shasha521 / LIIR_pytorch

9 stars 0 forks source link

Any idea to further boost the performance #4

Open miaoweixin opened 2 years ago

miaoweixin commented 2 years ago

I split the training into three stages and choose the best performance checkpoint at each stage as the init weight of the next stage. I also try freeze bn at the last stage. Currently I can get J&F_m = 71.3 on DAVIS17. Are there any suggestions to further boost the performance?

--------------------------- Global results for val --------------------------- J&F-Mean J-Mean J-Recall J-Decay F-Mean F-Recall F-Decay 0.712757 0.685753 0.817705 0.181915 0.739762 0.846216 0.221825

---------- Per sequence results for val ---------- Sequence J-Mean F-Mean bike-packing_1 0.618614 0.758005 bike-packing_2 0.826989 0.855601 blackswan_1 0.919175 0.968542 bmx-trees_1 0.325861 0.681219 bmx-trees_2 0.666273 0.819022 breakdance_1 0.778644 0.794687 camel_1 0.772447 0.817312 car-roundabout_1 0.873699 0.834908 car-shadow_1 0.882705 0.904354 cows_1 0.905036 0.934488 dance-twirl_1 0.631263 0.677474 dog_1 0.890696 0.919624 dogs-jump_1 0.669330 0.736038 dogs-jump_2 0.723436 0.728859 dogs-jump_3 0.864254 0.949477 drift-chicane_1 0.856312 0.955514 drift-straight_1 0.611742 0.566988 goat_1 0.851929 0.842650 gold-fish_1 0.753708 0.736321 gold-fish_2 0.752097 0.814130 gold-fish_3 0.817566 0.862553 gold-fish_4 0.807246 0.836457 gold-fish_5 0.838179 0.814157 horsejump-high_1 0.820492 0.933838 horsejump-high_2 0.731164 0.934194 india_1 0.624830 0.590144 india_2 0.723372 0.736465 india_3 0.691825 0.698357 judo_1 0.706619 0.827395 judo_2 0.779499 0.812317 kite-surf_1 0.284904 0.455871 kite-surf_2 0.239449 0.268279 kite-surf_3 0.675439 0.899571 lab-coat_1 0.000000 0.000000 lab-coat_2 0.000000 0.000000 lab-coat_3 0.916442 0.886862 lab-coat_4 0.877575 0.776149 lab-coat_5 0.864206 0.819380 libby_1 0.828319 0.934479 loading_1 0.851316 0.790240 loading_2 0.454991 0.574898 loading_3 0.751853 0.779707 mbike-trick_1 0.671071 0.758691 mbike-trick_2 0.588445 0.531598 motocross-jump_1 0.475037 0.579491 motocross-jump_2 0.647792 0.684268 paragliding-launch_1 0.810217 0.888886 paragliding-launch_2 0.702811 0.920165 paragliding-launch_3 0.131694 0.430396 parkour_1 0.872762 0.935050 pigs_1 0.776370 0.791333 pigs_2 0.697800 0.821662 pigs_3 0.917698 0.897705 scooter-black_1 0.652826 0.817819 scooter-black_2 0.706053 0.639942 shooting_1 0.216472 0.243083 shooting_2 0.756410 0.707598 shooting_3 0.850765 0.936752 soapbox_1 0.728054 0.743560 soapbox_2 0.512407 0.617597 soapbox_3 0.566734 0.663376

aytackanaci commented 2 years ago

Hi miao, I am also trying to reproduce the the paper results but can't match the paper. Sounds like you got it working! :) Could you send us the conda env file if that's not too much of a hassle. We also tried splitting the three stages of training but could't match the paper. Thanks in advance.