Paperspace Usage Information

Preliminary

You should have been provided with Paperspace credits - these will allow you to spawn paid instances with better GPUs than what colab or paperspace offer on their free-tier. Additionally, unless you enable auto-shutdown, these instances will stay awake permanently, and not abort half-way through training.

Instance Setup

Once you have signed up for Paperspace's gradient platform and applied your coupon:

1. Go to console.paperspace.com and create a under the "ICL Deep Learning 2023" team (selected at top left)
2. Create a notebook within that project. We recommend using the PyTorch 1.10 GPU runtime
3. Select a machine - you can use the free-tier CPU or GPU machines for most of the coursework questions, but when training your models we recommend using a paid P4000 instance. Choose your auto-shutdown period carefully, base on the network's training time, and whether you will need to go AFK
4. You can directly connect the instance to your gitlab repository by
   a. Going to Advanced Options and pasting your gitlab link into the workspace field (https://gitlab.doc.ic.ac.uk/<course>/<your_repo>.git) the .git is needed - the easiest way to get this link is to copy the HTTPS clone link from gitlab.
   b. Entering your doc username (e.g. afs219 - not your email!) and password into the workspace username and password information
   Note it seems there is a bug for passwords containing special characters (% & etc.) - the password field on paperspace will not properly pass these to the gitlab login, and you will get "Access denied". The current fix is to change your gitlab password.)

Using the Paperspace Gradient Environment

Once your instance has spun-up, you will be greeted by a .ipynb similar to Google Colab. You can choose to work here, or press the Jupyter icon to the side () to open JupyterLab.

If you are going to work in the web, I would strongly recommend using JupyterLab over their notebook environment. However, the best development experience is provided by working with VSCode and connecting to the remote kernel.

Connecting VSCode to Gradient

In order to make this work you will need the Jupyter extension installed on VSCode. You should also git clone your repo onto whichever machine you'll be working from.

Assuming you have opened the repository within VScode, and have the dl_cw_X.ipynb file open you can then follow the steps in https://docs.paperspace.com/gradient/notebooks/notebooks-remote-kernel/