Skip to content
forked from filick/scene

Scene classification (AI Challenger)

Notifications You must be signed in to change notification settings

YanWang2014/scene

 
 

Repository files navigation

For testing, remember to alter the file name when finishing a run to keep the best results.
Scores below are used to verify the code, only for reference. 
Also note that, when we use TTA like ten_crops, fcn, spp layer or dense sliding window (or their combinations), maybe we should also change the way we eval in train.py to boost the finally testing accuracy, which means the model needs to be slightly different when loaded in train and eval phase. (may not be a must, but we need to think about how to collocate train, eval and test across (may be various) train.py and various test.py)

test1    
resnet18_places_submit1_val.json
resnet18_places_submit1_test.json
0.5544943820224719

test2_* (average before softmax, actually model output is logits)
json marked by submit2
0.5726123595505618

test3_* (average after softmax, more common)
json marked by submit3
0.5728932584269663

test4_fcn (following resnet paper, fcn-style testing, except for a center_crop after Resize)
json marked by submit4
0.5589887640449438, a very poor setting due to no enough GPU

test5_ensemble (integrate outputs from test1,test2,test 3,test4)
json marked by submit5
0.5544 + 0.5589 -> 0.5631
注意:存了几个示例文件在submit下,如resnet18_places_softmax1_val.txt,比较大,添加到.gitignore了,应该不能直接pull下来

About

Scene classification (AI Challenger)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%