Upload folder using huggingface_hub

infer/__init__.py

@@ -0,0 +1,17 @@
+# -*- coding: utf-8 -*-
+"""
+推理模块
+"""
+from .f0_extractor import (
+    F0Extractor,
+    get_f0_extractor,
+    shift_f0,
+    F0Method
+)
+
+__all__ = [
+    "F0Extractor",
+    "get_f0_extractor",
+    "shift_f0",
+    "F0Method"
+]

infer/advanced_dereverb.py

@@ -0,0 +1,280 @@
+# -*- coding: utf-8 -*-
+"""
+高级去混响模块 - 基于二进制残差掩码和时域一致性
+参考: arXiv 2510.00356 - Dereverberation Using Binary Residual Masking
+"""
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Tuple, Optional
+
+
+class BinaryResidualMask(nn.Module):
+    """
+    二进制残差掩码网络 - 专注于抑制混响而非预测完整频谱
+
+    核心思想：
+    1. 学习识别并抑制晚期反射（late reflections）
+    2. 保留直达声路径（direct path）
+    3. 使用时域一致性损失隐式学习相位
+    """
+
+    def __init__(self, n_fft=2048, hop_length=512):
+        super().__init__()
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.freq_bins = n_fft // 2 + 1
+
+        # U-Net编码器
+        self.encoder1 = nn.Sequential(
+            nn.Conv2d(1, 32, kernel_size=3, padding=1),
+            nn.BatchNorm2d(32),
+            nn.ReLU(),
+            nn.Conv2d(32, 32, kernel_size=3, padding=1),
+            nn.BatchNorm2d(32),
+            nn.ReLU()
+        )
+
+        self.encoder2 = nn.Sequential(
+            nn.MaxPool2d(2),
+            nn.Conv2d(32, 64, kernel_size=3, padding=1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(),
+            nn.Conv2d(64, 64, kernel_size=3, padding=1),
+            nn.BatchNorm2d(64),
+            nn.ReLU()
+        )
+
+        self.encoder3 = nn.Sequential(
+            nn.MaxPool2d(2),
+            nn.Conv2d(64, 128, kernel_size=3, padding=1),
+            nn.BatchNorm2d(128),
+            nn.ReLU(),
+            nn.Conv2d(128, 128, kernel_size=3, padding=1),
+            nn.BatchNorm2d(128),
+            nn.ReLU()
+        )
+
+        # 瓶颈层 - 时序注意力
+        self.bottleneck = nn.Sequential(
+            nn.MaxPool2d(2),
+            nn.Conv2d(128, 256, kernel_size=3, padding=1),
+            nn.BatchNorm2d(256),
+            nn.ReLU()
+        )
+
+        # U-Net解码器
+        self.decoder3 = nn.Sequential(
+            nn.ConvTranspose2d(256, 128, kernel_size=2, stride=2),
+            nn.Conv2d(256, 128, kernel_size=3, padding=1),
+            nn.BatchNorm2d(128),
+            nn.ReLU()
+        )
+
+        self.decoder2 = nn.Sequential(
+            nn.ConvTranspose2d(128, 64, kernel_size=2, stride=2),
+            nn.Conv2d(128, 64, kernel_size=3, padding=1),
+            nn.BatchNorm2d(64),
+            nn.ReLU()
+        )
+
+        self.decoder1 = nn.Sequential(
+            nn.ConvTranspose2d(64, 32, kernel_size=2, stride=2),
+            nn.Conv2d(64, 32, kernel_size=3, padding=1),
+            nn.BatchNorm2d(32),
+            nn.ReLU()
+        )
+
+        # 输出层 - 二进制掩码
+        self.output = nn.Sequential(
+            nn.Conv2d(32, 1, kernel_size=1),
+            nn.Sigmoid()  # 输出0-1的掩码
+        )
+
+    def forward(self, x):
+        """
+        Args:
+            x: [B, 1, F, T] - 输入频谱幅度
+        Returns:
+            mask: [B, 1, F, T] - 二进制残差掩码
+        """
+        # 编码
+        e1 = self.encoder1(x)
+        e2 = self.encoder2(e1)
+        e3 = self.encoder3(e2)
+
+        # 瓶颈
+        b = self.bottleneck(e3)
+
+        # 解码 + 跳跃连接
+        d3 = self.decoder3(b)
+        d3 = torch.cat([d3, e3], dim=1)
+
+        d2 = self.decoder2(d3)
+        d2 = torch.cat([d2, e2], dim=1)
+
+        d1 = self.decoder1(d2)
+        d1 = torch.cat([d1, e1], dim=1)
+
+        # 输出掩码
+        mask = self.output(d1)
+        return mask
+
+
+def advanced_dereverb(
+    audio: np.ndarray,
+    sr: int = 16000,
+    n_fft: int = 2048,
+    hop_length: int = 512,
+    device: str = "cuda"
+) -> Tuple[np.ndarray, np.ndarray]:
+    """
+    高级去混响 - 分离干声和混响
+
+    Args:
+        audio: 输入音频 [samples]
+        sr: 采样率
+        n_fft: FFT大小
+        hop_length: 跳跃长度
+        device: 计算设备
+
+    Returns:
+        dry_signal: 干声（直达声）
+        reverb_tail: 混响尾巴
+    """
+    import librosa
+
+    # STFT
+    spec = librosa.stft(audio, n_fft=n_fft, hop_length=hop_length, win_length=n_fft)
+    mag = np.abs(spec).astype(np.float32)
+    phase = np.angle(spec)
+
+    # 基于能量的混响检测
+    # 1. 计算时域RMS能量
+    rms = librosa.feature.rms(y=audio, frame_length=n_fft, hop_length=hop_length, center=True)[0]
+    rms_db = 20.0 * np.log10(rms + 1e-8)
+    ref_db = float(np.percentile(rms_db, 90))
+
+    # 2. 检测晚期反射（late reflections）
+    # 晚期反射特征：能量衰减 + 时间延迟
+    late_reflections = np.zeros_like(mag, dtype=np.float32)
+
+    for t in range(2, mag.shape[1]):
+        # 递归估计：衰减的历史 + 延迟的观测
+        late_reflections[:, t] = np.maximum(
+            late_reflections[:, t - 1] * 0.92,  # 衰减系数
+            mag[:, t - 2] * 0.80  # 延迟观测
+        )
+
+    # 3. 计算直达声（direct path）
+    # 直达声 = 总能量 - 晚期反射
+    direct_path = np.maximum(mag - 0.75 * late_reflections, 0.0)
+
+    # 4. 动态floor：保护有声段
+    # 扩展RMS到频谱帧数
+    if len(rms) < mag.shape[1]:
+        rms_extended = np.pad(rms, (0, mag.shape[1] - len(rms)), mode='edge')
+    else:
+        rms_extended = rms[:mag.shape[1]]
+
+    # 有声段（高能量）：vocal_strength接近1
+    # 无声段（低能量/混响尾）：vocal_strength接近0
+    vocal_strength = np.clip((rms_db[:len(rms_extended)] - (ref_db - 35.0)) / 25.0, 0.0, 1.0)
+
+    # 动态floor系数
+    reverb_ratio = np.clip(late_reflections / (mag + 1e-8), 0.0, 1.0)
+    floor_coef = 0.08 + 0.12 * vocal_strength[np.newaxis, :]
+    floor = (1.0 - reverb_ratio) * floor_coef * mag
+    direct_path = np.maximum(direct_path, floor)
+
+    # 5. 时域平滑（避免音乐噪声）
+    kernel = np.array([1, 2, 3, 2, 1], dtype=np.float32)
+    kernel /= np.sum(kernel)
+    direct_path = np.apply_along_axis(
+        lambda row: np.convolve(row, kernel, mode="same"),
+        axis=1,
+        arr=direct_path,
+    )
+    direct_path = np.clip(direct_path, 0.0, mag)
+
+    # 6. 计算混响残差
+    reverb_mag = mag - direct_path
+    reverb_mag = np.maximum(reverb_mag, 0.0)
+
+    # 7. 重建音频
+    # 干声：使用原始相位（保留音色）
+    dry_spec = direct_path * np.exp(1j * phase)
+    dry_signal = librosa.istft(dry_spec, hop_length=hop_length, win_length=n_fft, length=len(audio))
+
+    # 混响：使用原始相位
+    reverb_spec = reverb_mag * np.exp(1j * phase)
+    reverb_tail = librosa.istft(reverb_spec, hop_length=hop_length, win_length=n_fft, length=len(audio))
+
+    return dry_signal.astype(np.float32), reverb_tail.astype(np.float32)
+
+
+def apply_reverb_to_converted(
+    converted_dry: np.ndarray,
+    original_reverb: np.ndarray,
+    mix_ratio: float = 0.8
+) -> np.ndarray:
+    """
+    将原始混响重新应用到转换后的干声上
+
+    Args:
+        converted_dry: 转换后的干声
+        original_reverb: 原始混响尾巴
+        mix_ratio: 混响混合比例 (0-1)
+
+    Returns:
+        wet_signal: 带混响的转换结果
+    """
+    # 对齐长度
+    min_len = min(len(converted_dry), len(original_reverb))
+    converted_dry = converted_dry[:min_len]
+    original_reverb = original_reverb[:min_len]
+
+    # 混合
+    wet_signal = converted_dry + mix_ratio * original_reverb
+
+    # 软限幅
+    from lib.audio import soft_clip
+    wet_signal = soft_clip(wet_signal, threshold=0.9, ceiling=0.99)
+
+    return wet_signal.astype(np.float32)
+
+
+if __name__ == "__main__":
+    # 测试
+    print("Testing advanced dereverberation...")
+
+    # 生成测试信号：干声 + 混响
+    sr = 16000
+    duration = 2.0
+    t = np.linspace(0, duration, int(sr * duration))
+
+    # 干声：440Hz正弦波
+    dry = np.sin(2 * np.pi * 440 * t).astype(np.float32)
+
+    # 混响：衰减的延迟
+    reverb = np.zeros_like(dry)
+    delay_samples = int(0.05 * sr)  # 50ms延迟
+    for i in range(3):
+        delay = delay_samples * (i + 1)
+        decay = 0.5 ** (i + 1)
+        if delay < len(reverb):
+            reverb[delay:] += dry[:-delay] * decay
+
+    # 混合信号
+    wet = dry + reverb * 0.5
+
+    # 去混响
+    dry_extracted, reverb_extracted = advanced_dereverb(wet, sr)
+
+    print(f"Input RMS: {np.sqrt(np.mean(wet**2)):.4f}")
+    print(f"Dry RMS: {np.sqrt(np.mean(dry_extracted**2)):.4f}")
+    print(f"Reverb RMS: {np.sqrt(np.mean(reverb_extracted**2)):.4f}")
+    print(f"Separation ratio: {np.sqrt(np.mean(dry_extracted**2)) / (np.sqrt(np.mean(reverb_extracted**2)) + 1e-8):.2f}")
+
+    print("\n[OK] Advanced dereverberation test passed!")

