Laurence Moroney Is this a book for all coders, regardless of their years of experience or programming languages they use?

Hi, Laurence Moroney thanks a lot for this wonderful opportunity! My question is: after studying properly the NLP section would it be possible to create a meaning-extract NLP algorithm that accepts a natural language phrase like “Show me sales by the company N” and ‘translates’ it to a query for the database?

Hey, Laurence Moroney! Thanks for doing this.
Question: how much math do you go into in your book? How much math do you expect your audience to know before reading it?

How much theory do you cover in your book?

What resources do you recommend after reading your book to go deeper into AI and ML?

Are there any supplemental books you recommend that go well with your writings?

Laurence Moroney Looks like a fun read! In particular, I’d be curious to hear more about how/why you chose the deployment options you did later in the book?

Krzysztof Ograbek – The first half of the book is about learning how to train models, and it will require you to know a bit of python. Not expert-level by any means, but it would be helpful. The second half of the book deals with deploying models to Android (Kotlin), iOS (Swift) and the browser (JavaScript), so will need you to understand a little of them.

Darya Petrashka – The NLP section is primarily focussed on the underlying basics of NLP, leading to two outcomes – classifying text for sentiment, and generating new text, so not NLP->SQL type querying, sorry. I think there are a number of products that can do that, but I teach generally the underlying principles in ML

Matthew Emerick – Almost no math! The point of the book is so that the usual math/calculus barrier that prevents people from learning ML is removed.
How much theory? – Only enough to help you get to the point of being able to code. For example, I’ll teach (lightly) what a convolutional filter is, but from a code perspective (i.e. multiply these values by these) and not a math perspective (no greek letters!)
Resources to go deeper: Aurelien Geron’s book and/or Andrew Ng’s Coursera courses
Supplemental books: Anything that will help you with the language parts (basic Python, basic Android/Kotlin, basic iOS/Swift etc)

David Cox – Deployment is what generally differentiates TF from other ML frameworks. It’s ecosystem allows you to deploy optimized ML to servers (via TF-Serving), iOS/Android (via TF-Lite), Browser/Node (via TF.js) and microcontrollers (via TF-Lite Micro). The goal was to give the developer a broad understanding and a grounding in what they need to know to get their models deployed to these runtimes.

David Cox – I ended up not being able to put TF-Lite Micro into the book for microcontrollers, as that would have required many chapters to cover properly, and Pete Warden / Daniel Situnayake’s book covers that brilliantly

Got it! This was very helpful. The know-how to deploy ML models is often a critical skill I see many data science students lacking when they join our team. Excited to see you talking about that in this book and to check out your implementation!

Welcome Laurence Moroney, 👋

Laurence Moroney, thank you for presenting your book. I wonder what is the specificity of this book compared to the number of books currently on the market, and is it intended for people with previous data science practice or new learner? Thank you for clarification.

Lamjed Debbich – Hi! The goal of this book was primarily to bring ML into the hands of the traditional software developer. As such half of it is an introduction to building common types of models for general, vision, nlp and sequence, and the other half is in getting those models into people’s hands with different deployment technologies. So, someone who previously has data science might find it useful.

I think it’s really useful for new learners – as it assumes no prior knowledge, particularly of data science.

Hey Laurence Moroney
Thanks for your time with us here
First question is about the title itself
“coders”
What do you mean by coders exactly?
How is it differ than developers?

Hello Laurence Moroney,
I had a question about fit_generator method in Tensorflow. What does steps_per_epoch argument do?
I have tried to look about it online but couldn’t understand them.
And secondly, are we only supposed to ask questions about the book ?

