Remember a deep learning game

  • As a loser ~, did not participate halfway, the end did not find why. Now copy the answers you ~
  • For Huawei cloud contest
    1. When using various models found that almost does not rise point of time, it should know that the problem is not a model, but of data, data should be expanded by up point.
    2. There are a lot of data expansion method, and draw the confusion matrix, it is difficult to distinguish the type of targeted augmented
    3. Recommended joined CBAM in the network.
    4. Other trick: mix_up (conducted regular on the data), weighted loss, increase picture resolution, tag smooth (to inhibit over-fitting), random erase, multi-scale training (increasing generalization model).
  • Reference: https://github.com/xyxy875/huawei_xian_image_classification , https://github.com/jackcywang/Xian-Huawei-Al-competion

Guess you like

Origin www.cnblogs.com/SsoZhNO-1/p/12459738.html