Train/test split calculator

How to use it

1. Enter the number of samples.
2. Enter the test and validation shares in percent (0 for no validation set).
3. Enter the number of folds and the batch size.
4. Read the set sizes, the fold sizes and the batches per epoch.

Formulas

• Test = ⌈N × test share⌉, validation = ⌈N × validation share⌉, training = N − test − validation
• Cross-validation pool P = N − test
• Fold sizes ⌊P ÷ k⌋ and ⌈P ÷ k⌉; P mod k folds have the larger size
• Training per round = P − fold size
• Batches per epoch = ⌈training ÷ batch size⌉; the last batch holds the remainder

Worked example

1000 samples, 20 % test, 10 % validation, 5 folds, batch size 32:

• test 200, validation 100, training 700;
• cross-validation pool: 1000 − 200 = 800, so each of the 5 folds has 160 samples and each round trains on 640;
• 700 ÷ 32 = 21.875, so an epoch takes 22 batches and the last one holds 700 − 21 × 32 = 28 samples.

Tips

• Shuffle before splitting, and for classification use a stratified split so every set keeps the class proportions.
• Split by group when samples are related — several rows per patient or user — or the test score will be too optimistic.
• If the last batch is tiny, consider drop_last or a batch size that divides the training set more evenly.