Fixed some errors to make it easier for first-time users.

[MingHui-Fang]

Actually, I ran into some issues when I used the official code, and I've seen other people raise the same issue, so maybe that's worth addressing.

Many users do not choose to use a 600M model due to GPU limitations, and will often use helpers/model_init_scripts/init_dummy_model_with_encodec.py or helpers/model_init_scripts/init_dummy_model.py.

However, the configurations in these two files have some potential problems for first-time users (Such as #63 (comment) and #66 (comment)), and it would probably be better if they could be fixed and comments added.

In addition, as users may change the model architecture, leading to errors, it might be possible to consider refinements training/run_parler_tts_training.py.

The above content is merely a personal suggestion. Please forgive any shortcomings.

[MingHui-Fang]

There is a bug with the handling of the variables ratand lensin the code. As mentioned in #73 (comment), If the batch size is 1, after thesqueeze() process, the rat and lens will become 0-d tensor. So replacing squeeze() with reshape(-1) solves this problem simply.

parler-tts/training/run_parler_tts_training.py

Lines 433 to 437 in 8b8c576

	if accelerator.is_main_process:
	lab = generate_labels["labels"].cpu().transpose(1, 2).to(torch.int16)
	rat = generate_labels["ratio"].cpu().squeeze()
	lens = generate_labels["len_audio"].cpu().squeeze()
	lab = [l[:, : int(ratio * length)] for (l, ratio, length) in zip(lab, rat, lens)]