Upload 10 files

hifi-gan/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Jungil Kong
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

hifi-gan/README.md

@@ -0,0 +1,105 @@
+# HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis
+
+### Jungil Kong, Jaehyeon Kim, Jaekyoung Bae
+
+In our [paper](https://arxiv.org/abs/2010.05646), 
+we proposed HiFi-GAN: a GAN-based model capable of generating high fidelity speech efficiently.<br/>
+We provide our implementation and pretrained models as open source in this repository.
+
+**Abstract :**
+Several recent work on speech synthesis have employed generative adversarial networks (GANs) to produce raw waveforms. 
+Although such methods improve the sampling efficiency and memory usage, 
+their sample quality has not yet reached that of autoregressive and flow-based generative models. 
+In this work, we propose HiFi-GAN, which achieves both efficient and high-fidelity speech synthesis. 
+As speech audio consists of sinusoidal signals with various periods, 
+we demonstrate that modeling periodic patterns of an audio is crucial for enhancing sample quality. 
+A subjective human evaluation (mean opinion score, MOS) of a single speaker dataset indicates that our proposed method 
+demonstrates similarity to human quality while generating 22.05 kHz high-fidelity audio 167.9 times faster than 
+real-time on a single V100 GPU. We further show the generality of HiFi-GAN to the mel-spectrogram inversion of unseen 
+speakers and end-to-end speech synthesis. Finally, a small footprint version of HiFi-GAN generates samples 13.4 times 
+faster than real-time on CPU with comparable quality to an autoregressive counterpart.
+
+Visit our [demo website](https://jik876.github.io/hifi-gan-demo/) for audio samples.
+
+
+## Pre-requisites
+1. Python >= 3.6
+2. Clone this repository.
+3. Install python requirements. Please refer [requirements.txt](requirements.txt)
+4. Download and extract the [LJ Speech dataset](https://keithito.com/LJ-Speech-Dataset/).
+And move all wav files to `LJSpeech-1.1/wavs`
+
+
+## Training
+```
+python train.py --config config_v1.json
+```
+To train V2 or V3 Generator, replace `config_v1.json` with `config_v2.json` or `config_v3.json`.<br>
+Checkpoints and copy of the configuration file are saved in `cp_hifigan` directory by default.<br>
+You can change the path by adding `--checkpoint_path` option.
+
+Validation loss during training with V1 generator.<br>
+![validation loss](./validation_loss.png)
+
+## Pretrained Model
+You can also use pretrained models we provide.<br/>
+[Download pretrained models](https://drive.google.com/drive/folders/1-eEYTB5Av9jNql0WGBlRoi-WH2J7bp5Y?usp=sharing)<br/> 
+Details of each folder are as in follows:
+
+|Folder Name|Generator|Dataset|Fine-Tuned|
+|------|---|---|---|
+|LJ_V1|V1|LJSpeech|No|
+|LJ_V2|V2|LJSpeech|No|
+|LJ_V3|V3|LJSpeech|No|
+|LJ_FT_T2_V1|V1|LJSpeech|Yes ([Tacotron2](https://github.com/NVIDIA/tacotron2))|
+|LJ_FT_T2_V2|V2|LJSpeech|Yes ([Tacotron2](https://github.com/NVIDIA/tacotron2))|
+|LJ_FT_T2_V3|V3|LJSpeech|Yes ([Tacotron2](https://github.com/NVIDIA/tacotron2))|
+|VCTK_V1|V1|VCTK|No|
+|VCTK_V2|V2|VCTK|No|
+|VCTK_V3|V3|VCTK|No|
+|UNIVERSAL_V1|V1|Universal|No|
+
+We provide the universal model with discriminator weights that can be used as a base for transfer learning to other datasets.
+
+## Fine-Tuning
+1. Generate mel-spectrograms in numpy format using [Tacotron2](https://github.com/NVIDIA/tacotron2) with teacher-forcing.<br/>
+The file name of the generated mel-spectrogram should match the audio file and the extension should be `.npy`.<br/>
+Example:
+    ```
+    Audio File : LJ001-0001.wav
+    Mel-Spectrogram File : LJ001-0001.npy
+    ```
+2. Create `ft_dataset` folder and copy the generated mel-spectrogram files into it.<br/>
+3. Run the following command.
+    ```
+    python train.py --fine_tuning True --config config_v1.json
+    ```
+    For other command line options, please refer to the training section.
+
+
+## Inference from wav file
+1. Make `test_files` directory and copy wav files into the directory.
+2. Run the following command.
+    ```
+    python inference.py --checkpoint_file [generator checkpoint file path]
+    ```
+Generated wav files are saved in `generated_files` by default.<br>
+You can change the path by adding `--output_dir` option.
+
+
+## Inference for end-to-end speech synthesis
+1. Make `test_mel_files` directory and copy generated mel-spectrogram files into the directory.<br>
+You can generate mel-spectrograms using [Tacotron2](https://github.com/NVIDIA/tacotron2), 
+[Glow-TTS](https://github.com/jaywalnut310/glow-tts) and so forth.
+2. Run the following command.
+    ```
+    python inference_e2e.py --checkpoint_file [generator checkpoint file path]
+    ```
+Generated wav files are saved in `generated_files_from_mel` by default.<br>
+You can change the path by adding `--output_dir` option.
+
+
+## Acknowledgements
+We referred to [WaveGlow](https://github.com/NVIDIA/waveglow), [MelGAN](https://github.com/descriptinc/melgan-neurips) 
+and [Tacotron2](https://github.com/NVIDIA/tacotron2) to implement this.
+

