xinference 1.10.0__py3-none-any.whl → 1.11.0__py3-none-any.whl

xinference/thirdparty/indextts/s2mel/modules/bigvgan/config.json

@@ -0,0 +1,63 @@
+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 32,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [4,4,2,2,2,2],
+    "upsample_kernel_sizes": [8,8,4,4,4,4],
+    "upsample_initial_channel": 1536,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "use_tanh_at_final": false,
+    "use_bias_at_final": false,
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "use_cqtd_instead_of_mrd": true,
+    "cqtd_filters": 128,
+    "cqtd_max_filters": 1024,
+    "cqtd_filters_scale": 1,
+    "cqtd_dilations": [1, 2, 4],
+    "cqtd_hop_lengths": [512, 256, 256],
+    "cqtd_n_octaves": [9, 9, 9],
+    "cqtd_bins_per_octaves": [24, 36, 48],
+
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+    
+    "use_multiscale_melloss": true,
+    "lambda_melloss": 15,
+
+    "clip_grad_norm": 500,
+
+    "segment_size": 65536,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 22050,
+
+    "fmin": 0,
+    "fmax": null,
+    "fmax_for_loss": null,
+
+    "normalize_volume": true,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

xinference/thirdparty/indextts/s2mel/modules/bigvgan/env.py

@@ -0,0 +1,18 @@
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+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))

xinference/thirdparty/indextts/s2mel/modules/bigvgan/meldataset.py

@@ -0,0 +1,354 @@
+# Copyright (c) 2024 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+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
+import pathlib
+from tqdm import tqdm
+
+MAX_WAV_VALUE = 32767.0  # NOTE: 32768.0 -1 to prevent int16 overflow (results in popping sound in corner cases)
+
+
+def load_wav(full_path, sr_target):
+    sampling_rate, data = read(full_path)
+    if sampling_rate != sr_target:
+        raise RuntimeError(
+            f"Sampling rate of the file {full_path} is {sampling_rate} Hz, but the model requires {sr_target} Hz"
+        )
+    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):
+    return dynamic_range_compression_torch(magnitudes)
+
+
+def spectral_de_normalize_torch(magnitudes):
+    return dynamic_range_decompression_torch(magnitudes)
+
+
+mel_basis_cache = {}
+hann_window_cache = {}
+
+
+def mel_spectrogram(
+    y: torch.Tensor,
+    n_fft: int,
+    num_mels: int,
+    sampling_rate: int,
+    hop_size: int,
+    win_size: int,
+    fmin: int,
+    fmax: int = None,
+    center: bool = False,
+) -> torch.Tensor:
+    """
+    Calculate the mel spectrogram of an input signal.
+    This function uses slaney norm for the librosa mel filterbank (using librosa.filters.mel) and uses Hann window for STFT (using torch.stft).
+
+    Args:
+        y (torch.Tensor): Input signal.
+        n_fft (int): FFT size.
+        num_mels (int): Number of mel bins.
+        sampling_rate (int): Sampling rate of the input signal.
+        hop_size (int): Hop size for STFT.
+        win_size (int): Window size for STFT.
+        fmin (int): Minimum frequency for mel filterbank.
+        fmax (int): Maximum frequency for mel filterbank. If None, defaults to half the sampling rate (fmax = sr / 2.0) inside librosa_mel_fn
+        center (bool): Whether to pad the input to center the frames. Default is False.
+
+    Returns:
+        torch.Tensor: Mel spectrogram.
+    """
+    if torch.min(y) < -1.0:
+        print(f"[WARNING] Min value of input waveform signal is {torch.min(y)}")
+    if torch.max(y) > 1.0:
+        print(f"[WARNING] Max value of input waveform signal is {torch.max(y)}")
+
+    device = y.device
+    key = f"{n_fft}_{num_mels}_{sampling_rate}_{hop_size}_{win_size}_{fmin}_{fmax}_{device}"
+
+    if key not in mel_basis_cache:
+        mel = librosa_mel_fn(
+            sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax
+        )
+        mel_basis_cache[key] = torch.from_numpy(mel).float().to(device)
+        hann_window_cache[key] = torch.hann_window(win_size).to(device)
+
+    mel_basis = mel_basis_cache[key]
+    hann_window = hann_window_cache[key]
+
+    padding = (n_fft - hop_size) // 2
+    y = torch.nn.functional.pad(
+        y.unsqueeze(1), (padding, padding), mode="reflect"
+    ).squeeze(1)
+
+    spec = torch.stft(
+        y,
+        n_fft,
+        hop_length=hop_size,
+        win_length=win_size,
+        window=hann_window,
+        center=center,
+        pad_mode="reflect",
+        normalized=False,
+        onesided=True,
+        return_complex=True,
+    )
+    spec = torch.sqrt(torch.view_as_real(spec).pow(2).sum(-1) + 1e-9)
+
+    mel_spec = torch.matmul(mel_basis, spec)
+    mel_spec = spectral_normalize_torch(mel_spec)
+
+    return mel_spec
+
+
+def get_mel_spectrogram(wav, h):
+    """
+    Generate mel spectrogram from a waveform using given hyperparameters.
+
+    Args:
+        wav (torch.Tensor): Input waveform.
+        h: Hyperparameters object with attributes n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax.
+
+    Returns:
+        torch.Tensor: Mel spectrogram.
+    """
+    return mel_spectrogram(
+        wav,
+        h.n_fft,
+        h.num_mels,
+        h.sampling_rate,
+        h.hop_size,
+        h.win_size,
+        h.fmin,
+        h.fmax,
+    )
+
+
+def get_dataset_filelist(a):
+    training_files = []
+    validation_files = []
+    list_unseen_validation_files = []
+
+    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
+        ]
+        print(f"first training file: {training_files[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
+        ]
+        print(f"first validation file: {validation_files[0]}")
+
+    for i in range(len(a.list_input_unseen_validation_file)):
+        with open(a.list_input_unseen_validation_file[i], "r", encoding="utf-8") as fi:
+            unseen_validation_files = [
+                os.path.join(a.list_input_unseen_wavs_dir[i], x.split("|")[0] + ".wav")
+                for x in fi.read().split("\n")
+                if len(x) > 0
+            ]
+            print(
+                f"first unseen {i}th validation fileset: {unseen_validation_files[0]}"
+            )
+            list_unseen_validation_files.append(unseen_validation_files)
+
+    return training_files, validation_files, list_unseen_validation_files
+
+
+class MelDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        training_files,
+        hparams,
+        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,
+        is_seen=True,
+    ):
+        self.audio_files = training_files
+        random.seed(1234)
+        if shuffle:
+            random.shuffle(self.audio_files)
+        self.hparams = hparams
+        self.is_seen = is_seen
+        if self.is_seen:
+            self.name = pathlib.Path(self.audio_files[0]).parts[0]
+        else:
+            self.name = "-".join(pathlib.Path(self.audio_files[0]).parts[:2]).strip("/")
+
+        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
+
+        print("[INFO] checking dataset integrity...")
+        for i in tqdm(range(len(self.audio_files))):
+            assert os.path.exists(
+                self.audio_files[i]
+            ), f"{self.audio_files[i]} not found"
+
+    def __getitem__(self, index):
+        filename = self.audio_files[index]
+        if self._cache_ref_count == 0:
+            audio, sampling_rate = load_wav(filename, self.sampling_rate)
+            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(
+                    f"{sampling_rate} SR doesn't match target {self.sampling_rate} SR"
+                )
+            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:  # Validation step
+                # Match audio length to self.hop_size * n for evaluation
+                if (audio.size(1) % self.hop_size) != 0:
+                    audio = audio[:, : -(audio.size(1) % self.hop_size)]
+                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,
+                )
+                assert (
+                    audio.shape[1] == mel.shape[2] * self.hop_size
+                ), f"audio shape {audio.shape} mel shape {mel.shape}"
+
+        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)

