Politicians, athletes, journalists, NGOs, companies... They all use
corporation-owned social networks to communicate with us. Citizens
shouldn't have to hand over their data to companies to read content
published by public entities. Public Feeds changes that by making posts
of popular accounts freely available.
Read more.
Fast.ai
fast.ai is a non-profit research group focused on deep learning and artificial intelligence. It was founded in 2016 by Jeremy Howard and Rachel Thomas with the goal of democratizing deep learning. They do this by providing a massive open online course (MOOC) named "Practical Deep Learning for Coders," which has no other prerequisites except for knowledge of the programming language Python. More on Wikipedia
It's a strange time to be a programmer—easier than ever to get started, but easier to let AI steer you into frustration. We've got an antidote that we've been using ourselves with 1000 preview users for the last year: "solveit"
My @fastdotai cofounder @math_rachel has been studying immunology for the last couple of years, & she's discovered some fascinating opportunities for AI in immunology.
Soooo proud to be @fastdotai alumni 👨🎓 I was stuck in math previously. @fastdotai guided me into right direction. Can’t imagine only 2 months after the course, I co-founded startup with fast.ai Alumni and won $10,000 check 🌻🚀
Second notebook "Detect Sleep States" @kaggle competition is done 🎃 it contains: ▫️ EDA & visuals of time series data ▫️ sleep representations using spectrograms ▫️ time feature engineering using @fastdotai ▫️ @wandb tracking for datasets and visuals
I am so excited to see @jeremyphoward as a guest on @latentspacepod. The @fastdotai book and course helped to get me started with deep learning and also brought together an excellent community of fellow learners.
"The right way to fine-tune language models... is to actually throw away the idea of fine-tuning. There's no such thing. There's only continued pre-training."
If you need a bit more time before diving deeper, finish the following Kaggle tutorials:
• Intro to Machine Learning • Intermediate Machine Learning
At this point, you are ready to finish your first project: The Titanic Challenge on Kaggle.
If Math is not your strong suit, don't worry. I don't recommend you spend too much time learning Math before writing code. Instead, learn the concepts on-demand: Find what you need when needed.
From here, take the Machine Learning specialization in Coursera. It's more advanced, and it will stretch you out a bit.
The top universities worldwide have published their Machine Learning and Deep Learning classes online. Here are some of them:
• MIT 6.S191 Introduction to Deep Learning • DS-GA 1008 Deep Learning • UC Berkeley Full Stack Deep Learning • UC Berkeley CS 182 Deep Learning • Cornell Tech CS 5787 Applied Machine Learning
Many different books will help you. The attached image will give you an idea of my favorite ones.
Finally, keep these three ideas in mind:
1. Start by working on solved problems so you can find help whenever you get stuck.
2. ChatGPT will help you make progress. Use it to summarize complex concepts and generate questions you can answer to practice.
3. Find a community here on 𝕏 and share your work. Ask questions, and help others.
During this time, you'll deal with a lot. Sometimes, you will feel it's impossible to keep up with everything happening, and you'll be right.
Here are the good news:
Most people understand a tiny fraction of the world of Machine Learning. You don't need more to build a fantastic career in the space.
Focus on finding your path, and Write. More. Code.
It’s cool to me how many @fastdotai alums are working on LLM stuff now. It’s a real testament to how the course didn’t just teach us specific knowledge but also a practical, problem-solving approach to ML rooted in open source (And I’ve been seeing @jeremyphoward in Discord too!)
A few months ago as part of Walk with @fastdotai, I did a presentation discussing @huggingface Accelerate. This talked about what the framework is, its layers, and how to utilize it best for your problems.
Our new course, "From Deep Learning Foundations to Stable Diffusion", is finally done after 8 months of work!!!
With >30 hours of video content (all free, no ads!), you'll learn how to create and train a Stable Diffusion model starting from pure Python 🧵 www.fast.ai/posts/part2-2023.html
Our new course, "From Deep Learning Foundations to Stable Diffusion", is finally done after 8 months of work!!!
