Problems with inconsistencies between experimental results and articles
Opened this issue · 3 comments
Your work is excellent, I have been learning from your work recently. However, I found that the experimental results in the articles are inconsistent with what I get when running it myself, using DepthTrack's training and testing as an example. I re-tested with the weights you provided on GitHub and got results of Pr, Re, F being 0.561, 0.571, 0.566 respectively; I re-trained and tested with the weights I got and the results were Pr, Re, F being 0.588, 0.586, 0.587 respectively; the results I analyzed according to the testing results you provided in the article were consistent with the article, which are Pr, Re, F being 0.592, 0.596, 0.594. Could you please let me know the reason for this inconsistency? Is it due to environmental reasons or some other reasons? If it's due to the environment, could you please provide the versions of the packages configured in your virtual environment? I'm looking forward to your response, thank you!
Hi~ I think the performance you get is unnormal (0.594 - 0.566).
If everything else is fine, I suggest you try the following versions of the packages:
pytorch==1.7.0 cudatoolkit=10.2
Great job, I will try it, thank you very much!
@hoqolo Is it valid? I got a result of 0.573 on my side.