Is there any difference in starting the additional training from the frozen graph vs. starting from the released checkpoints?
Is there any difference in starting the additional training from the frozen graph vs. starting from the released checkpoints?
Likely that @reuben or @kdavis might give more details, but as much as I can say, there should be none?
The only difference is that the Adam statistics are in the checkpoint, but not in the frozen model.
great, thanks for the clarification