Loss for train_from_scratch and continue_train

[DeckerDai]

Hey, I wonder what's the loss level when you train from scratch?

For me I followed the instruction in README to format KITTI data. When training from scratch the loss is around 0.77 to 0.83, while when training from the provided kitti_depth_model the loss is also at a similar level of 0.6 to 0.7.

It seems that there is no big difference in loss even after 150K iterations. Is this true? Or there are something wrong from my side?

Thank you very much!
Best,