hifi-gan/env.py

@@ -0,0 +1,15 @@
+import os
+import shutil
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+
+def build_env(config, config_name, path):
+    t_path = os.path.join(path, config_name)
+    if config != t_path:
+        os.makedirs(path, exist_ok=True)
+        shutil.copyfile(config, os.path.join(path, config_name))

hifi-gan/inference.py

@@ -0,0 +1,95 @@
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import glob
+import os
+import argparse
+import json
+import torch
+from scipy.io.wavfile import write
+from env import AttrDict
+from meldataset import mel_spectrogram, MAX_WAV_VALUE, load_wav
+from models import Generator
+
+h = None
+device = None
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print("Loading '{}'".format(filepath))
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print("Complete.")
+    return checkpoint_dict
+
+
+def get_mel(x):
+    return mel_spectrogram(x, h.n_fft, h.num_mels, h.sampling_rate, h.hop_size, h.win_size, h.fmin, h.fmax)
+
+
+def scan_checkpoint(cp_dir, prefix):
+    pattern = os.path.join(cp_dir, prefix + '*')
+    cp_list = glob.glob(pattern)
+    if len(cp_list) == 0:
+        return ''
+    return sorted(cp_list)[-1]
+
+
+def inference(a):
+    generator = Generator(h).to(device)
+
+    state_dict_g = load_checkpoint(a.checkpoint_file, device)
+    generator.load_state_dict(state_dict_g['generator'])
+
+    filelist = os.listdir(a.input_wavs_dir)
+
+    os.makedirs(a.output_dir, exist_ok=True)
+
+    generator.eval()
+    generator.remove_weight_norm()
+    with torch.no_grad():
+        for i, filname in enumerate(filelist):
+            wav, sr = load_wav(os.path.join(a.input_wavs_dir, filname))
+            wav = wav / MAX_WAV_VALUE
+            wav = torch.FloatTensor(wav).to(device)
+            x = get_mel(wav.unsqueeze(0))
+            y_g_hat = generator(x)
+            audio = y_g_hat.squeeze()
+            audio = audio * MAX_WAV_VALUE
+            audio = audio.cpu().numpy().astype('int16')
+
+            output_file = os.path.join(a.output_dir, os.path.splitext(filname)[0] + '_generated.wav')
+            write(output_file, h.sampling_rate, audio)
+            print(output_file)
+
+
+def main():
+    print('Initializing Inference Process..')
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--input_wavs_dir', default='test_files')
+    parser.add_argument('--output_dir', default='generated_files')
+    parser.add_argument('--checkpoint_file', required=True)
+    a = parser.parse_args()
+
+    config_file = os.path.join(os.path.split(a.checkpoint_file)[0], 'config.json')
+    with open(config_file) as f:
+        data = f.read()
+
+    global h
+    json_config = json.loads(data)
+    h = AttrDict(json_config)
+
+    torch.manual_seed(h.seed)
+    global device
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(h.seed)
+        device = torch.device('cuda')
+    else:
+        device = torch.device('cpu')
+
+    inference(a)
+
+
+if __name__ == '__main__':
+    main()
+