Amr Alaa – Haha , good question! I don’t think there’s a deeper meaning in it, other than I know many people who are dabbling in code, but don’t have a job title calling them a ‘developer’ yet, so I wanted to focus on the book being for people who code, as opposed to somebody with a particular job title 🙂

Rohan - I’m not sure if that argument is fully used any more, I should check. But the idea behind it was to specify the number of steps to take in an epoch based on the data size and the batch size. So, for, example, if you have 60,000 training records, and a batch size of 1,000, then you’d have 60 steps in each epoch to get through the whole data. AFAIK it’s automatically counted now, (and fit_generator has been deprecated in favor of just fit)

Rohan – Feel free to AMA. I may not be able to answer everything, but I’ll try 🙂

Hi Laurence Moroney. Thank you for this great opportunity. I got certified as a Tensorflow Developer few months ago and I wonder where would you put your book (knowledge dependency) in comparison to your “Tensorflow Developer Professional Certificate” specialization on Coursera? Which one would you recommend going through first?

Seed Badran – They are complementary, and both the book and teh course were developed from the same material.

I am always looking for resources where I can practice building an End-to-end Machine Learning project(s), does the book cover such scope?

Laurence Moroney is an honor to have this conversation directly with you! My million-dollar question is: is this book a good way for those looking to get their first job as data science? What are the fundamental points to be learned in the book to help me search for my first job in data science? I am currently looking for a migration from the software testing area to data science

Laurence Moroney since you mentioned the book focuses on Android, IOS and web deployments for ML algorithms, does this mean only Client Side ML Deployment is included in your book vs Server Side ML?

Seed Badran – I guess it depends on the project, but in general, yes.

Caíque Coelho - It really depends on the requirements for the job. If the job requires hands-on coding of models, as well as the usual data science stuff, then I think this book would be useful, but in and of itself, it’s probably not enough. A survey of employers looking for ML Programmers had many wanting people who knew Computer Vision and/or NLP, and this book digs into those.

Bayram Kapti They’re not really the “focus” of the book. The first half of the book is on building models. The second half is on deployments, including Android, iOS and Browser. It also has a chapter on how to use TF-Serving for server side models accessible via REST or GPRC

Got it! Thank you!
A follow up question on deployment, what would be your recommendation for a startup to try ML algorithms for the first time for some optimization purposes?
Client Side or Server Side? I think I’m more interested in the ease of deployment and maintainance for now.

Bayram Kapti – I think client side for sure, just because of the install base, and because it’s easier to wrap an interface around a model in Android/iOS/JavaScript than dealing with the REST/GRPC interface.

Hey hey Laurence Moroney!!! You know I already have a copy of the book 🙂 However, I’m interested in what was your favorite dataset from the book, and why?

Gant – Hey hey! We’ll have to have you do an AMA too! 🙂 I think my favorite dataset, being completely biased, is my horse-or-human one. It’s a really tough one to build a good model from, but when it works it’s really cool. The images are all CGI, I synthesised them myself, and they greatly prove the concept of feature detection in CNNs because the computer can use features learned from a CGI image (like a horse’s ear or a human hand) to classify real, non-synthesised images!

Laurence Moroney For a developer who wants to get into machine learning and AI, what change in mindset is required to do well? Is there a different fundamental approach to solving problems? Thanks!

Laurence Moroney Your book looks very interesting! I’m particularly interested in learning how to convert Python models to JavaScript. I want to get back to JavaScript. I tried learning it a few years ago before I knew any other languages, and it was just too over my head. I think I could handle it much better now.
The other thing I’m keen to understand is federated learning. How would you recommend even starting with it? Could I use a couple of local devices to create a tiny “federation” and experiment at my kitchen table? Or am I totally misunderstanding what I have read on Wikipedia and elsewhere? 😅

Kenny – I think so – It’s as much about good data selection to train a model, as good code selection for definining one. It’s definitely a new mindset, but IMHO, a good one!

Eric Sims Thanks! I do have a couple of chapters on TensorFlow.js in my book, but all of Gant’s book is about JS! 🙂 Right now TF-Federated is a bit difficult to get started with as it is very experimental and doesn’t yet have a mobile runtime (you simulate the devices using Python)

Hello again, Laurence Moroney
While working with Natural Language Processing problems with Tensorflow, I have been getting stuck at choosing the correct model architecture for the project.
I have used,
Bidirectional LSTM layer ,
Or A GRU layer,
Or a Convolution network,
But none of them have reached a good accuracy.
Am I missing out something?
In your perspective what model architecture would you go with ?

