**1. Colab**
You might want to only use Colab to develop and debug your code, and use Sherlock or the Jupyterhub cluster to fully train the model, as you can easily crash your Colab RAM with a full dataset and a large model.

In Colab, in the "Runtime" panel, you can go into "Change runtime type" to choose CPU, GPU, or TPU. With a GPU (T4 provided by Colab), the acceleration of training FNO may easily reach 50+ times speedups compared to CPU. It will be highly recommended to run with a GPU for this task. But make sure that if you choose GPU, you shut down the runtime when finishing the program to avoid wasting limited runtime resources.

**2. Jupyterhub cluster**
You can train for up to 10 hours at a time, as long as you can **keep your browser window open and the computer in non-sleep mode throughout the training periods**. It's not ideal, but it's similar to running code on your own laptop, just with a faster remote GPU. Use it with the following precautions:

- Prevent Your Computer from Sleeping:
You can adjust this in your system settings. On Mac, apps like KeepingYouAwake or Caffeine allow easy toggling to keep your device awake. For overnight jobs, make sure your computer is plugged in and screen brightness is turned down.

- Use Checkpoints:
Consider saving your model's state at regular intervals. Implement logic to automatically resume from the latest checkpoint to reduce loss of progress after any unexpected disconnection.

- Use a Secondary Browser:
To avoid accidentally closing your training session, consider running JupyterHub in a browser you don’t use for regular browsing (e.g., Safari if you normally use Chrome).

[This document](https://docs.google.com/document/d/10akwR9tztA04XMvKKqC7cPHiw8Fsl7j9iGs8p2FQylA/edit?usp=sharing) containing detailed instructions on accessing and using the compute interface. Accounts have been created for each of you, and we have added the username and password in [this spreadsheet](https://docs.google.com/spreadsheets/d/1NwMOehALXiTOO2UMynRGSOw89H80TtP4BkNbhIS2pdk/edit?gid=289040126#gid=289040126). 

**3. Sherlock cluster**
PhD students can ask their advisor to request Sherlock account for them. If you already have an Sherlock account, this is the best option as the GPUs are fast and it will not interrupt the local tasks on your laptop.