infer/f0_extractor.py

@@ -0,0 +1,265 @@
+# -*- coding: utf-8 -*-
+"""
+F0 (基频) 提取模块 - 支持多种提取方法
+"""
+import numpy as np
+import torch
+from typing import Optional, Literal
+
+# F0 提取方法类型
+F0Method = Literal["rmvpe", "pm", "harvest", "crepe", "hybrid"]
+
+
+class F0Extractor:
+    """F0 提取器基类"""
+
+    def __init__(self, sample_rate: int = 16000, hop_length: int = 160):
+        self.sample_rate = sample_rate
+        self.hop_length = hop_length
+        self.f0_min = 50
+        self.f0_max = 1100
+
+    def extract(self, audio: np.ndarray) -> np.ndarray:
+        """提取 F0，子类需实现此方法"""
+        raise NotImplementedError
+
+
+class PMExtractor(F0Extractor):
+    """Parselmouth (Praat) F0 提取器 - 速度快"""
+
+    def extract(self, audio: np.ndarray) -> np.ndarray:
+        import parselmouth
+
+        time_step = self.hop_length / self.sample_rate
+        sound = parselmouth.Sound(audio, self.sample_rate)
+
+        pitch = sound.to_pitch_ac(
+            time_step=time_step,
+            voicing_threshold=0.6,
+            pitch_floor=self.f0_min,
+            pitch_ceiling=self.f0_max
+        )
+
+        f0 = pitch.selected_array["frequency"]
+        f0[f0 == 0] = np.nan
+
+        return f0
+
+
+class HarvestExtractor(F0Extractor):
+    """PyWorld Harvest F0 提取器 - 质量较好"""
+
+    def extract(self, audio: np.ndarray) -> np.ndarray:
+        import pyworld
+
+        audio = audio.astype(np.float64)
+        f0, _ = pyworld.harvest(
+            audio,
+            self.sample_rate,
+            f0_floor=self.f0_min,
+            f0_ceil=self.f0_max,
+            frame_period=self.hop_length / self.sample_rate * 1000
+        )
+
+        return f0
+
+
+class CrepeExtractor(F0Extractor):
+    """TorchCrepe F0 提取器 - 深度学习方法"""
+
+    def __init__(self, sample_rate: int = 16000, hop_length: int = 160,
+                 device: str = "cuda"):
+        super().__init__(sample_rate, hop_length)
+        self.device = device
+
+    def extract(self, audio: np.ndarray) -> np.ndarray:
+        import torchcrepe
+
+        audio_tensor = torch.from_numpy(audio).float().unsqueeze(0)
+        audio_tensor = audio_tensor.to(self.device)
+
+        f0, _ = torchcrepe.predict(
+            audio_tensor,
+            self.sample_rate,
+            self.hop_length,
+            self.f0_min,
+            self.f0_max,
+            model="full",
+            batch_size=512,
+            device=self.device,
+            return_periodicity=True
+        )
+
+        f0 = f0.squeeze(0).cpu().numpy()
+        return f0
+
+
+class RMVPEExtractor(F0Extractor):
+    """RMVPE F0 提取器 - 质量最高 (推荐)"""
+
+    def __init__(self, model_path: str, sample_rate: int = 16000,
+                 hop_length: int = 160, device: str = "cuda"):
+        super().__init__(sample_rate, hop_length)
+        self.device = device
+        self.model = None
+        self.model_path = model_path
+
+    def load_model(self):
+        """加载 RMVPE 模型"""
+        if self.model is not None:
+            return
+
+        from models.rmvpe import RMVPE
+
+        self.model = RMVPE(self.model_path, device=self.device)
+        print(f"RMVPE 模型已加载: {self.device}")
+
+    def extract(self, audio: np.ndarray) -> np.ndarray:
+        self.load_model()
+
+        # RMVPE 需要 16kHz 输入
+        f0 = self.model.infer_from_audio(audio, thred=0.01)
+
+        return f0
+
+
+def get_f0_extractor(method: F0Method, device: str = "cuda",
+                     rmvpe_path: str = None, crepe_threshold: float = 0.05) -> F0Extractor:
+    """
+    获取 F0 提取器实例
+
+    Args:
+        method: 提取方法 ("rmvpe", "pm", "harvest", "crepe", "hybrid")
+        device: 计算设备
+        rmvpe_path: RMVPE 模型路径 (rmvpe/hybrid 方法需要)
+        crepe_threshold: CREPE置信度阈值 (仅hybrid方法使用)
+
+    Returns:
+        F0Extractor: 提取器实例
+    """
+    if method == "rmvpe":
+        if rmvpe_path is None:
+            raise ValueError("RMVPE 方法需要指定模型路径")
+        return RMVPEExtractor(rmvpe_path, device=device)
+    elif method == "hybrid":
+        if rmvpe_path is None:
+            raise ValueError("Hybrid 方法需要指定RMVPE模型路径")
+        return HybridF0Extractor(rmvpe_path, device=device, crepe_threshold=crepe_threshold)
+    elif method == "pm":
+        return PMExtractor()
+    elif method == "harvest":
+        return HarvestExtractor()
+    elif method == "crepe":
+        return CrepeExtractor(device=device)
+    else:
+        raise ValueError(f"未知的 F0 提取方法: {method}")
+
+
+class HybridF0Extractor(F0Extractor):
+    """混合F0提取器 - RMVPE主导 + CREPE高精度补充"""
+
+    def __init__(self, rmvpe_path: str, sample_rate: int = 16000,
+                 hop_length: int = 160, device: str = "cuda",
+                 crepe_threshold: float = 0.05):
+        super().__init__(sample_rate, hop_length)
+        self.device = device
+        self.rmvpe = RMVPEExtractor(rmvpe_path, sample_rate, hop_length, device)
+        self.crepe = None  # 延迟加载
+        self.crepe_threshold = crepe_threshold
+
+    def _load_crepe(self):
+        """延迟加载CREPE模型"""
+        if self.crepe is None:
+            try:
+                self.crepe = CrepeExtractor(self.sample_rate, self.hop_length, self.device)
+            except ImportError:
+                print("警告: torchcrepe未安装，混合F0将仅使用RMVPE")
+                self.crepe = False
+
+    def extract(self, audio: np.ndarray) -> np.ndarray:
+        """
+        混合提取F0：
+        1. 使用RMVPE作为主要方法（快速、稳定）
+        2. 在RMVPE不稳定的区域使用CREPE补充（高精度）
+        """
+        # 提取RMVPE F0
+        f0_rmvpe = self.rmvpe.extract(audio)
+
+        # 如果CREPE不可用，直接返回RMVPE结果
+        self._load_crepe()
+        if self.crepe is False:
+            return f0_rmvpe
+
+        # 提取CREPE F0和置信度
+        import torchcrepe
+        audio_tensor = torch.from_numpy(audio).float().unsqueeze(0).to(self.device)
+        f0_crepe, confidence = torchcrepe.predict(
+            audio_tensor,
+            self.sample_rate,
+            self.hop_length,
+            self.f0_min,
+            self.f0_max,
+            model="full",
+            batch_size=512,
+            device=self.device,
+            return_periodicity=True
+        )
+        f0_crepe = f0_crepe.squeeze(0).cpu().numpy()
+        confidence = confidence.squeeze(0).cpu().numpy()
+
+        # 对齐长度
+        min_len = min(len(f0_rmvpe), len(f0_crepe), len(confidence))
+        f0_rmvpe = f0_rmvpe[:min_len]
+        f0_crepe = f0_crepe[:min_len]
+        confidence = confidence[:min_len]
+
+        # 检测RMVPE不稳定区域
+        # 1. F0跳变过大（超过3个半音）
+        f0_diff = np.abs(np.diff(f0_rmvpe, prepend=f0_rmvpe[0]))
+        semitone_diff = np.abs(12 * np.log2((f0_rmvpe + 1e-6) / (np.roll(f0_rmvpe, 1) + 1e-6)))
+        semitone_diff[0] = 0
+        unstable_jump = semitone_diff > 3.0
+
+        # 2. CREPE置信度高但RMVPE给出F0=0
+        unstable_unvoiced = (f0_rmvpe < 1e-3) & (confidence > self.crepe_threshold)
+
+        # 3. RMVPE和CREPE差异过大（超过2个半音）且CREPE置信度高
+        f0_ratio = (f0_crepe + 1e-6) / (f0_rmvpe + 1e-6)
+        semitone_gap = np.abs(12 * np.log2(f0_ratio))
+        unstable_diverge = (semitone_gap > 2.0) & (confidence > self.crepe_threshold * 1.5)
+
+        # 合并不稳定区域
+        unstable_mask = unstable_jump | unstable_unvoiced | unstable_diverge
+
+        # 扩展不稳定区域（前后各2帧）以平滑过渡
+        kernel = np.ones(5, dtype=bool)
+        unstable_mask = np.convolve(unstable_mask, kernel, mode='same')
+
+        # 混合F0：不稳定区域使用CREPE，其他区域使用RMVPE
+        f0_hybrid = f0_rmvpe.copy()
+        f0_hybrid[unstable_mask] = f0_crepe[unstable_mask]
+
+        # 平滑过渡边界
+        for i in range(1, len(f0_hybrid) - 1):
+            if unstable_mask[i] != unstable_mask[i-1]:
+                # 边界处使用加权平均
+                w = 0.5
+                f0_hybrid[i] = w * f0_rmvpe[i] + (1-w) * f0_crepe[i]
+
+        return f0_hybrid
+
+
+def shift_f0(f0: np.ndarray, semitones: float) -> np.ndarray:
+    """
+    音调偏移
+
+    Args:
+        f0: 原始 F0
+        semitones: 偏移半音数 (正数升调，负数降调)
+
+    Returns:
+        np.ndarray: 偏移后的 F0
+    """
+    factor = 2 ** (semitones / 12)
+    f0_shifted = f0 * factor
+    return f0_shifted