With >30 hours of video content (all free, no ads!), you'll learn how to create and train a Stable Diffusion model starting from pure Python 🧵 www.fast.ai/posts/part2-2023.html
- PyTorch 2.0 compatibility, thanks to @benjamin_warner - Pytorch 2.0 compiler compatibility, thanks to ggosline - Better version support for `tensorbase.new_empty`, thanks to @benj…
- add agplv3 license to project options, thanks to joel-lbth - add 0 - Pre-planning status to project options, thanks to ddobrinskiy - add conda installation docs to end-to-end tutori…
At @fastdotai we’ve often been asked by orgs for help. So we're building a partner program, which allows businesses to become part of our global community, to help access the resources and advice they need to take advantage of what we are building. www.fast.ai/partners.html
@fastdotai callbacks have always looked like wizardry to me. They allow infinite flexibility and they just work. I am attending fastai course part2 now and I tried to demystify them. What is going on inside a callback? Hope you find the deep-dive useful! francescopochetti.com/fastai-course-part-2-2022-understan…
Has @fastdotai's DataBlock API confused or frustrated you? Me too.
All of Walk with fastai revisited puts a ~heavy~ focus on the Datasets API instead (with some DataBlock) which is much more flexible and easy to work with
nbdev-extensions, a collection of extensions for @fastdotai's nbdev is now on version 0.1.0. It's still the three main extensions, however just now they're stable enough for me to write the entire course with :) muellerzr.github.io/nbdev-extensions/
- utility that creates a `requirements.txt` file from `settings.ini`, thanks to @HamelHusain - user-friendly error if py file has `# %%` comments with unexpected format, thanks to @wasi…
- Add torch save and load kwargs, thanks to JonathanGrant - This lets us do nice things like set pickle_module to cloudpickle - PyTorch 1.13 Compatibility, thanks to @warner_benja…
- better error messages for `nbdev_migrate` - experimental: Users can provide extra processors via the `procs` key in `settings.ini`, thanks to @wasimlorgat - enable Documentation Only …
This week in the @fastdotai part 2 course we started the exciting work of implementing our own deep learning library from scratch (mostly). This involved @jeremyphoward getting the data from the source files, parsing them out, defining a pseudo-tensor (or Matrix) class and so on.
The impact of @fastdotai on the practical implementation of AI in the world since its first version is massively understated! This is the course that changes the course of everyone who does it!
After just 2 weeks of the new @fastdotai course, our students are already making research advances in Stable Diffusion.
@sebderhy developed a novel yet simple modification to classifier-free guidance that gives better results (previous approach on left, new approach on right)
1/ I've been acting as interim @ezrainc CTO for the past few months, so I'm spending a lot of my time working on ML projects with our incredible AI team.
Decided it's time to go deeper into building models myself, and I'll be documenting my journey here (inspired by @Suhail).
In my seven years experience leading data science teams, I am yet to see a problem in ML that I was not able to solve with this stack: 1. Text: @huggingface and @fastdotai 2. Tabular: XGBoost and LightGBM
- Set recursive `True` by default, thanks to @wasimlorgat - `exec_doc` supports re-rendering widgets, thanks to @wasimlorgat - If users update widgets with the `exec_doc` directive, t…
- fix: error in IPython while handling `HTTP4xxClientError`, thanks to @wasimlorgat - fix: `Config.get` returns `False` instead of `None` for missing `bool` keys, thanks to @wa…
- `_quarto.yml` no longer replaced with automatically generated version - You can now customise your `_quarto.yml` file and it will not be overridden by nbdev - This means that …
The @fastdotai Course is a fantastic place to start. Instead of going from the basics towards an application, it starts from an application and breaks it down piece by piece.
- Add `printit` arg to `nbdev_filter` so it can be called with `fname` and still print to stdout, thanks to @wasimlorgat - Run `nbdev_readme` in `nbdev_new`, thanks to @wasimlorgat - In…
- Build `_modidx.py` on demand in order to git conflicts - Add source link to index - Order left navigation sections using numeric prefix - Automatic rendering of python files with fron…
- Combine preprocs and postprocs into new `Processor` class - Rename `nbdev.read` to `nbdev.config` - Use H3 for functions and properties, instead of H4 - Remove `nbflags` directive…
- exclude `passed to {func}` from docments by passing `verbose=False` to `delegates`, thanks to @wasimlorgat - enable meta.delegates to enforce KEYWORD_ONLY for kwargs from, tha…
- Add `custom_quarto_yml` setting, thanks to benoit-cty - Display multiline docstrings - Include filename in `nbdev_export` warning when nbdev1 syntax is used - Streamline `nbdev_new`: …
- layered `Config` with `extra_files`; `Config` `repr`, thanks to @wasimlorgat - in-memory `Config` by passing `save=False`, thanks to @wasimlorgat - use `==` for non iterable a…
- use global defaults instead of respecifying each time, thanks to @wasimlorgat - `get_config` works without a settings file, thanks to @wasimlorgat - add site url, thanks to @HamelHusa…
The latest fastxtend release brings automatic Progressive Resizing to @fastdotai.
