching whether Slidebean has a presenter mode with separate screen (searched: “slidebean presenter screen”). Which it does not have. But, hey, I found your book, so overall I got a really good result from that search query ;) What ultimately sold me on book’s webpage was praise from Dan Martell. I participate in one of Dan’s entrepreneur-focused programs. If he recommends something it must have a real value, no questions asked :) -Mat B. Notice how it was first a search, followed by a blog post, and finally an influencer Dan Martell’s praise for the product. Why not offer a limited time discount to persuade people to purchase your product? That may work, but it’s important that you understand what is going on in the heads of customers. Let’s take a simple example: you’re launching your first book. You price it at $25, but during launch week you’ll offer a 20% discount. There are some customers on the fence. But do you really think that saving $5 is really going to convince them to buy the book now versus later? If they happen to be bargain shoppers who are always looking for a discount, then the $5 savings could tip them to purchase it immediately instead of waiting another week. Who doesn’t like a bargain? Those who value their time. If someone just doesn’t have the time to read your book right now, a $5 savings isn’t necessarily going to persuade them to purchase your book, because they know they will buy it and it will just sit on their shelf. Finding Those Who Need Your Product Right Away What you really want to be doing is attracting customers who are going to purchase and use your product immediately. Why? Because you need referrals! Word of mouth is the most powerful. In my book, a customer who just pays me for a product and then doesn’t use it is worthless. Sure I’ve made some money, but I don’t continue to make money off of them. The only way I’ll continue to make money off of them is if they continue to buy for me or they tell others to buy from me. marketing To attract this group, you need to understand what is their acute need. Have they tried other products? What do they love or hate about them? Why would they choose your product over others? For example, when I was launching my second self-published book a lot of people told me they didn’t have time to speak in public. Present! A Techie’s Guide To Public Speaking Why? Some told me they had other priorities like staying heads down or family obligations. Others worried that their boss’ or colleagues’ would find them to be too self-promotional or unfocused in their work. And of course, there was the group that was just afraid to speak up for various reasons: not feeling like an expert, being shy or introverted, or fearing an audience’s reaction. It seemed like no one needed my product right away… Photo by Jonathan Simcoe https://unsplash.com/@jdsimcoe Building Up Demand Through Education Knowing that people were afraid of public speaking, didn’t have time, and had a series of other objections, I could have just thrown in the towel. But I didn’t. Instead, I chose an alternative, to educate people on the importance of public speaking for their careers. Prior to launching my book, I spent nearly a year educating people. I did talks, I created posts, and I spread the word no matter People became curious, curiosity led to interest, and eventually, the message was clear: you need to invest the time to speak if you’re going to advance in your career. Prior to launching, , and it became clear that people were eager to purchase my product. I ran a pre-order campaign on Publishizer Why People Don’t Educate Customers About Their Product For two reasons: You have to invest the time to listen to people’s objections. Then you have to listen to what their goals or dreams are. Since they don’t know there is a disconnect from their objections to their goals and dreams, you have to figure out how your product can bridge the gap. 1. It takes time. Ads are seen as a quick way to attract people to a product, but are they really convincing people to buy? Think about the last product you saw an ad for, then immediately went out and purchased it and used it, and you’ll have your answer ;) 2. They’re looking for a quick solution. Re-Launch Your Product If your first launch wasn’t a success that’s OK. There is no shame is re-launching a product or doing an iterative launch. I’m not proud of them, but each launch taught me something valuable about how to appeal to customers. Those lessons have kept me moving forward and launching new products like books, courses and software apps. I’ve had some epic failed product launches. Whether you’re about to launch or you’ve already launched, think about the following questions: 1. Why aren’t people interested in this product (competition, time, need)? 2. What are they looking to achieve (goals and dreams)? 3. Where is the disconnect between their current behavior and their goals and dreams? 4. How could my product resolve that disconnect? 5. How can I present the disconnect in a way they’ll understand, and it will build up demand for my product? Did you recently launch a product or you’re getting ready to launch one? Let me know in the comments below what worked for you or what you’re thinking of doing. Enjoyed this post and want more? Help others enjoy it too by hitting the ♡ below! Sign up to receive additional lessons on how to ship your idea here.

How To Bring A Product To Market After Its Initial Launch Failed

Creating Staggered Animation in Flutter: A Micro-Interactions Guide

ফ্লটারে স্তব্ধ অ্যানিমেশন তৈরি করা: একটি মাইক্রো-ইন্টার্যাকশন গাইড

Erstellen von gestaffelten Animationen in Flutter: Ein Leitfaden für Mikrointeraktionen

Creación de animaciones escalonadas en Flutter: una guía de microinteracciones

Création d'une animation échelonnée dans Flutter : un guide des micro-interactions

