tf.signal is differentiable

[rryan]

Hi there,

I'm the author of Tensorflow's tf.signal package. Your paper says that tf.signal does not support gradients, however this is not true. All operations in tf.signal are fully differentiable and come with GPU and TPU support. Could you please update your paper on arXiv to correct this?

[rryan]

I think you were trying to say that it does not support a trainable basis? That's true, but it's still differentiable so you can incorporate it into a network and backpropagate through it (e.g. to add a spectral loss to model that generates time-domain waveforms -- as in Parallel WaveNet)

[rryan]

Nice job on this repo and the paper by the way :).

[KinWaiCheuk]

I think you were trying to say that it does not support a trainable basis? That's true, but it's still differentiable so you can incorporate it into a network and backpropagate through it (e.g. to add a spectral loss to model that generates time-domain waveforms -- as in Parallel WaveNet)

Thanks for pointing it out. You are right, I was talking about the trainable basis. I will fix the paper now.

[KinWaiCheuk]

Hi Ryan, I have removed the claim that tf.signal doesn't support backpropagation.
Can you check if it is better now?
https://arxiv.org/abs/1912.12055

[rryan]

Thanks a lot for updating it!

[antonwnk]

Hi. Can I ask what is meant by trainable basis?
I came across this issue while googling around to see whether I could optimize tf.signal.stft's frame_length and frame_step parameters along with all the other parameters of my network, through backprop. Would this be technically possible?