본문 바로가기

머신러닝

(13)
링크, google landmark recognition 설명 https://personal-record.onrender.com/post/google-landmark-recognition-2020/?fbclid=IwAR12P_Dd68wxYstWYgwi1EgUiPZsLDS2zYrP_PWYtObtQ6lMdNd8fyUjhQY
finetuning시 big transfer and efficientnet v2 보호되어 있는 글입니다.
캐글 전략 madgrad + cosine warmup 구림 0.75 lovasz + 0.25 bce 512*512, efficientnet v2-M(논문 보고 최적의 setting 필요) Hflip + RandomReSizedCrop + adjust saturation (training + inference) blurout 지우고 cutout -> 엄청난 boost (370->375) efvn 384 384 good yolo flip 넣으면 cv 급상승 FLIP th1 0.4, th2 0.4, 0.5 5072 5499 5339 5075 5165 vs 0.5 0.5 4941 5324 5148 4966 4949
캐글 SIIM-FISABIO-RSNA COVID-19 Detection -심화 전략 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)) l..
캐글 SIIM-FISABIO-RSNA COVID-19 Detection - 베이스 전략 study label을 통해서 image label의 병변별 클래스 존재 여부를 알 수 있다. image level의 2class cv와 multilbel class cv를 비교하여야 한다. opacity와 negative가 무조건 일치하는 것은 아니기 때문에 일단 none vs opacity를 구분하는 2 classifier를 만들자. study-label은 segmentation aux loss를 사용하는 방법과 그냥 4 class output prediction을 사용하는 방법을 고려해야 한다. object detection은 그냥 yolov5로 가자...
캐글 SIIM-FISABIO-RSNA COVID-19 Detection - Study image의 특징 Studies either contain a single image (the majority) or multiple images (up to 9) All of the images within the studies are also found in the image level dataset If a study contains multiple images, a maximum of one image and a minimum of zero images will have bounding boxes even though all of the images exist in the image level dataset. There are 2 typical scenarios that occur when multiple imag..
공부기록 - Optimal Bayesian Rule 공부한 것을 기록할 겸 남겨 놓는다. Bayesian Classifier가 오류 확률을 최소화 해 준다는 것이다. 당연한 것 같지만, 사소하게 넘겨서는 안 되는 내용이다.
파이썬 Gradient Boosting 바닥부터 구현하기 기말고사 공부 할 겸 Gradient Boosting을 바닥부터 구현해 보았다. 나는 이해하지 못하면 암기를 못하는 편이다... 찝찝함을 없애기 위해 한번 파이썬으로 Gradient Boosting을 구현해 보았다. 하악... 코드 >> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 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 9..