helmholtz-analytics / heat Goto Github PK

The unit test that checks whether ht.randn throws a ValueError depends on the fact that pytorch
shall detect malformed input. Unfortunately, this is not the case for my pytorch installation (trying to print the resulting tensor, however, results in an infinite loop, such that we can be sure that pytorch does not magically know how to handle negative dimensions ... )

For completeness the pytest output:

        with self.assertRaises(ValueError):
>           ht.randn(-1, 3, dtype=ht.float64)
E           AssertionError: ValueError not raised

heat/core/tests/test_tensor.py:278: AssertionError

There seems to be already some all_ints variable available in ht.randn, which could be augmented to also check for positive arguments?

Similar to numpy, it would be nice to have a array factory function to create tensors from local data.
Major implementation challenge is how to deal with imbalanced data distributions/initial passed arrays. Do we need to balance the data?

Need to set MACOSX_DEPLOYMENT_TARGET.

The function is a general purpose function and does not really belong into the clustering algorithm.

• Installation, requirements, etc.
• Goals
• License

• Allow rudimentary netCDF support akin to what the HDF5 layer can already do
• Offer a possibility to write netCDF files

Type: Classification, Ground Truth
Size: Should fit on GPU, max 6 GB

Apparently, our coverage report is not generated again on every build. It seems like the 'coverage' executable is not among the cached contents for build, but the Python library is.

https://numpy.org/doc/stable/reference/generated/numpy.convolve.html

Everytime I edit edit tensor.py and run pytest, pytest prints a warning that the docstring of ht.floor contains an unknown escape sequence \lfloor. Mysteriously, this warning does not show up in the second run of pytest, which might explain why this was not caught in CI yet.

We want to automatically test our code base for

• bugs
• performance (regressions)
• general functioning

As far as I know, travis-ci supports MPI (https://travis-ci.org/).

Possible design:

• An allocation parameter for all factory functions

import heat as ht
a = ht.zeros((10,2,), device='gpu<:id>')

• .cpu() and .gpu() methods for the tensor class that allow adhoc switching between devices
• global default device toggle, using a function call and/or a context manager

import heat as ht
ht.device('gpu<:id>')

#alternative
with ht.gpu(<':id'>):
    ht.zeros()

We need to support function astype

Required at

• https://github.com/NeuralEnsemble/elephant/blob/ed8ad8bf477e080c1b79c20a206b65131b85ffa0/elephant/asset.py#L280

PyPI requires both a description and long_description to be set, with the former being used for listing a project among others and the latter for the detailed project page.

http://www.sphinx-doc.org/en/master/

All factory function like asarray, zeros, ones etc must support a dtype argument as in

• https://github.com/NeuralEnsemble/elephant/blob/ed8ad8bf477e080c1b79c20a206b65131b85ffa0/elephant/asset.py#L873

This should be done from successful Travis CI build. Possibly only tagged versions. See https://docs.travis-ci.com/user/deployment/pypi/ for details.

inside Tensor.binop:

Case requried: Both arrays distributed along different axes.