Why the size of validation dataset is 5000 rather than 10,000

[tjmoon0104]

Why the size of validation dataset is 5000 rather than 10,000

[yangtao2019yt]

I guess the author remove the original test set (10,000 images). Then cut the original validation set (10,000 images) in half and get two subsets as a new test(5000 images) and validation(5000 images), respectively. The details are in val_format.py.

I also don't understand why only using half of the original validation set as the new test for verifying accuracy in test_model.py.

Could the author offer an explanation?

[tjmoon0104]

Original dataset test set does not include labels. Which is not usable for experiment purpose.
So I simply divided the validation set (10,000 images) into half 5,000 for validation and 5,000 for test (with labels).

Validation set was used during training process to calculate validation accuracy and loss.
At the very end I use the test set to calculate test accuracy.