search

prs-eth / Marigold

[CVPR 2024 - Oral, Best Paper Award Candidate] Marigold: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation
https://marigoldmonodepth.github.io
Apache License 2.0
2.36k stars 132 forks source link

Other training metric are abnormal #130

Closed zad37 closed 1 week ago

zad37 commented 1 week ago

After the first training, I found that the experimental result only had the value of loss. Then I added "abs_rel", "rmse_log", "rmse_linear", "log10", "delta1_acc", "delta2_acc", "delta3_acc", "i_rmse", Indicators such as "silog_rmse" are trained again. I found that several metrics were abnormal, mainly including "abs_rel", "rmse_log", "log10", "silog_rmse". abs_rel" indicator showed a large number of strange negative values, which is theoretically impossible; The "rmse_log", "log10", and "silog_rmse" indicators all contain nan. This makes me very confused. My data set uses a small part of hypersim, and the related functions use metirc. I would like to know what causes this, and hope to get some suggestions from you.

markkua commented 1 week ago

Hi. Sorry, I don't get your question. The evaluation metrics are only used in validation and should not affect training.

zad37 commented 1 week ago

Sorry, thanks for your help, I am a beginner, I seem to have made a low-level mistake, that is, trying to put the evaluation indicators into the training process, and observing the changes of these evaluation indicators on wanb, which made some of the results strange. Thanks again for your help!