xinference/thirdparty/indextts/s2mel/modules/bigvgan/utils.py

@@ -0,0 +1,99 @@
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import glob
+import os
+import matplotlib
+import torch
+from torch.nn.utils import weight_norm
+
+matplotlib.use("Agg")
+import matplotlib.pylab as plt
+from .meldataset import MAX_WAV_VALUE
+from scipy.io.wavfile import write
+
+
+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 plot_spectrogram_clipped(spectrogram, clip_max=2.0):
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(
+        spectrogram,
+        aspect="auto",
+        origin="lower",
+        interpolation="none",
+        vmin=1e-6,
+        vmax=clip_max,
+    )
+    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(f"Loading '{filepath}'")
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print("Complete.")
+    return checkpoint_dict
+
+
+def save_checkpoint(filepath, obj):
+    print(f"Saving checkpoint to {filepath}")
+    torch.save(obj, filepath)
+    print("Complete.")
+
+
+def scan_checkpoint(cp_dir, prefix, renamed_file=None):
+    # Fallback to original scanning logic first
+    pattern = os.path.join(cp_dir, prefix + "????????")
+    cp_list = glob.glob(pattern)
+
+    if len(cp_list) > 0:
+        last_checkpoint_path = sorted(cp_list)[-1]
+        print(f"[INFO] Resuming from checkpoint: '{last_checkpoint_path}'")
+        return last_checkpoint_path
+
+    # If no pattern-based checkpoints are found, check for renamed file
+    if renamed_file:
+        renamed_path = os.path.join(cp_dir, renamed_file)
+        if os.path.isfile(renamed_path):
+            print(f"[INFO] Resuming from renamed checkpoint: '{renamed_file}'")
+            return renamed_path
+
+    return None
+
+
+def save_audio(audio, path, sr):
+    # wav: torch with 1d shape
+    audio = audio * MAX_WAV_VALUE
+    audio = audio.cpu().numpy().astype("int16")
+    write(path, sr, audio)

