looking for some solutions? You are welcome.

SOLVED: Tensorflow: Custom trained model mislabels pre-trained classes


I am trying to build a hand detector using machine learning. For a start, I followed a tutorial and tried to train the Faster RCNN model with Tensorflow on a large hands dataset created by others. However, I did not train it from scratch but used a version of the model which is pre-trained on the coco dataset.

Now, having trained the model for some time, it does detect hands pretty well. Unfortunately, it also detects all other 90 coco dataset classes and labels them as hand. While it is no surprise that all other detected objects are labled as hands since the label file contained only this one label, I do need to find a way to make the model stop detecting those 90 classes it was pre-traiend to detect.

Everybody completing the tutorial which I followed seemed to be facing the same problem, but I could not read about any working solution in the comment section. Apparantly, changing the

fine_tune_checkpoint: "faster_rcnn_nas_coco_2018_01_28/model.ckpt"
from_detection_checkpoint: true

part in the .config file to "false" solves the problem, but also means to train from scratch (not sure I got this right). So, if anyone knows how to make the model forget about the pre-trained classes while being able to use transfer learning, I would be super grateful.

Posted in S.E.F
via StackOverflow & StackExchange Atomic Web Robots

No comments: