Are Feature Extractors Trained in the Tensorflow Object Detection API?

Question

In the Tensorflow Object Detection API, a typical neural network will have 2 components: A feature extractor and then a neural network which uses the output of the feature extract to further classify an image.

If you were to custom train a faster resnet50 neural network in the object detection api, to detect an extra 2 objects, during the training process is the feature extractor also trained? I.E. does the weights of the feature extractor component change?

Answer 1

Yes it is trained by default. You can turn that off using the freeze_variables field in your train config. However, TF people discourage to do so, as they find it yields worse results and the same training time. See this thread on github for more details.