hifi-gan/inference_e2e.py

@@ -0,0 +1,90 @@
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import glob
+import os
+import numpy as np
+import argparse
+import json
+import torch
+from scipy.io.wavfile import write
+from env import AttrDict
+from meldataset import MAX_WAV_VALUE
+from models import Generator
+
+h = None
+device = None
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print("Loading '{}'".format(filepath))
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print("Complete.")
+    return checkpoint_dict
+
+
+def scan_checkpoint(cp_dir, prefix):
+    pattern = os.path.join(cp_dir, prefix + '*')
+    cp_list = glob.glob(pattern)
+    if len(cp_list) == 0:
+        return ''
+    return sorted(cp_list)[-1]
+
+
+def inference(a):
+    generator = Generator(h).to(device)
+
+    state_dict_g = load_checkpoint(a.checkpoint_file, device)
+    generator.load_state_dict(state_dict_g['generator'])
+
+    filelist = os.listdir(a.input_mels_dir)
+
+    os.makedirs(a.output_dir, exist_ok=True)
+
+    generator.eval()
+    generator.remove_weight_norm()
+    with torch.no_grad():
+        for i, filname in enumerate(filelist):
+            x = np.load(os.path.join(a.input_mels_dir, filname))
+            x = torch.FloatTensor(x).to(device)
+            y_g_hat = generator(x)
+            audio = y_g_hat.squeeze()
+            audio = audio * MAX_WAV_VALUE
+            audio = audio.cpu().numpy().astype('int16')
+
+            output_file = os.path.join(a.output_dir, os.path.splitext(filname)[0] + '_generated_e2e.wav')
+            write(output_file, h.sampling_rate, audio)
+            print(output_file)
+
+
+def main():
+    print('Initializing Inference Process..')
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--input_mels_dir', default='test_mel_files')
+    parser.add_argument('--output_dir', default='generated_files_from_mel')
+    parser.add_argument('--checkpoint_file', required=True)
+    a = parser.parse_args()
+
+    config_file = os.path.join(os.path.split(a.checkpoint_file)[0], 'config.json')
+    with open(config_file) as f:
+        data = f.read()
+
+    global h
+    json_config = json.loads(data)
+    h = AttrDict(json_config)
+
+    torch.manual_seed(h.seed)
+    global device
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(h.seed)
+        device = torch.device('cuda')
+    else:
+        device = torch.device('cpu')
+
+    inference(a)
+
+
+if __name__ == '__main__':
+    main()
+

