Pretraining of the model

[hyukiggle]

I was trying to reproducing the equivalent result from the original paper.
Though it is natural to pretrain encoder of model on ImageNet, the decoder is task-specifically trained.
As of I know, since ImageNet pretraining is classification, it is difficult to train decoder.
I'd like to know how the pretrained decoder weight is initialized or the method of decoder's pretraining on ImageNet.
Thank you

[manupillai308]

@HuCaoFighting