I use it all the time without any problem.
tmux
jupyter notebook --no-browser
Once you return to your machine, you reconnect to your previous session:
tmux attach #0
update tmux attach-session -t 0
where 0 is the session_id from tmux list-sessions
I have a bunch of useful shortcuts in my snippets repo
Thank you very much. I wasn’t sure whether tmux should be run local or remote side.
So basically after following your steps, I turn my local PC back on, ssh back in to the remote server and then run tmux attach #0 (as an example) to rejoin that session.
How about rejoining/reloading the output from the Jupyter notebook. Should that be automatically updated when I browse to notebook url?
You’re welcome. That’s exactly right!
Once you do more with Tmux the repo contains some snippets for multiple windows, zooming, …
This used to work in Jupyter Notebooks. When you reconnected you could just see the progress. However, in the latest versions of Jupyter I see that the updates are no longer coming through.
The thing is:
Additionally, I just looked at my notes from v1 where I explained all the steps of what you are doing. (or at least most of them 
)
In one of the slides / page from the notebook called “Sylvain says”, it said that book.fast.ai website contains links to recommended tutorials for each chapter. Are these links available currently or will they be available only in July? If yes, where can I find them. I did not see them on fastbook Github
If I had to guess (going by the names) it’ll be on the fastai2 library.
Oh no, it mentioned python tutorials / tricks. For example, the Path library instead of using strings. So more such things.
It will only be available by July:
We’ll be adding more information to this site as the official book release in July 2020 gets closer.
from book.fast.ai
For math equations, yes they do.
But no support for inline equations, only standalone chunks.
@Lankinen I’ve also been experimenting with Notion. It’s super clean but I feel like it can’t beat the flexibility of Jupyter notebooks. I’d rather spend more time making Jupyter look clean than adopting Notion because of it’s inherent limitations.
That’s good point. Jupyer notebook could beat notion in many things if not all but the biggest problem is that it doesn’t look as good and I really like the structure of notion where I connect notes to other notes. I need to test the fastpages thing to see what it looks like and then maybe contribute some ideas.
That’s true, it’s tough to beat the clean look and the ability to link to pages is pretty handy (though I don’t care of it too much since I don’t use it). Curious to know what you find about about fastpages, I’ve never used it yet.
FWIW, Jupyter can be made to look a lot cleaner by customising the custom css file. Here’s what mine looks like:
@rsomani95 that looks great! Can you share the CSS file and instructions on how to achieve this?
There is one freeze step (by default) before the train epoch:
Learner.fine_tune(epochs, base_lr=0.002, freeze_epochs=1, lr_mult=100, pct_start=0.3, div=5.0, **kwargs)
Sure.
jupyter-themes by entering pip install jupyterthemes in your terminaljt -t chesterish (doesn’t matter which one you activate, because you’ll be replacing the css file)cd ~/.jupyter/custom/
custom.css, delete all the text and replace with the text from the custom.css here. It’s crucial that you don’t replace the file itself, but open it and replace its contents. I don’t know why but that’s the only way it works for me.fonts folder from here in ~/.jupyter/custom/
PS: My theme is modded from this repo, and the one area that needs some improvement is the font size of code inside headings, but overall it’s pretty functional.
I think you mean there is one fit_one_cycle by default ie freeze_epoch=1🙂
Do you mind if we re-post this article on our blog, with you listed as the author? I think it’ll get more distribution from there, than as a google doc.
The course was great from the very beginning four years ago, but it gets better and better every year. And so does the library too!
But @jeremy I hope I won’t appear disrespectful if I state my opinion about a minor aspect. Saying that one does not need university education to so something in general and DL in particular, imho, correct and commendable. Much could be said about that, but let’s just say that traditional university education cuts the society in two, those who can afford it and those who cannot, while at the same time failing to guarantee quality of education (Nicholas N. Taleb wrote a well-known essay about that, “The intellectual yet idiot”), not to mention that a lot of brilliant people don’t even get accepted. A. Einstein, just to cite an example, was mediocre at school.
A different thing, however, is stressing that one does not need math to do DL (again, in particular… Or, in general something else). I find this a bit less commendable, if you allow me.
And that’s because of two main reasons: firstly, and more generally, it sends the message (although I’m sure it’s not your intention) that math is somewhat useless and redundant if you want to make something useful. I could not disagree more with such point of view. Without being verbose, let’s jsut say that mathematics shapes the mind of people for greater good. It makes you understand the world around you better. It makes you listen to your mind more than your guts, even in everyday life. It bestows upon you the capacity of (more) abstract thought, which is arguably a defining tract of intelligence. In fewer words, it makes you a better person.
Second, I’m not so sure it won’t be useful in deep learning. It’s not just about really understanding how backprop works… It’s the ability of reading research papers that is of utmost importance, as stated many times by yourself. So imho it’s ok to do a course with fewer math than usual, but one should be encouraging towards studying it. Let us remember that A.E. said once that “Everything should be made as simple as possible, but no simpler ”.