Closed kwea123 closed 5 years ago
Only the encoder is initialized with ImageNet pretrained weights, while the decoder is randomly initialized. I hope to train the randomly initialized decoder for one epoch first, then the whole network. I think this could produce more robust model.
Oh, I misunderstood that as "training the decoder only for the first epoch, and stop training it from the second epoch and on." Now it's clear and good.
https://github.com/xy-guo/Learning-Monocular-Depth-by-Stereo/blob/3547a9e6f081177114e997e5986b2b546cee6109/main_distill_mono.py#L104
Here you train the decoder only for the first epoch. Why?