Encoding new file - use of `zero_pad` about descript-audio-codec HOT 2 CLOSED

I think this is the receptive field of the convolutions.
Each latent is based not just on its corresponding time range of audio, but also considers the before/after context. The first/last have nothing before/after, so zero padding is added. Here is an animation of 2D convolution, the dotted line is the zero padding.

from descript-audio-codec.

Yes, this makes sense. Also, it seems like the encode task is disabling all the paddings in the conv layers. And, finally, zero_pad is in place operation and hence the results of it does not need to be stored.

Thanks for the clarification.

from descript-audio-codec.