data_dirs: [./training/libritts-small/valle/] | |
spkr_name_getter: "lambda p: p.parts[-1].split('-')[0]" # I don't know what the importance of speaker IDs are in this implementation | |
models: [ 'ar-quarter', 'nar-quarter' ] # valid suffix: "-quarter", "-half", no suffix | |
optimizer: Adamw # set to Adam if you're a sadist | |
max_phones: 256 # overkill, as the max phoneme length in the data is smaller than this | |
batch_size: 16 # keep as a multiple of 4/8/16 | |
gradient_accumulation_steps: 2 # iffy on this, slower training but more "accurate" results | |
eval_batch_size: 8 # you can set this to your batch_size | |
save_ckpt_every: 100 # player preference | |
eval_every: 500 # player preference | |
max_iter: 10000 # increase as needed | |