infer/lib/audio.py

@@ -0,0 +1,60 @@
+import platform, os
+import ffmpeg
+import numpy as np
+import av
+from io import BytesIO
+import traceback
+import re
+
+
+def wav2(i, o, format):
+    inp = av.open(i, "rb")
+    if format == "m4a":
+        format = "mp4"
+    out = av.open(o, "wb", format=format)
+    if format == "ogg":
+        format = "libvorbis"
+    if format == "mp4":
+        format = "aac"
+
+    ostream = out.add_stream(format)
+
+    for frame in inp.decode(audio=0):
+        for p in ostream.encode(frame):
+            out.mux(p)
+
+    for p in ostream.encode(None):
+        out.mux(p)
+
+    out.close()
+    inp.close()
+
+
+def load_audio(file, sr):
+    try:
+        # https://github.com/openai/whisper/blob/main/whisper/audio.py#L26
+        # This launches a subprocess to decode audio while down-mixing and resampling as necessary.
+        # Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
+        file = clean_path(file)  # 防止小白拷路径头尾带了空格和"和回车
+        if os.path.exists(file) == False:
+            raise RuntimeError(
+                "You input a wrong audio path that does not exists, please fix it!"
+            )
+        out, _ = (
+            ffmpeg.input(file, threads=0)
+            .output("-", format="f32le", acodec="pcm_f32le", ac=1, ar=sr)
+            .run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True)
+        )
+    except Exception as e:
+        traceback.print_exc()
+        raise RuntimeError(f"Failed to load audio: {e}")
+
+    return np.frombuffer(out, np.float32).flatten()
+
+
+
+def clean_path(path_str):
+    if platform.system() == "Windows":
+        path_str = path_str.replace("/", "\\")
+    path_str = re.sub(r'[\u202a\u202b\u202c\u202d\u202e]', '', path_str)  # 移除 Unicode 控制字符
+    return path_str.strip(" ").strip('"').strip("\n").strip('"').strip(" ")

infer/lib/infer_pack/attentions.py

@@ -0,0 +1,459 @@
+import copy
+import math
+from typing import Optional
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from infer.lib.infer_pack import commons, modules
+from infer.lib.infer_pack.modules import LayerNorm
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        window_size=10,
+        **kwargs
+    ):
+        super(Encoder, self).__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = int(n_layers)
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+
+        self.drop = nn.Dropout(p_dropout)
+        self.attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    window_size=window_size,
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        zippep = zip(
+            self.attn_layers, self.norm_layers_1, self.ffn_layers, self.norm_layers_2
+        )
+        for attn_layers, norm_layers_1, ffn_layers, norm_layers_2 in zippep:
+            y = attn_layers(x, x, attn_mask)
+            y = self.drop(y)
+            x = norm_layers_1(x + y)
+
+            y = ffn_layers(x, x_mask)
+            y = self.drop(y)
+            x = norm_layers_2(x + y)
+        x = x * x_mask
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        proximal_bias=False,
+        proximal_init=True,
+        **kwargs
+    ):
+        super(Decoder, self).__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+
+        self.drop = nn.Dropout(p_dropout)
+        self.self_attn_layers = nn.ModuleList()
+        self.norm_layers_0 = nn.ModuleList()
+        self.encdec_attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.self_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    proximal_bias=proximal_bias,
+                    proximal_init=proximal_init,
+                )
+            )
+            self.norm_layers_0.append(LayerNorm(hidden_channels))
+            self.encdec_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                    causal=True,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, h, h_mask):
+        """
+        x: decoder input
+        h: encoder output
+        """
+        self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(
+            device=x.device, dtype=x.dtype
+        )
+        encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            y = self.self_attn_layers[i](x, x, self_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_0[i](x + y)
+
+            y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels,
+        out_channels,
+        n_heads,
+        p_dropout=0.0,
+        window_size=None,
+        heads_share=True,
+        block_length=None,
+        proximal_bias=False,
+        proximal_init=False,
+    ):
+        super(MultiHeadAttention, self).__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+        self.attn = None
+
+        self.k_channels = channels // n_heads
+        self.conv_q = nn.Conv1d(channels, channels, 1)
+        self.conv_k = nn.Conv1d(channels, channels, 1)
+        self.conv_v = nn.Conv1d(channels, channels, 1)
+        self.conv_o = nn.Conv1d(channels, out_channels, 1)
+        self.drop = nn.Dropout(p_dropout)
+
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+            self.emb_rel_k = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+            self.emb_rel_v = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+
+        nn.init.xavier_uniform_(self.conv_q.weight)
+        nn.init.xavier_uniform_(self.conv_k.weight)
+        nn.init.xavier_uniform_(self.conv_v.weight)
+        if proximal_init:
+            with torch.no_grad():
+                self.conv_k.weight.copy_(self.conv_q.weight)
+                self.conv_k.bias.copy_(self.conv_q.bias)
+
+    def forward(
+        self, x: torch.Tensor, c: torch.Tensor, attn_mask: Optional[torch.Tensor] = None
+    ):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, _ = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+    ):
+        # reshape [b, d, t] -> [b, n_h, t, d_k]
+        b, d, t_s = key.size()
+        t_t = query.size(2)
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+        if self.window_size is not None:
+            assert (
+                t_s == t_t
+            ), "Relative attention is only available for self-attention."
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(
+                query / math.sqrt(self.k_channels), key_relative_embeddings
+            )
+            scores_local = self._relative_position_to_absolute_position(rel_logits)
+            scores = scores + scores_local
+        if self.proximal_bias:
+            assert t_s == t_t, "Proximal bias is only available for self-attention."
+            scores = scores + self._attention_bias_proximal(t_s).to(
+                device=scores.device, dtype=scores.dtype
+            )
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                assert (
+                    t_s == t_t
+                ), "Local attention is only available for self-attention."
+                block_mask = (
+                    torch.ones_like(scores)
+                    .triu(-self.block_length)
+                    .tril(self.block_length)
+                )
+                scores = scores.masked_fill(block_mask == 0, -1e4)
+        p_attn = F.softmax(scores, dim=-1)  # [b, n_h, t_t, t_s]
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(
+                self.emb_rel_v, t_s
+            )
+            output = output + self._matmul_with_relative_values(
+                relative_weights, value_relative_embeddings
+            )
+        output = (
+            output.transpose(2, 3).contiguous().view(b, d, t_t)
+        )  # [b, n_h, t_t, d_k] -> [b, d, t_t]
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        """
+        x: [b, h, l, m]
+        y: [h or 1, m, d]
+        ret: [b, h, l, d]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        """
+        x: [b, h, l, d]
+        y: [h or 1, m, d]
+        ret: [b, h, l, m]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length: int):
+        max_relative_position = 2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+        pad_length: int = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        if pad_length > 0:
+            padded_relative_embeddings = F.pad(
+                relative_embeddings,
+                # commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
+                [0, 0, pad_length, pad_length, 0, 0],
+            )
+        else:
+            padded_relative_embeddings = relative_embeddings
+        used_relative_embeddings = padded_relative_embeddings[
+            :, slice_start_position:slice_end_position
+        ]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        """
+        x: [b, h, l, 2*l-1]
+        ret: [b, h, l, l]
+        """
+        batch, heads, length, _ = x.size()
+        # Concat columns of pad to shift from relative to absolute indexing.
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]])
+            [0, 1, 0, 0, 0, 0, 0, 0],
+        )
+
+        # Concat extra elements so to add up to shape (len+1, 2*len-1).
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = F.pad(
+            x_flat,
+            # commons.convert_pad_shape([[0, 0], [0, 0], [0, int(length) - 1]])
+            [0, int(length) - 1, 0, 0, 0, 0],
+        )
+
+        # Reshape and slice out the padded elements.
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+            :, :, :length, length - 1 :
+        ]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        """
+        x: [b, h, l, l]
+        ret: [b, h, l, 2*l-1]
+        """
+        batch, heads, length, _ = x.size()
+        # padd along column
+        x = F.pad(
+            x,
+            # commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, int(length) - 1]])
+            [0, int(length) - 1, 0, 0, 0, 0, 0, 0],
+        )
+        x_flat = x.view([batch, heads, int(length**2) + int(length * (length - 1))])
+        # add 0's in the beginning that will skew the elements after reshape
+        x_flat = F.pad(
+            x_flat,
+            #    commons.convert_pad_shape([[0, 0], [0, 0], [int(length), 0]])
+            [length, 0, 0, 0, 0, 0],
+        )
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length: int):
+        """Bias for self-attention to encourage attention to close positions.
+        Args:
+          length: an integer scalar.
+        Returns:
+          a Tensor with shape [1, 1, length, length]
+        """
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout=0.0,
+        activation: str = None,
+        causal=False,
+    ):
+        super(FFN, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+        self.causal = causal
+        self.is_activation = True if activation == "gelu" else False
+        # if causal:
+        #     self.padding = self._causal_padding
+        # else:
+        #     self.padding = self._same_padding
+
+        self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+        self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+        self.drop = nn.Dropout(p_dropout)
+
+    def padding(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+        if self.causal:
+            padding = self._causal_padding(x * x_mask)
+        else:
+            padding = self._same_padding(x * x_mask)
+        return padding
+
+    def forward(self, x: torch.Tensor, x_mask: torch.Tensor):
+        x = self.conv_1(self.padding(x, x_mask))
+        if self.is_activation:
+            x = x * torch.sigmoid(1.702 * x)
+        else:
+            x = torch.relu(x)
+        x = self.drop(x)
+
+        x = self.conv_2(self.padding(x, x_mask))
+        return x * x_mask
+
+    def _causal_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l: int = self.kernel_size - 1
+        pad_r: int = 0
+        # padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape(padding)
+            [pad_l, pad_r, 0, 0, 0, 0],
+        )
+        return x
+
+    def _same_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l: int = (self.kernel_size - 1) // 2
+        pad_r: int = self.kernel_size // 2
+        # padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape(padding)
+            [pad_l, pad_r, 0, 0, 0, 0],
+        )
+        return x

