학습방법
- 학습 모델로는 yolo v11을 커스텀하여 모델을 개발하였으며 각 파라미터는 아래와같음
min_lr: 0.000100000000 # initial learning rate
max_lr: 0.010000000000 # maximum learning rate
momentum: 0.9370000000 # SGD momentum/Adam beta1
weight_decay: 0.000500 # optimizer weight decay
warmup_epochs: 3.00000 # warmup epochs
box: 7.500000000000000 # box loss gain
cls: 0.500000000000000 # cls loss gain
dfl: 1.500000000000000 # dfl loss gain
hsv_h: 0.0150000000000 # image HSV-Hue augmentation (fraction)
hsv_s: 0.7000000000000 # image HSV-Saturation augmentation (fraction)
hsv_v: 0.4000000000000 # image HSV-Value augmentation (fraction)
degrees: 0.00000000000 # image rotation (+/- deg)
translate: 0.100000000 # image translation (+/- fraction)
scale: 0.5000000000000 # image scale (+/- gain)
shear: 0.0000000000000 # image shear (+/- deg)
flip_ud: 0.00000000000 # image flip up-down (probability)
flip_lr: 0.50000000000 # image flip left-right (probability)
mosaic: 1.000000000000 # image mosaic (probability)
mix_up: 0.000000000000 # image mix-up (probability)
names:
0: class_0
1: class_1+
2: class_2+
3: class_3+
- 학습시 multi scale로 진행하였으며 hover-net의 특징을 사용함. (이미지크기는 층개수^2 의 배수면 같이 학습이 진행됨)
학습결과
WSI analysis
