Add slogdet implementation

[eserie]

Fixes #27

[codecov]

Codecov Report

Merging #28 (1e4cbb9) into master (6ef7313) will not change coverage.
The diff coverage is 100.00%.

@@            Coverage Diff            @@
##            master       #28   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           16        16           
  Lines         1739      1753   +14     
=========================================
+ Hits          1739      1753   +14     

Impacted Files	Coverage Δ
eagerpy/tensor/tensor.py	100.00% <ø> (ø)
eagerpy/framework.py	100.00% <100.00%> (ø)
eagerpy/tensor/jax.py	100.00% <100.00%> (ø)
eagerpy/tensor/numpy.py	100.00% <100.00%> (ø)
eagerpy/tensor/pytorch.py	100.00% <100.00%> (ø)
eagerpy/tensor/tensorflow.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6ef7313...1e4cbb9. Read the comment docs.

[jonasrauber]

The implementation looks good.

We can almost merge this.

But I'd prefer if you could make a few changes to the tests:

• revert the changes to compare_allclose
• see the tests for topk regarding multiple return values (simply use a separate test for each return value; would also give a more specific error message)
• add a few more test cases that result in qualitatively different results (e.g., a matrix with a finite logdet)

[eserie]

Hi @jonasrauber , thanks for your feedback!
I modified the test as you required.
For the test on real values I considered the examples in the numpy documentation..
Let me know if you think something else should be accomodate.

[jonasrauber]

@eserie Looks good, but we had a small misunderstanding regarding the third change: no need to manually compare to the reference values (because the decorator does compare everything to numpy anyway). What I meant was that you should just duplicate the two tests (for sign and logdet) to have two for the original matrix you had and two for a matrix with finite logdet (so for simple test functions in the end).

[eserie]

@jonasrauber , thanks for the comment.
I factorized a bit the test using pytest fixtures. I hope that this keeps the code enough simple.
Also, I restored the infinite test case.

[jonasrauber]

@eserie thanks, that's also a nice solution!
I have to say, a bit surprised that our custom compare_all* decorator and pytest.mark.parametrize don't break each other (or, more precisely, that they work together in this order; would make more sense to me if it only works if the compare_all* decorator is the inner one (that turns the function with return value into a test function with asserts) … can you explain why it works the way you implemented it?

[eserie]

@jonasrauber, it's true that the test code seems to work for both orders of decorators. When they are used with pytest.mark.parametrize as inner decorator I would understand that pytest will generate a list of functions with arguments specified applied, like a functools.partial operation. However I do not fully understand the mechanism and we should go in the code of pytest to fully understand how decorators are managed.
Indeed, in the last commit I swapped the order of decorators and it works also. I can let the order you prefer.

[eserie]

Thanks for the integration @jonasrauber !