speecht5_tts model config contains unused and non-default parameter "max_length"
#37
by JackismyShephard - opened
The speecht5_tts model config contains a max_length parameter.
As far as I can tell, this parameter is only used by models which output text. In particular, SpeechT5ForTextToSpeech.generate overrides class transformers.GenerationMixin.generateand does not utilize max_length any place. Rather, it utilizes parameter maxlenratio for this purpose.
Additionally, having the max_length parameter in the speecht5_tts model config results in warnings during training using Trainer.train, so I propose that this parameter be removed.
JackismyShephard changed discussion status to closed