Inspired by @MosaicML's implementation for Composer, which in turn was inspired by the fastai course, Progressive Resizing can speed up training with a small decrease in model performance.
- add split value argument to ColSplitter, thanks to @DanteOzML - deterministic repr for PIL images - option to skip default callbacks in `Learner` - update for nbdev2
- git merge hooks: automatically resolve conflicts and render markers as separate cells, thanks to @wasimlorgat - Allow clean to keep some metadata keys, thanks to dleen - enable mac te…
- new jupyter save hook to clean NBs, thanks to @wasimlorgat - new directive `exec_doc` to auto-exec cell when building docs - automatically escape YAML strings for title and descriptio…
@HamelHusain & I have just recorded a step-by-step walk-thru. We show, from scratch, how to build a complete Python project including documentation website, modules, pip installer, README, etc.
Given there's so many new folks exploring the @fastdotai library with the new course, thought I'd push a tutorial I wrote for using @PyTorch in the framework from scratch, step by step. Hope it helps someone! 🤗
- add `delete_file` and `create_file` - Add a "Documentation" link to the sidebar on PyPI, thanks to @nedbat - Add support for media types, thanks to lfdebrux
I've just published over 20 hours of tutorials and live coding showing how to: install python the right way; set up a terminal; write shell scripts; use vim; use a remote Jupyter server; use git, github, tmux, and ssh; use the python debugger; and more! 🧵 youtube.com/playlist?list=PLfYUBJiXbdtSLBPJ1GMx-sQWf6iNhb8mM
"Practical Deep Learning for Coders" - latest version of this free course is released by @fastdotai Accompanied by a book with all the exercises available in Colab.
After 2 years, Practical Deep Learning for Coders v5 is finally ready! 🎊
This is a from-scratch rewrite of our most popular course. It has a focus on interactive explorations, & covers @PyTorch, @huggingface, DeBERTa, ConvNeXt, @Gradio & other goodies 🧵 course.fast.ai
@fastdotai flexibility keeps surprising me I manage to build and train a graph convolutional neuralnet (GCN) for lipophilicity prediction. The performance isn't that bad either, and it was simpler than pure torch or other frameworks! #compchem#deeplearning#machinelearning
A public submittable kernel to get you on your way to competing in the @kaggle "Feedback Prize" competition using #blurr, @fastdotai, and @huggingface.
I stumbled upon two AI projects which I found pretty unique and these are some excellent side projects. The first one is to build an AI to play fall guys build with @fastdotai . Also the creator Juan Gonzalez had created an excellent video.
We had really fun doing this with @manan_goel . TSAI is a breeze to use and as it is built on top of @fastdotai , you get W&B support by just passing a callback.
😎 Here's @manan_goel's and @capetorch's report on how you can use tsai to train a classification model on time series data. It also shows how to use W&B to compare experiments and track your models.
One of my fave chapters of "Practical Deep Learning for Coders", co-written with @GuggerSylvain, is chapter 8. I've just made the whole thing available as an executable notebook on Kaggle!
Nice study. TL;DR: if you want a resolution independent, pre-trained, image recognition model that can be fine-tuned with a small and different dataset, use ConvNext.
- Distributed training now uses Hugging Face Accelerate, rather than fastai's launcher. Distributed training is now supported in a notebook -- see [this tutorial](docs.f…
- Added webrick spec to Gemfile., thanks to MarkB2 - Change doc() default for docments, thanks to @TheZachMueller - Better checks for cls and self, thanks to @TheZachMueller - Use the…
- Custom raise message helper, thanks to @TheZachMueller - make `docments` handle delegates, thanks to @HamelHusain - Add a `nested_callable` and `getcallable` util, thanks to @Th…
FYI you'll see a few "new release" notifications today for releases that were actually a few weeks ago. Turns out we had a firewall rule that was blocking a webhook, so those new release tweets weren't getting sent.
Chapter 4 of our book (covering the foundations of deep learning) is, for most people, the most challenging. It covers a lot of math and coding concepts very quickly.
This series from @strickvl is an amazing resource to help you understand how a neural net really works:
This week I wanted to really make sure I grasped what was going on in ch 4 of the @fastdotai book, so I put my study hat on. 🤓 Writing helps me understand so I split the things I learned into separate blog posts. I reimplemented the methods with the Fashion MNIST dataset.
