How do you know a series of dimension transformations such as dimension exchange in decoder_inputs = decoder_inputs.transpose(1, 2)? why, on what basis
here is link. tacotron2/model.py at master · NVIDIA/tacotron2 · GitHub
@Gorgen In order to understand this, you can think about this logically
-
B denotes batch_size. This cannot change
-
n_mel_channels represent features. This can change
-
T_out also represent features and can change
The above operation breaks out n_channels into further features based on the n_frames_per_step. If the n_frames_per_step > 1 then n_channels will increase. vice versa will happen to the T_out