About the wavegru architecture about lyra HOT 4 CLOSED

@loopzy the number 3 comes from the calculation of RNN. You can see the following image about how torch.nn.LSTMCell is calculated.

where x is calculated by three matrix, Wii, Wif, Wig，which can fused together.

That makes the number 3.

from lyra.

@loopzy, kernel_size = 3 means that the first conv1d layer looks at 3 logmel frames: 1 logmel frame in the future (a frame look-ahead), 1 frame in the present and 1 frame in the past.

RNN layer takes 1 feature vector each time but this feature vector can be computed from 1 logmel frame look ahead and many frames at present and in the past

from lyra.

@loopzy the number 3 comes from the calculation of RNN. You can see the following image about how torch.nn.LSTMCell is calculated.

where x is calculated by three matrix, Wii, Wif, Wig，which can fused together.

That makes the number 3.

Thanks! I guessed that too. But in that way, the 3*1024 vector is a combination of 3 copies of 1024 vector and I am not sure.

Also, it is strange that he first conv1d of conditioning stack has kernel_size=3 while the model accepts only one logmel feature in each time. Do you have any ideas about that?

from lyra.

@loopzy, kernel_size = 3 means that the first conv1d layer looks at 3 logmel frames: 1 logmel frame in the future (a frame look-ahead), 1 frame in the present and 1 frame in the past.

RNN layer takes 1 feature vector each time but this feature vector can be computed from 1 logmel frame look ahead and many frames at present and in the past

Thanks a lot！

from lyra.