부스트캠프 11주차 주간학습 정리
0. 배운 것들 Advanced Topis in MRC Elasticsearch Sources of Bias in MRC
Pytorch Tensor operators Numpy + AutoGrad Tensor list (Python) -> Ndarray (Numpy) -> Tensor (PyTorch) Almost identical with Ndarr...
0. 들어가기에 앞서 아래 강의는 안수빈 서울대학교 컴퓨터공학과 석박과통합과정 학생의 부스트 캠프 강의를 요약하여 개인적인 첨언을 담은 포스트입니다
Tramsformer: To solve the problem in dealing with sequential data; the input often changes in sequence and length Recursive x attententive o Encoder (...
Sequential Model: Problem: does not know about the dimension of the input Naive Sequence Model: Consider all information from the past. Probl...
1. Mitigating Training Bias A lot of bias present; models can learn sexual bias
1. Closed-book QA Popular approaches such as MRC and Open-domain QA are allowed to use supporting documents outside of the model New approaches have tri...
1. Types of Retrieval Boolean Retrieval: retrieves a document if it includes specified word Rank Retrieval: specify a weight for terms. The weight for w...
1. What is MRC
Weight and Biases Easy to share the experiemnt with other people! can track hardware status install with pip !pip install wandb -q make ...
Tensorboard Supports PyTorch Shows scalars (accuracy, loss etc) Shows the computational graph Shows the histogram of weights
Contents
Boosting 은 모델을 순차적으로 결합하는 모델 합성 방식이다. 첫 부스팅 알고리즘인 AdaBoost가 등장한 지 20년이 넘었기 때문에 구체적인 알고리즘은 모델마다 크게 다르지만, 순차적으로 이전 모델들의 오답에 유의하여 새 모델을 학습하는 아이디어는 동일하다.
Objective Paper: AN IMAGE IS WORTH 16X16 WORDS: TRANSFORMERS FOR IMAGE RECOGNITION AT SCALE
Goal Classify 18 agumented classes 3 age intervals mask, no mask, and incorrect mask male or female