Validation and test set?

Do I need to have both validation and test set when my data is small (312 samples).
I’m googling a lot but still can’t find a clear answer.

Hello! Chapter 1 of the Fastai book is discussing in detail your question in the section " Validation Sets and Test Sets".

Quote from the book from that specific section:

The discipline of the test set helps us keep ourselves intellectually honest. That doesn’t mean we always need a separate test set—if you have very little data, you may need to just have a validation set—but generally it’s best to use one if at all possible.

But I guess using your judgment and some tests will help you decide what is best for this particular project :slightly_smiling_face:


I’m trying to impliment the result of a paper I read.
Can you explain why both valid and test set included in optimisation routine and cross validation ?

hi @HaoTieu,

I see @amalia has been very generous with her time to answer you, and advise where you will likely find the information you are seeking. I see that you replied only 20 minutes after @amalia and that your chart is not part of Chapter 1 and you don’t indicate you read the chapter, so I will ask you directly… did you read Chapter 1 as requested?

Looking forward to your questions from Chapter 1.
Also please read this.

btw, another really great resource is the 2022 Fastai Course Videos.

I completed chapter 1. As far as I know, test data is a seperate set for confirm the result of final model, but when you look at the image, test data also included in the dash box. This is what I’m confused.

Short answer is: if you train your model too often you can over-fit the validation set, so a separate test set is a guard against that. Jeremy explains this at 45:11 in Lesson 4.

1 Like