hifi-gan/meldataset.py

@@ -0,0 +1,168 @@
+import math
+import os
+import random
+import torch
+import torch.utils.data
+import numpy as np
+from librosa.util import normalize
+from scipy.io.wavfile import read
+from librosa.filters import mel as librosa_mel_fn
+
+MAX_WAV_VALUE = 32768.0
+
+
+def load_wav(full_path):
+    sampling_rate, data = read(full_path)
+    return data, sampling_rate
+
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    return np.exp(x) / C
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+def spectral_de_normalize_torch(magnitudes):
+    output = dynamic_range_decompression_torch(magnitudes)
+    return output
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def mel_spectrogram(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
+    if torch.min(y) < -1.:
+        print('min value is ', torch.min(y))
+    if torch.max(y) > 1.:
+        print('max value is ', torch.max(y))
+
+    global mel_basis, hann_window
+    if fmax not in mel_basis:
+        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
+        mel_basis[str(fmax)+'_'+str(y.device)] = torch.from_numpy(mel).float().to(y.device)
+        hann_window[str(y.device)] = torch.hann_window(win_size).to(y.device)
+
+    y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
+    y = y.squeeze(1)
+
+    spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[str(y.device)],
+                      center=center, pad_mode='reflect', normalized=False, onesided=True)
+
+    spec = torch.sqrt(spec.pow(2).sum(-1)+(1e-9))
+
+    spec = torch.matmul(mel_basis[str(fmax)+'_'+str(y.device)], spec)
+    spec = spectral_normalize_torch(spec)
+
+    return spec
+
+
+def get_dataset_filelist(a):
+    with open(a.input_training_file, 'r', encoding='utf-8') as fi:
+        training_files = [os.path.join(a.input_wavs_dir, x.split('|')[0] + '.wav')
+                          for x in fi.read().split('\n') if len(x) > 0]
+
+    with open(a.input_validation_file, 'r', encoding='utf-8') as fi:
+        validation_files = [os.path.join(a.input_wavs_dir, x.split('|')[0] + '.wav')
+                            for x in fi.read().split('\n') if len(x) > 0]
+    return training_files, validation_files
+
+
+class MelDataset(torch.utils.data.Dataset):
+    def __init__(self, training_files, segment_size, n_fft, num_mels,
+                 hop_size, win_size, sampling_rate,  fmin, fmax, split=True, shuffle=True, n_cache_reuse=1,
+                 device=None, fmax_loss=None, fine_tuning=False, base_mels_path=None):
+        self.audio_files = training_files
+        random.seed(1234)
+        if shuffle:
+            random.shuffle(self.audio_files)
+        self.segment_size = segment_size
+        self.sampling_rate = sampling_rate
+        self.split = split
+        self.n_fft = n_fft
+        self.num_mels = num_mels
+        self.hop_size = hop_size
+        self.win_size = win_size
+        self.fmin = fmin
+        self.fmax = fmax
+        self.fmax_loss = fmax_loss
+        self.cached_wav = None
+        self.n_cache_reuse = n_cache_reuse
+        self._cache_ref_count = 0
+        self.device = device
+        self.fine_tuning = fine_tuning
+        self.base_mels_path = base_mels_path
+
+    def __getitem__(self, index):
+        filename = self.audio_files[index]
+        if self._cache_ref_count == 0:
+            audio, sampling_rate = load_wav(filename)
+            audio = audio / MAX_WAV_VALUE
+            if not self.fine_tuning:
+                audio = normalize(audio) * 0.95
+            self.cached_wav = audio
+            if sampling_rate != self.sampling_rate:
+                raise ValueError("{} SR doesn't match target {} SR".format(
+                    sampling_rate, self.sampling_rate))
+            self._cache_ref_count = self.n_cache_reuse
+        else:
+            audio = self.cached_wav
+            self._cache_ref_count -= 1
+
+        audio = torch.FloatTensor(audio)
+        audio = audio.unsqueeze(0)
+
+        if not self.fine_tuning:
+            if self.split:
+                if audio.size(1) >= self.segment_size:
+                    max_audio_start = audio.size(1) - self.segment_size
+                    audio_start = random.randint(0, max_audio_start)
+                    audio = audio[:, audio_start:audio_start+self.segment_size]
+                else:
+                    audio = torch.nn.functional.pad(audio, (0, self.segment_size - audio.size(1)), 'constant')
+
+            mel = mel_spectrogram(audio, self.n_fft, self.num_mels,
+                                  self.sampling_rate, self.hop_size, self.win_size, self.fmin, self.fmax,
+                                  center=False)
+        else:
+            mel = np.load(
+                os.path.join(self.base_mels_path, os.path.splitext(os.path.split(filename)[-1])[0] + '.npy'))
+            mel = torch.from_numpy(mel)
+
+            if len(mel.shape) < 3:
+                mel = mel.unsqueeze(0)
+
+            if self.split:
+                frames_per_seg = math.ceil(self.segment_size / self.hop_size)
+
+                if audio.size(1) >= self.segment_size:
+                    mel_start = random.randint(0, mel.size(2) - frames_per_seg - 1)
+                    mel = mel[:, :, mel_start:mel_start + frames_per_seg]
+                    audio = audio[:, mel_start * self.hop_size:(mel_start + frames_per_seg) * self.hop_size]
+                else:
+                    mel = torch.nn.functional.pad(mel, (0, frames_per_seg - mel.size(2)), 'constant')
+                    audio = torch.nn.functional.pad(audio, (0, self.segment_size - audio.size(1)), 'constant')
+
+        mel_loss = mel_spectrogram(audio, self.n_fft, self.num_mels,
+                                   self.sampling_rate, self.hop_size, self.win_size, self.fmin, self.fmax_loss,
+                                   center=False)
+
+        return (mel.squeeze(), audio.squeeze(0), filename, mel_loss.squeeze())
+
+    def __len__(self):
+        return len(self.audio_files)

