ZYT

this problem happens when u highlight some words/painting in the pdf with Edge

"ValueError: `x0` is infeasible" means initial out of bound when using lstsq this may occur when input xy

``` File "atrack.py", line 170, in lines = asteroids.detect_lines(outdir, fitsdir) ``` Because lines=[] is empty, error turns out when send to pool.map

I suggest to change from ``` # Set rotation matrix (use PC instead of CROTA2 since PC is the # recommended approach) pc = np.array([[np.cos(angle), -np.sin(angle)], [np.sin(angle), np.cos(angle)]]) wcs_final.wcs.pc =...

lol... I'm misleaded by `WCS.to_header` when using SIP

Maybe multiprocessing pool scheduler conflicts with dask scheduler if `parallel` is set `True` in 'reproject_interp'

If anyone has tested `cupyx.scipy.ndimage.map_coordinates`? `scipy.ndimage.map_coordinates` seems to be much slower than `cv2.remap` in fits larger than 8k×8k maybe gpu acceleration is prefered here, but writing cp.array from gpu back...

I also tried `pytorch` but not using `map_coordinates`. I used `F.grid_sample` , found it a bit slower than `cupyx`. Well both tensor and cupy.array need to send back to CPU...

[array_utils.txt](https://github.com/user-attachments/files/19428725/array_utils.txt) [common.txt](https://github.com/user-attachments/files/19428724/common.txt) [core.txt](https://github.com/user-attachments/files/19428726/core.txt) I managed to write the `cupy` gpu 'acceleration' version of function `reproject_interp` above. But the speed still needs to carefully modified... Currently 8s for a 8K×8K image...

same problem with numpy 2.2.2