qimw

@Molly6 请问成功复现了嘛

没，update一下，R-50-FPN-ss复现结果是 mAP: 0.7519399442904969 ap of each class: plane:0.8917197611880395, baseball-diamond:0.8250308909092904, bridge:0.5270227177910475, ground-track-field:0.7044010147947455, small-vehicle:0.7671405328320966, large-vehicle:0.8228071178607328, ship:0.8801036472343933, tennis-court:0.9086764824965626, basketball-court:0.8419320047454201, storage-tank:0.8458570430628705, soccer-ball-field:0.6131911560864126, roundabout:0.68786374488329, harbor:0.7489686103336343, swimming-pool:0.6724137494789859, helicopter:0.5419706906599334

> 没，update一下，R-50-FPN-ss复现结果是 mAP: 0.7519399442904969 ap of each class: plane:0.8917197611880395, baseball-diamond:0.8250308909092904, bridge:0.5270227177910475, ground-track-field:0.7044010147947455, small-vehicle:0.7671405328320966, large-vehicle:0.8228071178607328, ship:0.8801036472343933, tennis-court:0.9086764824965626, basketball-court:0.8419320047454201, storage-tank:0.8458570430628705, soccer-ball-field:0.6131911560864126, roundabout:0.68786374488329, harbor:0.7489686103336343, swimming-pool:0.6724137494789859, helicopter:0.5419706906599334 不同gpu型号会影响结果吗

是的， fix 可以看作是 gliding_vertex 的一个alias

I found that the bns are set requires_grad = False, but it will still update the running_mean and running_var. So what's the meaning of doing this?

No, the result is super bad. Maybe this is due to the large domain gap. The last iter we will train model on target domain. So the batchnorm parameters(running_mean, running_var)...

Sorry to bother you, but I don't understand why it isn't trilinear. The code in the url above try to generate interpolation value by surrounding eight values. Am I misunderstand...

I try to train the net work on cityscape and foggy cityscape to remove the fog. The results appears to be much better after modifying. :) before:  after: 

yes! And we still need to make the trilinear clear!

@2019jack @zhoubeichen20 @luuuuuuy 可以提取 BAK_0_TEXT 和 BAK_0_MEDIA 和 DB 里面的内容了吗