Hi Laurence Moroney, thanks for doing this!
I’m definitely in favour of making this subject more explicable to a wider audience - however, I wonder if you could comment on potential risks. I work in data, and have often encountered legacy problems that originated from someone having insufficient understanding of creating/maintaining databases. Do you think there are similar dangers with AI/ML, and if so do you have any advice on safeguards/preventative measures, other than (obviously) a strong organisational culture?

Laurence Moroney I really liked your decision to make the book easy to be read by coders from different backgrounds, including me I think 🤔
Here’s my second question
In your book I see from the content table, i see that there are 4 chapters for TensorFlow.js and 3 for TensorFlow lite
Which one do you think will dominate the next phase of developing satisfying needs of rapid development and growth?

Laurence Moroney I wonder if the book covers contents from your course on Coursera “ TensorFlow Advanced Techniques”. 🤔
And does it also covers topics such as TensorFlow serving?

Laurence Moroney Does the book explore any ideas on building large scale NLP models like GPT3 using Tensorflow? What’s your take on it? Is it possible for small scale companies/groups to build and serve models with billions of parameters by scraping the internet for data in future. Thanks.

Laurence Moroney I wonder if book covers the deployment of DL model over a web based tool developed using tensorflow-python and tensorflow.js both. So that someone can better understand stand how to use them in production environment and when to use which one.
Thanks
Chirag

Hi Laurence Moroney - what are your favourite types of projects to work on with AI? E.g. such as working with images/ video etc . Which do you have the most fun with or feel most passionate about? 😊

A McCauley – I can’t say I have a particular favorite data type with AI projects, to me I’m more excited about particular difficult projects to solve, like medical diagnosis (be it based on image or NLP), but recently I’ve been exploring more about how models come up with their decisions, with a view to turning the black box into a glass box, and stuff around images here is fascinating. For example, when doing a ‘Class Activation Map’ for the famous Cats v Dogs dataset, I found the main thing my model was making a decision on was the eyes. I had hundreds (maybe thousands) of feature extractor filters, but in the end, the eyes have it (sic!), and being able to show and prove that was awesome.

Chirag Aggarwal - Yes. There’s a chapter on TF-Serving with a basic scenario of putting a model on a server and using TF-Serving to wrap it in a REST/gRPC interface. There are several chapters on JavaScript, making models, using models, and transfer learning with models.

Ramit Surana – No, I don’t focus on large scale NLP models, more on the fundamentals of how NLP works, using smaller datasets for classification. I do a text generation algorithm though, trained on Traditional Irish songs, so you can write your own! 🙂

Rohan – It predates ‘Advanced Techniqes’, so it doesn’t cover that. There is a chapter on TF-Serving though

Amr Alaa The 3 on TFLite are – 1 on the tech, 1 on an Android scenario, 1 on an iOS scenario. For TFJS, I needed some extra time to discuss it, because we don’t just do inference in JS, we also do training, and there are some special considerations you need to take into account when training in the browser. I think both can be huge for the future, but only time will tell.

Ricky McMaster – In some ways AI/ML doesn’t really change the ethics of responsibility with data and/or with how the work we create is used. I think the fact that AI can lead to more powerful solutions has brought the conversation to the forefront, and that’s a good thing, but issues with ethics, bias, responsibility etc have always been with us. Now’s the time to double down on them and take them more seriously.

Rohan – For NLP layers, it really depends. I did find, however, that for NLP solutions, good vocabulary management, good, clean, datasets etc had much more of an impact on my final results than a neural architecture search for the best layer types and hyperparameters. That’s one of the things I discuss in the book. 🙂

Laurence Moroney I think that building a model and working with data is an art. But data is present in different format and working with them can be very daunting sometimes. Transforming the given data into an insightful information is not everyones cup of tea.
What would you suggest if a person has difficulties working on a dataset? 😅
(The person cannot leave this field. 😂)

Hi again Laurence Moroney, on the subject of text generation… the stuff I’ve read from GPT-3 is fun, and impressive in its own way, but do you foresee a point where it could be used in a genuinely creative context? At least one use might be in literary works with elements of pastiche, but are there other contexts you can think of, or are aware of?

