Hi @diathesis, you did nothing wrong. Since, it is free, they only have limited resources. You can try again later. 
Cool ā understandable, just wanted to make sure that was the reason, so if I check later and itās still not available, I know to just keep trying or decide if i prefer to pay, but not still wonder if Iām ādoing it wrongā. Thanks for letting me know!
I find that my free Paperspace Gradient notebook is not usable. E.g. I have currently been waiting more than 1.5 hours for it to start: https://imgur.com/a/fKI0pUK. Waited 20 mins yesterday. Should I try a paid instance, or move to GCP?
Waiting 20 min for a notebook to be provisioned is not unusual.
But yesterday was unusual.
They were down for most of the day yesterday, this was the first time I have experience that on their site (with my spotty use, and subscription for which I pay $8/mnth).
Hi,
Iāve been taking the fast.ai course using a free tier instance on paper space. Now I have problems with running out of memory, and Iād like to upgrade to a GPU. But Iād also like to keep my existing notebook, as Iāve written in answers etc.
Is there a way I can start a notebook with a different machine instance? Or start a new instance and get my files copied over?
Iām trying to finish following the basic steps but Iām having a problem connecting to the terminal. This is what my screen looks like. Anyone know how I can connect to the terminal in Step 3?
Capture|690x333
If I understood correctly, then you should simply enter the instance via the āOpen V2 (Beta)ā option which starts a Jupyter Lab environment where youāll have the terminal to use. I hope this works for you.
Yup - got it to work. Thanks!
1 Like
Hello, Iām a new one in here.
I just wonder what I did is correct or not.
I already make a container guided by this page. And the machine is already running.
Is there something Iāve missed before?
I was following the instructions, but wasnāt able to get the code running (I tried both free and paid instance).
I can run the notebook, but the UI is different, thereās no top black header from the Gradient service (with Stop and Share buttons on the right) as it is in the guide, I get just pure Jupyter notebook.
After running the notebook, run the update command and git pull. I can run first few cells, but at the cell under Running Your First Notebook, shows this error
PIL.UnidentifiedImageError: cannot identify image file '/storage/data/oxford-iiit-pet/images/american_bulldog_95.jpg'
(thatās just last line of the stacktrace)
EDIT: It looks like the next cells run ok, despite this exception
Is it just me or does it take forever for a notebook on paperspace to open up? I typically have to wait 10 -15 mins when it used to be instantaneous.
I have the exact same problem. Sometimes it takes 30 mins to provision the server. Did you find a workaround? I might just switch to Colab.
See Paperspace start up extremely slow
Basically, make sure to clean your /notebooks folder before shutdown to reduce shutdown/spin-up time
Iāve switched over to JarvisLabs. Fast spin up.
1 Like
Unfortunately, non-free - their cheaper GPU is the RTX5000, at $0.49/hr. The moment youāre considering a non-free option, itās very hard to beat the AWS spot price of their g4dn.xlarge instances (T4 GPU) - thatās about 66% the speed of an RTX5000 for training, at $0.16/hr, so basically JarvisLabs is twice the cost-per-epoch. Oh, and once youāve configured your AMI, spin-up is instant.
*correction - JarvisLabs offer a 20% discount for FastAI on the RTX5000 (thanks guys!) - so theyāre only $0.39/hr. Thatās still 50% more per hour than AWS spot, but considerably cheaper than AWS on-demand.
To be clear: if you want free, do google colab or Paperspace M4000 Free-GPU. If you want to pay money, be very careful on the math by costing epochs-per-dollar. You can find an extensive list of GPUs and their relative training times for all sorts of workloads here: https://ai-benchmark.com/ranking_cpus_and_gpus_detailed.html and the cheapest global spot-price of all AWS instance types here: https://simonpbriggs.co.uk/amazonec2/
1 Like
In my experience spot instances are definitely difficult to get started with. Salamander did a great job in automating the entire setup on spot instances but later faced a lot of challenges during peak hours. There were times when you lose the instances due to the nature of Spot instances.
Definitely correct - expect to spend a day of your life learning how to set up EC2 spot instances and also EFS for persistent storage between your multiple EC2 instances (the reason youāre considering paying anything is because youāre needing clusters of GPUs, right?). But given that youāre going to be using multiple machines in parallel, that cost difference is going to more than pay for itself.
RE losing spot instances - that used to be a problem until 2018, when the spot pricing mechanism changed. Now just set your willingness to pay as higher than the spot price, and itāll be someone else who loses their machine instead (just setting it at the on-demand price practically guarantees that you donāt get kicked off, yet you pay only the spot price as it slowly fluctuates).
Iām just starting setup, and am slightly confused. I used the default notebook (Paperspace + Fast.AI). My root on Jupyter doesnāt show coursev4 or fastbook. I have three folders (clean, images, tools) and then the course chapters starting with 01_intro. Did I miss a step?
2 Likes
I just began Chapter 3 Ethics and this is paired with the Deep Learning for Coders with fastai & Pytorch book. Maybe you setup your Notebook for the wrong course?
Having the same problem after I updated my setup and created a new paperspace notebook! There is no more persistent storage (used to be /storage folder). Anyone know what is happening?
