Locally Linear Embedding manifold doesn't work with integer inputs #6033
Description
An error occurs when you try to use manifold.LocallyLinearEmbedding with an integer array as the input if you have certain values of the method parameter. I'm running the latest version of scikit-learn (I got this error, updated through conda, and the error still occurs).
Traceback (most recent call last):
File "C:/Users/jswanson/Downloads/plot_compare_methods.py", line 63, in <module>
method=method).fit_transform(X)
File "C:\Users\jswanson\AppData\Local\Continuum\Anaconda3\lib\site-packages\sklearn\manifold\locally_linear.py", line 652, in fit_transform
self._fit_transform(X)
File "C:\Users\jswanson\AppData\Local\Continuum\Anaconda3\lib\site-packages\sklearn\manifold\locally_linear.py", line 623, in _fit_transform
random_state=random_state, reg=self.reg)
File "C:\Users\jswanson\AppData\Local\Continuum\Anaconda3\lib\site-packages\sklearn\manifold\locally_linear.py", line 476, in locally_linear_embedding
Xi -= Xi.mean(0)
TypeError: Cannot cast ufunc subtract output from dtype('float64') to dtype('int64') with casting rule 'same_kind'
For some code that demonstrates the error, you can run the example here http://scikit-learn.org/stable/auto_examples/manifold/plot_compare_methods.html, but cast the data to int64 first. You can do this by inserting the following line below line 38
X = (1000 * X).astype('int64')
By glancing through similar issues, I think this is caused by an update to numpy.
I'll just work around the issue by casting my data to floats, but I thought I should still bring it up.