조사 결과 efficientNet-v2,
Mask를 활용한 AuxLoss 사용이 주요한 전략인 듯 싶다.
*mask 전략
32*32 mask
*augmentation 전략
for fn in np.random.choice([
lambda image, mask : do_random_scale(image, mas, mag=0.2) -> random crop
lambda image, mask : do_random_stretch_y(image, mask, mag=0.2)
lambda image, mask : do_random_stretch_x(image, mask, mag=0.2)
lambda image, mask : do_random_shift(image,mask,mag=int(0.2*image_size))
lambda image, mask : (image, mask)],1)
for fn in np.random.choice([
lambda image, mask : do_random_rotate(image, mask, mag=15),
lambda image, mask : do_random_hflip(image, mask),
labmda image, mask : (image, mask)],1)
for fn in np.random.choice([
lambda image : do_random_intensity_shift_contrast(image, mag=([0.5,0.5]))
lambda image : do_random_noise(image, mag=0.05)
lambda image : do_random_gaussian_blur(image),
lambda image : do_random_blurout(image, size=0.25, num_cut=2)
lambda image : do_random_clahe(image),
lambda image : do_histogram_norm(image),
lambda image : image
'머신러닝 > 캐글' 카테고리의 다른 글
finetuning시 big transfer and efficientnet v2 (0) | 2021.07.10 |
---|---|
캐글 전략 (0) | 2021.07.09 |
캐글 SIIM-FISABIO-RSNA COVID-19 Detection - 베이스 전략 (0) | 2021.07.01 |
캐글 SIIM-FISABIO-RSNA COVID-19 Detection - Study image의 특징 (0) | 2021.07.01 |