Piotr Żelasko
Piotr Żelasko
Are you able to reproduce this in some isolated example? E.g. using `from lhotse.testing.dummies import dummy_cut`?
That's correct, currently adding a new recording-level transform is a little complicated. This PR shows all necessary steps by adding volume perturbation: https://github.com/lhotse-speech/lhotse/pull/382 I'll try to add some docs how...
Looks good. > This can cause hands, but for now let's try it. @titu1994 can you elaborate? do you expect an uneven number of validation batches per node (why), or...
This is on validation epoch end so only a single sync per validation loop. That should be OK both hang-wise and speed-wise.
@krishnacpuvvada @titu1994 what's the verdict on this one? Do we need a Canary-specific mechanism to synchronize the validation metrics if it can't be done at the modelPT level?
> I have addressed everything except for `restore_orig_sr=True`. I am not sure how to achieve that! You are very close! Add a parameter `restore_orig_sr=True` in `def narrowband(self, ...)` for cut...
If you don't restore orig sr, you'll have to update both sampling_rate and num_samples property on the Recording object.
Thanks for the contribution, merging!
> But definitely the time for model forward and backward is still quite a lot. I do notice that your --max-duration is really quite small: 450. Unless your model is...
CC @csukuangfj @danpovey