This repository contains the `./scripts/prepare_dataset.sh` script which will automatically download and extract the whole dataset. By default, data will be extracted to the `./LJSpeech-1.1` directory. The dataset directory contains a `README` file, a `wavs` directory with all audio samples, and a file `metadata.csv` that contains audio file names and the corresponding transcripts.

To use datasets different than the default LJSpeech dataset:

2. Add two text files containing file lists: one for the training subset (`--training-files`) and one for the validation subset (`--validation files`).

The structure of the filelists should be as follows:

* Introduced batched inference

* Included warmup in the inference script

There are no known issues in this release.

parser.add_argument('-i', '--input', type=str, required=True,

parser.add_argument('-o', '--output', required=True,

help='output folder to save audio (file per phrase)')

parser.add_argument('--tacotron2', type=str,

return module

""" Code chooses a model based on name"""

model = None

if model_name == 'Tacotron2':

model = WaveGlow(**model_config)

else:

raise NotImplementedError(model_name)

if to_cuda:

model = model.cuda()

return model

dilation=1, w_init_gain='relu'),

nn.BatchNorm1d(encoder_embedding_dim))

convolutions.append(conv_layer)

self.lstm = nn.LSTM(encoder_embedding_dim,

int(encoder_embedding_dim / 2), 1,

return outputs

class Decoder(nn.Module):

def __init__(self, n_mel_channels, n_frames_per_step,

help='Enable cudnn')

training.add_argument('--cudnn-benchmark', action='store_true',

help='Run cudnn benchmark')

optimization = parser.add_argument_group('optimization setup')

optimization.add_argument(

model_config = models.get_model_config(model_name, args)

model = models.get_model(model_name, model_config,

if not args.amp_run and distributed_run:

model = DDP(model)

ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:19.06-py3

FROM ${FROM_IMAGE_NAME}

RUN pip install --upgrade --no-cache-dir pip \

&& pip install --no-cache-dir \

In order to test the accuracy of our implementation, we have run experiments with different seeds for 100 epochs with batch size 5120 per GPU and learning rate 6e-4 in the pytorch-18.12-py3 Docker container. The plot below shows the BLEU score changes.<br/>

Running this code with the provided hyperparameters will allow you to achieve the following results. Our setup is a DGX-1 with 8x Tesla V100 16GB. We've verified our results after training 32 epochs to obtain multi-GPU and mixed precision scaling results.

- initial commit, forked from [fairseq](https://github.com/pytorch/fairseq/commit/ac5fddfc691267285a84c81d39475411da5ed1c6)

May 2019:

June 2019

- new README

- Course of a training heavily depends on a random seed. There is high variance in the time required to reach a certain BLEU score. Also the highest BLEU score value observed vary between runs with different seeds.

"""Return a torch.optim.optimizer.Optimizer instance."""

if not hasattr(self, '_optimizer'):

raise NotImplementedError

raise ValueError('_optimizer must be an instance of torch.optim.Optimizer')

return self._optimizer

parser.add_argument('--seed', default=1, type=int, metavar='N',

help='pseudo random number generator seed')

parser.add_argument('--fp16', action='store_true', help='use FP16')

parser.add_argument('--profile', type=int, default=None)

# Task definitions can be found under fairseq/tasks/

parser.add_argument(