how to train

[Ericchen328]

I have my own dataset and want to train this model on my own dataset but i didn‘t see the detail of how to train

[lpiccinelli-eth]

We have not released the training code, yet, and we are trying to release it soon.

In the meantime, you can find the losses under unidepth/ops/losses.py (SILog as depth loss, MSE as camera prediction loss and SelfCons as invariance loss) and UniDepth as a single input-output block, and you can use your own dataloading and training pipeline. We will push changes such that the forward method is like an input-output function that can be used for further training.

Remember that UniDepth expects the input shape as 462x616.

[CDchenlin]

Thank you for your great work. I wonder if you have a timeline for the training code release?

[lpiccinelli-eth]

The code for training has been made public in the PR #76 , although still not 100% sure it will run smoothly, you can try it out.