AttributeError: 'HifiGAN' object has no attribute 'model'

[duheyu]

I'm following the instructions to train a model for DiffSinger, but there seem to be some issues with vocoders/hifigan.py:

"# first run fs2 infer;"

$ CUDA_VISIBLE_DEVICES=0 python tasks/run.py --config usr/configs/popcs_fs2.yaml --exp_name popcs_fs2_pmf0_1230 --reset --infer 
...
| model Trainable Parameters: 24.256M
  0%|                                                                                                                                                       | 0/1 [00:00<?, ?it/s]
Traceback (most recent call last):                                                                                                                          | 0/1 [00:00<?, ?it/s]
  File "tasks/run.py", line 15, in <module>
    run_task()
  File "tasks/run.py", line 10, in run_task
    task_cls.start()
  File "/.../DiffSinger/tasks/base_task.py", line 258, in start
    trainer.test(task)
  File "/.../DiffSinger/utils/pl_utils.py", line 586, in test
    self.fit(model)
  File "/.../DiffSinger/utils/pl_utils.py", line 489, in fit
    self.run_pretrain_routine(model)
  File "/.../DiffSinger/utils/pl_utils.py", line 545, in run_pretrain_routine
    self.run_evaluation(test=True)
  File "/.../DiffSinger/utils/pl_utils.py", line 1245, in run_evaluation
    eval_results = self.evaluate(self.model,
  File "/.../DiffSinger/utils/pl_utils.py", line 1185, in evaluate
    output = self.evaluation_forward(model,
  File "/.../DiffSinger/utils/pl_utils.py", line 1307, in evaluation_forward
    output = model.test_step(*args)
  File "/.../DiffSinger/tasks/tts/fs2.py", line 363, in test_step
    return self.after_infer(sample)
  File "/.../DiffSinger/tasks/tts/fs2.py", line 410, in after_infer
    wav_pred = self.vocoder.spec2wav(mel_pred, f0=f0_pred)
  File "/.../DiffSinger/vocoders/hifigan.py", line 56, in spec2wav
    device = self.device
AttributeError: 'HifiGAN' object has no attribute 'device'
Testing:   0%|          | 0/14 [00:01<?, ?batch/s]

So I've added self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") to the __init__ of HifiGAN, but then I get this:

| model Trainable Parameters: 24.256M
  0%|                                                                                                                                                       | 0/1 [00:00<?, ?it/s]
Traceback (most recent call last):                                                                                                                          | 0/1 [00:00<?, ?it/s]
  File "tasks/run.py", line 15, in <module>
    run_task()
  File "tasks/run.py", line 10, in run_task
    task_cls.start()
  File "/.../DiffSinger/tasks/base_task.py", line 258, in start
    trainer.test(task)
  File "/.../DiffSinger/utils/pl_utils.py", line 586, in test
    self.fit(model)
  File "/.../DiffSinger/utils/pl_utils.py", line 489, in fit
    self.run_pretrain_routine(model)
  File "/.../DiffSinger/utils/pl_utils.py", line 545, in run_pretrain_routine
    self.run_evaluation(test=True)
  File "/.../DiffSinger/utils/pl_utils.py", line 1245, in run_evaluation
    eval_results = self.evaluate(self.model,
  File "/.../DiffSinger/utils/pl_utils.py", line 1185, in evaluate
    output = self.evaluation_forward(model,
  File "/.../DiffSinger/utils/pl_utils.py", line 1307, in evaluation_forward
    output = model.test_step(*args)
  File "/.../DiffSinger/tasks/tts/fs2.py", line 363, in test_step
    return self.after_infer(sample)
  File "/.../DiffSinger/tasks/tts/fs2.py", line 410, in after_infer
    wav_pred = self.vocoder.spec2wav(mel_pred, f0=f0_pred)
  File "/.../DiffSinger/vocoders/hifigan.py", line 64, in spec2wav
    y = self.model(c, f0).view(-1)
AttributeError: 'HifiGAN' object has no attribute 'model'
Testing:   0%|          | 0/14 [00:01<?, ?batch/s]   

[MoonInTheRiver]

I never meet this problem. Is there anyone who could help him?

[ghost]

I confirm that I have the exactly same issue. I have not resolved it yet.

[MoonInTheRiver]

I deleted my 'checkpoints/0109_hifigan_bigpopcs_hop128' folder, then your bug-logs came. So, why not download the checkpoint file for the vocoder first, before conducting the experiments.

[duheyu]

Ah, placing the original checkpoint file in checkpoints/0109_hifigan_bigpopcs_hop128 does resolve this issue.

But this requires the use of the original phone_set to be used. The point of training the model is to train it on a completely new data set, with its own phone_set. Can you please indicate how to train a model from scratch?

[cantabile-kwok]

https://github.com/MoonInTheRiver/DiffSinger/blob/3d050f76aefb766d004fddcf52e9307affefd3c4/vocoders/hifigan.py#L41 There is no super call from the __init__ method of HifiGAN, which I think is a reason of this error. The PWG class has a device attribute. I have come through this also.