多类分类的模型推理问题
我培训了用于多元分类任务的CNN模型。借助培训，验证和测试准确性（全部达到总计约88％），没有过度拟合，每个班级的准确性都超过80％。但是，当我仅根据一个类输入数据时，该模型给出了完全不同的结果（错误标签）。假设我将大豆标记为“ 2”，预先培训的模型最终将全部'2'as'5'视为'5'，这是另一个类别：建筑区域。当我输入所有类，从“ 1'to'6”中，每个类的分类准确性都像以前一样高出80％。
我真的很困惑。所有类别或单个类的输入的数据的所有预处理步骤都是相同的。在这种情况下，我无法从模型中对整个地图进行预测，因为它导致了巨大的错误分类像素。
I trained a CNN model for multivariate classification task. With training, validation and testing accuracy (all reached to around 88% overall), there's not overfitted, and the accuracy for every class is above 80%. However, when I only input data based on one of classes, the model gives exactly different results (wrong label). Say I labelled soybean '2', and the pre-trained model finally considered all '2' as '5' which is another class: building area. When I input all classes, from '1'to'6', the classification accuracy for each class was above 80% as before.
I am really confused. All pre-processing steps of data for either input of all class or single class are same. In this case, I cannot make prediction for the whole map from the model since it led to huge misclassified pixels.
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论