When I run the training using a “reasonable” net (based on earlier labs and lectures), each epoch takes a few minutes and there are 15 epochs. This is fine, but when I stepped away from the computer it self canceled halfway through the training. How can I make it so that I don’t have to babysit Colab?
If you are not using GPU as the runtime environment, go to Runtime > Change runtime type > Notebook settings and set the Hardware accelerator to GPU.
Google colab offers pro account with a pay as you go option. If that option doesn’t work, you’ll have to rent a cloud instance or solve the problem on your personal desktop / laptop.
Thanks Balaji. I checked that it is using GPUs and it seems to be doing so by default. My question is twofold:
- Is there a way to make it go faster? (Seems like the answer is “pay additional money”) and
- Is there a way to make it so I don’t need to babysit it? This means moving my mouse around on the webpage so it knows I’m “there.” I failed to do this for just
a bit too long on one training, and it quit after finishing 14.4 of 15 epochs which had taken 45 minutes already. This course is generally efficient and pro-education, so I perceive this UI behavior as running contrary to the spirit of the course.
- All things remaining the same (i.e. python environment and network architecture), the paid plan is likely to speed things up. The reason I say likely is because I’m unsure if GPU allocation is random or if there’s a guaranteed speedup.
- Try:
a. NN with the same number of layers but with fewer parameters.
b. I’ve heard of Mouse Jiggler. Not sure if it’s the answer in keeping the colab session active.