search

noahzn / Lite-Mono

[CVPR2023] Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation
MIT License
527 stars 58 forks source link

How the reproduction of the results of the model achieves the results in the paper #141

Closed Wshenv closed 2 months ago

Wshenv commented 3 months ago

Hello, thanks for great code.I got the pretrained model and I ran the train.py to reproduce the lite-mono, I trained with the same dependencies on the A6000，but I got different results. 564

noahzn commented 3 months ago

Hi, please make sure you use the same versions of dependencies as we used. There are also some tickets discussing the reproduction of the results. Please search in this repo.

Wshenv commented 3 months ago

Thanks for your suggestion, I reconfigured the dependencies again and made the data set in.jpg format according to the way of monodepth2, but the result of retraining is still not satisfactory. This is my dependency, and only the version of tensorboardX dependency is inconsistent, but I think it will not affect the result of training 1 2

Wshenv commented 3 months ago

In addition, I found a torch>=1.8 in requirements.txt of linear_warmup_cosine_annealing_warm_restarts_weight_decay=1.0. Is that the reason?

noahzn commented 3 months ago

It should be fine because we could reproduce our results using pytorch 1.7.1 and 1.12.1 https://github.com/noahzn/Lite-Mono/issues/58

Wshenv commented 3 months ago

Thank you very much. I tried to adjust the dropout to 0.3 and conducted the experiment again, and got a result very close to the paper in the 24th round. I have another question for you. Is the data set in jpg format converted strictly according to the conversion command in monodepth2? Figure_1 Figure_2

noahzn commented 3 months ago

Hi, good to know that you have better results now! Yes, we directly converted the images to .jpg according to the command in the monodepth2's repo.