Bug with float128 and not with float64?

[doronbehar]

The following MWE:

from pwlf import PiecewiseLinFit
import numpy as np

fit = PiecewiseLinFit(np.linspace(0, 10, dtype=np.float128), np.random.random(50))
fit.fit(2)
fit.p_values()

Errors for me with:

Traceback (most recent call last):
  File "/home/doron/repos/lab-ion-trap-simulations/./pwfit_float128_bug.py", line 8, in <module>
    print(fit.p_values())
          ^^^^^^^^^^^^^^
  File "/nix/store/8p528p5cji81aidpbpwiyyz5s9nb1lz4-python3-3.12.8-env/lib/python3.12/site-packages/pwlf/pwlf.py", line 1484, in p_values
    p = 2.0 * stats.t.sf(np.abs(t), df=n - k - 1)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/nix/store/8p528p5cji81aidpbpwiyyz5s9nb1lz4-python3-3.12.8-env/lib/python3.12/site-packages/scipy/stats/_distn_infrastructure.py", line 2203, in sf
    place(output, cond, self._sf(*goodargs))
                        ^^^^^^^^^^^^^^^^^^^
  File "/nix/store/8p528p5cji81aidpbpwiyyz5s9nb1lz4-python3-3.12.8-env/lib/python3.12/site-packages/scipy/stats/_continuous_distns.py", line 7716, in _sf
    return sc.stdtr(df, -x)
           ^^^^^^^^^^^^^^^^
TypeError: ufunc 'stdtr' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''

However, without the dtype=np.float64, no error.

[cjekel]

@doronbehar thank you for the simple reproducer.

It looks like scipy does not fully support float128, and many of the scipy.stats methods are float64 only (e.g. scipy.stats.t.sf).

How would you like this calculation to work?

I could cast down to float64 to get an answer?

I just added smoke tests for float32 and float128 support due to this ba76791

[doronbehar]

Hmm maybe create a MWE for scipy and create a bug report there? Untill that will be solved, we can downcast to float64 in pwlf...

…

On 31 December 2024 23:19:45 GMT+02:00, Charles Jekel ***@***.***> wrote: @doronbehar thank you for the simple reproducer. It looks like scipy does not fully support float128, and many of the scipy.stats methods are float64 only (e.g. `scipy.stats.t.sf`). How would you like this calculation to work? I could cast down to float64 to get an answer? I just added smoke tests for float32 and float128 support due to this ba76791 -- Reply to this email directly or view it on GitHub: #119 (comment) You are receiving this because you were mentioned. Message ID: ***@***.***>

[cjekel]

here is a scipy reproducer

import numpy as np
import scipy
scipy.stats.t.sf(np.random.random(3).astype(np.float128), df=3)

[doronbehar]

Hmm I wonder if scipy devs will simply reply to this: "What do you want, the error message explicitly states that this input type is not supported". Maybe pwlf should simply downcast floats 128 to floats 64?

[doronbehar]

Or on a 2nd thought, perhaps Scipy should downcast floats 128 to floats 64 with a warning instead of failing.