admin管理员组

文章数量:1530018

Integrated learning in machine learning can improve prediction accuracy to a certain extent. Common integrated learning methods include Stacking, Bagging, and Boosting. At the same time, these integrated learning methods are closely related to the division of specific verification sets.

Since deep learning models generally require a long training period, if the hardware device does not allow it, it is recommended to choose the set aside method, and if you need to pursue accuracy, you can use the cross-validation method.

The following assumes that a 10-fold cross-validation is constructed and 10 CNN models are trained.

  • The probability values of the predicted results are averaged and then decoded into specific characters;
  • Voting on the predicted characters to get the final character.
    There may be different solutions in different tasks. Models with different ideas can not only learn from each other, but also modify the final prediction results.
Post Process

In this competition, the prediction results can be post-processed from the following ideas:

  • Count the frequency of characters in each position in the picture, and use the rules to modify the results;
  • Train a character length prediction model separately to predict the number of characters in the picture and correct the result.
Summary

In this chapter, we explained various methods of deep learning models for integrated learning, and explained some of the code with the title of this competition as an example.

-Integrated learning can only improve accuracy to a certain extent, and requires a large amount of training time, so it is recommended to use first to improve the accuracy of a single model before considering the integrated learning process;
-The specific integrated learning method needs to be combined with the verification set division method. Dropout and TTA can work in all scenarios.

END

本文标签： ViewPicturecompetitionStreetensemble

版权声明：本文标题：Competition Street view picture character recognition 5: ensemble model 内容由热心网友自发贡献，该文观点仅代表作者本人， 转载请联系作者并注明出处：https://m.elefans.com/dianzi/1726693498a1080932.html， 本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如发现本站有涉嫌抄袭侵权/违法违规的内容，一经查实，本站将立刻删除。