Why not encoder?

[Asuka0002]

I try th encoder-decoder model and I found my loss dosen't convergence. I wonder why is the extra information (the encoder input) results in a bad performance?

[jason9693]

@Asuka0002 Thank you for interest my repository. Some task needs both feature extraction and generation like NMT is necessary for encoder architecture, but not the case in this task. Actually, Music Generation Task is not focused on feature extraction, but "Generation Task", like GPT. this only needs next token prediction, and style is controlled by