when use pwlf, I encounter errors: TypeError: must be real number, not NoneType about piecewise_linear_fit_py HOT 11 CLOSED

You can now manually specify rcond when you initialize pwlf. bf67d7d

I will push pwlf==0.4.3 to pypi shortly.

my_pwlf = pwlf.PiecewiseLinFit(x, y, rcond=-1)

The default is still rcond=None.

from piecewise_linear_fit_py.

Can you give me the versions of numpy and pwlf?

pwlf.__version__
np.__version__

from piecewise_linear_fit_py.

pwlf.version 0.3.3
np.version 1.14.3

from piecewise_linear_fit_py.

I'll have to create an environment with numpy==1.14.3 later today. I know numpy<1.14 would break because they changed how rcond is handled.

In the meantime is it possible for you to update numpy to at least numpy>=1.15.4 and see if that fixes your problem?

I think the anaconda command within your environment would be

conda update numpy 

Can you run this simple least squares fit? Or do you get the same error?

# https://stackoverflow.com/questions/27128688/how-to-use-least-squares-with-weight-matrix-in-python
import numpy as np
A=np.array([[1,1,1,1],[1,1,1,1],[1,1,1,1],[1,1,1,1],[1,1,0,0]])
B = np.array([1,1,1,1,1])
W = np.array([1,2,3,4,5])
Aw = A * np.sqrt(W[:,np.newaxis])
Bw = B * np.sqrt(W)
X = np.linalg.lstsq(Aw, Bw, rcond=None)

from piecewise_linear_fit_py.

Thank you! By upgrading my numpy to 1.15.4, the problem disappears! 👍

from piecewise_linear_fit_py.

I'm glad you got it working!

I tried setting up a numpy==0.14.3 environment and couldn't reproduce. I think something was wrong with your numpy build =/

Edit. I should note that I was expecting this error to occur from numpy<=0.14. Perhaps I should I add rcond as an optional parameter in the class...

from piecewise_linear_fit_py.

FYI, I keep seeing the same error with the pwlf==0.4.2 and numpy==1.16.2.

Running the code from above produces

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-2-e4acef74a3ed> in <module>()
      5 Aw = A * np.sqrt(W[:,np.newaxis])
      6 Bw = B * np.sqrt(W)
----> 7 X = np.linalg.lstsq(Aw, Bw, rcond=None)

/opt/conda/lib/python3.6/site-packages/numpy/linalg/linalg.py in lstsq(a, b, rcond)
   1913         work = zeros((lwork,), t)
   1914         results = lapack_routine(m, n, n_rhs, a, m, bstar, ldb, s, rcond,
-> 1915                                  0, work, -1, iwork, 0)
   1916         lwork = int(work[0])
   1917         work = zeros((lwork,), t)

TypeError: must be real number, not NoneType

changing rcond=-1 a suggested here makes it work again.

from piecewise_linear_fit_py.

@stklik Are you absolute sure that you are running numpy==1.16.2? If so, you should raise an issue on numpy that recommended default of rcond=None does not work with all numpy backends.

Some old discussion on the topic:
numpy/numpy#11015
numpy/numpy#9582

In terms of pwlf, I could make a parameter in __init__ which sets and option for rcond, allowing for backwards compatibility.

from piecewise_linear_fit_py.

so, my pip freeze | grep numpy tells me it's 1.16.2.
I'm running inside a Docker image, so I'll try to check if there are any issues with the installations.

I will push pwlf==0.4.3 to pypi shortly.

Thanks, this will certainly solve my problem.

from piecewise_linear_fit_py.

@cjekel After further investigation, I am fairly convinced that my issue has to do with the Jupyter dockerstacks and the fact that they use conda to install numpy while pwlf setup triggers a pip install + some access rights issues.
I will investigate and raise it with the people over there if I find out what to do.

PS: Thanks for adding the rcond param, it works for me now, despite "my situation".

--> can be closed from my POV

from piecewise_linear_fit_py.

Having multiple Python instances is a pain. The pip in your docker setup is probably for the OS's python. I normally run

which pip
which pip3
which pip3.6

to give some idea where the pip files are located. It's possible that one of those is your conda pip. You may have more options... Feel free to email me if you can't seem to get it sorted out.

You can also try

python -m pip install pwlf

to make sure that this Python environment is the same as pip.

If anything rcond adds backwards compatibility for someone stuck on numpy <= 1.14. :)

from piecewise_linear_fit_py.