Wow. My last exposure to this excellent course series was in 2018, after which I’ve stayed peripherally connected but really have been “off doing my own thing” since then research-wise and learning from others elsewhere.
I recently returned to the
fastaiv2 library and
nbdev in the process of planning out the code & teaching for my own university-based DL course this fall, for which I’d decided to base it on using Colab and/or Gradient, coding in “PyTorch and a bit of fastai for convenience”, after eschewing Lightning for [you know] and noting that JAX, while gaining popularity among researchers, didn’t have yet have a mature-enough ecosystem of “tools” for for the things I wanted to do with the students.
And what I wanted to do were some things I’d done in my previous ML course (e.g. integrated ethics), but “new” things like:
- dataset building and cleaning (influenced by Justin Salamon of Adobe),
- visualization tools (influenced via contacts at Facebook) for weights & activations, and saliency/attribution/intepretability
- model deployment (via local ML developers in my city, & 2018-era fast.ai),
- specific advice for people in business (influenced by contacts at Seattle U., and the fact that my course is co-listed in the Business school)…
- embeddings and zero-shot classifiers (b/c been writing a book on classification!)
- Emphasize Transfer Learning (totally influenced by fast.ai alone!)
…Familiar, right? I did not even know the first 4 things were now such highly developed parts of the course—and I celebrated news of your book release but haven’t even looked at it yet—, until two days ago @NathanSepulveda hit me up on Twitter about the project he’s building as part of the fast.ai course, while I was in the middle of building my own graphical dataset-curation tool to use on Colab, and trying to build my own library via
nbdev for the first time & posting to these forums about learning
nbdev, but still not having viewed the 2020 course. I wasn’t even planning to watch the 2020 course, but after interacting with Nathan I started it yesterday, and…
OH MY GOODNESS. It is so good! Watching Jeremy is like…watching a concert by a virtuoso musician, only it’s a performance of “deep learning education”. Even the little “side” comments are right on time and apropos. So, soo good. Wow. Jeremy, Rachel, Sylvain: Be encouraged! I won’t “at” you, but I hope you see this post. Thanks for your work and service. Hope things go well in Australia & at HuggingFace.
So, then, the Disclaimer:
Talk about “ideas are rarely new or original”: My data-scraping and cleaning tools don’t look like like yours but are similarly Jupyter widget-based interactive GUIs. So…In teaching my own course, I will freely & liberally attribute & cite whatever I glean as I continue viewing the 2020 fast.ai course (& other courses by others), and won’t claim that things in my course would be “first”…and I certainly won’t do any weird slander like the Lightning gang… but also just in case: Please don’t “come after me” if I don’t cite every feature that fast.ai also includes, as some of these ideas I’d already had “independently” or via other people (again, I won’t be making any claims of originality or precedence). And I’ll continue contributing to the community here – particularly, I’ll be encouraging the audio grad students to come to me, to help with
Also my starting my own little nbdev-based DL-teaching library for packaging together “things I like” will in no sense be intended to compete with or snub fast.ai; it’ll just be more of a separate workspace for me & my students, where I don’t have to wait for Pull Requests to pass fast.ai’s amazingly-crafted set of standards while I’m developing.
Looking forward to lessons 4 onward!
P.S.- Oh hey: The “first draft” version of my course will be available for free late this summer; “freely ye have received, freely give.” Again, not to compete with fast.ai—that would be impossible LOL—but if anyone would want to check it out, I’d be happy to share a link when it’s ready. There will be an autograder.