फ़्लटर में स्टैगर्ड एनीमेशन बनाना: एक माइक्रो-इंटरैक्शन गाइड

Flutter でスタッガードアニメーションを作成する: マイクロインタラクションガイド

Flutter에서 시차 애니메이션 만들기: 마이크로 상호작용 가이드

Criando animação escalonada no Flutter: um guia de microinterações

Создание шахматной анимации во Flutter: руководство по микровзаимодействиям

Flutter'da Aşamalı Animasyon Oluşturma: Mikro Etkileşimler Kılavuzu

Tạo hoạt ảnh so le trong Flutter: Hướng dẫn tương tác vi mô

在 Flutter 中创建交错动画：微交互指南

This paper is available on arxiv under CC 4.0 license. Authors: (1) Terence Blésin, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (2) Wil Kao, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (3) Anat Siddharth, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (4) Alaina Attanasio, OxideMEMS lab, Purdue University; (5) Hao Tian, OxideMEMS lab, Purdue University; (6) Sunil A. Bhave, OxideMEMS lab, Purdue University; (7) Tobias J. Kippenberg, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL). Table of Links Introduction Results A. Physics and design B. Device characterization C. Bidirectional microwave-optical transduction D. Pulsed transduction Discussion Acknowledgments Appendix A: Theory Appendix B: Fabrication process flow Appendix C: Optical characterization Appendix D: Characterization of acoustic resonances Appendix E: Characterization of bidirectional transduction Appendix F: Optimization of optical extraction efficiency Appendix D: Characterization of acoustic resonances The microwave reflection measurements (also known as S11) shown here were measured at room temperature using custom probes (GGB industries nickel-alloy 40A-GS-135-PC-N Picoprobe) and a vector network analyzer (Rhode&Schwarz ZNB-20). The spectrum shown in Fig. 2d is corrected using a calibration substrate (GGB industries CS-8) to remove the phase delay and attenuation introduced by coaxial cables and probes. However, this calibration procedure was only used to get the S-parameter, and was not applied during the total conversion efficiency measurements shown in Fig. 3 and Fig. 4. In a previous version of the transducers and in Ref. [11] where the Band-Aid process (Appendix B) is not employed, the piezoelectric actuator extends beyond the suspended cladding and covers the electrical feedline, causing two detrimental effects for transduction. First, the resulting stray capacitance leads to excess microwave insertion loss. Given the small size of the suspended area (∼ 1000 µm2 ) compared to that of the feedline (≳ 10 000 µm2 ), this effect is readily seen in Supplementary Fig. 3, where the spectrum without the Band-Aid process exhibits reduced background reflection. Second, the part of the actuator on top of the unreleased cladding leads to resonant excitation of HBAR modes extending over the entire Si substrate. These modes have a smaller free spectral range (FSR) of around 19 MHz and a lower microwave extraction efficiency. The theory presented in Appendix A does not correspond to the case of such a multimode system as it assumes being in the sideband-resolved regime, implying the high density of substrate modes is inappropriate for low-noise frequency conversion. The feedline Band-Aid, discussed in Appendix B and shown in Supplementary Fig. 3, enables the removal of the piezoelectric layer above the substrate, leaving only HBAR modes well confined in the oxide cladding. We fit the measured complex reflection to obtain the mechanical quality factor Qm = 284 and microwave extraction efficiency κex,m/κm = 0.11 for the transduction mode [15]. In addition, we simulate microwave reflection of the designed stack geometry using finite-element method (FEM). Fitting the simulated spectrum leads to Qm = 205 and κex,m/κm = 0.61. The discrepancy is attributed to difference in clamping losses and piezoelectric coefficients between the fabricated and simulated devices. We extract the effective mass of the transduction HBAR mode from FEM simulations as

Bidirectional Microwave-optical Transduction Based on Integration: Appendix D: Characterization

This paper is available on arxiv under CC 4.0 license. Authors: (1) Terence Blésin, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (2) Wil Kao, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (3) Anat Siddharth, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (4) Alaina Attanasio, OxideMEMS lab, Purdue University; (5) Hao Tian, OxideMEMS lab, Purdue University; (6) Sunil A. Bhave, OxideMEMS lab, Purdue University; (7) Tobias J. Kippenberg, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL). Table of Links Introduction Results A. Physics and design B. Device characterization C. Bidirectional microwave-optical transduction D. Pulsed transduction Discussion Acknowledgments Appendix A: Theory Appendix B: Fabrication process flow Appendix C: Optical characterization Appendix D: Characterization of acoustic resonances Appendix E: Characterization of bidirectional transduction Appendix F: Optimization of optical extraction efficiency C. Bidirectional microwave-optical transduction We measure the off-chip efficiency η tot of the microwave-to-optical (up-conversion) and optical-tomicrowave (down-conversion) processes as a function of off-chip CW input optical pump power Pin. Compared to the on-chip efficiency η oc, this efficiency also accounts for the loss channels of the microwave and optical ports. First, we study the transducer in the triply resonant configuration. For up-conversion, the optical sideband generated from the microwave input, detuned by ωin = ωm from the pump frequency, is measured via a self-calibrated heterodyne detection method. The off-chip optical output containing both the pump (ωL) and the sideband (ωs) is combined with a local oscillator (LO; ωLO) and detected. By placing the LO at a frequency ωLO = (ωL + ωs) /2 + δ such that the photodetector response is constant over a frequency span of δ, we determine the sideband power relative to the pump. Picking off a fraction of the optical output then enables the determination of its absolute power and, by extension, that of the sideband with a power meter. The measured optical sideband power is compared to that of the microwave input to yield η tot. In the case of down-conversion, the optical input is generated by modulating the phase of the pump with an electrooptic modulator (EOM) driven by a microwave source of frequency ωEOM = ωm. Only one of the resulting sidebands is admitted into the photonic molecule and transduced, as the other is far off-resonance. The converted microwave power is directly probed with an electrical spectrum analyzer (ESA). Summarized in Figs. 3a and c, bidirectional transduction processes corresponding to both the effective beam-splitter (ωL = ω−; anti-Stokes) and two-mode-squeezing (ωL = ω+; Stokes) interactions are investigated. We reach a maximal η tot of −48 dB at 21 dBm pump for each configuration. Knowing the port losses, we estimate an on-chip efficiency η oc = −41 dB (Appendix E). Accounting for the extraction efficiency η ext, we further obtain an internal conversion efficiency η int = 2 × 10−3 from Eq. 2. Furthermore, we deviate from triple resonance to map out the transduction bandwidth. First, for a pump still resonant with one of the optical supermodes, the input microwave frequency ωin and input optical detuning (controlled by EOM drive frequency ωEOM) are varied for up- and down-conversion, respectively. Shown in Fig. 3b and d, transduction leveraging the main transduction mode exhibits a full width at half maximum (FWHM) of 25 MHz. The multimode nature of the transducer is further manifest in Fig. 3e, where the pump is in the beam-splitter configuration but slightly detuned from aˆ−. An additional transduction peak in η tot with a FWHM of 10 MHz is observed around ωin = 2π × 3.165 GHz, which corresponds to another HBAR one FSR away from the main transduction mode ωm = 2π × 3.480 GHz. The engineering degrees of freedom such as cladding and actuator thickness (HBAR FSR), and optical dispersion (supermode splitting as a function of optical wavelength) in the present system offer possibilities for frequencymultiplexed transduction.

