# Practical Deep Learning for Coders: nine lessons from code to a working model > Jeremy Howard's free course has you running a model on real data in the first lesson. The theory comes later. Bản gốc: https://fdetimes.net/en/books-courses/fast-ai-practical-deep-learning-for-coders/ The first lesson of Practical Deep Learning for Coders does not open with derivatives or matrices. It opens with a Kaggle notebook called "Is it a bird? Creating a model from your own data", in which you build a bird detector from data you have collected yourself. That ordering is the course's whole philosophy. For a developer who wants to become an FDE, it is also how the job works on a client site: get something running on real data first, then explain it and tune it. ## Who runs the course, and what does it ask of you? The course comes from fast.ai, which Jeremy Howard and Rachel Thomas co-founded. The course.fast.ai site describes it as free and aimed at people who already have coding experience. The version currently listed is Practical Deep Learning for Coders 2022, part 1: nine lessons of about 90 minutes each, roughly 13.5 hours of video in total. The entry bar is surprisingly low. You need to know how to program, and according to the course site a year of experience is enough. No university-level maths is required. You do not need to buy a GPU either. The main lesson 1 notebook runs on Kaggle, and the companion book's notebooks open in Google Colab with no environment to set up. ## Why does top-down learning suit FDEs? Jeremy Howard writes that he starts by having learners use a complete, working, state-of-the-art deep learning network that they can put to use straight away. Only then does he work gradually into how it is built. The course site states its second principle just as plainly: it always teaches through examples rather than algebraic notation. Consider a scenario. A factory wants to know whether AI can spot defective products from photographs. An engineer trained bottom-up will spend weeks revising theory before daring to touch the data. Someone who has done fast.ai's lesson 1 will recognise a familiar structure: two kinds of image, "defective" and "not defective", in place of "bird" and "not a bird". The first demo can therefore happen in days rather than weeks. And it runs on the client's own photos, not a sample dataset. **Điểm mấu chốt:** Get a working model first, then dig into each layer underneath. That is the difference between learning ML to pass an exam and learning it to deploy. At the prototype stage, a model running on the client's own images is usually more persuasive than any formula, and the course trains exactly that reflex. Teaching through examples also helps when you present results. For an operations manager, five images the model got right and five it got wrong are usually easier to grasp than any loss curve. ## In what order should you study to avoid being overwhelmed? A sensible path has three steps. Step one is Part 1: work through the nine lessons in order and rerun every notebook. Step two is the official companion book, "Deep Learning for Coders with Fastai and PyTorch: AI Applications Without a PhD" by Jeremy Howard and Sylvain Gugger, published by O'Reilly in 2020. The book's notebooks form the foundation of the course, and the authors publish them free on GitHub in the fastai/fastbook repository. The book arrived as part of a large release: on 21 August 2020 fast.ai launched the 2020 course, the fastai v2 library and a 600-page book at the same time. Reading the book alongside the videos gives you a written version to look things up in, which a 90-minute video struggles to provide. Step three, only once you are genuinely solid, is Part 2: "From Deep Learning Foundations to Stable Diffusion". It was released free on 4 April 2023, with more than 30 hours of video, and teaches you to implement the Stable Diffusion algorithm from scratch. fast.ai is explicit that to get the most from Part 2 you should be a reasonably confident deep learning practitioner. Part 1 is therefore a stepping stone you should not skip. A suggested schedule for people with a day job: spend the first week on lesson 1 alone, swapping your own data into the notebook. Use the next three weeks for lessons 2 to 9, two or three a week. Spend the following two weeks rereading the matching book chapters on Colab, and the final two on a small project. If your goal is an FDE role within six months, Part 1 plus the book is enough for the foundations. Part 2 suits people who want a deep understanding of generative models, or who are targeting companies building products around image models. ## Mistakes that waste the course The most common mistake is jumping to Part 2 too early because Stable Diffusion sounds more exciting than classifying birds. fast.ai itself sets the condition that you should be fairly confident with deep learning; ignore it and you will be copying code without understanding why it runs. The second mistake is running a notebook top to bottom and calling it learned. The first notebook's title says it plainly: create a model from your own data. If you have never swapped in your own data and looked at where the model fails, you have watched a demo, not built a prototype. Client data is rarely as clean as a sample set. Practising how to collect, filter and label real data while you are still learning is an investment worth making. The third mistake is watching the videos and skipping the book. The videos give you pace; the book's notebooks give you a place to stop, change a line and see what happens. ## Turning the course into evidence on your CV A completion certificate says very little about your ability to prototype on real data. A small project with a clear story says much more. When an FDE job description asks for rapid prototyping or work with customer data, that project is a direct answer. Take the lesson 1 notebook and replace the birds with a problem close to a real business, such as classifying photos of invoices, or photos of farm produce that does or does not meet grading standards. On your CV, write one line saying you collected the data yourself, built the model and analysed the cases where it went wrong. That line shows you have completed a full prototype loop on real data, from raw images to a working model whose failures you understand. The course is not stronger on theory than a university textbook. Its value is that it teaches the habit deployment work needs most: no talk of theory until something runs. **Thử ngay tuần này:** - Open the lesson 1 Kaggle notebook 'Is it a bird?', run it end to end, then replace the two image classes with two kinds of image from a problem you care about - Write one page covering where the data came from, which images the model got wrong, and what you would ask the client next - Open the fastbook chapter 1 notebook in Google Colab and compare it with the lesson you have just watched ## Nguồn - [Practical Deep Learning](https://course.fast.ai/) - [Practical Deep Learning – Lesson 1](https://course.fast.ai/Lessons/lesson1.html) - [From Deep Learning Foundations to Stable Diffusion](https://www.fast.ai/posts/2023-04-04-part2-2023.html) - [fastai/fastbook: The fastai book, published as Jupyter Notebooks](https://github.com/fastai/fastbook) - [fast.ai – Making neural nets uncool again](https://www.fast.ai/)