infer/lib/infer_pack/attentions_onnx.py

@@ -0,0 +1,459 @@
+############################## Warning! ##############################
+#                                                                    #
+#           Onnx Export Not Support All Of Non-Torch Types           #
+#           Include Python Built-in Types!!!!!!!!!!!!!!!!!           #
+#                   If You Want TO Change This File                  #
+#                  Do Not Use All Of Non-Torch Types!                #
+#                                                                    #
+############################## Warning! ##############################
+import copy
+import math
+from typing import Optional
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from infer.lib.infer_pack import commons, modules
+from infer.lib.infer_pack.modules import LayerNorm
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        window_size=10,
+        **kwargs
+    ):
+        super(Encoder, self).__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = int(n_layers)
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+
+        self.drop = nn.Dropout(p_dropout)
+        self.attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    window_size=window_size,
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        zippep = zip(
+            self.attn_layers, self.norm_layers_1, self.ffn_layers, self.norm_layers_2
+        )
+        for attn_layers, norm_layers_1, ffn_layers, norm_layers_2 in zippep:
+            y = attn_layers(x, x, attn_mask)
+            y = self.drop(y)
+            x = norm_layers_1(x + y)
+
+            y = ffn_layers(x, x_mask)
+            y = self.drop(y)
+            x = norm_layers_2(x + y)
+        x = x * x_mask
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        proximal_bias=False,
+        proximal_init=True,
+        **kwargs
+    ):
+        super(Decoder, self).__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+
+        self.drop = nn.Dropout(p_dropout)
+        self.self_attn_layers = nn.ModuleList()
+        self.norm_layers_0 = nn.ModuleList()
+        self.encdec_attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.self_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    proximal_bias=proximal_bias,
+                    proximal_init=proximal_init,
+                )
+            )
+            self.norm_layers_0.append(LayerNorm(hidden_channels))
+            self.encdec_attn_layers.append(
+                MultiHeadAttention(
+                    hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout
+                )
+            )
+            self.norm_layers_1.append(LayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FFN(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                    causal=True,
+                )
+            )
+            self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, h, h_mask):
+        """
+        x: decoder input
+        h: encoder output
+        """
+        self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(
+            device=x.device, dtype=x.dtype
+        )
+        encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            y = self.self_attn_layers[i](x, x, self_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_0[i](x + y)
+
+            y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels,
+        out_channels,
+        n_heads,
+        p_dropout=0.0,
+        window_size=None,
+        heads_share=True,
+        block_length=None,
+        proximal_bias=False,
+        proximal_init=False,
+    ):
+        super(MultiHeadAttention, self).__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+        self.attn = None
+
+        self.k_channels = channels // n_heads
+        self.conv_q = nn.Conv1d(channels, channels, 1)
+        self.conv_k = nn.Conv1d(channels, channels, 1)
+        self.conv_v = nn.Conv1d(channels, channels, 1)
+        self.conv_o = nn.Conv1d(channels, out_channels, 1)
+        self.drop = nn.Dropout(p_dropout)
+
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+            self.emb_rel_k = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+            self.emb_rel_v = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+
+        nn.init.xavier_uniform_(self.conv_q.weight)
+        nn.init.xavier_uniform_(self.conv_k.weight)
+        nn.init.xavier_uniform_(self.conv_v.weight)
+        if proximal_init:
+            with torch.no_grad():
+                self.conv_k.weight.copy_(self.conv_q.weight)
+                self.conv_k.bias.copy_(self.conv_q.bias)
+
+    def forward(
+        self, x: torch.Tensor, c: torch.Tensor, attn_mask: Optional[torch.Tensor] = None
+    ):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, _ = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+    ):
+        # reshape [b, d, t] -> [b, n_h, t, d_k]
+        b, d, t_s = key.size()
+        t_t = query.size(2)
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+        if self.window_size is not None:
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(
+                query / math.sqrt(self.k_channels), key_relative_embeddings
+            )
+            scores_local = self._relative_position_to_absolute_position(rel_logits)
+            scores = scores + scores_local
+        if self.proximal_bias:
+            assert t_s == t_t, "Proximal bias is only available for self-attention."
+            scores = scores + self._attention_bias_proximal(t_s).to(
+                device=scores.device, dtype=scores.dtype
+            )
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                assert (
+                    t_s == t_t
+                ), "Local attention is only available for self-attention."
+                block_mask = (
+                    torch.ones_like(scores)
+                    .triu(-self.block_length)
+                    .tril(self.block_length)
+                )
+                scores = scores.masked_fill(block_mask == 0, -1e4)
+        p_attn = F.softmax(scores, dim=-1)  # [b, n_h, t_t, t_s]
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(
+                self.emb_rel_v, t_s
+            )
+            output = output + self._matmul_with_relative_values(
+                relative_weights, value_relative_embeddings
+            )
+        output = (
+            output.transpose(2, 3).contiguous().view(b, d, t_t)
+        )  # [b, n_h, t_t, d_k] -> [b, d, t_t]
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        """
+        x: [b, h, l, m]
+        y: [h or 1, m, d]
+        ret: [b, h, l, d]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        """
+        x: [b, h, l, d]
+        y: [h or 1, m, d]
+        ret: [b, h, l, m]
+        """
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length):
+        max_relative_position = 2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+
+        pad_length = torch.clamp(length - (self.window_size + 1), min=0)
+        slice_start_position = torch.clamp((self.window_size + 1) - length, min=0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        padded_relative_embeddings = F.pad(
+            relative_embeddings,
+            # commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
+            [0, 0, pad_length, pad_length, 0, 0],
+        )
+        used_relative_embeddings = padded_relative_embeddings[
+            :, slice_start_position:slice_end_position
+        ]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        """
+        x: [b, h, l, 2*l-1]
+        ret: [b, h, l, l]
+        """
+        batch, heads, length, _ = x.size()
+        # Concat columns of pad to shift from relative to absolute indexing.
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]])
+            [0, 1, 0, 0, 0, 0, 0, 0],
+        )
+
+        # Concat extra elements so to add up to shape (len+1, 2*len-1).
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = F.pad(
+            x_flat,
+            [0, length - 1, 0, 0, 0, 0],
+        )
+
+        # Reshape and slice out the padded elements.
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+            :, :, :length, length - 1 :
+        ]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        """
+        x: [b, h, l, l]
+        ret: [b, h, l, 2*l-1]
+        """
+        batch, heads, length, _ = x.size()
+        # padd along column
+        x = F.pad(
+            x,
+            [0, length - 1, 0, 0, 0, 0, 0, 0],
+        )
+        x_flat = x.view([batch, heads, length*length + length * (length - 1)])
+        # add 0's in the beginning that will skew the elements after reshape
+        x_flat = F.pad(
+            x_flat,
+            [length, 0, 0, 0, 0, 0],
+        )
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length):
+        """Bias for self-attention to encourage attention to close positions.
+        Args:
+          length: an integer scalar.
+        Returns:
+          a Tensor with shape [1, 1, length, length]
+        """
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout=0.0,
+        activation: str = None,
+        causal=False,
+    ):
+        super(FFN, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+        self.causal = causal
+        self.is_activation = True if activation == "gelu" else False
+        # if causal:
+        #     self.padding = self._causal_padding
+        # else:
+        #     self.padding = self._same_padding
+
+        self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+        self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+        self.drop = nn.Dropout(p_dropout)
+
+    def padding(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+        if self.causal:
+            padding = self._causal_padding(x * x_mask)
+        else:
+            padding = self._same_padding(x * x_mask)
+        return padding
+
+    def forward(self, x: torch.Tensor, x_mask: torch.Tensor):
+        x = self.conv_1(self.padding(x, x_mask))
+        if self.is_activation:
+            x = x * torch.sigmoid(1.702 * x)
+        else:
+            x = torch.relu(x)
+        x = self.drop(x)
+
+        x = self.conv_2(self.padding(x, x_mask))
+        return x * x_mask
+
+    def _causal_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = self.kernel_size - 1
+        pad_r = 0
+        # padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape(padding)
+            [pad_l, pad_r, 0, 0, 0, 0],
+        )
+        return x
+
+    def _same_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = (self.kernel_size - 1) // 2
+        pad_r = self.kernel_size // 2
+        # padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(
+            x,
+            #   commons.convert_pad_shape(padding)
+            [pad_l, pad_r, 0, 0, 0, 0],
+        )
+        return x