xinference/thirdparty/indextts/s2mel/modules/campplus/DTDNN.py

@@ -0,0 +1,115 @@
+# Copyright 3D-Speaker (https://github.com/alibaba-damo-academy/3D-Speaker). All Rights Reserved.
+# Licensed under the Apache License, Version 2.0 (http://www.apache.org/licenses/LICENSE-2.0)
+
+from collections import OrderedDict
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+from indextts.s2mel.modules.campplus.layers import DenseLayer, StatsPool, TDNNLayer, CAMDenseTDNNBlock, TransitLayer, BasicResBlock, get_nonlinear
+
+
+class FCM(nn.Module):
+    def __init__(self,
+                block=BasicResBlock,
+                num_blocks=[2, 2],
+                m_channels=32,
+                feat_dim=80):
+        super(FCM, self).__init__()
+        self.in_planes = m_channels
+        self.conv1 = nn.Conv2d(1, m_channels, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(m_channels)
+
+        self.layer1 = self._make_layer(block, m_channels, num_blocks[0], stride=2)
+        self.layer2 = self._make_layer(block, m_channels, num_blocks[1], stride=2)
+
+        self.conv2 = nn.Conv2d(m_channels, m_channels, kernel_size=3, stride=(2, 1), padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(m_channels)
+        self.out_channels =  m_channels * (feat_dim // 8)
+
+    def _make_layer(self, block, planes, num_blocks, stride):
+        strides = [stride] + [1] * (num_blocks - 1)
+        layers = []
+        for stride in strides:
+            layers.append(block(self.in_planes, planes, stride))
+            self.in_planes = planes * block.expansion
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = x.unsqueeze(1)
+        out = F.relu(self.bn1(self.conv1(x)))
+        out = self.layer1(out)
+        out = self.layer2(out)
+        out = F.relu(self.bn2(self.conv2(out)))
+
+        shape = out.shape
+        out = out.reshape(shape[0], shape[1]*shape[2], shape[3])
+        return out
+
+class CAMPPlus(nn.Module):
+    def __init__(self,
+                 feat_dim=80,
+                 embedding_size=512,
+                 growth_rate=32,
+                 bn_size=4,
+                 init_channels=128,
+                 config_str='batchnorm-relu',
+                 memory_efficient=True):
+        super(CAMPPlus, self).__init__()
+
+        self.head = FCM(feat_dim=feat_dim)
+        channels = self.head.out_channels
+
+        self.xvector = nn.Sequential(
+            OrderedDict([
+
+                ('tdnn',
+                 TDNNLayer(channels,
+                           init_channels,
+                           5,
+                           stride=2,
+                           dilation=1,
+                           padding=-1,
+                           config_str=config_str)),
+            ]))
+        channels = init_channels
+        for i, (num_layers, kernel_size,
+                dilation) in enumerate(zip((12, 24, 16), (3, 3, 3), (1, 2, 2))):
+            block = CAMDenseTDNNBlock(num_layers=num_layers,
+                                   in_channels=channels,
+                                   out_channels=growth_rate,
+                                   bn_channels=bn_size * growth_rate,
+                                   kernel_size=kernel_size,
+                                   dilation=dilation,
+                                   config_str=config_str,
+                                   memory_efficient=memory_efficient)
+            self.xvector.add_module('block%d' % (i + 1), block)
+            channels = channels + num_layers * growth_rate
+            self.xvector.add_module(
+                'transit%d' % (i + 1),
+                TransitLayer(channels,
+                             channels // 2,
+                             bias=False,
+                             config_str=config_str))
+            channels //= 2
+
+        self.xvector.add_module(
+            'out_nonlinear', get_nonlinear(config_str, channels))
+
+        self.xvector.add_module('stats', StatsPool())
+        self.xvector.add_module(
+            'dense',
+            DenseLayer(channels * 2, embedding_size, config_str='batchnorm_'))
+
+        for m in self.modules():
+            if isinstance(m, (nn.Conv1d, nn.Linear)):
+                nn.init.kaiming_normal_(m.weight.data)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+    def forward(self, x):
+        x = x.permute(0, 2, 1)  # (B,T,F) => (B,F,T)
+        x = self.head(x)
+        x = self.xvector(x)
+        return x