DeepLearningExamples icon indicating copy to clipboard operation
DeepLearningExamples copied to clipboard

Published 20 hours ago • verified publisher icon NVIDIA

[HiFiGAN] Missing class variable in model ResBlock2

Open itakatz opened this issue 1 year ago • 0 comments

Related to PyTorch/SpeechSynthesis/HiFiGAN

Describe the bug The class variable lrelu_slope is used in the class method forward but it is not defined. which raises an exception if HiFiGAN is trained with a parameter --resblock 2. The variable is referenced here: https://github.com/NVIDIA/DeepLearningExamples/blob/d56fe703b034bf70d5e3aab4e1fec7bbe3d7735b/PyTorch/SpeechSynthesis/HiFiGAN/hifigan/models.py#L130

To Reproduce Train HiFiGAN using the regular script (scripts/train.sh), adding the input parameter --resblock 2

Expected behavior Training should run with no errors

Solution Define the class variable in the __init__ method, similarly as in class ResBlock1, that is add: self.lrelu_slope = LRELU_SLOPE

Environment

  • Container version: nvcr.io/nvidia/pytorch:22.12-py3
  • GPUs in the system: 8X A40
  • CUDA driver version: 520.61.05

itakatz avatar May 18 '23 20:05 itakatz