Bidirectional Microwave-optical Transduction Based on Integration: C. Bidirectional

This paper is available on arxiv under CC 4.0 license. Authors: (1) Terence Blésin, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (2) Wil Kao, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (3) Anat Siddharth, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL); (4) Alaina Attanasio, OxideMEMS lab, Purdue University; (5) Hao Tian, OxideMEMS lab, Purdue University; (6) Sunil A. Bhave, OxideMEMS lab, Purdue University; (7) Tobias J. Kippenberg, Institute of Physics, Swiss Federal Institute of Technology Lausanne (EPFL) & Center of Quantum Science and Engineering (EPFL). Table of Links Introduction Results A. Physics and design B. Device characterization C. Bidirectional microwave-optical transduction D. Pulsed transduction Discussion Acknowledgments Appendix A: Theory Appendix B: Fabrication process flow Appendix C: Optical characterization Appendix D: Characterization of acoustic resonances Appendix E: Characterization of bidirectional transduction Appendix F: Optimization of optical extraction efficiency B. Device characterization We perform metrological characterization to affirm the integrity of the stack. Shown in Fig. 2a, the region of suspended cladding is first identified through optical micrography; the released substrate below leads to a contrast in the image. To further access the layer structure, we use focused ion beam milling to create an opening on the device surface at the suspension site. This opening provides sufficient clearance to directly image the cladding acoustic resonator, pictured in Fig. 2b, confirming the removal of silicon. We then characterize the transducer as an optoelectronic network with one microwave port, two optical ports, and one auxiliary electrical DC port. Each optical port comprises an edge-coupled lensed fiber and on-chip bus waveguide terminated by 1D nanotapers designed for TE-polarized light, yielding a fiber-to-fiber coupling efficiency of −8 dB. For adjusting the frequency splitting between optical supermodes, a bias voltage is applied to the DC port to drive either the piezoelectric actuator [44] or an integrated thermo-optic heater placed in the vicinity of the micro-rings (Appendix C 2). The former affords no additional heat load and is hence cryogenic-compatible, whereas the latter provides an easy alternative for fast room-temperature characterization. The avoided mode crossing characteristic of a photonic molecule is exemplified in Fig. 2c. We access the HBARs through the microwave port. Calibrated high-frequency probes are used to contact the top and bottom electrode pads of the piezoelectric actuator, effectively linking the acoustic resonator with a transmission line. The reflection spectrum (Fig. 2d) reveals predominantly one single series of HBARs with a free spectral range (FSR) of 320 MHz, which corresponds to the acoustic length of the cladding. The transduction acoustic mode at 3.48 GHz exhibits a typical total linewidth of 13 MHz with a microwave extraction efficiency of κex,m/κm ≈ 11%. The microwave response is distinct from that of an identical stack composition with unreleased substrate. There, a periodic envelope corresponding to the cladding modes that are more strongly coupled to the microwave is superimposed over the full-stack HBAR response (Appendix D and Ref. [40]). The absence of these full-stack modes provide another piece of evidence of successful cladding suspension. Finally, we study the device acousto-optic response in the triply resonant configuration, where the transducer effectively operates as a resonant single-sideband modulator. The optical output containing both the pump at the symmetric-mode frequency and generated sideband at the antisymmetric-mode frequency are mixed by a photodetector. Such a beat-note spectrum shown in Fig. 2e displays response peaks aligned with the HBAR frequencies in Fig. 2d, demonstrating three-wave mixing.

Bidirectional Microwave-optical Transduction Based on Integration: B. Device Characterization