hifi-gan/models.py

@@ -0,0 +1,283 @@
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
+from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+from utils import init_weights, get_padding
+
+LRELU_SLOPE = 0.1
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.h = h
+        self.convs1 = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
+                               padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
+                               padding=get_padding(kernel_size, dilation[1]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
+                               padding=get_padding(kernel_size, dilation[2])))
+        ])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1)))
+        ])
+        self.convs2.apply(init_weights)
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.h = h
+        self.convs = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
+                               padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
+                               padding=get_padding(kernel_size, dilation[1])))
+        ])
+        self.convs.apply(init_weights)
+
+    def forward(self, x):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            xt = c(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Generator(torch.nn.Module):
+    def __init__(self, h):
+        super(Generator, self).__init__()
+        self.h = h
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+        self.conv_pre = weight_norm(Conv1d(80, h.upsample_initial_channel, 7, 1, padding=3))
+        resblock = ResBlock1 if h.resblock == '1' else ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(weight_norm(
+                ConvTranspose1d(h.upsample_initial_channel//(2**i), h.upsample_initial_channel//(2**(i+1)),
+                                k, u, padding=(k-u)//2)))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel//(2**(i+1))
+            for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
+                self.resblocks.append(resblock(h, ch, k, d))
+
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+    def forward(self, x):
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i*self.num_kernels+j](x)
+                else:
+                    xs += self.resblocks[i*self.num_kernels+j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList([
+            norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
+            norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
+            norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
+            norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
+            norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
+        ])
+        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0: # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self):
+        super(MultiPeriodDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList([
+            DiscriminatorP(2),
+            DiscriminatorP(3),
+            DiscriminatorP(5),
+            DiscriminatorP(7),
+            DiscriminatorP(11),
+        ])
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList([
+            norm_f(Conv1d(1, 128, 15, 1, padding=7)),
+            norm_f(Conv1d(128, 128, 41, 2, groups=4, padding=20)),
+            norm_f(Conv1d(128, 256, 41, 2, groups=16, padding=20)),
+            norm_f(Conv1d(256, 512, 41, 4, groups=16, padding=20)),
+            norm_f(Conv1d(512, 1024, 41, 4, groups=16, padding=20)),
+            norm_f(Conv1d(1024, 1024, 41, 1, groups=16, padding=20)),
+            norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+        ])
+        self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiScaleDiscriminator(torch.nn.Module):
+    def __init__(self):
+        super(MultiScaleDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList([
+            DiscriminatorS(use_spectral_norm=True),
+            DiscriminatorS(),
+            DiscriminatorS(),
+        ])
+        self.meanpools = nn.ModuleList([
+            AvgPool1d(4, 2, padding=2),
+            AvgPool1d(4, 2, padding=2)
+        ])
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            if i != 0:
+                y = self.meanpools[i-1](y)
+                y_hat = self.meanpools[i-1](y_hat)
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+def feature_loss(fmap_r, fmap_g):
+    loss = 0
+    for dr, dg in zip(fmap_r, fmap_g):
+        for rl, gl in zip(dr, dg):
+            loss += torch.mean(torch.abs(rl - gl))
+
+    return loss*2
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+    loss = 0
+    r_losses = []
+    g_losses = []
+    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+        r_loss = torch.mean((1-dr)**2)
+        g_loss = torch.mean(dg**2)
+        loss += (r_loss + g_loss)
+        r_losses.append(r_loss.item())
+        g_losses.append(g_loss.item())
+
+    return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+    loss = 0
+    gen_losses = []
+    for dg in disc_outputs:
+        l = torch.mean((1-dg)**2)
+        gen_losses.append(l)
+        loss += l
+
+    return loss, gen_losses
+

hifi-gan/requirements.txt

@@ -0,0 +1,7 @@
+torch==1.4.0
+numpy==1.17.4
+librosa==0.7.2
+scipy==1.4.1
+tensorboard==2.0
+soundfile==0.10.3.post1
+matplotlib==3.1.3

