search

vincentherrmann / pytorch-wavenet

An implementation of WaveNet with fast generation
MIT License
968 stars 225 forks source link

Doubts regarding layers and blocks #35

Open fmorenopino opened 4 years ago

fmorenopino commented 4 years ago

Hi Vincent,

Congrats for for code. I have a doubt regarding two of the hyperparameters: as far as I know, the layers of your model would be equivalent to each of the rows that we can see on the Figure 2 of the paper, but I don´t really see why do we need the blocks for the implementation. Could you clarify this to me?

Thanks. Kind regards.

rpatrik96 commented 4 years ago

Congrats for for code. I have a doubt regarding two of the hyperparameters: as far as I know, the layers of your model would be equivalent to each of the rows that we can see on the Figure 2 of the paper, but I don´t really see why do we need the blocks for the implementation. Could you clarify this to me?

Hi @fmorenopino! I think I know the answer to your question. The number of layers means how many dilated convolutions are stacked after each other with increasing receptive fields; while the number of blocks means how many of those layer sequences are in the model.

With a concrete example: if you have 3 layers in 2 block, then the dilations are:

1, 2, 4, 1, 2, 4