infer/lib/infer_pack/commons.py

@@ -0,0 +1,172 @@
+from typing import List, Optional
+import math
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+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 get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+# def convert_pad_shape(pad_shape):
+#     l = pad_shape[::-1]
+#     pad_shape = [item for sublist in l for item in sublist]
+#     return pad_shape
+
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+    """KL(P||Q)"""
+    kl = (logs_q - logs_p) - 0.5
+    kl += (
+        0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q)
+    )
+    return kl
+
+
+def rand_gumbel(shape):
+    """Sample from the Gumbel distribution, protect from overflows."""
+    uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
+    return -torch.log(-torch.log(uniform_samples))
+
+
+def rand_gumbel_like(x):
+    g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
+    return g
+
+
+def slice_segments(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = ids_str[i]
+        idx_end = idx_str + segment_size
+        ret[i] = x[i, :, idx_str:idx_end]
+    return ret
+
+
+def slice_segments2(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = ids_str[i]
+        idx_end = idx_str + segment_size
+        ret[i] = x[i, idx_str:idx_end]
+    return ret
+
+
+def rand_slice_segments(x, x_lengths=None, segment_size=4):
+    b, d, t = x.size()
+    if x_lengths is None:
+        x_lengths = t
+    ids_str_max = x_lengths - segment_size + 1
+    ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+    ret = slice_segments(x, ids_str, segment_size)
+    return ret, ids_str
+
+
+def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
+    position = torch.arange(length, dtype=torch.float)
+    num_timescales = channels // 2
+    log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (
+        num_timescales - 1
+    )
+    inv_timescales = min_timescale * torch.exp(
+        torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment
+    )
+    scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+    signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
+    signal = F.pad(signal, [0, 0, 0, channels % 2])
+    signal = signal.view(1, channels, length)
+    return signal
+
+
+def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return x + signal.to(dtype=x.dtype, device=x.device)
+
+
+def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
+
+
+def subsequent_mask(length):
+    mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
+    return mask
+
+
+@torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+# def convert_pad_shape(pad_shape):
+#     l = pad_shape[::-1]
+#     pad_shape = [item for sublist in l for item in sublist]
+#     return pad_shape
+
+
+def convert_pad_shape(pad_shape: List[List[int]]) -> List[int]:
+    return torch.tensor(pad_shape).flip(0).reshape(-1).int().tolist()
+
+
+def shift_1d(x):
+    x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+    return x
+
+
+def sequence_mask(length: torch.Tensor, max_length: Optional[int] = None):
+    if max_length is None:
+        max_length = length.max()
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def generate_path(duration, mask):
+    """
+    duration: [b, 1, t_x]
+    mask: [b, 1, t_y, t_x]
+    """
+    device = duration.device
+
+    b, _, t_y, t_x = mask.shape
+    cum_duration = torch.cumsum(duration, -1)
+
+    cum_duration_flat = cum_duration.view(b * t_x)
+    path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
+    path = path.view(b, t_x, t_y)
+    path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+    path = path.unsqueeze(1).transpose(2, 3) * mask
+    return path
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    parameters = list(filter(lambda p: p.grad is not None, parameters))
+    norm_type = float(norm_type)
+    if clip_value is not None:
+        clip_value = float(clip_value)
+
+    total_norm = 0
+    for p in parameters:
+        param_norm = p.grad.data.norm(norm_type)
+        total_norm += param_norm.item() ** norm_type
+        if clip_value is not None:
+            p.grad.data.clamp_(min=-clip_value, max=clip_value)
+    total_norm = total_norm ** (1.0 / norm_type)
+    return total_norm