hifi-gan/train.py

@@ -0,0 +1,271 @@
+import warnings
+warnings.simplefilter(action='ignore', category=FutureWarning)
+import itertools
+import os
+import time
+import argparse
+import json
+import torch
+import torch.nn.functional as F
+from torch.utils.tensorboard import SummaryWriter
+from torch.utils.data import DistributedSampler, DataLoader
+import torch.multiprocessing as mp
+from torch.distributed import init_process_group
+from torch.nn.parallel import DistributedDataParallel
+from env import AttrDict, build_env
+from meldataset import MelDataset, mel_spectrogram, get_dataset_filelist
+from models import Generator, MultiPeriodDiscriminator, MultiScaleDiscriminator, feature_loss, generator_loss,\
+    discriminator_loss
+from utils import plot_spectrogram, scan_checkpoint, load_checkpoint, save_checkpoint
+
+torch.backends.cudnn.benchmark = True
+
+
+def train(rank, a, h):
+    if h.num_gpus > 1:
+        init_process_group(backend=h.dist_config['dist_backend'], init_method=h.dist_config['dist_url'],
+                           world_size=h.dist_config['world_size'] * h.num_gpus, rank=rank)
+
+    torch.cuda.manual_seed(h.seed)
+    device = torch.device('cuda:{:d}'.format(rank))
+
+    generator = Generator(h).to(device)
+    mpd = MultiPeriodDiscriminator().to(device)
+    msd = MultiScaleDiscriminator().to(device)
+
+    if rank == 0:
+        print(generator)
+        os.makedirs(a.checkpoint_path, exist_ok=True)
+        print("checkpoints directory : ", a.checkpoint_path)
+
+    if os.path.isdir(a.checkpoint_path):
+        cp_g = scan_checkpoint(a.checkpoint_path, 'g_')
+        cp_do = scan_checkpoint(a.checkpoint_path, 'do_')
+
+    steps = 0
+    if cp_g is None or cp_do is None:
+        state_dict_do = None
+        last_epoch = -1
+    else:
+        state_dict_g = load_checkpoint(cp_g, device)
+        state_dict_do = load_checkpoint(cp_do, device)
+        generator.load_state_dict(state_dict_g['generator'])
+        mpd.load_state_dict(state_dict_do['mpd'])
+        msd.load_state_dict(state_dict_do['msd'])
+        steps = state_dict_do['steps'] + 1
+        last_epoch = state_dict_do['epoch']
+
+    if h.num_gpus > 1:
+        generator = DistributedDataParallel(generator, device_ids=[rank]).to(device)
+        mpd = DistributedDataParallel(mpd, device_ids=[rank]).to(device)
+        msd = DistributedDataParallel(msd, device_ids=[rank]).to(device)
+
+    optim_g = torch.optim.AdamW(generator.parameters(), h.learning_rate, betas=[h.adam_b1, h.adam_b2])
+    optim_d = torch.optim.AdamW(itertools.chain(msd.parameters(), mpd.parameters()),
+                                h.learning_rate, betas=[h.adam_b1, h.adam_b2])
+
+    if state_dict_do is not None:
+        optim_g.load_state_dict(state_dict_do['optim_g'])
+        optim_d.load_state_dict(state_dict_do['optim_d'])
+
+    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=h.lr_decay, last_epoch=last_epoch)
+    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=h.lr_decay, last_epoch=last_epoch)
+
+    training_filelist, validation_filelist = get_dataset_filelist(a)
+
+    trainset = MelDataset(training_filelist, h.segment_size, h.n_fft, h.num_mels,
+                          h.hop_size, h.win_size, h.sampling_rate, h.fmin, h.fmax, n_cache_reuse=0,
+                          shuffle=False if h.num_gpus > 1 else True, fmax_loss=h.fmax_for_loss, device=device,
+                          fine_tuning=a.fine_tuning, base_mels_path=a.input_mels_dir)
+
+    train_sampler = DistributedSampler(trainset) if h.num_gpus > 1 else None
+
+    train_loader = DataLoader(trainset, num_workers=h.num_workers, shuffle=False,
+                              sampler=train_sampler,
+                              batch_size=h.batch_size,
+                              pin_memory=True,
+                              drop_last=True)
+
+    if rank == 0:
+        validset = MelDataset(validation_filelist, h.segment_size, h.n_fft, h.num_mels,
+                              h.hop_size, h.win_size, h.sampling_rate, h.fmin, h.fmax, False, False, n_cache_reuse=0,
+                              fmax_loss=h.fmax_for_loss, device=device, fine_tuning=a.fine_tuning,
+                              base_mels_path=a.input_mels_dir)
+        validation_loader = DataLoader(validset, num_workers=1, shuffle=False,
+                                       sampler=None,
+                                       batch_size=1,
+                                       pin_memory=True,
+                                       drop_last=True)
