Hacker News Reader

Comment by metafunctor | Hacker News Reader

Part 1 was great.

However, the first lesson took a bit of stamina to go through. Much of it was introducing basic Unix/AWS/shell/Python things I know intimately and have strong opinions and deeply set ways about. Shell aliases, how to use AWS, what Python distribution to run, running Python from some crazy web tool called notebooks (and not Emacs), etc. felt like I was forced to learn a random selection of randomly flavored tools for no good reason.

Yes, it's a random selection of tools. The good reason to bear them is that you'll learn how to implement state of the art deep learning solutions for a lot of common problems.

So, I ended up viewing the lessons not as "this is how you should do it", but rather as "here's one way to do it". And it does get much easier after internalizing the tools in Lesson 1.

Just something to keep in mind when branding this as "deep learning for coders". Coders have deep opinions about the tools they use :)

View on HN

You're partially right. But I would like to mention that in many ways, their way is the "standard" way to do things.

Just as an example, the "crazy web tool called notebooks" is pretty much the standard way of working in many different areas of machine learning/deep learning (and others). It doesn't mean you necessarily have to do it yourself, but tons of material out there will be in this format, so it's valuable in and of itself to know it.

Also, you really should give notebooks a chance, they're a game-changer for productivity, IMO. (Although I'm a vim user, not emacs, so maybe you shouldn't listen to me ;).

Also the python distribution in question (anaconda) is becoming standard for data science - pytorch, for instance, (covered in part 2) assumes you're using anaconda for its default install method.

I come from the opposite side of the spectrum. I know nothing about UNIX/AWS/Shell and have no intimate or strong opinions on any of the tools. I am confused by basics such as how to make a special request for P2 access, since the set up video is out of date. The git cloning recommendations on the wiki is also confusing to me. I'd rather see a video on how to do the git cloning or step-by-step instructions (with pictures) on how to do this and what to expect if it has been properly done.

I felt like some of the material is outdated and I find it not clear how to get around it. I ended up getting frustrated with the set up videos and git cloning instructions, so I skipped it in their entirety. I hope the future iterations of this class spell things out for beginners like me. Otherwise, I can't even begin the class since I don't know how to request access to P2, clone git, etc.

I am not uneducated. I come from a pure mathematics background and don't know anything about this set up business. I can code, know theoretical CS, but when it comes to setting up the tools (along with the outdated material) I am utterly lost.

At this point I'm just watching the videos. I can't actually do any of the coding stuff since I don't have the tools set up, but I like their top-down approach and am learning a lot despite these obstacles.

We (including the rest of the learning community) would love to help you - could you post on the forums the problems you are having and we'll try to help sort them out? If you've already posted, please at-mention me there so I don't miss it.

It's a fine line between helping too much vs too little in lesson 1! For the next time we run part 1, we're thinking we'll have an optional weekend workshop to teach the necessary (non deep learning specific) software pieces - python, numpy, AWS, shell, etc. That would end up being a separate mini-MOOC I suspect.

Based on the discussions on the forums it seems most students are between these two extremes, and largely follow the video advice, branching off sometimes where they need to do some additional research, or are already familiar with some alternative approach.

Edit: actually thinking more about it - probably your best bet is to simply use http://crestle.com . All the data, notebooks, and software is pre-installed, so you can start coding right away.

Thanks for the reply. I had not (yet) posted on the forum as I was methodically searching through the previous replies to ensure my question had not already been answered. Thanks for crestle recommendation. I will now look into it. This should simplify my tool setup. If I still have questions around tool setup after posting, then I will at-mention you. I think the mini-MOOC (for people like me) is a great idea. I get lost in basic tooling (that many experienced devs can skip). So, if you guys include a mini-MOOC on tool set up, then this would benefit people like me tremendously and help us get up to speed with everyone else.

Will the revamped course in October be offered online (like the current version is) for public viewing?

Yes it'll be online a couple of months after completion, or you can apply to join remotely as an international fellow.

Excellent. Thank you.

I'm doing part 1 at the moment, and the outdated AWS setup instructions on the site frustrated me a bit as well.

Maybe an idea would be to leverage the community a bit more? Make it easier to edit the wiki for MOOC-students which are the ones watching/reading the material later on. And maybe refer to that wiki more on the site.

I end up using the forums a lot for trying to find info on stuff that is not working, or I don't understand. But it's not an easy platform to find information since it isn't structured well.

