search

autonomousvision / unimatch

[TPAMI'23] Unifying Flow, Stereo and Depth Estimation
https://haofeixu.github.io/unimatch/
MIT License
980 stars 102 forks source link

loss curve #50

Closed ZhouQianang closed 4 months ago

ZhouQianang commented 4 months ago

Very good job! Could you provide a loss curve from a regular training session? I'm using the same model on another dataset, but the EPE is decreasing slowly.

ZhouQianang commented 4 months ago

Hello? This is the first 2000 steps of my training loss. Is this decline rate normal? I am using gmflow_scale1_train.sh with batch_size=8, and my dataset is a sparse flow dataset. step: 000100 epe: 32.581 step: 000200 epe: 17.243 step: 000300 epe: 12.278 step: 000400 epe: 10.447 step: 000500 epe: 9.192 step: 000600 epe: 7.654 step: 000700 epe: 6.763 step: 000800 epe: 6.227 step: 000900 epe: 5.693 step: 001000 epe: 5.022 step: 001100 epe: 4.580 step: 001200 epe: 4.256 step: 001300 epe: 3.908 step: 001400 epe: 3.654 step: 001500 epe: 3.544 step: 001600 epe: 3.247 step: 001700 epe: 3.109 step: 001800 epe: 2.928 step: 001900 epe: 2.851 step: 002000 epe: 2.692

haofeixu commented 4 months ago

Hi, this looks reasonable, you can check the flow visualization results as well.