torch-audiomentations
torch-audiomentations copied to clipboard
asteroid-team
No fast pitch-shift ratios could be computed for the given sample rate and transpose range.
I use PitchShift wih min=-0.9 and max = 1.1, then it raise No fast pitch-shift ratios could be computed for the given sample rate and transpose range.. I dont face with this error when use cpu version github. Do you have any plan to fix it?
And what do you mean with cpu version github? Are you refering to https://github.com/iver56/audiomentations?
And what do you mean with cpu version github? Are you refering to https://github.com/iver56/audiomentations?
Yep, that what i mean.
Oh, that is indeed a low sample rate. Maybe you can use audiomentations for now then, since it works there
@KentoNishi Have you tried torch-pitch-shift with sr=8000?
Good :) torch-audiomentations actually depends on torch-pitch-shift, but uses its get_fast_shifts feature. So yeah, if you use torch-pitch-shift directly without get_fast_shifts, it'll probably work.
I have the same issue (version=0.11.0) with sample_rate = 16000 and (min, max)=(-0.2, 0,2). However, when the range changed to (-0.5, 0.5), the error didn't appear. In my case, the original dataset is piano sounds, so too much altering pitches can produce invalid data. Therefore, increasing the range is not a solution.
I have also tested torch_pith_shift. It works sometimes, but due to extensive memory consumption, the program crashed randomly for some shift values (for example, -0.16093115439744646).
Also, the running time dramatically increased after using PitchShift transform. In my laptop and using six other transformations, the average augmentation time per example (10s) is
| setting | average augmentation time per example (s) |
|---|---|
without PitchShift |
0.026 |
with PitchShift |
0.232 |
with pitch_shift function of torch_pith_shift |
3.951 |
Pitch shifts are generally pretty intensive operations so I'm not too surprised about the increase in execution time. Regarding speed issues with some pitch shift factors, unfortunately some factors aren't ideal for speed, so the rate of data transfer to the GPU becomes the bottleneck. In your specific case, it may be better to apply cpu-based pitch shift instead.
That's an interesting use case, AliKarimi95
Note: Although the pitch shift transform in torch-audiomentations can be comparatively fast on GPU, it is slow on CPU. When running pitch shift on CPU, the one in audiomentations is roughly 3x as fast.