TypeError on import TypeError: data type "float128" not understood

[karnn]

On Windows, Anaconda 4.8.3. Created conda environment with conda env create -f environment.yml inside Rethinking_2 directory.

Doing import pymc3 gives the following error

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-1-75554ce6e5eb> in <module>
      2 import matplotlib.pyplot as plt
      3 import numpy as np
----> 4 import pymc3 as pm
      5 import scipy.stats as stats

~\Miniconda3\envs\stat-rethink2-pymc3\lib\site-packages\pymc3\__init__.py in <module>
     49 
     50 from .blocking import *
---> 51 from .distributions import *
     52 from .distributions import transforms
     53 from .glm import *

~\Miniconda3\envs\stat-rethink2-pymc3\lib\site-packages\pymc3\distributions\__init__.py in <module>
     13 #   limitations under the License.
     14 
---> 15 from . import timeseries
     16 from . import transforms
     17 from . import shape_utils

~\Miniconda3\envs\stat-rethink2-pymc3\lib\site-packages\pymc3\distributions\timeseries.py in <module>
     20 
     21 from pymc3.util import get_variable_name
---> 22 from .continuous import get_tau_sigma, Normal, Flat
     23 from .shape_utils import to_tuple
     24 from . import multivariate

~\Miniconda3\envs\stat-rethink2-pymc3\lib\site-packages\pymc3\distributions\continuous.py in <module>
     30 from .special import log_i0
     31 from ..math import invlogit, logit, logdiffexp
---> 32 from .dist_math import (
     33     alltrue_elemwise, betaln, bound, gammaln, i0e, incomplete_beta, logpow,
     34     normal_lccdf, normal_lcdf, SplineWrapper, std_cdf, zvalue,

~\Miniconda3\envs\stat-rethink2-pymc3\lib\site-packages\pymc3\distributions\dist_math.py in <module>
     35 f = floatX
     36 c = - .5 * np.log(2. * np.pi)
---> 37 _beta_clip_values = {
     38     dtype: (np.nextafter(0, 1, dtype=dtype), np.nextafter(1, 0, dtype=dtype))
     39     for dtype in ["float16", "float32", "float64", "float128"]

~\Miniconda3\envs\stat-rethink2-pymc3\lib\site-packages\pymc3\distributions\dist_math.py in <dictcomp>(.0)
     36 c = - .5 * np.log(2. * np.pi)
     37 _beta_clip_values = {
---> 38     dtype: (np.nextafter(0, 1, dtype=dtype), np.nextafter(1, 0, dtype=dtype))
     39     for dtype in ["float16", "float32", "float64", "float128"]
     40 }

TypeError: data type "float128" not understood

[AlexAndorra]

Hi,
That's strange -- it works for me 🤔
Do you know which versions of python and numpy are on your env?

[karnn]

Hi
Thanks for your reply. I just checked and my Python is at 3.8.2. I suppose it's too recent? I directly installed from the yml file in the repo. It's strange. Numpy is 1.18.4

[lucianopaz]

Oh damn! It's a windows bug. Numpy under windows doesn't have a float128, we should use the longdouble dtype instead. That will be equal to float128 on Linux but it will equal float64 on windows. It we could even just remove float128. @karnn, could you please open an issue on pymc3? I'll fix it in the next few days.

[AlexAndorra]

These OS bugs keep coming 🤦 It's indeed related to your latest PR @lucianopaz.
Assign me to review the fix when you submit it -- or tell me if you don't have time in the coming days and I can take a stab (looks pretty simple).
Closing this in favor of the issue on PyMC repo.

[fonnesbeck]

I'm getting this on MacOS (M1 hardware) using miniforge. This is with NumPy 1.21.0.

ImportError while loading conftest '/Users/cfonnesbeck/GitHub/pymc3/pymc3/tests/conftest.py'.
pymc3/__init__.py:82: in <module>
    from pymc3 import gp, ode, sampling
pymc3/gp/__init__.py:16: in <module>
    from pymc3.gp.gp import TP, Latent, LatentKron, Marginal, MarginalKron, MarginalSparse
pymc3/gp/gp.py:25: in <module>
    from pymc3.distributions import draw_values
pymc3/distributions/__init__.py:15: in <module>
    from pymc3.distributions import shape_utils, timeseries, transforms
pymc3/distributions/timeseries.py:21: in <module>
    from pymc3.distributions import distribution, multivariate
pymc3/distributions/multivariate.py:35: in <module>
    from pymc3.distributions.continuous import ChiSquared, Normal
pymc3/distributions/continuous.py:30: in <module>
    from pymc3.distributions.dist_math import (
pymc3/distributions/dist_math.py:49: in <module>
    np.nextafter(0, 1, dtype="float128"),
E   TypeError: data type 'float128' not understood

[ogrisel]

For reference, on macOS ARM64 (Apple Silicon M1), float128 is not supported on numpy (at least when installed from conda-forge):

>>> import numpy as np
>>> np.dtype("float128")
Traceback (most recent call last):
  File "<ipython-input-5-10c53c28af62>", line 1, in <module>
    np.dtype("float128")
TypeError: data type 'float128' not understood

So it would be great to make float128 support an optional requirement for PyMC3 and pymc.

[ricardoV94]

Which version are you using? If I recall correctly our next release dos not use float128 anywhere, so it should not be a problem.

[elbamos]

I'm seeing this also. The most recent release of pymc3 that's available through conda on m1 is 3.9.3, because the versions of theano-pymc required for later versions of pymc3 aren't available for m1. At least that's what conda is telling me.

I was able to workaround this by downgrading pymc3 to 3.8.

Edit: theano-pymc and therefore pymc3 3.11.4 are now available on conda for m1, so I think this issue is resolved.