The old-school phone whistling is back! Well, almost, but this audio below isn’t that far off. https://securityinnovation.hubs.vidyard.com/watch/5b3Yk8ga9noiqKJ2rTNnSS?embedable=true On July 13th, we discovered an interesting vulnerability. A particular webpage was tracking digital radio chatter and it would display a running log of what its station was hearing. This is a handy tool if you’re into this sort of thing; after all, it’s great to know if your calls are being heard or if there are ongoing conversations that you might want to join. This tool captured whatever it was hearing and inserted it into the webpage. Send a message and a few seconds later it would show up there. If the message included HTML, well that would end up on the webpage too. Classic HTML Injection. To deliver an attack like this, you don’t really need the internet at all -you just need a long piece of wire. In our Attacks Over the Air series, in February 2023, we discussed the wireless issues presented by a basketball scoreboard with a controller that operates at 913 MHz. In this post, we’ll move way, way down in frequency, down into the shortwave radio band. This attack works at 7.078 and 14.078 MHz. So how does this work? published There are many interesting digital signals down here, and JS8 is one of them. JS8 is a popular keyboard-to-keyboard chatting tool for amateur radio. It absolutely excels at enabling direct conversations over long distances and through very poor link conditions. JS8 is itself based on FT8, an extremely popular tool for making simple contacts between stations that can hear each other. Night and day, FT8 and JS8 are continually in use. barely Several weird things happen to signals down in the shortwave band. It’s such a low frequency that at 100 miles above us, the E and F layers of Earth’s ionosphere begin to reflect the radio signal back down to the ground again. The signals bounce between the Earth and the ground several times, like skipping a stone off water. It allows us to hear commercial AM stations on the other side of the planet at night. During the day, this bounce works at higher frequencies too, enabling communication over vast distances. My backyard transmitter and the website’s station are 674 miles apart. One bounce covers that distance easily. JS8 messages are constructed in ASCII text, all uppercase characters. Most messages are sent to wide groups, such as Skywarn, JS8Chess, or, to reference everyone, simply ALLCALL. Following this, the message can be just about anything. First we build out HTML payload in the editor. We must legally provide our amateur radio callsign for identification (a step that will be ignored by a real attacker) and in the end our final message looks like this: So how do you attack a website with a shortwave signal? We are now ready to transmit. The JS8 software converts this text to binary, adds forward error correction, and modulates the result using Gaussian Frequency Shift Keying, (GFSK). The final product audibly sounds like strange whistling noises. If we open the sound using Audacity, and open a spectrogram, it looks like this: KJ7YLS: @TEST  <DIV STYLE="WIDTH: 500PX;HEIGHT:100PX;COLOR:RED;">HTML INJECTION!</DIV> Taking a closer look, you can see the GFSK modulation slowly transitioning from one frequency to another. It also holds at each frequency for a specific length of time. There are eight different frequencies in use here and thus each symbol has eight distinct states. As a result, each symbol can carry three bits at a time. This is illustrated with some labels: We can then send the whistling noise into the microphone of a radio transmitter. Like most amateur and military signals, the audio is delivered using upper-sideband modulation, which is a section of an AM broadcast type of signal. JS8 messages are standardized at several frequencies, and our target listens at 7.078 MHz and 14.078 MHz. We tune the radio to either frequency and send the audio using 20 watts of power into a 16-AWG wire about 30 feet in the air. The signal radiates out into the air, bounces off the ionosphere, and their antenna receives it. Their JS8 software reverses the process and pulls out the ASCII message. It’s at this point that we see the problem. The website (link withheld in this article out of respect to its author) logs JS8 traffic. Under the hood, it’s powered by , a useful Python tool that talks directly to the API of the JS8 software. In the repository, the script grabs the JS8 messages and places them on the website. At time of writing, it contains the following lines of code at line 859: JS8Net monitor.py # Text
j['text']=j['stuff']['params']['TEXT']
key=str(j['stuff']['time'])
j['id']=':'.join([fmcall,tocall,key,str(j['freq'])])
traffic[key]=json.dumps(j) is the ASCII JS8 message pulled from , which pulls from the API of the JS8 software. After this step, the text in is placed onto the page using Javascript. These steps are fine. The issue occurs in the change in context because there is an assumption that the JS8 traffic, already fully processed by the JS8 software, is ready to go and will not affect the context of the webpage. This is the behavior that we can then exploit. j['stuff']['params']['TEXT'] collector.py traffic[key] From start to finish, our attack looks like this: https://securityinnovation.hubs.vidyard.com/watch/dEaxwfcKAopxUVnVeMQtbf?embedable=true A few minutes later, our message is logged on the website... The end result is mostly cosmetic. This is a public page, so there’s no session cookies to steal or controls to spoof. Stored XSS is a real possibility too, but after a frustrating morning sending many, many time-consuming payloads (each JS8 message takes several minutes to send) we couldn’t get working Javascript that would execute on the page. I suspect that the JS8net software accidentally blocked many of the key characters and words and our workarounds didn’t seem to fire. Nevertheless, Burp Suite came in handy here. The ionosphere was cooperative that particular day and bouncing signals pretty well. I knew from other tools that stations in other countries were hearing JS8 traffic from the U.S.. So my next payload was This URL points to a DNS and HTTP listener run by Burp Suite’s Collaborator tool, which we often use during penetration tests for finding out-of-band information leaks. This was no different. Within a few minutes, DNS and HTTP traffic was recorded by the Collaborator Server from users visiting the page in Chile. I also viewed the page to inspect the injection and was logged as well. KJ7YLS: @ALLCALL <img src="http://727k2w1hfoamqewpm7rpiocu0l6cu2ir.oastify.com" alt="@HB DN13"> This was about as far as I wanted to ethically go with the injection. I reported the issue to the website’s owner, Jeff, with suggestions on remediation. He got back to me the next day, saying: While I waited for the patch to go live, I reflected on a few lines from the JS8Net README: I work in security myself (I do R&D for a network security company in the Bay Area), but I'll have to admit that submitting JS via radio to hack a web site was totally off my radar. It's important to note that this part of the software is still very much in the development stage, and may have critical vulnerabilities that make exposing the exposed services to the open Internet a Very Bad Idea. While it certainly will work, it's intended for protected, internal LAN use at this time. Honestly this is an excellent warning in general, especially for projects that are in progress and have not been thoroughly tested. Vulnerabilities at this point are an “unknown unknown” but the author was primarily thinking about the exposed listeners and APIs in his software. However, we exploited a slightly different type of problem. The README warns that the service should be run inside a LAN and this would indeed protect others from HTML injections and defilement of the webpage. However, a payload with Collaborator payloads would leak the public IP address of anyone viewing the page within their private LAN as well, so that’s a little harder to address. In short, this is a radio signal attacking a webpage on the Internet. An attacker doesn’t need to fool around with VPNs, Tor, or botnets to try to hide their identity on the Internet. JS8 is so effective at extracting messages -24dB below the noise floor (the whistling literally cannot be audibly heard at this level) that very little power is needed to reach distant stations. One watt of transmission power, on good ionospheric conditions, can go coast-to-coast across the U.S.. An attacker could be anywhere in the world. A burst message is all that is needed, so direction-finding would also be difficult. I think that’s about as close to untraceable as you can get with something like this. Timeline: July 13: discovered the issue, reported to Jeff, applied to MITRE for a CVE July 14: Jeff responds. July 22: Jeff applies a patch to JS8Net on Github. July 25: The patch goes live on the website. Aug 18: the blog post is published.

