Issues about training pipeline

[Cuirundi]

Should image reconstruction and depth estimation be carried out alternately for each batch during training,such as doing depth estimation for one batch and image reconstruction for the next batch.Or should the two tasks be carried out separately in two unet processes for one batch?

[haodong2000]

Hi @Cuirundi , thanks for your interests!

During training, the image reconstruction and dense estimation are carried out in adjacent and continuous batches. There is only one UNet.

[starsong98]

So, does it mean that, with the same UNet, you switch back and forth between image reconstruction and dense estimation for different batches?

e.g., batch 1 trains with image reconstruction, batch 2 trains with dense prediction, batch 3 trains with image reconstruction again, and so on.