Platform: Paperspace and Gradient ✅

FYI we added support for GCP preemptible instances to our 1-click Jupyter Notebooks:

So, I had started the previous course on a Paperspace instance, but when I saw v3 announced I decided to re-start the course with the new version of the library. I’m following the Gradient set-up instructions.

When I try to open a terminal via the Notebook, the resulting terminal doesn’t have ‘normal’ terminal behavior. E.g., when I start typing a folder name and hit tab, it doesn’t autocomplete the folder – it gives me an actual tab space.

Is this typical? Is there any way to enable the behaviors I’m used to in my Mac terminal?

Also, with Gradient is there a way to transfer files between my machine and whatever I’ve got in the cloud? Say, I wanna run a model with my own data. How do I get the data into the Gradient environment to run on my own?

Or alternately, is there any way to set up the environment for v3 of the course in a VM instance so I can have all of the behaviors I’m used to? I couldn’t find instructions for that anywhere.

2 Likes

Also, what are people’s experiences with the “preemptible” sessions? Not that interruptions are regular and predictable like a train schedule, but about how often (besides the 24-hour maximum) do they happen?

Hi,

I am currently working on GPU+ on Paperspace. I have recently transitioned to v3, however I would like to retain the GPU+. I see for this course a separate repo was created course-v3/. Could I just clone this repo? Do I need to update any other libraries?

Thanks in advance!

Hi @dillon , I dediced to go through the Gradient approach using a P4000. However, I am getting an error
Error creating volume: Error response from daemon: create jsmksqrc7fmgp6: error while creating volume path ‘/var/lib/docker/volumes/jsmksqrc7fmgp6/_data’: mkdir /var/lib/docker/volumes/jsmksqrc7fmgp6: no space left on device.

It seems like the notebook is running, however cannot access it.

1 Like

Hey I got the same error, deleted the notebook and did the setup again but didn’t work:

2 Likes

I’m a linux/jupyter notebook beginner who is completely lost on the process of installing software like google_images_download to get datasets together for level 1. I followed this thread…
Tips for building large image datasets, but wget doesn’t work in the jupyter terminal.

  1. Where do I go to install chromium/chrome so that I can download images from the command line?
  2. How do I SSH into my gradient instance? I searched for this and found a thread, but it seems that paperspace has since changed the format.

Thanks in advance for the help.

@adric This terminal behavior is expected/normal for Jupyter :frowning: The terminal in Jupyter Lab, the newest version of Jupyter, is a bit better but it’s still not nearly as functional as say the default terminal on a Mac/Linux system.

Jupyter has an upload button and you can easily upload data to /storage, a persistent directory for storing data. For large datasets, you could either set up a VM and transfer files to /storage using SCP (or just use the VM itself). To clarify, both VMs and Gradient can access the /storage directory.

Hope that helps!

@crayoneater Great question. This article says “preemption rate varies between 5% and 15% per day…on a seven day average, occasionally spiking higher depending on time and zone”.

I definitely recommend checkpointing the model so if your session is interrupted, you can pick up where you left off/don’t lose your work.

1 Like

@dkobran Thanks! I’d kinda wanna just set up the VM and run that. But it’s not clear how to get the VM setup for the new course. From what I can tell, the public template for fast.ai installs the v0.7 for the old course – I’m a noob, so not quite sure how to update the libraries and notebooks for the new course. Any pointers?

@adric Totally understand. We did set up the new template ourselves so let me see if I can gather some info from our team. Be right back :slight_smile:

1 Like

@MadeUpMasters You can use wget in the Jupyter terminal – it’s just not available by default. Just run:

apt-get update
apt-get install wget
  1. If this ^ doesn’t work, you might just want to use a VM which has a desktop and Chromium installed by default. I would start with our ML in a Box template and just update the Fast.ai components. I’ll try to test this process for you.
  2. Need to look into this. I know you can SSH into a Job but not sure if this is possible (yet) with a Notebook.
1 Like

@Alican1 @Jeffrey Really sorry about this, we found a bug with the most recent Fast.ai container and it has been fixed. Happy to credit you for any lost time dealing with this.

I started an instance of a P4000 notebook and immediately stopped it. It has been in the process of “stopping” with a spinning circle for over an hour. What should I do? Am I being billed? Thank you.

No problem. Daniel it seems like the notebook backed instance never ran. Thanks!

I am trying to ssh into my paperspace gradient machine. But when I enter username and password on my local machine terminal, I get ‘Permission Denied’ error.

I need to download the model file to local PC so that I can upload it to my Google Drive for deploying on Web.

Please help.

For those interested, we updated the Fast.ai VM template with the latest course material. You can access it on the Public Templates tab on the machine provisioning page:

I’m working on getting a script together so you can update an existing machine (instead of creating a new one).
@MadeUpMasters @adric

1 Like

I created a new machine, pulled from git and updated fastai library.
I’m trying to reproduce lesson5 in a jupyter notebook. I’m on this line:

learn = Learner(data, Mnist_NN(), loss_func=loss_func, metrics=accuracy)

And I get the same error as this guy did:

Something to do with the loss_func not recognizing the nn.CrossEntropyLoss() function

Which GPU notebook should I choose?
Is the K80 preemptible a good choice?
With resnet50 I still get 20 min learning times where I see our tutor has 4mins. Sadly slows things down quite a bit.

@dkobran Can we get auto-shutdown options of less than 8 hours? It seems trivial to add and most of us are using this for working with the fast.ai coursework which generally is a few hours of work.

Also is there a reason you don’t offer auto-shutdown for gradient-0 customers? There could be a proper justification but it comes off like you’re basing your business model around people forgetting to shutdown their instances…

1 Like