Attacks Over The Air — Phreaking HTML Injection

বাতাসে আক্রমণ - এইচটিএমএল ইনজেকশন

Angriffe über Funk – Phreaking-HTML-Injection

Ataques por aire: inyección de Phreaking HTML

Attaques en direct — Phreaking HTML Injection

हवा से हमले - भयभीत करने वाला HTML इंजेक्शन

無線攻撃 - 異常な HTML インジェクション

공중파 공격 - HTML 인젝션을 위협하다

Ataques Over The Air — Phreaking HTML Injection

Havadan Saldırılar — Phreaking HTML Enjeksiyonu

Tấn công qua mạng — Phreaking HTML injection

空中攻击 — Phreaking HTML 注入

Catching signals from others is how we started communicating as human beings. It all started, of course, with our vocal cords. Then we moved to smoke signals for long-distance communication. At some point, we discovered radio waves and are still using them for contact. This article will describe how you can tune in using and an SDR dongle. Fedora Linux My journey I got interested in radio communication as a hobby when I was a kid, while my local club, LZ2KRS, was still a thing. I was so excited to be able to listen and communicate with people worldwide. It opened a whole new world for me. I was living in a communist country back then and this was a way to escape just for a bit. It also taught me about ethics and technology. Year after year my hobby grew and now, in the Internet era with all the cool devices you can use, it’s getting even more exciting. So I want to show you how to do it with Fedora Linux and a hardware dongle. What is Ham Radio Amateur Radio (ham radio) is a popular hobby and service that brings people, electronics, and communication together. People use ham radio to talk across town, worldwide, or even into space, without the Internet or cell phones. What’s SWLing? To broadcast with your ham radio or SDR system, you need to obtain a license from a governmental body. But to intercept signals and listen to the open communication between two amateur radio stations, you don’t need one. The term SWLing comes from the abbreviation of Short Wave Listener, where you listen to stations communicating in the shortwave bands between 3 and 30 MHz. This can be used for long-distance communication using the , a layer of the Earth’s atmosphere. ionosphere To get started, you don’t need a license. Still, I recommend getting yourself an SWL sign to identify yourself in a listening contest. These are competitions for categories like who will discover the most connections in a month or who can listen to contacts from each country in the world. How to get an SWL Sign? There are two options: Contact your national radio club and ask them to issue one for you. I got my Czech one, OK1-36568, after a few weeks. Join the community and request a sign there. Short Wave Amateur Radio Listening You will get more information and help from either of these locations if you get stuck in some fashion! QSL Cards You can also use your sign to send QSL cards via post or electronically. This is a great way to communicate with people worldwide and make friends. Per Wikipedia, A QSL card is a written confirmation of either a two-way radio communication between two amateur radio or citizens band stations; a one-way reception of a signal from an AM radio, FM radio, television, or shortwave broadcasting station; or the reception of a two-way radio communication by a third party listener (in our case). A typical QSL card is the same size and made from the same material as a regular postcard; most are sent through snail mail. Replace the radio receiver with your Fedora Linux The focal point of the ham radio hobby is the radio transmitter/receiver. Most of the time, enthusiasts build their radio from scratch, but this differs from what I will write about here. SDR A software-defined radio (SDR) system is a radio communication system that uses software for the modulation and demodulation of radio signals. In other words, a piece of hardware and software replaces a radio transmitter/receiver. This helps you discover more in a way you are familiar with – a User Interface with built-in functions instead of the limited interface of a radio receiver. My explanation oversimplifies things, so if you want to go deep and read more about SDR, here is an excellent start. SDR Set Up under Fedora Linux Choosing the proper hardware If you search the Internet for an SDR dongle, you’ll find many ideas depending on your budget. In this tutorial, I’ll work with the one I have, which works well under Fedora 37. A note: The dongle covers frequencies from 25MHz to 1750MHz, which doesn’t cover the Short Wave bands. You would need an additional device to listen to them. This is included in the package I linked above. Some other hardware providers offer all-in-one products. Check if the dongle is visible Before installing anything, detect whether Fedora Linux recognizes your USB dongle. I hope you didn’t buy a fake one :-). Use the following command to list the USB devices on your system. lsusb One of the output lines (in the case of Nooelec) should be Realtek Semiconductor Corp. RTL2838 DVB-T Now proceed by installing the software you need Fedora offers a set of tools and drivers packaged as a group. Even though you would not use all the components in this package from the beginning, I recommend installing it. You’ll have more software to play with. sudo dnf group install 'Electronic Lab' I advise you to explore what’s in the group by running this command: sudo dnf group info 'Electronic Lab' Now check if you have everything set up correctly by running: rtl_test You should see something like this: Do not forget to kill this process because the device will be busy and cannot be used in the next step. A simple Ctrl + c works. Gqrx You have the dongle already in your device’s USB port and all the software you need to get started. Now it’s time to intercept your first signal. Start the program called . Don’t be alarmed by the strange interface. You’ll get used to it. Gqrx Configure the I/O Device Screen From the “Device” Dropdown, select the ‘ .’ RealtekRTL2838.. Leave the rest untouched for the moment. If you don’t see your device there, click the “ ” button at the bottom of the screen. Device Scan When your device is selected, click “OK,” and the dialogue will close. Configure the frequency screen Before you start intercepting signals, ensure there is something out there that proves everything works correctly. Since the dongle covers the FM radio band as well, do this: Locate your favorite radio station’s frequency. Mine is 105MHZ Set it in the Frequency field Select WFM (stereo) in the “Mode” dropdown. If you don’t do this, you will hear a sound. not Play And now, you need to start the reception by clicking the “play button” in your main menu. You will see the frequency visualized like this: If you hear a sound, everything will move to the next step. Check if everything is set up correctly if you don't hear anything. You may ask a question in the comments for this article; I can direct you to the proper forum to solve this. Feel free to play with some more FM broadcasts. You have the antenna for it in your pack. Let’s go Short Wave In the case of the Nooelec, you need to add one more device to the USB dongle and turn it on. Instructions on how to do that are included in the package you receive. In short, you plug the “Up Converter” into your USB dongle and make sure the switch is in the “convert” position. Some videos are available on how to do it if you get stuck. You will need an antenna and a good location Now things get trickier. If you live in an area where you don’t see an open space out your window or other buildings surrounding your building, you might have trouble catching a Short Wave radio amateur signal. Let’s try this to see if it works Try to be in the open. I usually listen from my terrace, which could be better but works under particular conditions. You would need a long wire to act as your antenna apart from the hardware. Try the antenna with the hardware initially – the telescoping one from , but it will catch only powerful signals. Nooelec Let’s go back to Gqrx Now with the converter, you need to make some changes to your device screen: Please note the – for the . This is required for the Up Converter to work. 125Mhz LNB LO field Tune your frequency to 14.100 Mhz and make sure your Mode is (standing for ) because this is this band’s main demodulation option. USB Upper sideband Then go to your screen, use the zoom slider, and set it to see about . In our case, you should have between 14.05 to 14.15 Mhz on your screen. FFT Settings 100 kHz Also, click the “ ” to see the information about the SW bands you are exploring. Enable Band Plan Then hit the play button and explore the space between 14.0 and 14.3 Mhz to get any amateur radio transmission. When intercepting a transmission, adjust your settings to improve your listening experience. It’s a journey that you have already started. Most probably, you will hear something like this: “CQ CQ CQ this is ..(followed by the radio license number spelled with the ). ham radio phonetic alphabet Listen very carefully, and by the call sign, you can determine the location of the radio amateurs’ country. You can visit the to learn more about them and even send them a QSL card confirming their connection. QRZCQ website . Keep the momentum going Now you have some tools and ideas for starting Short Wave Listening. This is the first step of an incredible and exciting journey you can have together with your Fedora Linux OS. You will discover the pleasure of , reading more about , how to be a part of a listening competition, and what mean. building your antenna for the specific band how the ionosphere helps those Q-codes This article was initially published in Fedora Magazine