There's currently a problem on @kaggle that in their latest @PyTorch upgrade they didn't compile in the MAGMA library (required for linear algebra support). So GPU-based transforms don't work. Here's how to fix it.
-From learning to code at 29 -through learning ML @fastdotai -winning a @kaggle competition -jobs at 🔥 startups -moving continents thx to AI -to joining the illustrious Merlin team ❤️
Excited to announced the release of fastai 2.6, which now supports @wightmanr's brillant PyTorch Image Models library (timm). That includes support for any Transformers-based models in timm!
To use, just pass the name of a model to `vision_learner`.
The second part of my work on #resnet for traffic sign recognition using #pytorch is out! In this, I use #knowledgedistillation to get closer to the previous results with a smaller model. Will I make it?🤔
I'm pleased to (𝘧𝘪𝘯𝘢𝘭𝘭𝘺) announce the official 𝘃.𝟭 𝗿𝗲𝗹𝗲𝗮𝘀𝗲 of 🄱🄻🅄🅁🅁 🎉, a library designed for @fastdotai developers that want to train and deploy @huggingface transformers.
Tip: if you're looking for help getting through a data science interview, @bhutanisanyam1 and the @fastdotai community have put together a great list of resources here:
With another @fastdotai course a month away, I'm working hard to finish my commentary on at least all of the fastbook chapters that will be covered in it! Hoping the series will be a helpful resource to both new and veteran fastai folks alike 🙏
Relevant metric: number of single-sample gradient evaluations.
Countless students have told me "I used 10% of the training set because 100 epochs on the full set takes too long" My answer: "there is *never* *any* *reason* to drop training samples."
(rant) "epochs" are a bug-inducing concept in neural net training and should be avoided in favor of number of iterations. Use of epochs silently functionally distorts training (as datasets change/grow) and decreases code portability (when one wishes to train on different dataset)
📣 Folks, after much anticipation, we're THRILLED to announce that enrolments for the 7-week short course "Practical deep Learning for Coders (a UQ collab w/ fast.ai@jeremyphoward)" are now OPEN!
"Renowned #ai expert joins #uq to help develop local talent"
1/ 📣 A big welcome to our #uqitee Honorary Professor @jeremyphoward on his recent appointment 👏 We're thrilled to have him as part of our community! More info below 👇
I started my ML internship @TambuaHealth this week🥳. I met such a great team and I am already learning a lot from them. I'm delighted to be part of a team that is creating value and having fun at the same time. www.t-health.ai
- New method: `Path.delete()` to remove a file or directory ([#336](github.com/fastai/fastcore/issues/336)) - Add `urlretrieve`, which as same as `urllib.request.url…
- remove fastai-specific pieces from `Config` ([#334](github.com/fastai/fastcore/issues/334)) - in `untar_dir` create a directory if there is more than one item in r…
Have you ever wanted an application-based @fastdotai course, throwing as many different DataBlock situations as possible, while learning how to customize the API to it's fullest extent? I have a free course at walkwithfastai.com!
Sneak preview at an upcoming blog post (jwuphysics.substack.com)! Converting SDSS (left) to DECaLS galaxy imaging (right), via a U-Net (center left), and a U-Net + GAN (center right).
It’s a large part of why I’m so active. The amount of knowledge you soak in by just helping small bugs, or reading ideas of others, allows for you to get a firm grasp of a library (over time), while also introducing you to ideas and concepts you would have never explored
We've just released fastai 2.2. This contains some fairly minor breaking changes, which should only impact you if you write your own callbacks, or use custom RNN code.
We've just released fastai 2.2. This contains some fairly minor breaking changes, which should only impact you if you write your own callbacks, or use custom RNN code.
Each of the breaking changes in the release notes above includes a link to more details about the change. For instance, here's the link for the changes to mixed precision training 2/ github.com/fastai/fastai/issues/3127
Writing tests with pytest and figuring out how to capture logs, output, err correctly with custom loggers 😠 Makes me like nbdev from @fastdotai even more
Join the School of Professional Studies's Master's in Data Science Student Leadership Council for an event next week featuring Hamel Husain, the Staff Machine Learning Engineer at @github. Register now. www.eventbrite.com/e/msds-slc-speaker-series-fall-event-o…
Looking for the @fastdotai of quantum computing : @pennylaneai is the way to go. It’s clear with all you need at one place : 1/ a simple wrapper with associates boiler plate 2/ a strong community 3/ code oriented first rather than fully theorical 4/ very clear educational content
If you're training deep learning models, @fastdotai should definitely be your tool.
