[Keras Ops] view_as_complex() and view_as_real()

[DavidLandup0]

Convenience operation, used to convert between complex and real tensors.
Useful for doing operations in architectures like DeepSeek V3/R1 (keras-team/keras-hub#2171).

Equivalent to PyTorch's:

• https://pytorch.org/docs/stable/generated/torch.view_as_complex.html
• https://pytorch.org/docs/stable/generated/torch.view_as_real.html

Usage:

view_as_real() converts a complex-type tensor into a real tensor:

>>> from keras import ops
>>> import torch
>>> x=torch.randn(4, dtype=torch.cfloat)
>>> ops.view_as_real(x)
>>> ops.view_as_real(x).numpy()
array([[-0.6658956 , -1.199816  ],
       [ 0.6203052 , -0.647082  ],
       [ 0.6746265 , -0.75243765],
       [ 0.02358519, -0.8950158 ]], dtype=float32)

view_as_complex() does the opposite:

>>> x=torch.randn(4, 2)
>>> x
tensor([[-0.6625, -1.4961],
        [-0.0678, -0.6881],
        [ 0.1599,  0.1223],
        [ 1.8277,  0.1058]])
>>> ops.view_as_complex(x).numpy()
array([-0.6624739 -1.496139j  , -0.06781768-0.6881176j ,
        0.15992023+0.12225848j,  1.8276887 +0.10576057j], dtype=complex64)

/cc @fchollet for review

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 73.68421% with 10 lines in your changes missing coverage. Please review.

Project coverage is 82.60%. Comparing base (37eacb0) to head (4ef4b65).
Report is 12 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/src/ops/math.py	76.47%	8 Missing ⚠️
keras/api/_tf_keras/keras/ops/__init__.py	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #21221      +/-   ##
==========================================
- Coverage   82.61%   82.60%   -0.02%     
==========================================
  Files         564      564              
  Lines       54476    54581     +105     
  Branches     8470     8476       +6     
==========================================
+ Hits        45005    45084      +79     
- Misses       7395     7412      +17     
- Partials     2076     2085       +9     

Flag	Coverage Δ
keras	82.41% <73.68%> (-0.02%)	⬇️
keras-jax	63.67% <73.68%> (-0.05%)	⬇️
keras-numpy	58.80% <73.68%> (-0.04%)	⬇️
keras-openvino	32.99% <31.57%> (+<0.01%)	⬆️
keras-tensorflow	64.09% <73.68%> (-0.03%)	⬇️
keras-torch	63.75% <73.68%> (-0.05%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[DavidLandup0]

The torch tests don't like the dtype comparisons

AssertionError: torch.float32 != dtype('float32')

Is there a Keras-preferred way to handle this case or should I just find a reasonable way to do the comparison instead?

[fchollet]

Is there a Keras-preferred way to handle this case or should I just find a reasonable way to do the comparison instead?

Please use standardize_dtype before the comparison

[fchollet]

Thanks for the PR!

These functions do not appear to be NumPy operations, so at least they should not go in the keras.ops.numpy API (and should not be part of numpy.py).

Is there a standard name we can use? Maybe something in NumPy?

[DavidLandup0]

Strictly speaking, not Numpy operations, but they make use of real() and imag(), which are currently under numpy.py in Keras, hence the position in the PR. Open to putting them anywhere else - perhaps math.py?

Is there a standard name we can use? Maybe something in NumPy?

Numpy offers real() and imag() to extract components of complex arrays, but doesn't have a utility for converting between real and complex. The most standard name I can think of is the PyTorch one, since that's where the utility originated.

We could name it something like as_complex() or as_real(), but a more common name is probably going to be better for dev UX.

[fchollet]

Sounds good, let's move them to math.py.

numpy.py is strictly for actual numpy APIs.

[DavidLandup0]

Sounds good, let's move them to math.py.

numpy.py is strictly for actual numpy APIs.

Thanks for weighing in! Moved it to math.py.

[fchollet]

There is not math namespace -- just export to ops

[DavidLandup0]

Whoops - good point. Accidentally left it in due to the original keras.ops.numpy location.
Removed!