Should I custom split my image data?

[Ayoola_Fakoya]

Even with auto Ml, should carefully custom split my data to my satisfaction or just leave it to AutoML?

And what difference does it make?

[tuba · Accepted solution]

Hi Ayoola

If your data is large enough and have wide representation of each category, you may go with the automated split in AutoML. That would save time and perform well.

If you have some specific needs, such as the representation of certain observations in a specific category is important and limited within the data, you may want to make sure that it is well distributed for validation and test. And custom split would help for that. Another reason of using custom split could be for comparison of your model performance with external models so you use exactly the same training/test datasets and make an apples to apples comparison.

Here are some tips I find useful in this doc:

https://cloud.google.com/vision/automl/docs/beginners-guide#distribute_examples_equally_across_categ...

Cheers

Tuba.

[Ayoola_Fakoya]

Thanks so much for the response.