And if you want to additionally monitor your training (believe us, you should want that), try Neptune-fastai integration. It's just one additional callback.
Updated blurr to be @PyTorch 1.7 and @fastdotai 2.1.x complaint. Also a new "Examples" section in the docs with hopes users will submit more examples via PRs ... the first I included is for doing multi-label sequence classification (see ohmeow.github.io/blurr/examples-multilabel/)
Thread of some posts about diversity & inclusion I've written over the years. I still stand behind these.
(I'm resharing bc a few folks are suggesting Jeremy's CoC experience ➡️ partially our fault for promoting diversity, we should change our values, etc. Nope!)
The new `fastcore.basics` provides a lot of great functionality, but is *really* lightweight (only 0.02 secs to import, and the entire `fastcore` has no deps other than python itself)
I'm about halfway through @fastdotai#deeplearning for #coders and just wanted to say thank you so much! I have learned a bit about deep learning from @iamtrask 's amazing book, but taking the theory and seeing it applied in your course has been a truly incredible experience.
Quick Python debugging tip: if you're debugging a command-line app which isn't run by typing "python ..." you can drop into a debugger if it hits an exception like this:
New blog post! Fine-tuning ResNet-18 for Audio Classification: Reproducible experiments using fastai, fastaudio, and wandb sweeps to explore hyperparameters for achieving high accuracy on the ESC-50 dataset. wandb.ai/jhartquist/fastaudio-esc-50/reports/Fine-tuning-… Special thanks to @fastdotai@wandb
We'll be experimenting with the Twitter API a bit today, so apologies if you see some test tweets appearing in disappearing in our stream over the next couple of hours...
The reason is that we got sucked into accidentally using paid-only features in Zapier, so when the free trial ran out they disabled our auto-posting thing.
(...which is the reason I normally avoid these kinds of services. I much prefer writing my own scripts anyway.)
- new callback event: `after_create` ([#2842](github.com/fastai/fastai/issues/2842)) - This event runs after a `Learner` is constructed. It's useful for initial setup which isn't needed for
Knowing @fastdotai feels like a cheat code. I can't resist sharing this - a piece of code golf that currently gets 5th place on @ZindiAfrica's @CGIAR@CGIAR_Data Wheat Growth Stage Challenge (~0.47). I'm working on a nice starter notebook so you don't need to parse this mess 🙃⛳️
`dict2obj`, just added to fastcore, is a handy (yet tiny) function that makes nested dict access more convenient, including providing tab completion of nested dicts.
- Undo breaking num_workers fix ([#2804](github.com/fastai/fastai/pull/2804)) - Some users found the recent addition of `num_workers` to inference functions was causing problems, partic
Heads up: we will be removing #nbdev callbacks and nbdev ipython magics, in the next minor release (v1.1.0). At that time, I'll provide little scripts to auto-convert nbdev magics to nbdev comments.
(NB: this is not in any way related to fastai callbacks)
nbdev magics and callbacks, whilst cool, were a recent experimental feature that just added too much maintenance overhead and code complexity, and didn't have enough benefits to justify that.
.@jeremyphoward will deliver a keynote at #jupytercon2020! Jeremy is a Co-Founder of @fastdotai, a Researcher in Residence at @usfca, and the Chief Scientist at @platform_ai.
We've just enabled tweeting an announcement of each new release to the #fastai, #nbdev, and #fastcore libraries, to this account. Just about to test it out, so watch this space! (We're using @zapier for this.)
A wonderful post from @fastdotai on their philosophy of teaching. It's very much in line with our own and a great summation of the essence of good teaching.
Our in-person deep learning course starting March 17 at @DataInstituteSF will be our biggest opportunity ever to experience the cutting edge of this field.
We've made a draft of our deep learning book available as @ProjectJupyter notebooks.
Important: This is for personal use, and not licensed for any redistribution or change of format or medium. Eg. don't make a PDF and share it, please! github.com/fastai/fastbook
Thank you everyone for your patience - our new deep learning MOOC is here! Includes 5 lessons diving into the foundations with @PyTorch, and 2 lessons (co-taught with @clattner_llvm) building a new deep learning library in #swift www.fast.ai/2019/06/28/course-p2v3/