Hi Raghav Bali - fastai is a great platform, particularly for beginners. That being said, I see two noteworthy drawbacks: (1) compared to TF/Keras or vanilla PyTorch, there are not as many examples of fastai being deployed in production, (2) regressions. Compared to Keras, it is easier to hit regressions with the fastai platform (that is, something that used to work that stops working after a platform update), and you need to be prepared to work around regressions if you come across them in fastai. An example of a regression in fastai is the code to set the random seed, which stopped working as expected after an update. In addition to these drawbacks, the documentation for fastai is not as comprehensive as Keras documentation, but the fastai course forum https://forums.fast.ai/ helps to make up for that.

Awesome… Thanks for the detailed answer

Hi WingCode - the advantage of the abstracted data structures in fastai is that they are integrated into the deep learning workflow and make it easy, for example, to look at a sample batch or examine individual elements in a dataset. Simply put, they help make it easy to focus on the essential tasks of creating and training a deep learning model without having to write a lot of additional code.

Thank you Mark for the answer 🙂

Hi John Trengrove - that is a great question. There are examples in the book of vision, tabular, text and recommender system models, but there aren’t examples that combine these data types.

Hi WingCode - the recommender system with the small dataset was an attempt to demonstrate an MVP of fastai’s recommender system support. The resulting model works but it is of limited practical value because the columns in the dataset don’t mean much to humans (it’s a movie rating dataset with no movie titles). Once the reader has been through that example, it motivates the second example with the large dataset (which is more than 10 times bigger than the small dataset and has a much more complex structure). The second example takes many more steps to get it to work, but the result is more satisfying - you see the recommender system making predictions about the rating a particular user will give for a movie that is generally thought to be excellent (L.A. Confidential) vs. a movie that is generally thought to be terrible (Showgirls).

The reason for doing an MVP first and then a bigger example is that some of the existing material on fastai doesn’t take the time to do an MVP and jumps right into the full-blown solution with a big, complex dataset. The problem with skipping the MVP is that it’s hard to figure out the reason for the additional steps for the big, complex dataset if you haven’t already grasped the minimum set of steps.

Yes that makes sense Mark. Thank you again 🙂

Mark Ryan I think the approach of showing the MVP approach first makes a lot of sense, for the reasons you described. It also gives the reader a good example of how things might be done for a real-world problem or project they might be working on… start first with something simple and small to evaluate/validate the approach before moving on and investing the time on something more realistic and complex.

Hi Carlos Orjuela - the book by Jeremy Howard and Sylvain Gugger is a master work. It not only covers exhaustive details about fastai, it also takes the reader through many aspects of the theory of deep learning. It also includes a whole chapter on AI ethics. By contrast, my book is focused on using fastai to solve practical problems. It includes details on setting up an environment to use fastai and then has a chapter each on the 4 major application areas supported by fastai: tabular data, text data, recommender systems, and image data. In each of these chapters, you’re led through a working code example for a fastai curated dataset and then working through an example for a standalone dataset. My book wraps up with a chapter on model deployment and then a chapter on more advanced topics (including callbacks, augmented data, and more additional deployment tips).

So, Jeremy Howard’s book can teach you many general deep learning concepts and covers all kinds of nuances and details about fastai. It is also a complement to the deep learning for coders course https://course.fast.ai/. My book is focused on step-by-step application of fastai to solve practical problems.

Thanks Mark Ryan for the detailed response 🙂

Hi Haseeb Arshad - thanks for your question. As for the best part of the book, it is probably not right for me to say so as the author, but if you look at the reviews on Amazon you will see some of the things that people liked about the book: https://www.amazon.com/Deep-Learning-fastai-Cookbook-easy/dp/1800208103#customerReviews. I do believe that the book is good for beginners to learn about fastai - it takes you from the very start of setting up an environment all the way to deploying a trained model.

Hi Carlos Orjuela - the book covers practical examples and talks about ways to work around some of the drawbacks of fastai. It also describes the many strong points of fastai, such as (1) the way fastai provides comprehensive support for tabular datasets and (2) how fastai makes it easier than Keras to ensure that data goes through the same transformations when the model makes a prediction as the data went through when the model was trained. The book has a whole chapter on deployment. Deployment is a huge topic, so this chapter doesn’t cover all deployment options. In particular, many production deployments would be via a cloud platform like AWS, Google Cloud, or Azure, and the book does not explain how to do that.

