unrea1-sama

It seems that checkpoint is loaded twice. The checkpoint is loaded according to path specified by hparams['load_checkpoint']. https://github.com/NATSpeech/NATSpeech/blob/238165e8cd430531b69c484cabb032c1313ee73b/utils/commons/trainer.py#L150 However, the checkpoint is loaded again https://github.com/NATSpeech/NATSpeech/blob/238165e8cd430531b69c484cabb032c1313ee73b/utils/commons/trainer.py#L153-L155 This can overwrite model parameters...

**Is your feature request related to a problem? Please describe.** Right now MFA requires different audios from different speakers have different file names. For example, this requires a corpus directory...