draft: linear -> spline interp for waveform data #162
Conversation
Add spline interpolation
|
@deepchatterjeeligo can you comment on the tolerance increases for the tests |
| @@ -225,16 +225,16 @@ def test_phenom_d( | |||
| hc_torch = hc_torch[torch_mask] | |||
|
|
|||
| assert np.allclose( | |||
| 1e21 * hp_lal_data.real, 1e21 * hp_torch.real.numpy(), atol=2e-4 | |||
| 1e21 * hp_lal_data.real, 1e21 * hp_torch.real.numpy(), atol=3e-4 | |||
There was a problem hiding this comment.
I could not make the tests pass without bumping the threshold a bit, which is strange given that spline interpolation should work better than linear. Any thought whether tweaking the kx, sx values can help? CC @wbenoit26
There was a problem hiding this comment.
Honestly, I might steer clear of using the spline interpolation for this application for now. There are significant deviations from scipy's interpolation if you interpolate too close to the edges of your input range. This is workable for something like the Q-transform, where we can crop out those effects, but that's not something that you can do here. I'd be curious to see if the match is better for values well within the range, though.
There was a problem hiding this comment.
OK. I'll leave this marked as draft until the sanity is met.
deepchatterjeeligo
changed the title
lalsimulation uses spline interpolation for the phenom data. We have been using linear interpolation due to lacking implementation which we have now.