+
+        sw = SummaryWriter(os.path.join(a.checkpoint_path, 'logs'))
+
+    generator.train()
+    mpd.train()
+    msd.train()
+    for epoch in range(max(0, last_epoch), a.training_epochs):
+        if rank == 0:
+            start = time.time()
+            print("Epoch: {}".format(epoch+1))
+
+        if h.num_gpus > 1:
+            train_sampler.set_epoch(epoch)
+
+        for i, batch in enumerate(train_loader):
+            if rank == 0:
+                start_b = time.time()
+            x, y, _, y_mel = batch
+            x = torch.autograd.Variable(x.to(device, non_blocking=True))
+            y = torch.autograd.Variable(y.to(device, non_blocking=True))
+            y_mel = torch.autograd.Variable(y_mel.to(device, non_blocking=True))
+            y = y.unsqueeze(1)
+
+            y_g_hat = generator(x)
+            y_g_hat_mel = mel_spectrogram(y_g_hat.squeeze(1), h.n_fft, h.num_mels, h.sampling_rate, h.hop_size, h.win_size,
+                                          h.fmin, h.fmax_for_loss)
+
+            optim_d.zero_grad()
+
+            # MPD
+            y_df_hat_r, y_df_hat_g, _, _ = mpd(y, y_g_hat.detach())
+            loss_disc_f, losses_disc_f_r, losses_disc_f_g = discriminator_loss(y_df_hat_r, y_df_hat_g)
+
+            # MSD
+            y_ds_hat_r, y_ds_hat_g, _, _ = msd(y, y_g_hat.detach())
+            loss_disc_s, losses_disc_s_r, losses_disc_s_g = discriminator_loss(y_ds_hat_r, y_ds_hat_g)
+
+            loss_disc_all = loss_disc_s + loss_disc_f
+
+            loss_disc_all.backward()
+            optim_d.step()
+
+            # Generator
+            optim_g.zero_grad()
+
+            # L1 Mel-Spectrogram Loss
+            loss_mel = F.l1_loss(y_mel, y_g_hat_mel) * 45
+
+            y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = mpd(y, y_g_hat)
+            y_ds_hat_r, y_ds_hat_g, fmap_s_r, fmap_s_g = msd(y, y_g_hat)
+            loss_fm_f = feature_loss(fmap_f_r, fmap_f_g)
+            loss_fm_s = feature_loss(fmap_s_r, fmap_s_g)
+            loss_gen_f, losses_gen_f = generator_loss(y_df_hat_g)
+            loss_gen_s, losses_gen_s = generator_loss(y_ds_hat_g)
+            loss_gen_all = loss_gen_s + loss_gen_f + loss_fm_s + loss_fm_f + loss_mel
+
+            loss_gen_all.backward()
+            optim_g.step()
+
+            if rank == 0:
+                # STDOUT logging
+                if steps % a.stdout_interval == 0:
+                    with torch.no_grad():
+                        mel_error = F.l1_loss(y_mel, y_g_hat_mel).item()
+
+                    print('Steps : {:d}, Gen Loss Total : {:4.3f}, Mel-Spec. Error : {:4.3f}, s/b : {:4.3f}'.
+                          format(steps, loss_gen_all, mel_error, time.time() - start_b))
+
+                # checkpointing
+                if steps % a.checkpoint_interval == 0 and steps != 0:
+                    checkpoint_path = "{}/g_{:08d}".format(a.checkpoint_path, steps)
+                    save_checkpoint(checkpoint_path,
+                                    {'generator': (generator.module if h.num_gpus > 1 else generator).state_dict()})
+                    checkpoint_path = "{}/do_{:08d}".format(a.checkpoint_path, steps)
+                    save_checkpoint(checkpoint_path, 
+                                    {'mpd': (mpd.module if h.num_gpus > 1
+                                                         else mpd).state_dict(),
+                                     'msd': (msd.module if h.num_gpus > 1
+                                                         else msd).state_dict(),
+                                     'optim_g': optim_g.state_dict(), 'optim_d': optim_d.state_dict(), 'steps': steps,
+                                     'epoch': epoch})
+
+                # Tensorboard summary logging
+                if steps % a.summary_interval == 0:
+                    sw.add_scalar("training/gen_loss_total", loss_gen_all, steps)
+                    sw.add_scalar("training/mel_spec_error", mel_error, steps)
+
+                # Validation
+                if steps % a.validation_interval == 0:  # and steps != 0:
+                    generator.eval()
+                    torch.cuda.empty_cache()
+                    val_err_tot = 0
+                    with torch.no_grad():
+                        for j, batch in enumerate(validation_loader):
+                            x, y, _, y_mel = batch
+                            y_g_hat = generator(x.to(device))
+                            y_mel = torch.autograd.Variable(y_mel.to(device, non_blocking=True))
