Support stft and inverse stft ops.

[nsthorat]

See:

https://www.tensorflow.org/api_docs/python/tf/signal/stft
https://www.tensorflow.org/api_docs/python/tf/signal/inverse_stft

[Lewuathe]

I tried to implement this and found we first need to implement window functions like hann or hamming. I'll work on them first.

https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/signal/hann_window
https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/signal/hamming_window

[Gurpreetsingh9465]

@Lewuathe can we work together in this module and add support to more ops as mentioned in tf.signals similar PR tensorflow/tensorflow#27075

[Lewuathe]

Thanks! I think you can add necessary ops in tf.signal. Some other ops like fft are still in spectral_ops.ts, though.

[Gurpreetsingh9465]

thanks! 😃
I would like to implement.

1. fftfreq
2. rfftfreq
3. fftshift(required tf.roll)
4. ifftshift(required tf.roll)

reference np.fft

[nsthorat]

@Gurpreetsingh9465 since those don't exist in TensorFlow we probably won't implement them in TensorFlow.js for now.

[Gurpreetsingh9465]

@nsthorat can i work on dct and idct

[nsthorat]

Great! Just curious if you have a concrete use case for those ops?

[Gurpreetsingh9465]

I am not sure about that but i think discrete cosine transform converts one function into sum of cosine functions of different amplitude which can be used to compress mp3 or image files.

[nsthorat]

For now let's try to keep it to the ops we have a concrete use case (one that someone wants to build), otherwise we bloat the bundle size without a real need for the op :)

[Gurpreetsingh9465]

Ok @nsthorat 😃