Question about LSTM input shape about stylegestures HOT 3 CLOSED

Hi @TaeilJin,

Apologies for the late response. The last few weeks have been busy.

If I understand correctly, there's a simple reason for the shape difference you are wondering about. The dimension of interest here (dimension 1, which measures 70 during training and 1 during synthesis) corresponds to time. During training, the entire input and output sequences are available to us, so we can easily run the LSTM (and the network) forwards over all 70 time steps in one go. This can be done without using any explicit for-loop over time, by applying the LSTM to the input sequence as normal.

Generation time is different. Here we do not know the output sequence, but must instead create it. However, because the MoGlow model is autoregressive, it takes the output of recent time steps as an input to the flow when generating the next pose. This means that we must generate output poses one at a time, so that each new pose can be manually fed back in to the network to generate the next one. We also need an explicit loop over time. This should explain why the input shape only contains 1 single time frame at any given point during synthesis.

from stylegestures.

@TaeilJin Has your question in this issue been answered?
If so, I would recommend that you close the issue.

from stylegestures.

Closing this since the question appears to have been answered.

from stylegestures.