Alternative environment to AWS


(Rich) #1

Hi,

Currently working my way through the course (great stuff so far). I’ve reached a point where part of the homework was to ask a question. I don’t really have a question, however I’d like to share an alternative environment that I’ve been using.

The AWS environment works pretty well, and I was happy with it. However I have moved to Paperspace (https://www.paperspace.com), they have specific instances built for machine learning.

Their instances are faster (better GPU) than the AWS instance referred to in the course (about 50% faster) and cheaper $0.65/hr vs $0.90.

The ML-in-a-box has everything you need pre-installed to do this course.

I had to fix three things

  1. Jupyter notebook was not bound to the public network adapter (so I could access it from my browser)
  2. The notebook port needed adding to the firewall.
  3. Downgrade the version of Keras to 1.2.2 (seems to be some sort of bug in v2)

One final cool feature is the ‘Auto off after X hours’. This saves you from being charged if you accidentally forget to turn it off.

Cheers
Rich


Which steps would change if I use Papersace or Floydhub?
(yochanan scharf) #2

Just for USA.


(lateralplacket) #3

Does it allow monthly usage caps or something similar, to limit costs?

AWS and other mainstream cloud providers are a bit scary because of the potential for incurring unlimited usage costs if your account is hacked.


(Nikhil Dharap) #4

Are you using ubuntu or windows ?I dont see python installed :frowning:


(Nikhil Dharap) #5

Nevermind. They approved me for the wrong box.

For others who might go this route, if you sign up for the .65/hour plan, they will ask for a reason to provision this box. Mention your preference for a ML box.


(Nikhil Dharap) #6

Hi Rich, Would you mind helping me with adding the notebook port to the firewall. I am new to linux and can figure out how to do that. I was able to download the data into my paperspace machine, and was able to start jupyter notebook there.

I have a public IP setup but I cant access the remote jupyter notebook from my local machine using http://public_ip:8888

Thanks,
Nikhil


(vee) #7

try this
sudo ufw allow 8888


(Nikhil Dharap) #8

Thank you Vee.
I followed the steps here - http://amber-md.github.io/pytraj/latest/tutorials/remote_jupyter_notebook
and this worked for me.

Regards,
Nikhil


(Nikhil Dharap) #9

I dont see a cap, but there is a way to specify how long before the machine shutsdown automatically … I set it to 8 hours (there is an option to auto shutdown in an hour as well)

another advantage is that it bills per second you use so the hour is prorated …


(Daniel) #10

Hi there, Paperspace team here. We are not live in Europe (Amsterdam) and will continue to expand our global coverage. Hope that helps!


(Mike Moloch) #11

Hey There, any idea when the p5000 machines will be approved for fast.ai folks? I put in a request last week and was told you guys are backlogged but seems the backlog is quite large?

Any timelines on when one can expect a p5000 machine? currently it is locked for me.


(Daniel) #12

Sorry about the delay here, we should be finished approving everyone today. I think I found you in the system and approved you. If you don’t have an approval email, just DM me :slight_smile:


(Daniel) #13

Thanks for the feedback @xedarius! (Paperspace team here)

A couple people have asked me about storage options on and I wanted to let people know that we offer something called Drives. If you need a simple storage solution that you can mount on any of your VMs, this should be a great option. This feature is only available for Paperspace for Teams so it’s not obvious at all (sorry about that!). Anyway, Teams is totally free so just ping me if you need this and I’ll unlock it for you.


(Mike Moloch) #14

Thanks Daniel, I see it in my options now! much appreciated!

-M


(Astha Garg) #15

Hi,

I followed through this link to set up the jupyter notebook.: https://by-the-water.github.io/posts/2017/05/16/setting-up-a-jupyter-notebook-server-on-paperspace.html