Mastering Shortwave Listening (SWL) with Fedora Linux and Software Defined Radio

Training an Image Classification Model on Azure without storing data on-prem using Dagshub Direct Data Access and Azure ML SDK Introduction Machine Learning and Artificial Intelligence have become ubiquitous and one of the key skills of a Data Scientist is to be able to build models that are scalable, deal with sensitive data and also build reliable data and model pipelines — that are reproducible. While Cloud Technologies allow us to build scalable pipelines- we need to store all the files on our cloud before we can use them for training models incurring GPU cost and time to pull the data. In this article, we are going to train a model on Azure with data stored on Dagshub Repo, using the (DDA) feature by Dagshub which allows us to pull data in batches when required— reducing the time taken to pull and store the data on Azure thereby reducing GPU cost as well. Direct Data Access We will use Azure ML SDK to create training job right from our local machine and push it to Azure. We will also see how we can monitor the job right from our jupyter-notebook. We will use the data from Kaggle's, Mayo Clinic — STRIP AI challenge, where the focus is on differentiating between the two major acute ischemic strokes (AIS) etiology subtypes: cardiac and large artery atherosclerosis. Detecting the etiology of strokes can help physicians mitigate recurrent strokes. The code and the data used in this analysis can be found . here About the Data The data originally present for the challenge is in tiff format and is 356GB in size. For the purpose of this analysis, we downscale the images and store them in png format. The code for this downscaling of data can be found . The downscaled images can be found . To map the images to the classes they belong to, the competition data provides us with a CSV file that contains information about the data. This data can be found . here here here The “label” column in the training dataset consists of the class images belongs. The train data folder consists of 754 images out of which 547 belong to class CE (cardiac) and 207 belong to class LAA (large artery atherosclerosis). For our analysis, we have created a column called “int_labels” that assigns a value of 1 if the image belongs to class “LAA” else it is given a value of 0. Approach To use Azure to train our models, we will use Azure ML SDK and run our code on Azure. This will require us to create an ML Workspace and a GPU Compute on Azure. Generally, we upload our data as well onto Azure Storage and create Datastores to allow ML Workspace to access our data. We are not going to store the data on Azure but instead stream the data from Dagshub Repository. To set up Azure Workspace and create a compute using Azure ML SDK refer . here Dagshub is referred to as “Github for ML” as it allows us not only to maintain our code but also enables version control of our data. Data Version Control (DVC) allows us to track changes in large data files similar to how Git helps us to track changes in our code. To enable DVC we can use any of the cloud storage like AWS, GCP and Azure which can be expensive. Similar to how we do a git pull to get the latest code from Git, we do a dvc pull to get the data that are versioned using DVC. Similar to Git commands, we have dvc push, pull and commit commands as well. For this project, I created a Repo on Dagshub and uploaded all my data files and code. All the files and folders that use DVC are marked as DVC in the repo. Once we have the data pushed to Dagshub, we will create scripts that will use the streaming functionality from DDA to fetch data from the repository without having to do a which can take longer. dvc pull Data Streaming using Daghub’s Direct Data Access Traditionally, if we want to build a model or read data that is present in any Data Version Control system — we take a pull and extract all the data files that are present. But, many times we do not need all the files that are present and pulling large files can be time-consuming and incur CPU/GPU time as well. To avoid this, Dagshub introduced the Streaming API which is part of Dagshub's Direct Data Access (DDA) feature. pip install dagshub # Install the Dagshub package for DDA. We then clone the DagshubRepo using Gitpython library. ## Cloning the Repo
git_url="https://"+DAGSHUB_USER_NAME+":"+DAGSHUB_TOKEN+"@dagshub.com/"
        +DAGSHUB_USER_NAME+"/"+DAGSHUB_REPO_NAME+".git"