Rohan – I think as tooling evolves, where one can see how tuning input data affects output performance, that this part of the job will get a bit easier. Often there’s a lot of burden now where you’ll have a thesis that a particular format, slice, function, on data can impact a model positively, but then you still have to go through the process of building the model, adjusting the architecture to how the data is formatted, training, testing etc, before you can get any results to prove/disprove your thesis. As that process gets faster and easier, and then becomes automated and scriptable (think AutoML for data management too), then I think that part of the process will get much easier to do, and you can focus more on solutions and less on massaging data.

Ricky McMaster – I’ll say ‘never say never’, but when you do text generation, or indeed any type of generation, it rapidly descends into gibberish, when you realize that it’s a prediction on a prediction on a prediction on a prediction etc. That’s a hard problem to solve for it to be really useful. Large scale models like the one you mention attempt to solve that by being trained on more data, but I think there’s a ceiling to what’s possible with that approach. That being said, I’m sure somebody will someday have a breakthrough that makes text (or any other type) generation more realistic.

Laurence Moroney often while training a model, we encounter a case of stalling. The network stops learning at all. We do use early stopping technique to stop the training but, what do you think the person should keep in mind while building the model?
What are your observations while tweaking those hyper- parameters 🤓

Rohan – Stalls can be a result of too small a learning rate, so maybe it’s good to adjust that. Or, it could be because of sparse or too little data. Have you tried Keras Tuner to do a hyperparamter search?

Laurence Moroney you used Python to train models but other languages to deploy them. Why don’t you use Python to deploy? Is it because of performance? Are there more reasons?

Laurence Moroney I’m just curious about your predictions, what’s the next 3-5 years will look like. Is there any language that may soon take over training models from Python? Is Python going anywhere or its position is strong enough?

Krzysztof Ograbek – Because of where they are deployed. For example. the primary languages to build Android apps are Java and Kotlin. Or for iOS it’s Objective-C or Swift. Thus models, created with TF in Python need to be executed elsewhere, and we need to be able to interface with them.

Krzysztof Ograbek – I think Python will be the primary one, but keep an eye on languages like Go or Kotlin.

Laurence Moroney I have a few more questions that are out of the scope of your book. It would be awesome if you share your thoughts 😉
How do you explain to people that are completely non technical, what AI is?
Can Software Developer’s job be automated by AI? Are Developers gonna be replaced with AI? What are the skills that Developers should improve in order to future proof their careers?

Laurence Moroney hadn’t heard about Keras Tuner. Will try it out! The problem which I often face is of Vanishing Gradient. I tried SGD with momentum to tackle this. But that too didn’t have much of an effect.
The second thing which I wanted to know was, how do you decide the dimensions of a picture while feeding it into the network, given that the dataset has pictures of different dimensions which are very high!

Krzysztof Ograbek – I generally like to explain that AI is the result of using programming techniques that have computers act on data the way the people would. For example, Computer Vision, a person would look at a picture of a cat and say it’s a cat. When a computer can do the same, you have the beginnings of AI. Are developers going to be replaced by AI – absolutely not, in the same way as developers weren’t replaced by compilers, IDEs, intellisense, emulators etc. 🙂 Skills that they should improve on: ML, Data Science, Python etc.

Rohan – Definitely give it a try. KT might be able to help find a model that doesn’t have VG. For deciding dimensions of a picture, there’s a lot of trial and error of training time against accuracy…or, I’d do something like a Class Activation Model to determine what the features are that determine classificaiton of the image. For example, with the classic Cats v Dogs model, a good classifier, when passed through a CAM ended up showing me that the eyes were the single most important feature to distinguish…and if that was the case, maybe the eyes could have sufficient fidelity in tiny images so as not to make a difference, so why train on larger ones? 🙂