+                            y_g_hat_mel = mel_spectrogram(y_g_hat.squeeze(1), h.n_fft, h.num_mels, h.sampling_rate,
+                                                          h.hop_size, h.win_size,
+                                                          h.fmin, h.fmax_for_loss)
+                            val_err_tot += F.l1_loss(y_mel, y_g_hat_mel).item()
+
+                            if j <= 4:
+                                if steps == 0:
+                                    sw.add_audio('gt/y_{}'.format(j), y[0], steps, h.sampling_rate)
+                                    sw.add_figure('gt/y_spec_{}'.format(j), plot_spectrogram(x[0]), steps)
+
+                                sw.add_audio('generated/y_hat_{}'.format(j), y_g_hat[0], steps, h.sampling_rate)
+                                y_hat_spec = mel_spectrogram(y_g_hat.squeeze(1), h.n_fft, h.num_mels,
+                                                             h.sampling_rate, h.hop_size, h.win_size,
+                                                             h.fmin, h.fmax)
+                                sw.add_figure('generated/y_hat_spec_{}'.format(j),
+                                              plot_spectrogram(y_hat_spec.squeeze(0).cpu().numpy()), steps)
+
+                        val_err = val_err_tot / (j+1)
+                        sw.add_scalar("validation/mel_spec_error", val_err, steps)
+
+                    generator.train()
+
+            steps += 1
+
+        scheduler_g.step()
+        scheduler_d.step()
+        
+        if rank == 0:
+            print('Time taken for epoch {} is {} sec\n'.format(epoch + 1, int(time.time() - start)))
+
+
+def main():
+    print('Initializing Training Process..')
+
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument('--group_name', default=None)
+    parser.add_argument('--input_wavs_dir', default='LJSpeech-1.1/wavs')
+    parser.add_argument('--input_mels_dir', default='ft_dataset')
+    parser.add_argument('--input_training_file', default='LJSpeech-1.1/training.txt')
+    parser.add_argument('--input_validation_file', default='LJSpeech-1.1/validation.txt')
+    parser.add_argument('--checkpoint_path', default='cp_hifigan')
+    parser.add_argument('--config', default='')
+    parser.add_argument('--training_epochs', default=3100, type=int)
+    parser.add_argument('--stdout_interval', default=5, type=int)
+    parser.add_argument('--checkpoint_interval', default=5000, type=int)
+    parser.add_argument('--summary_interval', default=100, type=int)
+    parser.add_argument('--validation_interval', default=1000, type=int)
+    parser.add_argument('--fine_tuning', default=False, type=bool)
+
+    a = parser.parse_args()
+
+    with open(a.config) as f:
+        data = f.read()
+
+    json_config = json.loads(data)
+    h = AttrDict(json_config)
+    build_env(a.config, 'config.json', a.checkpoint_path)
+
+    torch.manual_seed(h.seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(h.seed)
+        h.num_gpus = torch.cuda.device_count()
+        h.batch_size = int(h.batch_size / h.num_gpus)
+        print('Batch size per GPU :', h.batch_size)
+    else:
+        pass
+
+    if h.num_gpus > 1:
+        mp.spawn(train, nprocs=h.num_gpus, args=(a, h,))
+    else:
+        train(0, a, h)
+
+
+if __name__ == '__main__':
+    main()

hifi-gan/utils.py

@@ -0,0 +1,58 @@
+import glob
+import os
+import matplotlib
+import torch
+from torch.nn.utils import weight_norm
+matplotlib.use("Agg")
+import matplotlib.pylab as plt
+
+
+def plot_spectrogram(spectrogram):
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower",
+                   interpolation='none')
+    plt.colorbar(im, ax=ax)
+
+    fig.canvas.draw()
+    plt.close()
+
+    return fig
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def apply_weight_norm(m):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        weight_norm(m)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size*dilation - dilation)/2)
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print("Loading '{}'".format(filepath))
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print("Complete.")
+    return checkpoint_dict
+
+
+def save_checkpoint(filepath, obj):
+    print("Saving checkpoint to {}".format(filepath))
+    torch.save(obj, filepath)
+    print("Complete.")
+
+
+def scan_checkpoint(cp_dir, prefix):
+    pattern = os.path.join(cp_dir, prefix + '????????')
+    cp_list = glob.glob(pattern)
+    if len(cp_list) == 0:
+        return None
+    return sorted(cp_list)[-1]
+