import git

git.Git().clone(git_url) Once you clone the repo , you can see the Repo in your local system - but you will not find the folders that are versioned using DVC (data and models folder) in the repo. There are two ways of using the streaming client. which can automatically detect calls to Python’s built-in file operations like open(), listdir() and is compatible with most python ML Libraries. using install_hooks(), one can simply access the file as if it was stored on the local machine itself. Python-only “Lite” hooks from dagshub.streaming import install_hooks
import pandas as pd

install_hooks(DAGSHUB_REPO_NAME,password=DAGSHUB_TOKEN)


train_data=pd.read_csv(DAGSHUB_REPO_NAME+"/data/raw/train.csv")
train_data.shape Using this method, when we use os.listdir() we can see the data and the models folder which are DVC versioned as well. On calling install_hooks() we can see that the “data” and “models” folders are visible using os.listdir() This method may not work with Tensorflow and OpenCV — since they have input/output frameworks written in low level language like C/C++ which needs to be handled in a different way as compared to library like Pytorch . In such cases, you can use File Streaming by loading the Dagshub FileSystem. File Streaming using DDA is as simple as loading the Dagshub FileSystem. from dagshub.streaming import DagsHubFilesystem
fs = DagsHubFilesystem(project_root=DAGSHUB_REPO_NAME,username=DAGSHUB_USER_NAME,
                        password=DAGSHUB_TOKEN) We can then replace any use of open(), os.stat(), os.listdir(), and os.scandir() with fs.open(), fs.stat(), fs.listdir(), and fs.scandir() respectively. Using fs.listdir() we can see now that the data and models folders are present, even though they are not there on our machine. Just as a check, let us use both fs and os to list files in the directory. As we can see below, we cannot access the data folders when using os — which means that we are accessing files that are on Dagshub Repo using the streaming functionality To read, say train.csv that is present in the data/raw/ folder we can use the open function Once you open a file, the file is stored in the cache in your local machine and hence will be visible when you use os.listdir() function. The train.csv file that was opened is stored in the cache and becomes visible in the Repo in Local System The main advantage of using the streaming functionality is that you do not download all the files at once instead only access the files that you require. This means we do not have to wait for all the files to download before we start training our models saving GPU cost and time. Visualize the Images using DDA We create a show_image function that takes the image and the streaming client as input and reads the images — converts them into (512,512) images and also flips them horizontally. Here, we use fs.open() to cache the file onto our local system and then use im.read() function to read the image. ## Code to visualise Images - by reading the image from Dagshub
import cv2  
import matplotlib.pyplot as plt


def show_img(fs,img_id):
    IMAGE_FOLDER=DAGSHUB_REPO_NAME+"/data/raw/train/"
    img_path = os.path.join(IMAGE_FOLDER, '{}.png'.format(img_id))
    print(img_path)
    #slide = Image.open(img_path)
    fs.open(img_path) ## This will cache the image
    slide = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)
    img=cv2.resize(slide,(512,512))
    print(type(slide))
    size = slide.shape
    print(size)
    #region = slide.read_region((0, 0), 0, size)
    flipHorizontal = cv2.flip(img, 1)
    

    
    plt.figure(figsize=(8, 8))
    plt.imshow(flipHorizontal)
    plt.show()
show_img(fs,"008e5c_0") Creating Scripts for Training Image Classification Models on Azure The first step to train a model on Azure is to create your own workspace and compute. This can be done using Azure ML SDK (refer ) or you can go to and create a workspace. To ensure reproducibility — I have created my workspace and compute clusters using Azure ML SDK. here portal.azure.com The next step is to create python scripts that will be used for training. All the scripts that are to be used are to be put in the same folder as that folder will be uploaded to Azure. ## Creating Script Folder to store all the scripts that are needed to run job on Azure
import pathlib
script_folder="../src/azure_ml_scripts"
pathlib.Path(script_folder).mkdir(parents=True,exist_ok=True)
print(script_folder) In the scripts folder — since I need to be able to access Dagshub Repo I create dagshub_config.py which contains the Dagshub username, token, and repo name. The token is to be kept secret and should not be shared. file that contains the code to create a streaming client, clone the git repo, list the images in the training folder using data streaming, read the train.csv data frame, and split the data into train and test sets and function to download EfficientNet model using data streaming. util.py Now, the last and the final script ( ) — where all the magic happens. To train an image classification model, we use the ImageDataGenerator class but this requires the files to be present in the system. Since, we are going to use the streaming client to read the images as well as the train.csv file to get the image filename and the class, and create a that will read the images in batches from the Dagshub Repo, using the streaming functionality. train.py Custom DataGenerator We will use the EfficientNet B5 model for our Blood Clot Prediction. EfficientNet allows provides a systematic way to scale CNN models by balancing network depth, width, and resolution. We have stored this model in “data/raw/Pretrained_Efficient_Models” in the repo which contains multiple EfficientNet Models. But, since we only want the B5 Model, we read the weights of only that model. This is possible because of the streaming API — else we will have to download the entire folder and then read that particular file. def download_EfficientNet(fs,model_filename):
    '''
      This function will download the model file into local system and return the path.
    '''
    fs.open(os.path.join(EFFICIENT_NET_MODEL_PATH,model_filename)) 

