[Bug] VITS recipe breaks with "local variable 'outputs' referenced before assignment"

Describe the bug When I try to train VITS in any way, e.g. with the recipe, it crashes with local variable 'outputs' referenced before assignment.

To Reproduce Steps to reproduce the behavior:

• Make a directory runs2
• Copy the VITS recipe recipes/ljspeech/vits_tts/train_vits.py into there
• Place LJSpeech dataset into there
• Run it: CUDA_VISIBLE_DEVICES=0 python runs2/train_vits.py
• Wait for the first > EVALUATION
• It immediately crashes

Expected behavior It doesn’t crash.

Environment (please complete the following information):

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04
• PyTorch or TensorFlow version (use command below): PyTorch 1.9.0 w/ CUDA 11.1
• Python version: 3.7.11
• CUDA/cuDNN version: 11.1
• GPU model and memory: RTX 3090 24GB
• Exact command to reproduce: See “To Reproduce” section.

Additional context

Here’s the last part of the output:

 > DataLoader initialization
 | > Use phonemes: True
   | > phoneme language: en-us
 | > Number of instances : 131
 | > Computing phonemes ...
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 131/131 [00:06<00:00, 19.03it/s] | > Max length sequence: 445286
 | > Min length sequence: 69478
 | > Avg length sequence: 288066.8244274809
 | > Num. instances discarded by max-min (max=5000, min=0) seq limits: 131
 | > Batch group size: 0.

 > EVALUATION

Traceback (most recent call last):
  File "/home/fijipants/repo/coqui-0.2.1/TTS/trainer.py", line 911, in fit
    self._fit()
  File "/home/fijipants/repo/coqui-0.2.1/TTS/trainer.py", line 898, in _fit
    self.eval_epoch()
  File "/home/fijipants/repo/coqui-0.2.1/TTS/trainer.py", line 839, in eval_epoch
    figures, audios = self.model.eval_log(self.ap, batch, outputs)
UnboundLocalError: local variable 'outputs' referenced before assignment

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "runs2/train_vits.py", line 52, in <module>
    trainer.fit()
  File "/home/fijipants/repo/coqui-0.2.1/TTS/trainer.py", line 930, in fit
    remove_experiment_folder(self.output_path)
  File "/home/fijipants/repo/coqui-0.2.1/TTS/utils/generic_utils.py", line 74, in remove_experiment_folder
    print(" ! Run is removed from {}".format(experiment_path))
  File "/home/fijipants/repo/coqui-0.2.1/TTS/trainer.py", line 986, in write
    with open(self.log_file, "a", encoding="utf-8") as f:
FileNotFoundError: [Errno 2] No such file or directory: '/home/fijipants/repo/coqui-0.2.1/runs2/vits_ljspeech-August-31-2021_11+23PM-5793dca7/trainer_0_log.txt'

[lexkoro]

you filtered out all of your data with the seq_limits.

 | > Number of instances : 131
 | > Min length sequence: 69478
 | > Avg length sequence: 288066.8244274809
 | > Num. instances discarded by max-min (max=5000, min=0) seq limits: 131

With the latest update the vits recipe uses the length of the audio instead of text for filtering.

[fijipants]

@erogol Oh no, I’ll try again from scratch and see if I can.