Augmentation 실험

[jun9choi]

실험 목록

실험	실험자	상태	Valid Sentence Accuracy	Valid Score	BASE 대비 개선 여부
BASE	준구	완료	0.2497	0.2959	-
RGB Shift	형민	완료	0.1487	0.196	X
Normalization(IMGNET)	지형	완료	0.2538	0.3004	O
ChannelShuffle	형민	완료	0.2455	0.2927	X
ShiftScaleRotate	지형	완료	0.2956	0.3401	O
Blur set	준철	완료	0.2372	0.2833	X
MedianBlur	준철	완료	0.2372	0.2833	X
CLAHE + Normalization	준구	완료	0.2645	0.3107	O
GridDistortion(8,0.5)	형민	완료	0.2867	0.3316	O
RandomBrightnessContrast + Norm	형민	완료	0.3275	0.3725	O
Flip + Normalization	준구	완료	0.2227	0.2693	X
ShiftScaleRotate + Norm(IMGNET)	지형	완료	0.3022	0.3467	O
SSR + GD + Norm	형민	진행중	0.3935	0.4346	O
CLAHE + SSR + GD + Norm	지형	진행중	-	-	-
GaussNoise + Norm	형민	진행중	-	-	-
AidaData(2500장) +SSR + GD + Norm	주영	완료	0.3933	0.4346	-

BASE 세팅

• Split Strategy: 준철 스플릿
• Train Dataset: Fold 0, 1
• Valid Dataset: Fold 4
• Naive Rotation: 세로 길이가 가로 길이에 비해 2배 이상 길 때 90도 회전
• RGB: 3
• batch_size: 64
• scheduler= 'CustomCosine'
  • T_0 = total_steps
  • T_up = int(T_0*0.1)
  • optimizer: 'Adam'
• Base augmentation
  • Tool: albumentations
  • Train

  A.Compose([A.Resize(height, width), ToTensorV2(p=1.0)],

  • Valid

  A.Compose([A.Resize(height, width), ToTensorV2(p=1.0)],

Augmentations

• ShiftScaleRotate

A.ShiftScaleRotate(shift_limit=0.0, scale_limit=0.1, rotate_limit=0, p=0.5)

• Normalization

A.Normalize(mean=[0.6280586 , 0.61502952, 0.58616558], std=[0.16464177, 0.16915324, 0.1757833])

• ChannelShuffle

A.ChannelShuffle(always_apply=False, p=0.5)

• Flip

A.Compose([A.HorizontalFlip(p=1), A.VerticalFlip(p=1)], p=0.5)

• GridDistortion

A.GridDistortion(p=0.5, num_steps=8, distort_limit=(-0.5, 0.5), interpolation=0, border_mode=0)

• RandomBrightnessContrast

A.RandomBrightnessContrast(p=0.5, brightness_limit=(-0.2, 0.2), contrast_limit=(-0.2, 0.2), brightness_by_max=True)

• CLAHE

A.CLAHE(p=0.2)

• GaussNoise

A.GaussNoise(var_limit=(10.0, 50.0), mean=0, p=0.5)

[doritos0812]

A.ChannelShuffle(always_apply=False, p=0.5)