What's in a "Domain" Mathematically, joint distribution over inputs and outputs differs over domains 1 and 2 $P_{d1}(X,Y) \ne P_{d2}(X,Y)$ 예를 들어, Content, whit is being discussed Style, the way in which it is being discussed Labeling Standards, the way thtat the same data is labeled Types of Domian Shift Covariate Shift: The input changes but not the labeling $P_{d1}(X) \ne P_{d2}(X)$ $P_{d1}(Y|..
Zero-shot learning: 인스턴스별 레이블 없이 학습 Few-shot learning: 몇 가지 훈련 예제로 학습 Unsupervised learning: 레이블 없는 학습. 제로샷 학습과의 차이점 Auto-encoders, contrastive learning Generative models: GANs, VAEs Semi-supervised laerning 라벨 정보가 없는 이미지의 큰 데이터셋이 존재한다면, 이것을 어떻게 다룰 것이닞? Pseudo-labeling 방법이 있다. Pseudo-labels Iterative algorithm: Train a classifier $f$ on $D_1$ Pseudo-labelling: build dataset $D_2$ where we used ..
제1고지 - 다차원 배열을 tensor라고 하며, 방향을 차원(dimension) 혹은 축(axis)라고 합니다. - 컴퓨터 과학에서 그래프는 노드(node)와 에지(edge)로 구성된 데이터 구조를 말함 - __call__ 메서드는 파이썬의 특수 메서드. 이 메메서드를 정의하면 f = Function() 형태로 함수의 인스턴스 변수를 변수 f에 대입해두고, 나중에 f(...) 형태로 __call__ 메서드를 호출 - 'NotImplementedError'는 ㅇ이 메서드는 상속하여 구현해야 한다는 의미 - assert 문은 결과가 True가 아니면 예외가 발생함 class Variable: def __init__(self, data): self.data = data self.grad = None self..
이번 포스팅은 cs231n 강의의 Lecture 12 Video Understanding, EECS Lecture 24 Videos 자료를 참고하였습니다. 또한, Videos 관련 강의로는 해당 영상을 참고하였습니다. (Lecture 18: Videos (UMich EECS 498-007) Video = 2D + Tensor 비디오는 이미지 4D 텐서의 시퀀스입니다. (T x 3 x H x W) T: the time or temporal dimension 3: the channel dimension which is three colors RGB channels for the raw input video H and W: two spatial dimensions 비디오에는 motivating task가 있습..
이번 포스팅은 cs231n 강의의 Lecture 11 Attention and Transformers, EECS Lecture 17 Attention 자료를 참고하였습니다. 또한, Transformer 관련 내용으론 해당 블로그의 번역 글을 참고하였습니다. (The Illustrated Transformer) Sequence to Sequence with RNNs Problem: Input sequence bottlenecked through fixedsized vector. What if T=1000? Idea: use new context vector at each step of decoder! Sequence to Sequnece의 아이디어로는 컨텍스트 벡터를 인코더와 디코더 사이에 연결하는 것입니..
이번 포스팅은 cs231n 강의의 Lecture 10 Recurrent Neural Networks, EECS 498.008 / 598.008 강의의 Lecture 16 Recurrent Neural Network을 참고하여 정리하였습니다. 이번 강의는 cs231n의 자료가 더 좋습니다. 이번 강의에서 다루는 내용은 RNN, LSTM, GRU Language modeling Sequence-to-sequence Image captioning 입니다. RNN의 Process Sequence를 보면 이렇습니다. 1) One to one Image classification: Image → Label 2)One to many Image Captioning: Image → Sequence of words 3) ..
이번 포스팅은 standford university의 cs231 lecture 9와 EECS 498.008 / 598.008 강의의 Lecture 13 Object Detection, Lecture 14 Object Detectors, Lecture 15 Image Segmentation을 참고했습니다. cs231n 강의를 몇 개월째 공부하고 있지만(언제 끝낼거니^^), 개인적으로 후반부터는 EECS의 강의와 강의 자료가 더 상세히 설명되어 있고, 좋다고 느껴졌습니다. cs231n의 기원과 원천은 결국 Justin의 영혼에서 비롯된 것이 아닌가? 하는 생각. 아무튼 위 4개의 슬라이드를 바탕으로 정리한 글이며, 특히 이 글에서 첨부된 이미지는 강의 자료를 바탕으로 직접 제작한 PPT 이미지라서 이미지 사..
이번 포스팅은 standford university의 cs231 lecture 8를 공부하고, 강의와 슬라이드를 바탕으로 정리한 글임을 밝힙니다. What's going on inside ConvNets? 이전 lecture에서 CNN Architecture를 살펴보고, 지금까지는 어떻게 CNN을 학습시킬 것인지를 배웠고, 다양한 문제를 풀기 위해서 CNN Architecture를 어떻게 설계하고 조합해야 하는지를 배웠습니다. 그렇다면 Convolution Neural Networks 안에서는 무슨 일이 일어나고 있을까요? CNN의 내부는 어떻게 생겼을까요? CNN이 어떤 종류의 feature를 찾고 있는 것일까요? 이번 Lec8의 내용은 Visualizing and Understanding입니다. 컨볼..
- Total
- Today
- Yesterday
- 구글드라이브다운
- Prompt
- prompt learning
- few-shot learning
- python
- 도커 컨테이너
- 구글드라이브연동
- NLP
- vscode 종료
- 프롬프트
- stylegan
- CNN
- 서버구글드라이브연동
- 데이터셋다운로드
- vscode 업데이트
- vscode 저장
- 딥러닝
- docker
- vscode 자동 저장
- support set
- Unsupervised learning
- clip
- 구글드라이브서버연동
- 서버에다운
- cs231n
- style transfer
- 퓨샷러닝
- 도커
- 구글드라이브서버다운
- 파이썬
일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
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 |