Retraining the model with your own examples

You have trained an image classifier that manages your examples but misses certain cases — the scissors when they are closed, say.

There are two ways of fixing that:

1. Change the model (larger, longer training) — it rarely helps.
2. Change the data — add 15 pictures of closed scissors and retrain.

Nearly always it is the data. The model can only do what it has seen examples of.

Do it properly — otherwise you do not know whether it helped:

1. Measure before. Run your test set (which the model has not been trained on). Write it down: 26/30 = 87 %.
2. Add examples of what is failing — and only that.
3. Retrain.
4. Measure after on the same test set: 29/30 = 97 %.
5. Check that nothing else got worse. Did the pencils go from 10/10 to 7/10? Then you have traded one problem for another.

	before	after
scissors	6/10	9/10
pencil	10/10	10/10
eraser	10/10	10/10

This table is the whole point: one figure hides what happened, three figures show it.