infer/lib/infer_pack/models.py

@@ -0,0 +1,1223 @@
+import math
+import logging
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import AvgPool1d, Conv1d, Conv2d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, spectral_norm, weight_norm
+from infer.lib.infer_pack import attentions, commons, modules
+from infer.lib.infer_pack.commons import get_padding, init_weights
+
+has_xpu = bool(hasattr(torch, "xpu") and torch.xpu.is_available())
+
+
+class TextEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super(TextEncoder, self).__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = float(p_dropout)
+        self.emb_phone = nn.Linear(in_channels, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            float(p_dropout),
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(
+        self,
+        phone: torch.Tensor,
+        pitch: torch.Tensor,
+        lengths: torch.Tensor,
+        skip_head: Optional[torch.Tensor] = None,
+    ):
+        if pitch is None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        if skip_head is not None:
+            assert isinstance(skip_head, torch.Tensor)
+            head = int(skip_head.item())
+            x = x[:, :, head:]
+            x_mask = x_mask[:, :, head:]
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super(ResidualCouplingBlock, self).__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse: bool = False,
+    ):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in self.flows[::-1]:
+                x, _ = flow.forward(x, x_mask, g=g, reverse=reverse)
+        return x
+
+    def remove_weight_norm(self):
+        for i in range(self.n_flows):
+            self.flows[i * 2].remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for i in range(self.n_flows):
+            for hook in self.flows[i * 2]._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(self.flows[i * 2])
+
+        return self
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super(PosteriorEncoder, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(
+        self, x: torch.Tensor, x_lengths: torch.Tensor, g: Optional[torch.Tensor] = None
+    ):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for hook in self.enc._forward_pre_hooks.values():
+            if (
+                hook.__module__ == "torch.nn.utils.weight_norm"
+                and hook.__class__.__name__ == "WeightNorm"
+            ):
+                torch.nn.utils.remove_weight_norm(self.enc)
+        return self
+
+
+class Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        n_res: Optional[torch.Tensor] = None,
+    ):
+        if n_res is not None:
+            assert isinstance(n_res, torch.Tensor)
+            n = int(n_res.item())
+            if n != x.shape[-1]:
+                x = F.interpolate(x, size=n, mode="linear")
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.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 __prepare_scriptable__(self):
+        for l in self.ups:
+            for hook in l._forward_pre_hooks.values():
+                # The hook we want to remove is an instance of WeightNorm class, so
+                # normally we would do `if isinstance(...)` but this class is not accessible
+                # because of shadowing, so we check the module name directly.
+                # https://github.com/pytorch/pytorch/blob/be0ca00c5ce260eb5bcec3237357f7a30cc08983/torch/nn/utils/__init__.py#L3
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+
+        for l in self.resblocks:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        return self
+
+    def remove_weight_norm(self):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+class SineGen(torch.nn.Module):
+    """Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(torch.pi) or cos(0)
+    """
+
+    def __init__(
+        self,
+        samp_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        noise_std=0.003,
+        voiced_threshold=0,
+        flag_for_pulse=False,
+    ):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = torch.ones_like(f0)
+        uv = uv * (f0 > self.voiced_threshold)
+        if uv.device.type == "privateuseone":  # for DirectML
+            uv = uv.float()
+        return uv
+    
+    def _f02sine(self, f0, upp):
+        """ f0: (batchsize, length, dim)
+            where dim indicates fundamental tone and overtones
+        """
+        a = torch.arange(1, upp + 1, dtype=f0.dtype, device=f0.device)
+        rad = f0 / self.sampling_rate * a
+        rad2 = torch.fmod(rad[:, :-1, -1:].float() + 0.5, 1.0) - 0.5
+        rad_acc = rad2.cumsum(dim=1).fmod(1.0).to(f0)
+        rad += F.pad(rad_acc, (0, 0, 1, 0), mode='constant')
+        rad = rad.reshape(f0.shape[0], -1, 1)
+        b = torch.arange(1, self.dim + 1, dtype=f0.dtype, device=f0.device).reshape(1, 1, -1)
+        rad *= b
+        rand_ini = torch.rand(1, 1, self.dim, device=f0.device)
+        rand_ini[..., 0] = 0
+        rad += rand_ini
+        sines = torch.sin(2 * np.pi * rad)
+        return sines
+        
+    def forward(self, f0: torch.Tensor, upp: int):
+        """sine_tensor, uv = forward(f0)
+        input F0: tensor(batchsize=1, length, dim=1)
+                  f0 for unvoiced steps should be 0
+        output sine_tensor: tensor(batchsize=1, length, dim)
+        output uv: tensor(batchsize=1, length, 1)
+        """
+        with torch.no_grad():
+            f0 = f0.unsqueeze(-1)
+            sine_waves = self._f02sine(f0, upp) * self.sine_amp
+            uv = self._f02uv(f0)
+            uv = F.interpolate(
+                uv.transpose(2, 1), scale_factor=float(upp), mode="nearest"
+            ).transpose(2, 1)
+            noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+            noise = noise_amp * torch.randn_like(sine_waves)
+            sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(
+        self,
+        sampling_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        add_noise_std=0.003,
+        voiced_threshod=0,
+        is_half=True,
+    ):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+        self.is_half = is_half
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(
+            sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
+        )
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+        # self.ddtype:int = -1
+
+    def forward(self, x: torch.Tensor, upp: int = 1):
+        # if self.ddtype ==-1:
+        #     self.ddtype = self.l_linear.weight.dtype
+        sine_wavs, uv, _ = self.l_sin_gen(x, upp)
+        # print(x.dtype,sine_wavs.dtype,self.l_linear.weight.dtype)
+        # if self.is_half:
+        #     sine_wavs = sine_wavs.half()
+        # sine_merge = self.l_tanh(self.l_linear(sine_wavs.to(x)))
+        # print(sine_wavs.dtype,self.ddtype)
+        # if sine_wavs.dtype != self.l_linear.weight.dtype:
+        sine_wavs = sine_wavs.to(dtype=self.l_linear.weight.dtype)
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+        return sine_merge, None, None  # noise, uv
+
+
+class GeneratorNSF(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels,
+        sr,
+        is_half=False,
+    ):
+        super(GeneratorNSF, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=math.prod(upsample_rates))
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sr, harmonic_num=0, is_half=is_half
+        )
+        self.noise_convs = nn.ModuleList()
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+            if i + 1 < len(upsample_rates):
+                stride_f0 = math.prod(upsample_rates[i + 1 :])
+                self.noise_convs.append(
+                    Conv1d(
+                        1,
+                        c_cur,
+                        kernel_size=stride_f0 * 2,
+                        stride=stride_f0,
+                        padding=stride_f0 // 2,
+                    )
+                )
+            else:
+                self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+        self.upp = math.prod(upsample_rates)
+
+        self.lrelu_slope = modules.LRELU_SLOPE
+
+    def forward(
+        self,
+        x,
+        f0,
+        g: Optional[torch.Tensor] = None,
+        n_res: Optional[torch.Tensor] = None,
+    ):
+        har_source, noi_source, uv = self.m_source(f0, self.upp)
+        har_source = har_source.transpose(1, 2)
+        if n_res is not None:
+            assert isinstance(n_res, torch.Tensor)
+            n = int(n_res.item())
+            if n * self.upp != har_source.shape[-1]:
+                har_source = F.interpolate(har_source, size=n * self.upp, mode="linear")
+            if n != x.shape[-1]:
+                x = F.interpolate(x, size=n, mode="linear")
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+        # torch.jit.script() does not support direct indexing of torch modules
+        # That's why I wrote this
+        for i, (ups, noise_convs) in enumerate(zip(self.ups, self.noise_convs)):
+            if i < self.num_upsamples:
+                x = F.leaky_relu(x, self.lrelu_slope)
+                x = ups(x)
+                x_source = noise_convs(har_source)
+                x = x + x_source
+                xs: Optional[torch.Tensor] = None
+                l = [i * self.num_kernels + j for j in range(self.num_kernels)]
+                for j, resblock in enumerate(self.resblocks):
+                    if j in l:
+                        if xs is None:
+                            xs = resblock(x)
+                        else:
+                            xs += resblock(x)
+                # This assertion cannot be ignored! \
+                # If ignored, it will cause torch.jit.script() compilation errors
+                assert isinstance(xs, torch.Tensor)
+                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):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for l in self.ups:
+            for hook in l._forward_pre_hooks.values():
+                # The hook we want to remove is an instance of WeightNorm class, so
+                # normally we would do `if isinstance(...)` but this class is not accessible
+                # because of shadowing, so we check the module name directly.
+                # https://github.com/pytorch/pytorch/blob/be0ca00c5ce260eb5bcec3237357f7a30cc08983/torch/nn/utils/__init__.py#L3
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        for l in self.resblocks:
+            for hook in self.resblocks._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        return self
+
+
+sr2sr = {
+    "32k": 32000,
+    "40k": 40000,
+    "48k": 48000,
+}
+
+
+class SynthesizerTrnMs256NSFsid(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr,
+        **kwargs
+    ):
+        super(SynthesizerTrnMs256NSFsid, self).__init__()
+        if isinstance(sr, str):
+            sr = sr2sr[sr]
