cannot start training from a pre-trained model about radtts HOT 5 OPEN

Hi, thank you for the pretrained model.
What type of config.json I should use for training? _agap, _bgap, or _dap?
I got this error:
line 187, in load_data
'emotion': d[3],

Thanks

from radtts.

nice catch! it's because checkpoint_path requires a checkpoint dictionary with an optimizer and iteration number, which we don't provide.
please use warmstart_checkpoint_path instead of checkpoint_path and let us know

python train.py -c config.json -p train_config.ignore_layers_warmstart=["speaker_embedding.weight"] train_config.warmstart_checkpoint_path=model_path.pt

from radtts.

can you please pull and try again? the previous dataloader expected emotion and durations in the filelist. the current works if the filelist has only filename, text and speaker label.

Are you planning to train on new data?

from radtts.

I too would be interested in seeing a sample config for warm starting. I imagine there may be differences in the learning scheduling?

from radtts.

I've had more success training from scratch and warm starting when not ignoring the speaker embedding. Warm starting a multispeaker does not work for me.

from radtts.