FastSpeech-Pytorch

The Implementation of FastSpeech Based on Pytorch.

Start

Before Training

  1. Download and extract LJSpeech dataset.
  2. Put LJSpeech dataset in data.
  3. Run preprocess.py.
  4. If you want to get the target of alignment before training(It will speed up the training process greatly), you need download the pre-trained Tacotron2 model published by NVIDIA here.
  5. Now I provide LJSpeech's alignments calculated by Tacotron2 in alignment_targets.zip. If you want to use it, just unzip it and don't need the steps of 6, 7.
  6. Put the pre-trained Tacotron2 model in Tacotron2/pre_trained_model
  7. Run alignment.py, it will take 7 hours on NVIDIA RTX2080ti.
  8. Change pre_target = True in hparam.py.

Training

Normal Mode

Run train.py.

Turbo Mode

Run train_accelerated.py.

Dependencies

  • python 3.6
  • CUDA 10.0
  • pytorch 1.1.0
  • numpy 1.16.2
  • scipy 1.2.1
  • librosa 0.6.3
  • inflect 2.1.0
  • matplotlib 2.2.2

Notes

  • If you don't prepare the target of alignment before training, the process of training would be very long.
  • In the paper of FastSpeech, authors use pre-trained Transformer-TTS to provide the target of alignment. I didn't have a well-trained Transformer-TTS model so I use Tacotron2 instead.
  • If you want to use another model to get targets of alignment, you need rewrite alignment.py.
  • The returned value of alignment.py is a tensor whose value is the multiple that encoder's outputs are supposed to be expanded by.
  • For example:
import torch

test_target = torch.stack([torch.Tensor([0, 2, 3, 0, 3, 2, 1, 0, 0, 0]),
                           torch.Tensor([1, 2, 3, 2, 2, 0, 3, 6, 3, 5])])
  • In the turbo mode, prefetcher prefetches training data and this operation may cost more memory.
  • The output of LengthRegulator's last linear layer passes through the ReLU activation function in order to remove negative values. It is the outputs of this module. During the inference, the output of LengthRegulator pass through torch.exp() and subtract one, as the multiple for expanding encoder output. During the training stage, duration targets add one and pass through torch.log() and then calculate loss.
  • For example:
duration_predictor_target = duration_predictor_target + 1
duration_predictor_target = torch.log(duration_predictor_target)

duration_predictor_output = torch.exp(duration_predictor_output)
duration_predictor_output = duration_predictor_output - 1
  • This Implementation supports data-parallel now.
  • The tacotron2 outputs of mel spectrogram and alignment are shown as follow:
    tacotron2_outputs

GitHub