PState3 실험 노트의 사본

	A	B	C	D	E	F	G	H	I	J	K	L	M	N	O	P
1	캠퍼	리더보드	Val mAp	Backbone	Neck	아키텍쳐	CLS Loss	Bbox Loss	Optimizer	LR (scheduler)	Augment (p=)	Fold	TTA	batch	epoch	해석

2	김현우	0.3115	0.1620	resnet50	FPN	faster rcnn	CrossEntropyLoss	L1Loss	SGD	0.02(Linear)	baseline			4	12	baseline model
3	서준배	0.1756	0.1530	darknet		yolov3	CrossEntropyLoss		SGD	0.001	baseline				273	epoch로그확인하니 큰 성능향상은 없음
4	윤준호	0.3310	0.1720	resnet50	FPN	cascade_rcnn	CE		SGD	0.02	baseline			8	40	epoch 30 보다 40이 더 잘 나옴
5	김현우	0.4149	0.2270	resnet50	RFP + SAC	cascade_rcnn	CrossEntropyLoss	L1Loss	SGD	0.02(Linear)	baseline			4	12	RFP + cascade FPN ( epochs 늘리면 더 올라감)
6	김종호	0.0603		resnet50		gfl	QualityFocalLoss	GIoULoss	SGD	0.02	baseline			4	24
7	김현우	0.3976	0.2170	resnet50	RFP + SAC	cascade_rcnn + HTC	CrossEntropyLoss	L1Loss	SGD	0.02(Linear)	baseline			4	60	HTC 는 semantic segmentation 과 함께 동작함 : 성능은 계속 올라가나 epochs 엄청 늘려야 할듯
8	김현우	0.3464	0.1970	resnet50	RFP + SAC	cascade_rcnn	CrossEntropyLoss	L1Loss	SGD	0.02(Linear)	HueSaturationValue,RandomGamma,CLAHE,RandomBrightnessContrast,Blur, MotionBlur, GaussNoise, ImageCompression, RandomRotate90			4	24	argumentation 검증하고 넣지 않으면 성능이 오히려 떨어짐
9	윤준호	0.3950	0.2210	ResNeXt101	FPN	VFNet	VarifocalLoss	GIoULoss	SGD		baseline			8	22
10	윤준호	0.3924	0.2210	ResNeXt101	FPN	VFNet	VarifocalLoss	GIoULoss	SGD		baseline			8	30	epoch 22 >= epoch 30 > epoch 40
11	김현우	0.4492	0.2720	resnet50	RFP + SAC	cascade_rcnn	CrossEntropyLoss	L1Loss	SGD	0.01 (Step LR)	RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift			4	48	argumentation + 48 epochs (10시간 학습) 43 epochs 부터 성능이 계속 올라감
12	윤준호	0.4024	0.2370	ResNeXt101	FPN	VFNet	VarifocalLoss	GIoULoss	SGD	0.015 (Step LR)	RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift			8	40
13	김현우	0.4539	11 번 + soft mns 적용
14	김종호	0.4311	0.2390	ResNeXt101	FPN	cascade_rcnn	CrossEntropyLoss	SmoothL1Loss	SGD	0.02	baseline			16	30	epoch 19까지 증가 이후 동일
15	윤준호	0.4037	0.2360	ResNeXt101	FPN	VFNet	VarifocalLoss	GIoULoss	SGD	0.015 (Step LR)	RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift			8	49
16	김종호	0.4630	0.2720	ResNeXt101	FPN	cascade_rcnn	CrossEntropyLoss	SmoothL1Loss	SGD	0.02	RandomRotate90, RandomSizedBBoxSafeCrop			16	35
17	김종호	0.4658	0.2700	ResNeXt101	FPN	cascade_rcnn	CrossEntropyLoss	SmoothL1Loss	SGD	0.02	RandomRotate90, RandomResizedCrop(512,512,0.5,0.8)			16	35	RandomSizedBBoxSafeCrop와 RandomResizedCrop 차이
18	김현우	0.5121	0.3100	resnet50	RFP + SAC	cascade_rcnn	SoftCrossEntropyLoss	SmoothL1Loss	SGD	0.01 (Step LR)	Multi Scale (512 ~ 768) , HueSaturationValue,RandomGamma,CLAHE,RandomBrightnessContrast,Blur, MotionBlur, GaussNoise, ImageCompression, RandomRotate90, RandomBrightnessContrast, RGBShift, Pad		MultiScaleFlipAug 512, 768 + horizontal, vertical flip, Pad	4	48
19	김현우	0.4606	18 번 + base TTA								Multi Scale (512 ~ 768) , HueSaturationValue,RandomGamma,CLAHE,RandomBrightnessContrast,Blur, MotionBlur, GaussNoise, ImageCompression, RandomRotate90, RandomBrightnessContrast, RGBShift, Pad		Base	4	48
