Train, check, test
Before we trust a model, it needs a fair test. Where do the test questions come from?
Step 1 of 5
- 200 labelled photos in one pile.
- Split it in three: 140, 30 and 30.
- Train on the 140. The model learns from these.
- Check on the first 30. Use them to improve the model.
- Test on the last 30. The model has never seen them.
They come from photos kept aside and hidden from training.
A dataset is usually split into three parts. The training set is what the model studies. The validation set, or check set, helps you compare model versions. The test set gives the last fair score.
The model saw the training photos, so they cannot test it fairly.
The sizes are a choice. Here they are 70%, 15% and 15%.
Train on one part. Improve with a second. Test on a third it never saw.
Choose the split for each task.
Loading lab…