TensorFlowTTS/examples/multiband_melgan_hf/README.md


# Multi-band MelGAN: Faster Waveform Generation for High-Quality Text-to-Speech
Based on the script [`train_multiband_melgan_hf.py`](https://github.com/dathudeptrai/TensorflowTTS/tree/master/examples/multiband_melgan_hf/train_multiband_melgan_hf.py).

## Training Multi-band MelGAN from scratch with LJSpeech dataset.
This example code show you how to train MelGAN from scratch with Tensorflow 2 based on custom training loop and tf.function. The data used for this example is LJSpeech Ultimate, you can download the dataset at  [link](https://machineexperiments.tumblr.com/post/662408083204685824/ljspeech-ultimate).

### Step 1: Create Tensorflow based Dataloader (tf.dataset)
Please see detail at [examples/melgan/](https://github.com/dathudeptrai/TensorflowTTS/tree/master/examples/melgan#step-1-create-tensorflow-based-dataloader-tfdataset)

### Step 2: Training from scratch
After you re-define your dataloader, pls modify an input arguments, train_dataset and valid_dataset from [`train_multiband_melgan_hf.py`](https://github.com/dathudeptrai/TensorflowTTS/tree/master/examples/multiband_melgan_hf/train_multiband_melgan_hf.py). Here is an example command line to training melgan-stft from scratch:

First, you need training generator with only stft loss: 

```bash
CUDA_VISIBLE_DEVICES=0 python examples/multiband_melgan_hf/train_multiband_melgan_hf.py \
  --train-dir ./dump/train/ \
  --dev-dir ./dump/valid/ \
  --outdir ./examples/multiband_melgan_hf/exp/train.multiband_melgan_hf.v1/ \
  --config ./examples/multiband_melgan_hf/conf/multiband_melgan_hf.lju.v1.yml \
  --use-norm 1 \
  --generator_mixed_precision 1 \
  --resume ""
```

Then resume and start training generator + discriminator:

```bash
CUDA_VISIBLE_DEVICES=0 python examples/multiband_melgan_hf/train_multiband_melgan_hf.py \
  --train-dir ./dump/train/ \
  --dev-dir ./dump/valid/ \
  --outdir ./examples/multiband_melgan_hf/exp/train.multiband_melgan_hf.v1/ \
  --config ./examples/multiband_melgan_hf/conf/multiband_melgan_hf.lju.v1.yml \
  --use-norm 1 \
  --resume ./examples/multiband_melgan_hf/exp/train.multiband_melgan_hf.v1/checkpoints/ckpt-200000
```

IF you want to use MultiGPU to training you can replace `CUDA_VISIBLE_DEVICES=0` by `CUDA_VISIBLE_DEVICES=0,1,2,3` for example. You also need to tune the `batch_size` for each GPU (in config file) by yourself to maximize the performance. Note that MultiGPU now support for Training but not yet support for Decode. 

In case you want to resume the training progress, please following below example command line:

```bash
--resume ./examples/multiband_melgan_hf/exp/train.multiband_melgan_hf.v1/checkpoints/ckpt-100000
```

If you want to finetune a model, use `--pretrained` like this with the filename of the generator and discriminator, separated by comma.
```bash
--pretrained ptgenerator.h5,ptdiscriminator.h5
```
It is recommended that you first train text2mel model then extract postnets so that vocoder learns to compensate for flaws, if you do so, append `--postnets 1` to arguments



**IMPORTANT NOTES**:

- If Your Dataset is 16K, upsample_scales = [2, 4, 8] worked.
- If Your Dataset is > 16K (22K, 24K, ...), upsample_scales = [2, 4, 8] didn't worked, used [8, 4, 2] instead.
- Mixed precision make Group Convolution training slower on Discriminator, both pytorch (apex) and tensorflow also has this problems. So, **DO NOT USE** mixed precision when discriminator enable.

### Step 3: Decode audio from folder mel-spectrogram
To running inference on folder mel-spectrogram (eg valid folder), run below command line:

```bash
CUDA_VISIBLE_DEVICES=0 python examples/multiband_melgan_hf/decode_mb_melgan.py \
  --rootdir ./dump/valid/ \
  --outdir ./prediction/multiband_melgan_hf.v1/ \
  --checkpoint ./examples/multiband_melgan_hf/exp/train.multiband_melgan_hf.v1/checkpoints/generator-920000.h5 \
  --config ./examples/multiband_melgan_hf/conf/multiband_melgan_hf.lju.v1.yml \
  --batch-size 32 \
  --use-norm 1
```

## Finetune MelGAN STFT with ljspeech pretrained on other languages
Just load pretrained model and training from scratch with other languages. **DO NOT FORGET** re-preprocessing on your dataset if needed. A hop_size should be 512 if you want to use our pretrained.

## Learning Curves
Here is a learning curves of melgan based on this config [`multiband_melgan_hf.v1.yaml`](https://github.com/dathudeptrai/TensorflowTTS/tree/master/examples/multiband_melgan_hf/conf/multiband_melgan_hf.v1.yaml)

<img src="fig/eval.png" height="300" width="850">

<img src="fig/train.png" height="300" width="850">

## Pretrained Models and Audio samples
| Model                                                                                                          | Conf                                                                                                                        | Lang  | Fs [Hz] | Mel range [Hz] | FFT / Hop / Win [pt] | # iters | Notes |
| :------                                                                                                        | :---:                                                                                                                       | :---: | :----:  | :--------:     | :---------------:    | :-----:   | :-----: |
| [multiband_melgan_hf.lju.v1](https://drive.google.com/drive/folders/1tOMzik_Nr4eY63gooKYSmNTJyXC6Pp55?usp=sharing)             | [link](https://github.com/tensorspeech/TensorFlowTTS/tree/master/examples/multiband_melgan_hf/conf/multiband_melgan_hf.lju.v1.yml)          | EN    | 44.1k  | 20-11025        | 2048 / 512 / 2048     | 920K    |  -|


## Reference

1. https://github.com/kan-bayashi/ParallelWaveGAN
2. [Parallel WaveGAN: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram](https://arxiv.org/abs/1910.11480)
3. [Multi-band MelGAN: Faster Waveform Generation for High-Quality Text-to-Speech](https://arxiv.org/abs/2005.05106)