20	윤준호	0.4040	12번 + 15번 (WBF로 ensemble) -> 같은 모델에 epoch만 근소하게 다른 두 모델을 앙상블했음에도 성능 근소하게 상승
21	김현우	0.5247	0.3360	ResNeXt101	RFP + SAC	cascade_rcnn	SoftCrossEntropyLoss	SmoothL1Loss	SGD	0.01 (Step LR)	18 번+ ShiftScaleRotate + RandomSizedBBoxSafeCrop + MixUP	0 (default)		4	50	결과는 1 day, 10:54:50, 후에
22	배철환	0.3155	0.3460	darknet		yolov5	CrossEntropy		SGD	0.01				32	120
23	서준배	0.3709	0.3600	darknet		yolov5	CrossEntropy		SGD	0.01	lr0: 0.01 lrf: 0.2 momentum: 0.937 weight_decay: 0.0005 warmup_epochs: 3.0 warmup_momentum: 0.8 warmup_bias_lr: 0.1 box: 0.05 cls: 0.5 cls_pw: 1.0 obj: 1.0 obj_pw: 1.0 iou_t: 0.2 anchor_t: 4.0 fl_gamma: 0.0 hsv_h: 0.015 hsv_s: 0.7 hsv_v: 0.4 degrees: 0.0 translate: 0.1 scale: 0.5 shear: 0.0 peRFPective: 0.0 flipud: 0.0 fliplr: 0.5 mosaic: 1.0 mixup: 0.0			32	57	mosaic(1.0)
24	김종호	0.4896	16번 + 17번 (WBF로 ensemble) -> 0.02 이상 상승
25	배철환	0.3637	0.4080	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	0		32	150
26	배철환	0.3916	0.4160	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	1		32	150
27	배철환	0.3919	0.3700	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	4		32	150
28	서준배	0.3706	0.3780	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	2		32	150
29	서준배	0.3684	0.3880	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	3		32	150
30	김종호	0.4656	0.5530	ResNeXt101	FPN	cascade_rcnn	CrossEntropyLoss	SmoothL1Loss	SGD	0.02	RandomRotate90, RandomSizedBBoxSafeCrop			16	25	Json2
31	김종호	0.4621		ResNeXt101	FPN	cascade_rcnn	CrossEntropyLoss	SmoothL1Loss	SGD	0.02	RandomRotate90, RandomSizedBBoxSafeCrop			16	30	Json1
32	배철환	0.4610	nms	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	ensemble		32	150
33	배철환	0.4714	wbf	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	ensemble		32	150
34	김종호	0.4668	0.5500	ResNeXt101	FPN	cascade_rcnn	CrossEntropyLoss	SmoothL1Loss	SGD	0.02	RandomRotate90, RandomSizedBBoxSafeCrop			16	25	Json3
35	배철환	0.4671	nms + tta	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	ensemble		32	150
36	윤준호	0.4315	0.246	ResNeXt101	FPN	GFL	QualityFocalLoss	GIoULoss	SGD	0.01 (Step LR)	RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift		RandomFlip	16	45	Generalized Focal Loss 모델 실험
37	배철환	0.3786	0.3300	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomRotate90, OneOf([HueSaturationValue, RandomGamma, CLAHE]), OneOf([RandomBrightnessContrast, RGBShift, ToGray]), OneOf([Blur, MotionBlur, GaussianBlur, ImageCompression])	4		16	150
38	김종호	0.4781	0.5460	ResNeXt101	FPN	cascade_rcnn	CrossEntropyLoss	SmoothL1Loss	SGD	0.02	RandomRotate90, RandomSizedBBoxSafeCrop			16	25	Json4
39	김종호	0.5130	json1 , json2, json3, json4 WWWBBBFFF								RandomRotate90, RandomSizedBBoxSafeCrop
40	배철환	0.1970	0.2460	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	4	o	32	40	no multi-scale
