Will the training process converge to a local minimum?

[yihsinchen]

Hi,
I am trying to use your concept for adapting a semantic segmentor (FCN).
But I found that after the error rate of the domain classifier began to increase,
the source domain test error rate began to increase, too.

Finally, the source domain test error rate increase to almost 100%. It seemed to be that the whole training process crashed.

Have you encountered similar phenomena?

Thank you so much for your reply!!