Practical Advice for Using ConvNets
Using Open-Source Implementation
open-source implementation을 잘 활용하여라.
skeleton code를 가져와서 implementation을 배울수도, pre-trained model을 가져와서 transfer learning을 할 수도 있고 어쨌든 open-source implementation을 잘 활용하면 좋다!
Transfer Learning
쌩으로 model training을 하는 것은 시간도 많이 걸리고 hyperparameter을 tuning하는것도 힘들다.
다른 분들이 이미 training 시킨 model을 가져와서 weight를 고정시키고 transfer learning을 하자.
뒤에 있는 layer를 조금 바꿔도 되고, 이미 training된 weight를 가지고 initialization을 할 수도 있을 것이다.
Data Augmentation
mirroring, random cropping, ratation, shearing, local warping, color shifting, PCA color augmentation 등등
multithreading을 이용해서 다양하게 data augmentation을 하면서 training을 동시에 할 수 있을 것이다.
State of Computer Vision
little data -> more hand-engineering ("hacks")
lots of data -> simpler algorithms (less hand-engineering)
Two sources of knowlege: labeled data, hand engineered features / netwrok architecture / other components
Tips for doing well on benchmarks / winning competitions: ensembling (voting, ...), multi-crop at test time (convert one input to multiple inputs)
Use architectures of networks published in the literature, open source implementation if possible, pretrained models and fine-tune on your dataset
'Google Machine Learning Bootcamp 2022 > Convolutional Neural Networks' 카테고리의 다른 글
4. Special Applications: Face recognition & Neural Style Transfer #2 (0) | 2022.08.02 |
---|---|
4. Special Applications: Face recognition & Neural Style Transfer #1 (0) | 2022.08.02 |
3. Object Detection (0) | 2022.08.02 |
2. Deep Convolutional Models: Case Studies #1 (0) | 2022.08.01 |
1. Foundations of Convolutional Neural Networks (0) | 2022.07.28 |
댓글