search

takaiyuk / probspace-religious-art

https://prob.space/competitions/religious_art
0 stars 0 forks source link

readme

Probspace - 宗教画テーマの分類

https://prob.space/competitions/religious_art

Run

$ ./scripts/docker/run.sh
$ ./scripts/docker/exec.sh
root@xxxxx:/workspace# venv-acitvate
(venv) root@xxxxx:/workspace# ./scrirpts/run.sh expXXX

Jupyter

$ ./scripts/docker/run.sh
$ ./scripts/docker/exec.sh
root@xxxxx:/workspace# venv-acitvate
(venv) root@xxxxx:/workspace# ./scripts/jupyter.sh

Submission

Late Sub

Name CV Public Score Private Score Base Description
submission_036.csv 0.66055 0.625 0.630 exp034 stacking: exp008, exp010, exp012, exp015, exp018 (2nd stage: vote)

Result

Name CV Public Score Private Score Base Description
submission_035.csv ------- ----- - - stacking: exp018, exp015, exp012, exp010 (2nd stage)
submission_034.csv 0.66055 0.625 0.633 exp027 stacking: exp018, exp015, exp012, exp010 (weighten with cv score)
submission_033.csv 0.56061 0.578 0.607 exp022 Larger image size
submission_032.csv 0.65749 0.625 0.637 - stacking: exp028, exp027, exp026 (average)
submission_031.csv 0.65291 0.641 0.630 - stacking: exp028, exp027, exp026, exp025 (average)
submission_030.csv 0.64220 0.641 0.624 - stacking: exp028, exp027, exp026, exp025, exp024 (average)
submission_029.csv 0.63914 0.625 0.624 - stacking: exp028, exp027, exp026, exp025, exp024, exp023 (average)
submission_028.csv 0.64832 0.609 - - same as submission_021
submission_027.csv 0.66055 0.656 - - same as submission_020
submission_026.csv 0.65291 0.641 0.630 - same as submission_019
submission_025.csv 0.61927 0.656 - - same as submission_016
submission_024.csv 0.61927 0.609 - - same as submission_013
submission_023.csv 0.59786 0.625 0.612 - same as submission_011
submission_022.csv ------- ----- - exp018 CutMix
submission_021.csv 0.64832 0.609 0.630 - stacking: exp012, exp015, exp018 (average)
submission_020.csv 0.66055 0.656 0.635 - stacking: exp010, exp012, exp015, exp018 (average)
submission_019.csv 0.65291 0.656 0.630 - stacking: exp008, exp010, exp012, exp015, exp018 (average)
submission_018.csv 0.64394 0.609 0.640 exp017 leak fix (avoid evaluating psuedo labeled dataset)
submission_017.csv 0.72126 0.578 0.635 exp015 psuedo labeling
submission_016.csv 0.61927 0.656 0.619 - stacking: exp008, exp010, exp012, exp015 (average)
submission_015.csv 0.60550 0.609 0.630 exp014 resnext50 -> resnext101
submission_014.csv 0.59939 0.562 0.612 exp012 larger image_size, more epochs
submission_013.csv 0.61927 0.609 0.619 - stacking: exp008, exp010, exp012 (average)
submission_012.csv 0.60550 0.609 0.594 exp008 resnext50, more augmentations
submission_011.csv 0.59786 0.625 0.612 - stacking: exp008, exp010 (average)
submission_010.csv 0.58716 0.578 0.582 exp008 resnest50
submission_009.csv 0.50765 - - exp008 mobilenetv3
submission_008.csv 0.57034 0.594 0.561 exp006 image サイズを大きく
submission_007.csv 0.56422 0.484 0.497 exp006 efficient_b2 に変更
submission_006.csv 0.55657 0.516 0.545 exp005 損失関数に weights を追加（少数クラスほど weight を大きく）
submission_005.csv 0.50917 - - exp002 model を resnet50 に変更
submission_004.csv 0.48165 0.484 0.547 exp002 重複画像のラベルデータを後処理で埋める
submission_003.csv 0.46636 - - exp002 輝度（Brightness）の統一
submission_002.csv 0.48165 0.484 0.545 - -

Probing

y CV train sample size Public public sample size
0 0.092 60 0.078 5
1 0.064 42 0.078 5
2 0.202 132 0.156 10
3 0.064 42 0.109 7
4 0.064 42 0.078 5
5 0.092 60 0.047 3
6 0.073 48 0.078 5
7 0.046 30 0.062 4
8 0.046 30 0.031 2
9 0.101 66 0.078 5
10 0.046 30 0.047 3
11 0.064 42 0.094 6
12 0.046 30 0.062 4