search

autonomousvision / unimatch

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

Reproduce Middlebury results #25

Closed javierBarandiaran closed 1 year ago

javierBarandiaran commented 1 year ago

Hello, I am trying to reproduce the results in the training set, but I am obtaining slightly worse results image I am using these parameters main_stereo.py --submission --resume pretrained/gmstereo-scale2-regrefine3-resumeflowthings-middleburyfthighres-a82bec03.pth --inference_size 1024 1536 --val_dataset middlebury --middlebury_resolution F --middlebury_submission_mode training --output_path /hal/pytorch/MiddEval3/trainingF --padding_factor 32 --upsample_factor 4 --num_scales 2 --attn_type self_swin2d_cross_swin1d --attn_splits_list 2 8 --corr_radius_list -1 4 --prop_radius_list -1 1 --reg_refine --num_reg_refine 3

Could you please tell me if I am doing something wrong with the parameters?

haofeixu commented 1 year ago

Hi, the parameters look good. This subtle difference might be caused by different library versions I guess. Could you double check whether your python environement is exactly the same as ours in readme?

Here are our submission files on the training set: https://mailustceducn-my.sharepoint.com/:u:/g/personal/xhf_mail_ustc_edu_cn/EWVPp1ulHvlMuOlNC_MNIZ4BOpP9JsGVFlTkLshWSmH1xg?e=caCGah I have double checked the results are consistent.

image
javierBarandiaran commented 1 year ago

Thank you, yes, that could be the cause, I am using Pytorch 2.0.1 with CUDA 11.8