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = float(p_dropout)
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        self.enc_p = TextEncoder(
+            256,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            float(p_dropout),
+        )
+        self.dec = GeneratorNSF(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+            sr=sr,
+            is_half=kwargs["is_half"],
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        logger.debug(
+            "gin_channels: "
+            + str(gin_channels)
+            + ", self.spk_embed_dim: "
+            + str(self.spk_embed_dim)
+        )
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        if hasattr(self, "enc_q"):
+            self.enc_q.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for hook in self.dec._forward_pre_hooks.values():
+            # The hook we want to remove is an instance of WeightNorm class, so
+            # normally we would do `if isinstance(...)` but this class is not accessible
+            # because of shadowing, so we check the module name directly.
+            # https://github.com/pytorch/pytorch/blob/be0ca00c5ce260eb5bcec3237357f7a30cc08983/torch/nn/utils/__init__.py#L3
+            if (
+                hook.__module__ == "torch.nn.utils.weight_norm"
+                and hook.__class__.__name__ == "WeightNorm"
+            ):
+                torch.nn.utils.remove_weight_norm(self.dec)
+        for hook in self.flow._forward_pre_hooks.values():
+            if (
+                hook.__module__ == "torch.nn.utils.weight_norm"
+                and hook.__class__.__name__ == "WeightNorm"
+            ):
+                torch.nn.utils.remove_weight_norm(self.flow)
+        if hasattr(self, "enc_q"):
+            for hook in self.enc_q._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(self.enc_q)
+        return self
+
+    @torch.jit.ignore
+    def forward(
+        self,
+        phone: torch.Tensor,
+        phone_lengths: torch.Tensor,
+        pitch: torch.Tensor,
+        pitchf: torch.Tensor,
+        y: torch.Tensor,
+        y_lengths: torch.Tensor,
+        ds: Optional[torch.Tensor] = None,
+    ):  # 这里ds是id，[bs,1]
+        # print(1,pitch.shape)#[bs,t]
+        g = self.emb_g(ds).unsqueeze(-1)  # [b, 256, 1]##1是t，广播的
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+        z_slice, ids_slice = commons.rand_slice_segments(
+            z, y_lengths, self.segment_size
+        )
+        # print(-1,pitchf.shape,ids_slice,self.segment_size,self.hop_length,self.segment_size//self.hop_length)
+        pitchf = commons.slice_segments2(pitchf, ids_slice, self.segment_size)
+        # print(-2,pitchf.shape,z_slice.shape)
+        o = self.dec(z_slice, pitchf, g=g)
+        return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
+
+    @torch.jit.export
+    def infer(
+        self,
+        phone: torch.Tensor,
+        phone_lengths: torch.Tensor,
+        pitch: torch.Tensor,
+        nsff0: torch.Tensor,
+        sid: torch.Tensor,
+        skip_head: Optional[torch.Tensor] = None,
+        return_length: Optional[torch.Tensor] = None,
+        return_length2: Optional[torch.Tensor] = None,
+    ):
+        g = self.emb_g(sid).unsqueeze(-1)
+        if skip_head is not None and return_length is not None:
+            assert isinstance(skip_head, torch.Tensor)
+            assert isinstance(return_length, torch.Tensor)
+            head = int(skip_head.item())
+            length = int(return_length.item())
+            flow_head = torch.clamp(skip_head - 24, min=0)
+            dec_head = head - int(flow_head.item())
+            m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths, flow_head)
+            z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+            z = self.flow(z_p, x_mask, g=g, reverse=True)
+            z = z[:, :, dec_head : dec_head + length]
+            x_mask = x_mask[:, :, dec_head : dec_head + length]
+            nsff0 = nsff0[:, head : head + length]
+        else:
+            m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+            z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+            z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec(z * x_mask, nsff0, g=g, n_res=return_length2)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+class SynthesizerTrnMs768NSFsid(SynthesizerTrnMs256NSFsid):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr,
+        **kwargs
+    ):
+        super(SynthesizerTrnMs768NSFsid, self).__init__(
+            spec_channels,
+            segment_size,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            spk_embed_dim,
+            gin_channels,
+            sr,
+            **kwargs
+        )
+        del self.enc_p
+        self.enc_p = TextEncoder(
+            768,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            float(p_dropout),
+        )
+
+
+class SynthesizerTrnMs256NSFsid_nono(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr=None,
+        **kwargs
+    ):
+        super(SynthesizerTrnMs256NSFsid_nono, self).__init__()
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = float(p_dropout)
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        self.enc_p = TextEncoder(
+            256,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            float(p_dropout),
+            f0=False,
+        )
+        self.dec = Generator(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        logger.debug(
+            "gin_channels: "
+            + str(gin_channels)
+            + ", self.spk_embed_dim: "
+            + str(self.spk_embed_dim)
+        )
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        if hasattr(self, "enc_q"):
+            self.enc_q.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for hook in self.dec._forward_pre_hooks.values():
+            # The hook we want to remove is an instance of WeightNorm class, so
+            # normally we would do `if isinstance(...)` but this class is not accessible
+            # because of shadowing, so we check the module name directly.
+            # https://github.com/pytorch/pytorch/blob/be0ca00c5ce260eb5bcec3237357f7a30cc08983/torch/nn/utils/__init__.py#L3
+            if (
+                hook.__module__ == "torch.nn.utils.weight_norm"
+                and hook.__class__.__name__ == "WeightNorm"
+            ):
+                torch.nn.utils.remove_weight_norm(self.dec)
+        for hook in self.flow._forward_pre_hooks.values():
+            if (
+                hook.__module__ == "torch.nn.utils.weight_norm"
+                and hook.__class__.__name__ == "WeightNorm"
+            ):
+                torch.nn.utils.remove_weight_norm(self.flow)
+        if hasattr(self, "enc_q"):
+            for hook in self.enc_q._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(self.enc_q)
+        return self
+
+    @torch.jit.ignore
+    def forward(self, phone, phone_lengths, y, y_lengths, ds):  # 这里ds是id，[bs,1]
+        g = self.emb_g(ds).unsqueeze(-1)  # [b, 256, 1]##1是t，广播的
+        m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
+        z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+        z_p = self.flow(z, y_mask, g=g)
+        z_slice, ids_slice = commons.rand_slice_segments(
+            z, y_lengths, self.segment_size
+        )
+        o = self.dec(z_slice, g=g)
+        return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
+
+    @torch.jit.export
+    def infer(
+        self,
+        phone: torch.Tensor,
+        phone_lengths: torch.Tensor,
+        sid: torch.Tensor,
+        skip_head: Optional[torch.Tensor] = None,
+        return_length: Optional[torch.Tensor] = None,
+        return_length2: Optional[torch.Tensor] = None,
+    ):
+        g = self.emb_g(sid).unsqueeze(-1)
+        if skip_head is not None and return_length is not None:
+            assert isinstance(skip_head, torch.Tensor)
+            assert isinstance(return_length, torch.Tensor)
+            head = int(skip_head.item())
+            length = int(return_length.item())
+            flow_head = torch.clamp(skip_head - 24, min=0)
+            dec_head = head - int(flow_head.item())
+            m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths, flow_head)
+            z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+            z = self.flow(z_p, x_mask, g=g, reverse=True)
+            z = z[:, :, dec_head : dec_head + length]
+            x_mask = x_mask[:, :, dec_head : dec_head + length]
+        else:
+            m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
+            z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+            z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec(z * x_mask, g=g, n_res=return_length2)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+class SynthesizerTrnMs768NSFsid_nono(SynthesizerTrnMs256NSFsid_nono):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr=None,
+        **kwargs
+    ):
+        super(SynthesizerTrnMs768NSFsid_nono, self).__init__(
+            spec_channels,
+            segment_size,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            spk_embed_dim,
+            gin_channels,
+            sr,
+            **kwargs
+        )
+        del self.enc_p
+        self.enc_p = TextEncoder(
+            768,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            float(p_dropout),
+            f0=False,
+        )
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11, 17]
+        # periods = [3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    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)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiPeriodDiscriminatorV2(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminatorV2, self).__init__()
+        # periods = [2, 3, 5, 7, 11, 17]
+        periods = [2, 3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    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)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            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, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, 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, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+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
+        self.use_spectral_norm = use_spectral_norm
+        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(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 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)
+            if has_xpu and x.dtype == torch.bfloat16:
+                x = F.pad(x.to(dtype=torch.float16), (0, n_pad), "reflect").to(
+                    dtype=torch.bfloat16
+                )
+            else:
+                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, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap

