Implement eagerpy.diag

[Holt59]

Closes #21

Implements eagerpy.diag, following numpy / torch convention:

• Given a one-dimensional tensor, returns a two-dimensional tensor whose diagonal contains the input tensor values, and all other entries are 0. The second (optional) parameter indicates which diagonal to set.
• Given a two-dimensional tensor, returns a one-dimensional tensor containing the values in the diagonal of the input tensor. The second (optional) parameter indicates which diagonal to get.

This cannot handle batched inputs like Tensorflow linalg.diag or linalg.diag_part, but I think it's cleaner that way, let me know what you think.

[Holt59]

So apparently there is a bug in the version of tensorflow used for the test (2.1.2?). The k parameter for tf.linalg.diag and tf.linalg.diag_part is completely neglected.

• https://github.com/tensorflow/tensorflow/issues/38434
• https://github.com/tensorflow/tensorflow/issues/37492
• https://github.com/tensorflow/tensorflow/issues/36493
• https://github.com/tensorflow/tensorflow/issues/35761

[jonasrauber]

Let's call the method diag not _diag. No need limit diag to ep.diag. Might be nice to chain diag like other methods.

[Holt59]

Let's call the method diag not _diag. No need limit diag to ep.diag. Might be nice to chain diag like other methods.

Done. I've also reverted the change to test_transpose_1d I made by mistake.

[jonasrauber]

Thanks! Seems a couple of tests fail because the frameworks return inconsistent results.

[Holt59]

Thanks! Seems a couple of tests fail because the frameworks return inconsistent results.

Yes, as I said in a previous comment, the tensorflow version used in the CI tests has a bug regarding the k parameter in tf.linalg.diag (see the link in the previous comment).

[Holt59]

I've rebased the PR from master to get the upgraded TF in tests, hopefully this should fix the bug previously mentioned.