If you’re trying to participate in a Kaggle Competition, or build machine learning models for your passion project, you’re likely to come across colab and/or Google Cloud Platform (GCP), among others. GCP offers a free introductory credit, sounds enticing? Which one is better for your passion project?
TLDR: Colab is better than GCP with free credit.
Here is the full story. I recently participated in a Kaggle Competition, that has a very large dataset (30+GB!). My team decided to give both a go and here’s what we found:
On Google Cloud free credit, you are quite limited in the types of instances you can use (so far, you can’t choose the ones with TPU). With the types of instances available, it can take a very long time to train and often run into memory error. In fact, I lost an entire instance (error 524) and had to create a new instance to start from scratch. In summary, with free credit, it can take a long time to train but no result is retrievable.
On the other hand, you can choose TPU in Colab and I trained the same and larger size model without issue. Granted, there are several OOM errors. But it is recoverable.
I would definitely go with Colab.