英文:
What is the cause of "RuntimeWarning: invalid value encountered in matmul ret = a @ b" when using sklearn?
问题
在使用 Ridge().fit(X_train, y_train) 时，我遇到了 RuntimeWarning，这发生在一个数据集上，其中 X_train.shape = (9440, 1900)。
当我减小数据集的大小，使得 X_train.shape = (1000, 1900) 时，RuntimeWarning 就消失了。
可能导致这个警告的原因是什么，我该如何避免它？我确保没有出现 np.nan、np.inf 和 -np.inf。
英文:
I encountered RuntimeWarning when using Ridge().fit(X_train, y_train) on a dataset where X_train.shape = (9440, 1900)
/Users/username/micromamba/envs/data-science/lib/python3.11/site-packages/sklearn/utils/extmath.py:189: RuntimeWarning: invalid value encountered in matmul
ret = a @ b
When I reduce the size of dataset X_train.shape = (1000, 1900) the RuntimeWarning goes away.
What might be causing this warning and how can I avoid it? I made sure there are no np.nan, np.inf, and -np.inf.
答案1
得分: 4
这个错误有点知名度，不幸的是，它与架构和环境相关。这在numpy中是一个纯粹的矩阵操作问题。我在我的Colab中使用以下代码复现了这个问题：
# 在numpy==1.25下有警告，在numpy==1.22.4下正常
M = np.random.random((30000, 270)) # 在Colab上运行
M.T @ M
@Stef测试了导致numpy版本崩溃的问题。
请查看这个最近的GitHub讨论，链接为 this recent GitHub discussion，我在其中提供了这个导致问题的示例。
我创建了这个Colab笔记本来帮助找出问题的根本原因（在Colab的选择中，numpy==1.22.4正常运行）。您可以在Colab中进行训练（可以上传您的数据）。
总的建议是降级numpy版本或等待维护者修复此问题。
英文:
This error is somewhat known, and is sadly architecture / environment dependent. This is a purely matrix op issue in numpy. I reproduced it in my Colab with the following code:
# warning under numpy==1.25, OK under numpy==1.22.4
M = np.random.random((30000, 270)) # on Colab
M.T @ M
and @Stef tested the breaking numpy version.
See this recent GitHub discussion under numpy, where I contributed this breaking example.
I created this Colab notebook to help root-causing the issue (works fine under numpy==1.22.4 which is Colab's choice). Try your training under Colab (you can upload your data).
Overall my suggestion would be to downgrade the numpy version or wait for a fix from the maintainers.
通过集体智慧和协作来改善编程学习和解决问题的方式。致力于成为全球开发者共同参与的知识库，让每个人都能够通过互相帮助和分享经验来进步。
评论