It's been a while since I watched the course, but IIRC the setup video had an overlay that said to simply put "fast.ai MOOC" (or something similar) into the P2 request form. Worked perfectly for me. Did that not show up for you?

Yeah, that showed up for me. It was more about how to use the setup scripts, they had changed a bit it seemed.

I thought if the resources around the course evolve over time, then the material should as well. It would be easier to keep a wiki up-to-date and refer to that, instead of a video (and site). For setup and "sidenotes" specifically, I'm not saying the lessons should be re-created in the wikis.

Although the homework and lesson notes in the wikis was indispensable.

If I'm understanding EC2 pricing, this looks a lot cheaper. What's the downside of using crestle instead of AWS for the course?

(I built Crestle.)

The downside is you don't get as much control over your environment (you get a terminal in the browser, but no SSH or sudo access, for example). If you have the inclination and ability to manage your own instances, EC2 is more flexible.

Just wondering: what kind of infrastructure are you running on? If I were to build something like this I'd probably run it atop of EC2 or GCE, so pricing would never be lower than them.

It's running on EC2 and uses spot pricing which is much lower than $0.90/hour.

If people aren't fully utilizing their instances, you can offer lower prices since you can have more than one user per underlying instance.

Not sure that's how he does it, but it's one way.

This looks great, can you give any idea of what your invite request queue looks like/how long I might be waiting?

You'll get an invite as soon as you sign up.

Most of the time I also ended up just watching the videos, picked up the concepts I didn't yet know about, and experimented on them in my own way. I did perform all the Kaggle submissions, though, to kind of calibrate my level.

I wonder how big the crowd between our two extremes is — the people who actually do run the commands exactly as explained in the videos? If the material is outdated, beginners cannot. Experts will do their own thing. How many people will follow the actual instructions?

I started watching the videos in Part 1 and was disappointed to find out that I was going to have to create an AWS account. I really don't want to deal with accidentally racking up a large bill on AWS.

For GPU instance, I would suggest https://www.hetzner.com/dedicated-rootserver/ex51-ssd-gpu?co....

It is only 99 euro/month.

The 99 euro setup fee is a dealbreaker if you'd want to use it just for some courses.

If your data is in S3 then data transfer fees can suck but def cheaper than $ 0.90 per hour if you're regularly leaving models to train overnight :)

If you have the hardware, you can do without an AWS account (you might need to be an "expert" to interpret the instructions to run on your own hardware instead of AWS). I used my gaming rig with a 1080 GTX Ti card.

I've been trying to get it up and running using the latest versions of the libs. I have theanos finding my GPU now and running basic cuda code but cannot get the cuDNN libraries to be recognized even after installing a few versions of the Nvidia site. Ugh. Giving up for now.

Why? Why can't I just do it with my crappy hardware, and it take longer?

You can do it on a poorer graphics card, but that's only if it has enough GPU RAM, and only if its supports a relatively recent compute version.

You _can_ run it on a CPU, and that can work fine for very simple models (like the Tensorflow demos with MNIST) but anything much larger and things really start to take forever. Like, you'll be waiting for months.

I'm giving serious thought to buying or building a rig for these courses. What's "enough GPU RAM"?

A GTX 1060 with 6 GB of RAM is a good intro card. You can find them for < $350.

The AWS P2 instances recommended for the course ($0.90/hour) have half a K80, which has access to 12 GB of VRAM. The list price for that card is $5000.

The GTX 1080 Ti has almost as much VRAM as half a K80, 11GB, and can be found for around $700.

I'm finding Acer rigs with GTX 1070, 8GB RAM. I'll probably get one of those, thanks.

You can. Tensorflow has a CPU implementation. But, it's not a little longer, it's incredibly longer.

Like 5h on an upper-class GPU vs 50 days on the fastest CPU.

Can this be done in Windows or is it better to install a Linux?

Better to install linux. Not all tools support Windows, and for those that do it is still a second class citizen and kind of annoying, e.g. look at installing CUDA/cuDNN on Windows vs Linux.

I agree that the focus on specific tooling was tricky. That said, I really do like the notebook environment. It's mostly the AWS setup that I found a bit fragile.

Completely uninformed and unlearned question: what exactly is the role of AWS in deep learning of this nature? Why do we need to use it?

Pay-as-you-go GPU instances.

Reply on news.ycombinator.com