Original article can be found here (source): Deep Learning on Medium
10. Specify TensorFlow version
Very soon, the default TF version in Colab will be TF 2. What this means is that if you have Notebooks running in TF 1, they will probably fail. In order to ensure that your code doesn’t break, it’s recommended that you specify the TF version in all your Notebooks. This way, if the default version changes, your Notebooks still work.
9. Use TensorBoard
Google Colab provides support for TensorBoard by default. This is a great tool for visualizing the performance of your model. Use it!
8. Train TFLite Models on Colab
When building mobile machine learning models, you can take advantage of Colab’s resources to train your models. The alternative to this is training your model using other expensive cloud solutions or on your laptop, which might not have the needed compute power.
7. Use TPUs
Should you need more powerful processing for your model, then change your default runtime to the TPU. However, you should only use TPUs when you really need them, because their availability is on a limited basis, given how resource-intensive they are.
6. Use Local Runtimes
In the event that you have your own powerful hardware accelerator, then Colab allows you to connect to it. Just click on the connect drop down and select your preferred runtime.
5. Use Colab Scratchpad
Sometimes you might find yourself in a situation where you want to test things out quickly. Instead of creating a new Notebook that you don’t intend to save on your drive, you can use an empty Notebook that doesn’t save to your drive.