search

itijyou / ademxapp

Code for https://arxiv.org/abs/1611.10080
Other
340 stars 98 forks source link

version of MXNet #10

Open xinyuh opened 7 years ago

xinyuh commented 7 years ago

I have tested three models, ade20k_rna-a1_cls150_s8_ep-0001, voc_rna-a1_cls21_s8_coco_ep-0001, and voc_rna-a1_cls21_s8_ep-0001. All of them produced very very low mean ious. Please let me know the verion of MXNet you are using currently. I would like to re-test these models. Thanks!

itijyou commented 7 years ago

v0.8.0 should do. So do the newer versions (although not tested).

xinyuh commented 7 years ago

which version of cudnn for 0.8.0? thanks

itijyou commented 7 years ago

cudnn 7.5

xinyuh commented 7 years ago

for ade20k, I only get 43.35% mean iou for the whole val set, which is similar to another person's result in your comments. Is that possible the version of cudnn or cuda cause the problem? I am using cudnn 5 and cuda 8.

itijyou commented 7 years ago

Read the reply https://github.com/itijyou/ademxapp/issues/9#issuecomment-274437532

xinyuh commented 7 years ago

I used MXNet 0.8.0, Cuda 8, and CuDNN 5, and re-test the whole val set of ade20k using ade20k_rna-a1_cls150_s8_ep-0001.params. The results are the same, the mean iou is 43.35%. Any ideas?

Since it is ade20k model A1, based on my understanding, it should not be affected by the newest MXNet?

itijyou commented 7 years ago

43.35% is the very result Model A1 should achieve. That one in the report (~0.4% higher) was obtained using Model A2 (which is not released for compatibility consideration).