ENH: Add Scipy and Cupy as fft interfaces #8
Conversation
|
Hey @paulmueller, let me know if you think this fits in the |
|
|
||
| sphinx | ||
| sphinxcontrib.bibtex | ||
| sphinx_rtd_theme |
There was a problem hiding this comment.
I removed the pinned versions, as some sphinx dependency did not like the old sphinx version.
| n_transforms = 100 | ||
|
|
||
| # one transform | ||
| results_1 = {} |
There was a problem hiding this comment.
I need to clean this script up a bit to make it simpler
| @@ -104,8 +104,10 @@ def get_filter_array(filter_name, filter_size, freq_pos, fft_shape): | |||
| # TODO: avoid the np.roll, instead use the indices directly | |||
| alpha = 0.1 | |||
| rsize = int(min(fx.size, fy.size) * filter_size) * 2 | |||
| tukey_window_x = signal.tukey(rsize, alpha=alpha).reshape(-1, 1) | |||
| tukey_window_y = signal.tukey(rsize, alpha=alpha).reshape(1, -1) | |||
| tukey_window_x = signal.windows.tukey( | |||
There was a problem hiding this comment.
scipy's new versions imports tukey from signal.window, not signal
| FFTFilterPyFFTW, | ||
| FFTFilterNumpy, | ||
| FFTFilterScipy, | ||
| FFTFilterCupy, |
There was a problem hiding this comment.
This isn't necessarily the "perferred order" we want, but I'd like to keep it as is due to old default pipelines.
| @@ -70,8 +70,11 @@ def __init__(self, data, subtract_mean=True, padding=2, copy=True): | |||
| else: | |||
| # convert integer-arrays to floating point arrays | |||
| dtype = float | |||
| if not copy: | |||
| # numpy v2.x behaviour requires asarray with copy=False | |||
There was a problem hiding this comment.
numpy has made copying a bit awkward. When an array can't be copied (with either np.array or np.asarray), and the user has copy=False, then it throws an error!
| data_gpu = cp.asarray(data) | ||
| # likely an inefficiency here, could use `set_global_backend` | ||
| with sp.fft.set_backend(cufft): | ||
| fft_gpu = sp.fft.fft2(data_gpu) |
There was a problem hiding this comment.
Not ideal, this is something to work on.
| "FFTW": "pyfftw>=0.12.0", | ||
| # manually install 'cupy-cuda11x' if you have older CUDA. | ||
| # See https://cupy.dev/ | ||
| "CUPY": "cupy-cuda12x", |
There was a problem hiding this comment.
Will have to check this version.
|
Hello, This is absolutely in the scope of qpretrieve! It would probably make sense to address all CI issues (GH actions, readthedocs) and dependencies (e.g. I would be fine with bumping numpy to >= 2.0) in a separate PR. I have converted this PR to a draft for now, so you can work on it in stealth mode. |
Sounds good, I will revert the actions changes. However, the docs requirements just don't build on my side because they are too restrictive.
🥇 |
| @@ -70,9 +75,13 @@ def __init__(self, data, subtract_mean=True, padding=2, copy=True): | |||
| else: | |||
There was a problem hiding this comment.
Here we check and allow complex data. But in the docstring above (originally), we specify real-valued inputs. In what situation would we take complex data?
About
I'd like to be able to use Cupy for Fourier transforms. Cupy is a python package that uses numpy-like syntax for using CUDA and GPU.
To do
FFTFilterCupy3D, so that the default input of a 2D image is kept by the default FFTFilter classes, and the user has to know what they are doing to useFFTFilterCupy3D. Combine the functionality with 2d because it uses numpy.run_pipelinesteps should also work with 3D arrays. Seems to fall down at line 86 inif_oah.pyfor now, as this is a section that must use 2D arrays.test_qlsi_unwrap_phase_2d_3d