Small fix for the checkpoint dropdown menu in finetune gradio

[hcsolakoglu]

A small fix for checkpoint listing functionality in finetune gradio.

[SWivid]

Hi @hcsolakoglu ,
will PR #729 always add "pretrained_" prefix? e.g. result in "pretrained_pretrained_xxx" and also #750 if file_checkpoint_train selected?

F5-TTS/src/f5_tts/train/finetune_cli.py

Line 115 in 738d502

file_checkpoint = os.path.join(checkpoint_path, "pretrained_" + os.path.basename(ckpt_path))

[hcsolakoglu]

Hi @hcsolakoglu , will PR #729 always add "pretrained_" prefix? e.g. result in "pretrained_pretrained_xxx" and also #750 if file_checkpoint_train selected?

F5-TTS/src/f5_tts/train/finetune_cli.py

Line 115 in 738d502

file_checkpoint = os.path.join(checkpoint_path, "pretrained_" + os.path.basename(ckpt_path))

Yes, I forgot to add the check for the pretrained suffix, but you took care of it, thank you! As for the other question, we were adding the "pretrained" suffix in the Gradio interface because we assumed that the provided model was a pre-trained one trained by authors (aka. original checkpoints). Is there a preferred behavior you would like? I will implement it.

By the way, I discovered a bug in the Gradio interface that is unrelated to this issue. Vocab file path and model file path are filled with irrelevant text. File path for the vocabulary is filled with the "pinyin", while the checkpoint file path is set to True, which prevents the training from starting if the correct file paths are not provided. These two fields should be empty by default. @SWivid

[SWivid]

By the way, I discovered a bug in the Gradio interface that is unrelated to this issue. Vocab file path and model file path are filled with irrelevant text. File path for the vocabulary is filled with the "pinyin", while the checkpoint file path is set to True, which prevents the training from starting if the correct file paths are not provided. These two fields should be empty by default. @SWivid

Yep, could just make a PR if you got that and when you are free.
We are not that familiar or sensitive to this part of code cuz we generally use train.py, so we appreciate your fixes~