Repository navigation
Cannot reproduce the results on DomainNet when K = 1 #5
Description
Activity
Hi, are you using pytorch 0.4.0 to reproduce the results?
Hi, are you using pytorch 0.4.0 to reproduce the results?
Yes. And I splited the data based on the txt files in SSDA_MME/data/txt/multi you provided.
That is some kind of weird that in the experiments of one-shot setting, I can obtain similar results (i.e., the difference is smaller than 0.5%) in some adaptation scenarios (e.g., C to S, S to P), while in other cases (like, P to C, R to S), the differences are larger than 1.0%.
Thank for sharing your project~
Based on the data split and settings of hyper-parameters in the paper, I can obtain similar classification performance on DomainNet when K = 3, but not when K = 1 (e.g., in P to R, I get 73.65 accuracy, instead of 76.1 claimed in the paper; in R to S, I get 59.655 accuracy, rather than 61.0 claimed in the paper).
Do I miss some important details ?