Thanks again for your reply Mark Ryan

Hi Allan - I think it’s fair to say that Jeremy Howard’s philosophy about learning deep learning is that it’s best to start by doing. His course (and the fastai framework) emphasize getting something to work in code first and then go back and dig into the theory / math. He contrasts the fastai approach with the traditional approach to learning deep learning, which starts with abstractions and theory that can be offputting to somebody who is trying to use deep learning to solve a real-world problem. I think that Howard wants fastai to be used by people who don’t have an academic background in machine learning but do have deep subject area expertise and access to interesting data sets.

Hi Tim Becker - in the book I did a simple “from the ground up” web deployment of some fastai models using Flask. In this scenario, fastai was easier than Keras. To do the same kind of deployment with Keras for a tabular model, I needed to worry about the pipeline and I needed to load the model repeatedly. With fastai, the pipeline was taken care of automatically and I only had to load the model when the web page initially loaded, leading to a smoother experience.

The question of Keras vs. fastai is a good one. Keras definitely has some advantages over fastai, including a much larger user community and better documentation. If that’s the case, why would somebody who has already used Keras go through the effort of learning to use fastai? I can think of two reasons (a) gateway to PyTorch for somebody who has been working in the TF sphere. fastai provides a low-impact way to ease into PyTorch for somebody who is familiar with the TF ecosystem and wants to take a peek at the “other side”. (b) full-throttle support for tabular datasets. It’s certainly possible to create models trained on tabular data using Keras (I spent another book on that topic: https://www.amazon.com/Deep-Learning-Structured-Data-Mark/dp/1617296724/ref=sr_1_1?keywords=deep+learning+with+structured+data&qid=1638993209&sr=8-1 ) but fastai makes it much easier.

Mark Ryan Thank you for your answers, very interesting! Could you elaborate a little bit what you mean by additional support for tabular data?

This is very helpful!

Hi Allan - there is a chapter in the book on deployment. This covers how to deploy fastai models trained on tabular and image datasets in a simple web application using the Python Flask library. These deployments include everything - the Flask module, the HTML and CSS files, and the trained models, as well as detailed instructions on how to set them up. The final chapter of the book gives additional detail on web deployment - for example, showing you how to show thumbnails of the images the model is making predictions on and showing you how to make a deployment on your local machine available to the external web. The final chapter also covers (a) fastai callbacks - these make the fastai training process efficient by avoiding non-productive training epochs and ensuring you exit the training run with the optimally trained model (b) fastai’s built-in support for augmenting image data (c) various features for getting additional details about models trained with fastai, such as confusion matrices and data points where the model did worst.

Thanks Mark Ryan , sounds like a very practical couple of chapters that adds a lot of value!

Hi Allan - the conventional wisdom is that TF is used more by industry and PyTorch is used more by academics. The Stack Overflow 2021 survey backs this up: https://insights.stackoverflow.com/survey/2021#section-most-popular-technologies-other-frameworks-and-libraries. I’d say that it’s hard to compare PyTorch and TF without bringing Keras into the discussion. After all, it is now the official high-level API for TF, and it’s used extensively, even by specialists. Search the Stack Overflow survey for Keras and you will see it, while fastai and Lightening (another high-level framework for PyTorch) are both missing. This could be because professional developers use PyTorch directly and don’t have use for a higher level framework, or it could be that developers working in industry are still predominantly on TF/Keras.

I can’t predict the future, but I think it would be a safe bet to say that while PyTorch will be adopted more by industry, TF/Keras isn’t going anywhere and will still be a significant part of the landscape in 10 years.

Thanks Allan - I really enjoyed the chance to answer the questions this week. I appreciate the time that people took to share these great questions.

Are you working on another book right now? =) I bet you are and we can invite you again soon =)

Hi Alexey Grigorev - I’m not actively writing right now, but I have a couple of ideas in the works, so I look forward to coming back again to talk about the next book.