    return os.path.join(EFFICIENT_NET_MODEL_PATH,model_filename)


efficientWeight=download_EfficientNet(fs,'efficientnet-b5_tf24_imagenet_1000_notop.h5') Resizing the image to (512,512) and creating the validation and training data generator using our CustomDataGenerator is also defined in our . We monitor the model metrics using MLFlow in Azure to track the metrics using training script mlflow.autolog(). Creating Environment and Running Script on Azure Now, that we have our scripts ready, we need to have an environment on Azure where we can run our script — this means installing our libraries and other dependencies. While we can create our own environment with the conda packages and pip packages, we can also use curated environments. In this case, I created my own Environment using an existing docker image that has CUDA installed. ## Creating Environment using CUDA docker image and installing dependencies
from azureml.core.environment import Environment
from azureml.core.conda_dependencies import CondaDependencies


env = Environment('blood-clot-env')

conda_dep = CondaDependencies()
conda_dep.add_conda_package("scikit-learn==0.22.2.post1")

for package in ['azureml-dataset-runtime[pandas,fuse]', 
               'azureml-defaults',
              'dagshub',
              'GitPython','opencv-python-headless','Pillow','mlflow']:
    conda_dep.add_pip_package(package)

other_pip_packages=['azureml-core==1.45.0' ,'azureml-defaults==1.45.0' ,
                                          'azureml-mlflow==1.45.0' ,
                                          'azureml-telemetry==1.45.0' ,
                                          'tensorboard~=2.7.0' ,
                                          'tensorflow-gpu~=2.7.0' ]
for package in other_pip_packages:
    conda_dep.add_pip_package(package)

                                
env.python.conda_dependencies=conda_dep

## Loading the Dockr Image with CUDA
env.docker.base_image = (
    "mcr.microsoft.com/azureml/openmpi4.1.0-cuda11.2-cudnn8-ubuntu20.04:20220902.v1"
)
# Register environment to re-use later
env.register(workspace = ws)

# Save the environment
env.save_to_directory("Environment/",overwrite=True) The next step is to create an experiment. Think of an experiment as a folder within which your jobs are organized. We then create ScriptRunConfig Object. You can think of ScriptRunConfig as a configuration for saving all the details required to run the job on Azure. ## Creating an Experiment and ScriptRunConfig to save details of the job to run on Azure
from azureml.core import ScriptRunConfig
from azureml.core import Experiment

experiment_name = 'blood-clot-classification-mlflow'


exp = Experiment(workspace=ws, name=experiment_name)

src = ScriptRunConfig(source_directory=script_folder,
                      script='train.py', 
                      arguments=None,
                      compute_target=gpu_cluster,
                      environment=env) To run the job on Azure, we submit the experiment with the ScriptRunConfig object and monitor the progress right from our notebook using RunDetails Widget. We can track the metrics from our notebook using run.get_metrics() We can see that the model achieved a validation accuracy of 0.75 and a training accuracy of 0.76. We can also download the saved model using run.download_files(). ## Download the model folder saved after training into our local system 

run.download_files(output_directory="outputs/efficientNet_Model") Oh!! But what if I wanted to write the metrics to a file and push it to my Dagshub Repo — do I need to pull all the data and then push it?? SURPRISE!!! Dagshub introduced the Upload functionality as part of the DDA feature that will allow us to upload our files to our Repo without having to do a pull. ## Uploading files to Dagshub Repo using upload() functionality of DDA
from dagshub.upload import Repo

repo = Repo(DAGSHUB_USERNAME,DAGSHUB_REPO_NAME,branch="main",
             username=DAGSHUB_USERNAME,token=DAGSHUB_TOKEN)

repo.upload(file="../metrics.txt",path="efficientNet_Metrics.txt",
            commit_message="Updating Metrics File",versioning="dvc") VOILA!!! The file is uploaded to Dagshub Repo with DVC versioning. Wrapping it Up… In this article, we saw how we can deploy our code on Azure and also train a model without having to spend a lot of time transferring the data to Azure. While I have taken the example of Image Classification to show how we can merge cloud technologies like Azure with Dagshub’s DDA feature, this can be extended to other data types as well — like Audio or Video Data Types as well. Generally, Video Files are large in size and often take up a lot of storage space, and pulling them every single time can be a time-consuming activity. Interestingly, using the streaming functionality is not limited to deep-learning models alone but can also be used when you have data from multiple tables stored as files that are frequently updated and your analysis just requires a few files — this way using data streaming, one can avoid downloading all the files and also simply re-run the script to get the updated results. Using Azure, we can create our own environment with their dependencies without a lot of familiarity with docker containers. Also, for those who are accustomed to python — Azure ML SDK gives us a way to run our models on Azure without even having to leave the familiar jupyter-notebook environment.

Stories, News & Mentions about SLIDEBEAN

Slidebean’s Stories on HackerNoon

Slidebean's News & Mentions