By the way, small factoid for everybody reading – one of the best things you can do for any book, to support it, is to give a review on amazon. A star rating is good, but a review is best. I’ve had books where sales were boosted even after a poor review! There’s clearly something in their ranking/recommendation engine that values written reviews. So, while I’m not asking for you to review mine (although I would love it), I would ask you to remember this to support authors. Margins on books are razor thin, and many writers simply give up because it’s not financially worth it – I’ve observed an almost inverse Moore’s law, where royalties halve every 18 months. I didn’t write this book for profit (all proceeds go to charity per Google employment agreement), but there are authors out there who need the income, so please remember this for any books you buy and/or read! Thanks 🙂

Laurence Moroney When doing an ML Project, how do you know, which algorithm to use? What are the factors? Is Deep Learning always a better choice than the traditional ML algorithms?

Krzysztof Ograbek – Not always…it really depends on the problem. Things like CNNs for images (mostly), LSTM/RNN/GRU for text etc. Generally, I would need to see what the problem I’m trying to solve is, and explore solutions for similar ones to use as a starting point.

Laurence Moroney it’s amazing to have the opportunity to talk to you. I have taken all of your courses with deeplearning.ai and I love the way you taught and broke down the API into small chunks.
My question is: What according to you is the ideal “Full-Stack” pipeline for building deep learning models ? How do we incorporate methodologies like CI/CD, Testing and Deployment with the Tensorflow Ecosystem. Does Post-Training Optimization methods such as Pruning and Quantisation come under “Full-Stack” and what according to you is the future of the tfmot package ?

Laurence Moroney I have asked several technical questions. But I was curious about journey in this field of AI. What excites you the most?
You are an inspiration to students like us. Any suggestions you can provide to people starting their career in the vast universe of AI.
Thank you. 😄

Laurence Moroney thanks again for this amazing opportunity. I didn’t know you don’t get any profit for your book. This leaves me speechless

Rohan – My inspiration. I guess there’s a number of places. First – I graduated college in 1991, in the UK, in the middle of the biggest economic recession since WW2, and matched only by the current one, I think. I was unemployed, and it was really hard to find work. I spent some time as a clerk for London Underground (and got laid off on Christmas Eve, very Dickensian), and doing odd-jobs here and there. I remember working in a recycling plant, being the guy that collects the tin cans that people come in off the street to hand us, expecting payment, weighing them, and seeing people’s disappointment at the few pennies they got. Then, the UK government started a scheme where they wanted to find 20 people, who are currently unemployed, to train them in AI to be a cohort of future consultants in industry. It was 1992 at this point, and I had been out of work for about a year, and really struggling. I went to the testing center, along with thousands of others, and was the first person selected to enter the program, with the highest scores. It was a massive boost to my flagging confidence! Unfortunately the program failed miserably after about 3 months, but the contract I had signed had entitled me to 2 years of education, so it got parlayed into a Masters degree, fully paid, which I graduated from in 1993, as the economy was emerging, and my career got on track. Then…Fast forward to 2017, I’m now at Google, and we had an initiative to train all engineers in ML and AI. I remember sitting in a conference room in Kirkland (near Seattle), with a few hundred other folks, excited as anything….and….after 3 hours of calculus, and most of the room bored out of their minds, I realized something was wrong with how we teach ML as an industry. I approached the TF team showing how a code-first approach would work better, and how the material is aimed at the 300K current AI practitioners globally (measured by academic publications) instead of the 30M+ software developers.
So they hired me, and asked me to fix it.
And here I am :)

Saurav Maheshkar – Thanks! 🙂 As for ‘Full Stack’ – I don’t think it’s well defined yet, but I will take a stab at definining it similarly to how we define full stack web developers – and that is someone who is involved, at least in part, in every part of the overall stack to solve a problem. A full stack web developer, for example, is involved in coding the part of the solution that lives on the server, but they’re not necessarily an expert in infosec. They’ll interface with the people who are. They’ll also be involved in coding the front end, and the UX, but they’re not necessarily a designer or a UXE. So, similarly, when it comes to ML, if you explore the life cycle from data->feature engineering->model training->deployment->management, I think a full stack ML Engineer will be involved in all of them to some extent, but the closer you get to the center of that chart, the more they do.