When I was training my model with a large dataset (1,200,000 x 1) on my mac, it took 19 hours for the model to finish 1 epoch.
I knew that google colab provided free access to their GPU, but I preferred using jupyter lab and saving notebooks on my local directory.
As I was searching for a way to run the code on a google colab server but save it on my laptop, I found out that google colab recently released a new vsc extension that allowed us to use its server through vsc.
I did have to setup a google drive folder and use vsc instead of jupyter lab, but come on, they’re letting us use their GPU from my laptop!
I created a drive with folder structure as similar as my local directory, then added some codes that connected my ipynb to their server.
import os
from google.colab import drive
drive.mount('/content/drive')
base_path = "/content/drive/MyDrive/Colab Notebooks/Quant"
os.chdir(base_path)Now when I try to train my model with the same dataset, it only takes 2 hours per epoch.
Model weights and checkpoints are all downloaded to Google drive instead of local directory, but I can simply download them when needed.
