Retraining a NeuralTools Net Gives Different Results with the Same Data

Applies to: NeuralTools 5.x and Up

Question:
I trained a net and saved the training report.

Later, I changed some settings and retrained, then reverted the settings back to the originals and retrained again. Even though the data are identical, the results are different. Why?

Answer:
By default, NeuralTools selects a different testing subset each time you train a network. Because the training and testing cases vary between runs, the results will not be identical—even when the data and settings are the same.

This effect is more noticeable when you have fewer cases, since small changes in the testing subset can have a larger impact on results.

---

How to Get Repeatable Results

If you want consistent results across training runs, you can freeze the testing subset:

1. Open the Training dialog.

2. Go to the Train tab.

3. In the middle of the dialog, enable:
   “Select Same Cases as Long as This Number is the Same.”

As long as this value remains unchanged, NeuralTools will use the same testing cases each time, producing repeatable training results.

 

Related Article:

Recommended Amount of Data