search

neel-dey / robust-nmf

Python PyTorch (GPU) and NumPy (CPU)-based port of Févotte and Dobigeon's robust-NMF algorithm appearing in "Nonlinear hyperspectral unmixing with robust nonnegative matrix factorization."
GNU General Public License v3.0
59 stars 13 forks source link

关于端元个数的设置 #1

Closed Scr-w closed 3 years ago

Scr-w commented 3 years ago

请问关于端元个数是怎么设置的呢

neel-dey commented 3 years ago

Assuming that this translates to "how do you choose the number of endmembers" (I can't speak/read Mandarin, sorry): that question is really specific to the dataset and is best decided in collaboration with domain-experts.

If it's truly blind/unknown, several automated heuristics to determine the number of components in a low-rank factorization or soft-clustering problem exist. E.g. HySime for hyperspectral imaging, minimum description length, automatic relevance determination if using bayesian methods, the elbow method, etc. Hope that helps.

Scr-w commented 3 years ago

I see. Thank you very much

------------------ 原始邮件 ------------------ 发件人: "Neel @.>; 发送时间: 2021年4月17日(星期六) 凌晨2:51 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [neel-dey/robust-nmf] 关于端元个数的设置 (#1)

假设这转换为“如何选择最终成员的数量”（抱歉，我不会说/读普通话）：这个问题实际上是针对数据集的，最好与领域专家合作决定。

如果它是真正的盲/未知，一些自动启发式，以确定在低秩分解或软聚类问题的组件的数量存在。例如，用于高光谱成像的 Hy Sime ，最小描述长度，如果使用贝叶斯方法，肘法等的自动相关性确定。希望对你有帮助

— 您正在接收此消息，因为您编写了线程。 直接回复此邮件，在GitHub上查看，或取消订阅.