(4.30, Jaeha) Experiments

Parameters

• model = DeepLabv3+ ResNeXt50 32x4d
• loss function = cross entropy
• optimizer = adam
• scheduler = CosineAnnealingWarmupRestarts
• batch = 8
• input_size = 256x256
• seed = 77c

---

• max_lr, min_lr = 1e-4, 1e-6
• Normalize는 train_all_dataset에 대하여 mean, std를 구하고 그 값으로 적용

모델	input_size	배치	epoch	loss	val_loss	val_mIoU	LB score	비고
DLV3+, resnext50[ssl]	256	8	18	0.0536	0.3546	0.5189	0.5695	base
DLV3+, resnext50[ssl]	256	8	17	0.1058	0.3337	0.5177	0.5865	+normalize

---

• pretrained 'swsl'
• Histogram Equalization, CLAHE는 train_set, val_set, test_set에 모두 p=1.0으로 적용함

모델	input_size	배치	epoch	loss	val_loss	val_mIoU	LB score	비고
DLV3+, resnext50[swsl]	256	8	20	0.0781	0.3448	0.5127	0.5924
DLV3+, resnext50[swsl]	256	8	18	0.0536	0.3812	0.4739		Histogram Eqaulization
DLV3+, resnext50[swsl]	256	8	18	0.0940	0.3402	0.4960	0.5591	CLAHE
DLV3+, resnext50[swsl]	256	8	17	0.1019	0.3427	0.5127	0.5775	+CLAHE -Normalize

---

• loss가 너무 튀어서 LR을 낮춤
• max_lr, min_lr = 5e-5, 5e-7
• optimizer를 AdamP를 사용해 봄

모델	input_size	배치	epoch	loss	val_loss	val_mIoU	LB score	비고
DLV3+, resnext50[swsl]	256	8	15	0.0892	0.3190	0.5159
DLV3+, resnext50[swsl]	256	8	17	0.0946	0.3282	0.5344	0.6023	AdamP

---

• Augmentation
  • HorizentalFlip(p=0.5)
  • VerticalFlip(p=0.5)
  • RandomRotate90(p=0.5)

모델	input_size	배치	epoch	loss	val_loss	val_mIoU	LB score	비고
DLV3+, resnext50[swsl]	256	8	18	0.0843	0.2863	0.5447	0.6002