AttributeError: module 'numpy' has no attribute 'float128'

[hjort]

Describe the bug

Got the following issue:

/opt/anaconda3/lib/python3.6/site-packages/autosklearn/automl.py in subsample_if_too_large(X, y, logger, seed, memory_limit, task)
820 elif X.dtype in (np.float64, np.float):
821 multiplier = 8
--> 822 elif X.dtype == np.float128:
823 multiplier = 16
824 else:
AttributeError: module 'numpy' has no attribute 'float128'

To Reproduce

1. On a Debian Linux 32-bit, with Python 3.6.7 (Anaconda 32-bit)
   Linux 3.13.0-170-generic Error on Refit with ProjLogit #220-Ubuntu SMP Thu May 9 12:41:17 UTC 2019 i686 athlon i686 GNU/Linux
2. The following code:
   import autosklearn.classification
   automl = autosklearn.classification.AutoSklearnClassifier(
   time_left_for_this_task=2060,
   per_run_time_limit=260,
   n_jobs=2,
   memory_limit=1024,
   seed=42,
   include_preprocessors=["no_preprocessing"]
   )
   automl.fit(X_train, y_train)

Expected behavior

The code should work regardless of 32 or 64-bit architectures, as Scikit-Learn does.

Environment and installation:

Please give details about your installation:

• OS: Ubuntu 32-bit
• Is your installation in a virtual environment or conda environment? Anaconda
• Python version: 3.6.7
• Auto-sklearn version: 0.12.5

[mfeurer]

Thanks for reporting this issue @hjort. Sorry for the trouble, I was not aware of the fact that np.float128 is not available on 32 bit systems. We would appreciate a PR fixing this issue.

[mfeurer]

Hey @hjort we just merged a fix to the development branch. Please reopen this issue if the bug persists with that cange.