/kaggle-aptos-taejjing

Primary LanguagePython

실험

randomstate 42

processing_v1

  • img_size 256
  • efn b4
  • Median subtraction
  • Only APTOS 2019
  • 5fold
  • coef = [0.57, 1.37, 2.57, 3.57] # 1.37...
  • steplr

(mse loss / kappa score)

fold1 fold2 fold3 fold4 fold5 lb
processing_v1 (tta 5) 0.2584/0.9074 0.2614/0.9127 0.2552/0.9110 0.2716/0.9032 0.2831/0.8943 0.762
processing_v1 (no tta) 0.781

processing_v2

  • img_size 256
  • efn b4
  • Median subtraction
  • Only APTOS 2019
  • 5fold, 20ep, lr:1e-4
  • coef = [0.57, 1.57, 2.57, 3.57]
  • v2 transform
  • steplr(5, 0.1)

(mse loss / kappa score)

fold1 fold2 fold3 fold4 fold5 lb
processing_v2 (no tta) 0.2208/0.9241 0.2279/0.9230 0.2095/0.0185 0.2327/0.9170 0.2338/0.9175 0.768

processing_v3

  • img_size 256
  • efn b4
  • Median subtraction
  • 2019 + prev(sample)
  • 5fold, 30ep, lr:1e-3
  • coef = [0.57, 1.57, 2.57, 3.57]
  • v2 transform
  • steplr(3, 0.2)
  • dropout(0.3)

(mse loss / kappa score)

fold1 fold2 fold3 fold4 fold5 lb
processing_v3 (no tta) 0.4826/0.8523 0.3268/0.9049 0.2545/0.9328 0.2363/0.9368 0.2513/0.9339 0.798

processing_v4

  • img_size 256
  • efn b4, 5fold, 20ep, lr 1e-3
  • 2019 + prev(sample) add_prev_v1
  • coef = [0.5, 1.5, 2.5, 3.5]
  • steplr(3, 0.2)
  • dropout(0.2)

site

https://www.kaggle.com/jeru666/aptos-preprocessing-update-histogram-matching

TODO

  • add_prev_v1 : 데이터가 늘어났긴 하지만 acc, kappa가 생갹보다 많이 늘어남.
  • add_prev_v2 : v1에 의해 v2도 실험해볼 필요가 있음.
  • circle_v3
  • coef weight
  • regression -> classification