41	배철환	0.1961	0.2430	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	4	o	32	40	multi-scale
42	배철환	0.4730	wbf + tta	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	ensemble		32	150	after_ensemble
43	배철환	0.4901	wbf + tta	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5	ensemble		32	150	before_ensemble
44	김현우	0.5203	0.3270	resnet50	RFP + SAC	cascade_rcnn	SoftCrossEntropyLoss	SmoothL1Loss	SGD	0.01	Multi Scale (512 ~ 768) , Argument (18번), Mixup	1	TTA	4	52
45	배철환	0.3631	0.3360	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	4		32	225
46	서준배		0.3298	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	2		32	225
47	윤준호	0.3426	0.2650	ResNet50	RFP + SAC	cascade_rcnn	FocalLoss	SmoothL1Loss	SGD	0.01	RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift, 현우님CutMix, multiscale	3	RandomFlip	16	39
48	윤준호	0.3863	0.2980	ResNet50	RFP + SAC	cascade_rcnn	FocalLoss	SmoothL1Loss	SGD	0.01	RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift, 현우님CutMix, multiscale	3	RandomFlip	16	60
49	배철환	0.3657	0.3630	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	0	o	32	224
50	김현우	0.5175	0.3300	resnet50	RFP + SAC	cascade_rcnn	SoftCrossEntropyLoss	SmoothL1Loss	SGD	0.01	Multi Scale (512 ~ 768) , Argument (18번), Mixup	1			72
51	김종호	0.4590	0.3270	resnet50	RFP + SAC	cascade_rcnn	SoftCrossEntropyLoss	SmoothL1Loss	SGD	0.01	Multi Scale (512 ~ 768) , RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift	2	o	8	60
52	배철환	0.3928	0.3530	darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	1	o	32	224
53	서준배	0.3807		darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	2	o	32	224	last.pt사용
54	윤준호	0.4570	0.3090	ResNet50	RFP + SAC	cascade_rcnn	LabelSmoothCrossEntropyLoss	SmoothL1Loss	SGD	0.01	RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift, multiscale	3	RandomFlip	16	70
55	서준배	0.3767		darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	2	o	32	150	last.pt
56	서준배	0.3636		darknet		yolov5	CrossEntropy		SGD	0.01	mosaic = 1, mixup = 0.5, RandomGamma, CLAHE, RandomBrightnessContrast	2	o	32	150	best.pt
57	윤준호	0.4963	0.3320			swin	LabelSmoothCrossEntropyLoss					3	o	6	45
58	윤준호	0.5476	0.3530			swin (재학습)	LabelSmooth + CE + Focal (box_head 별로)				RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift, multiscale	3	o	3	45 + 9
59	서준배	0.4483	0.401			swin						4		8	39
60	배철환	0.5355	0.3310			swin (재학습)	CE				RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift, multiscale	2	o	3	45 + 9
61
62
63	김종호	0.5023				swin			adamW	0.0001		0		6	30	base train json
64	김종호	0.5296				swin(재학습)			adamW	0.0001	RandomRotate90, RandomSizedBBoxSafeCrop			6	21	train_all
65	김종호	0.5302				swin(재학습)			adamW	0.0001	RandomRotate90, RandomSizedBBoxSafeCrop			6	24	train_all
66	배철환	0.5486	0.3390			swin (재학습)	CE				RandomRotate90, HueSaturationValue, RandomGamma, CLAHE,RandomBrightnessContrast, RGBShift, multiscale	2	o	3	45 + 9
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100