Tensorflow Object Detection - mAP steady, but Total Loss is dropping. Keep training?

Question

Training an Object Detection classifier with Tensorflow 1.12.2 and Python 3.6. Using about 20,000 images over 4 classes, and after 160K steps the mAP is pretty steady, but the Total Loss is still dropping (very slowly though). Should I keep training until the Total Loss ceases to drop or turns upward ?

Answer 1

When Total Loss turns upward, you have most likely went into the overfitting.

A good idea would be to check your current snapshot against a test data set and see if you get satisfactory results. If the results are good, I'd stop training.