infer/lib/infer_pack/models_onnx.py

@@ -0,0 +1,818 @@
+############################## Warning! ##############################
+#                                                                    #
+#           Onnx Export Not Support All Of Non-Torch Types           #
+#           Include Python Built-in Types!!!!!!!!!!!!!!!!!           #
+#                   If You Want TO Change This File                  #
+#                  Do Not Use All Of Non-Torch Types!                #
+#                                                                    #
+############################## Warning! ##############################
+
+import math
+import logging
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import AvgPool1d, Conv1d, Conv2d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, spectral_norm, weight_norm
+
+from infer.lib.infer_pack import commons, modules
+import infer.lib.infer_pack.attentions_onnx as attentions
+from infer.lib.infer_pack.commons import get_padding, init_weights
+
+
+class TextEncoder256(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(256, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class TextEncoder768(nn.Module):
+    def __init__(
+        self,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        f0=True,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.emb_phone = nn.Linear(768, hidden_channels)
+        self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+        if f0 == True:
+            self.emb_pitch = nn.Embedding(256, hidden_channels)  # pitch 256
+        self.encoder = attentions.Encoder(
+            hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, phone, pitch, lengths):
+        if pitch == None:
+            x = self.emb_phone(phone)
+        else:
+            x = self.emb_phone(phone) + self.emb_pitch(pitch)
+        x = x * math.sqrt(self.hidden_channels)  # [b, t, h]
+        x = self.lrelu(x)
+        x = torch.transpose(x, 1, -1)  # [b, h, t]
+        x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.encoder(x * x_mask, x_mask)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return m, logs, x_mask
+
+
+class ResidualCouplingBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ResidualCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.Flip())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+    def remove_weight_norm(self):
+        for i in range(self.n_flows):
+            self.flows[i * 2].remove_weight_norm()
+
+
+class PosteriorEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None):
+        x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+
+class Generator(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(Generator, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.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):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+class SineGen(torch.nn.Module):
+    """Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(
+        self,
+        samp_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        noise_std=0.003,
+        voiced_threshold=0,
+        flag_for_pulse=False,
+    ):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = torch.ones_like(f0)
+        uv = uv * (f0 > self.voiced_threshold)
+        if uv.device.type == "privateuseone":  # for DirectML
+            uv = uv.float()
+        return uv
+    
+    def _f02sine(self, f0, upp):
+        """ f0: (batchsize, length, dim)
+            where dim indicates fundamental tone and overtones
+        """
+        a = torch.arange(1, upp + 1, dtype=f0.dtype, device=f0.device)
+        rad = f0 / self.sampling_rate * a
+        rad2 = torch.fmod(rad[:, :-1, -1:].float() + 0.5, 1.0) - 0.5
+        rad_acc = rad2.cumsum(dim=1).fmod(1.0).to(f0)
+        rad += F.pad(rad_acc, (0, 0, 1, 0), mode='constant')
+        rad = rad.reshape(f0.shape[0], -1, 1)
+        b = torch.arange(1, self.dim + 1, dtype=f0.dtype, device=f0.device).reshape(1, 1, -1)
+        rad *= b
+        rand_ini = torch.rand(1, 1, self.dim, device=f0.device)
+        rand_ini[..., 0] = 0
+        rad += rand_ini
+        sines = torch.sin(2 * np.pi * rad)
+        return sines
+        
+    def forward(self, f0: torch.Tensor, upp: int):
+        """sine_tensor, uv = forward(f0)
+        input F0: tensor(batchsize=1, length, dim=1)
+                  f0 for unvoiced steps should be 0
+        output sine_tensor: tensor(batchsize=1, length, dim)
+        output uv: tensor(batchsize=1, length, 1)
+        """
+        with torch.no_grad():
+            f0 = f0.unsqueeze(-1)
+            sine_waves = self._f02sine(f0, upp) * self.sine_amp
+            uv = self._f02uv(f0)
+            uv = F.interpolate(
+                uv.transpose(2, 1), scale_factor=float(upp), mode="nearest"
+            ).transpose(2, 1)
+            noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+            noise = noise_amp * torch.randn_like(sine_waves)
+            sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(
+        self,
+        sampling_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        add_noise_std=0.003,
+        voiced_threshod=0,
+        is_half=True,
+    ):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+        self.is_half = is_half
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(
+            sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
+        )
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x, upp=None):
+        sine_wavs, uv, _ = self.l_sin_gen(x, upp)
+        if self.is_half:
+            sine_wavs = sine_wavs.half()
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+        return sine_merge, None, None  # noise, uv
+
+
+class GeneratorNSF(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels,
+        sr,
+        is_half=False,
+    ):
+        super(GeneratorNSF, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sr, harmonic_num=0, is_half=is_half
+        )
+        self.noise_convs = nn.ModuleList()
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+            if i + 1 < len(upsample_rates):
+                stride_f0 = np.prod(upsample_rates[i + 1 :])
+                self.noise_convs.append(
+                    Conv1d(
+                        1,
+                        c_cur,
+                        kernel_size=stride_f0 * 2,
+                        stride=stride_f0,
+                        padding=stride_f0 // 2,
+                    )
+                )
+            else:
+                self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(init_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+        self.upp = np.prod(upsample_rates)
+
+    def forward(self, x, f0, g=None):
+        har_source, noi_source, uv = self.m_source(f0, self.upp)
+        har_source = har_source.transpose(1, 2)
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            x_source = self.noise_convs[i](har_source)
+            x = x + x_source
+            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):
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+
+
+sr2sr = {
+    "32k": 32000,
+    "40k": 40000,
+    "48k": 48000,
+}
+
+
+class SynthesizerTrnMsNSFsidM(nn.Module):
+    def __init__(
+        self,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        spk_embed_dim,
+        gin_channels,
+        sr,
+        version,
+        **kwargs,
+    ):
+        super().__init__()
+        if type(sr) == type("strr"):
+            sr = sr2sr[sr]
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.gin_channels = gin_channels
+        # self.hop_length = hop_length#
+        self.spk_embed_dim = spk_embed_dim
+        if version == "v1":
+            self.enc_p = TextEncoder256(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+            )
+        else:
+            self.enc_p = TextEncoder768(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+            )
+        self.dec = GeneratorNSF(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+            sr=sr,
+            is_half=kwargs["is_half"],
+        )
+        self.enc_q = PosteriorEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        self.flow = ResidualCouplingBlock(
+            inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+        )
+        self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+        self.speaker_map = None
+        logger.debug(
+            f"gin_channels: {gin_channels}, self.spk_embed_dim: {self.spk_embed_dim}"
+        )
+
+    def remove_weight_norm(self):
+        self.dec.remove_weight_norm()
+        self.flow.remove_weight_norm()
+        self.enc_q.remove_weight_norm()
+
+    def construct_spkmixmap(self, n_speaker):
+        self.speaker_map = torch.zeros((n_speaker, 1, 1, self.gin_channels))
+        for i in range(n_speaker):
+            self.speaker_map[i] = self.emb_g(torch.LongTensor([[i]]))
+        self.speaker_map = self.speaker_map.unsqueeze(0)
+
+    def forward(self, phone, phone_lengths, pitch, nsff0, g, rnd, max_len=None):
+        if self.speaker_map is not None:  # [N, S]  *  [S, B, 1, H]
+            g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1))  # [N, S, B, 1, 1]
+            g = g * self.speaker_map  # [N, S, B, 1, H]
+            g = torch.sum(g, dim=1)  # [N, 1, B, 1, H]
+            g = g.transpose(0, -1).transpose(0, -2).squeeze(0)  # [B, H, N]
+        else:
+            g = g.unsqueeze(0)
+            g = self.emb_g(g).transpose(1, 2)
+
+        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+        z_p = (m_p + torch.exp(logs_p) * rnd) * x_mask
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
+        return o
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2, 3, 5, 7, 11, 17]
+        # periods = [3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    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)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiPeriodDiscriminatorV2(torch.nn.Module):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminatorV2, self).__init__()
+        # periods = [2, 3, 5, 7, 11, 17]
+        periods = [2, 3, 5, 7, 11, 17, 23, 37]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [
+            DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
+        ]
+        self.discriminators = nn.ModuleList(discs)
+
+    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)
+            # for j in range(len(fmap_r)):
+            #     print(i,j,y.shape,y_hat.shape,fmap_r[j].shape,fmap_g[j].shape)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            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, 16, 15, 1, padding=7)),
+                norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+                norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+                norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+                norm_f(Conv1d(1024, 1024, 41, 4, groups=256, 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, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+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
+        self.use_spectral_norm = use_spectral_norm
+        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(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(kernel_size, 1), 0),
+                    )
+                ),
+                norm_f(
+                    Conv2d(
+                        1024,
+                        1024,
+                        (kernel_size, 1),
+                        1,
+                        padding=(get_padding(kernel_size, 1), 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, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap

infer/lib/infer_pack/modules.py

@@ -0,0 +1,615 @@
+import copy
+import math
+from typing import Optional, Tuple
+
+import numpy as np
+import scipy
+import torch
+from torch import nn
+from torch.nn import AvgPool1d, Conv1d, Conv2d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, weight_norm
+
+from infer.lib.infer_pack import commons
+from infer.lib.infer_pack.commons import get_padding, init_weights
+from infer.lib.infer_pack.transforms import piecewise_rational_quadratic_transform
+
+LRELU_SLOPE = 0.1
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super(LayerNorm, self).__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+class ConvReluNorm(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        out_channels,
+        kernel_size,
+        n_layers,
+        p_dropout,
+    ):
+        super(ConvReluNorm, self).__init__()
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = float(p_dropout)
+        assert n_layers > 1, "Number of layers should be larger than 0."
+
+        self.conv_layers = nn.ModuleList()
+        self.norm_layers = nn.ModuleList()
+        self.conv_layers.append(
+            nn.Conv1d(
+                in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
+            )
+        )
+        self.norm_layers.append(LayerNorm(hidden_channels))
+        self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(float(p_dropout)))
+        for _ in range(n_layers - 1):
+            self.conv_layers.append(
+                nn.Conv1d(
+                    hidden_channels,
+                    hidden_channels,
+                    kernel_size,
+                    padding=kernel_size // 2,
+                )
+            )
+            self.norm_layers.append(LayerNorm(hidden_channels))
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask):
+        x_org = x
+        for i in range(self.n_layers):
+            x = self.conv_layers[i](x * x_mask)
+            x = self.norm_layers[i](x)
+            x = self.relu_drop(x)
+        x = x_org + self.proj(x)
+        return x * x_mask
+
+
+class DDSConv(nn.Module):
+    """
+    Dialted and Depth-Separable Convolution
+    """
+
+    def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
+        super(DDSConv, self).__init__()
+        self.channels = channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = float(p_dropout)
+
+        self.drop = nn.Dropout(float(p_dropout))
+        self.convs_sep = nn.ModuleList()
+        self.convs_1x1 = nn.ModuleList()
+        self.norms_1 = nn.ModuleList()
+        self.norms_2 = nn.ModuleList()
+        for i in range(n_layers):
+            dilation = kernel_size**i
+            padding = (kernel_size * dilation - dilation) // 2
+            self.convs_sep.append(
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    groups=channels,
+                    dilation=dilation,
+                    padding=padding,
+                )
+            )
+            self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
+            self.norms_1.append(LayerNorm(channels))
+            self.norms_2.append(LayerNorm(channels))
+
+    def forward(self, x, x_mask, g: Optional[torch.Tensor] = None):
+        if g is not None:
+            x = x + g
+        for i in range(self.n_layers):
+            y = self.convs_sep[i](x * x_mask)
+            y = self.norms_1[i](y)
+            y = F.gelu(y)
+            y = self.convs_1x1[i](y)
+            y = self.norms_2[i](y)
+            y = F.gelu(y)
+            y = self.drop(y)
+            x = x + y
+        return x * x_mask
+
+
+class WN(torch.nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+        p_dropout=0,
+    ):
+        super(WN, self).__init__()
+        assert kernel_size % 2 == 1
+        self.hidden_channels = hidden_channels
+        self.kernel_size = (kernel_size,)
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = float(p_dropout)
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = nn.Dropout(float(p_dropout))
+
+        if gin_channels != 0:
+            cond_layer = torch.nn.Conv1d(
+                gin_channels, 2 * hidden_channels * n_layers, 1
+            )
+            self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
+
+        for i in range(n_layers):
+            dilation = dilation_rate**i
+            padding = int((kernel_size * dilation - dilation) / 2)
+            in_layer = torch.nn.Conv1d(
+                hidden_channels,
+                2 * hidden_channels,
+                kernel_size,
+                dilation=dilation,
+                padding=padding,
+            )
+            in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
+            self.in_layers.append(in_layer)
+
+            # last one is not necessary
+            if i < n_layers - 1:
+                res_skip_channels = 2 * hidden_channels
+            else:
+                res_skip_channels = hidden_channels
+
+            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+            res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
+            self.res_skip_layers.append(res_skip_layer)
+
+    def forward(
+        self, x: torch.Tensor, x_mask: torch.Tensor, g: Optional[torch.Tensor] = None
+    ):
+        output = torch.zeros_like(x)
+        n_channels_tensor = torch.IntTensor([self.hidden_channels])
+
+        if g is not None:
+            g = self.cond_layer(g)
+
+        for i, (in_layer, res_skip_layer) in enumerate(
+            zip(self.in_layers, self.res_skip_layers)
+        ):
+            x_in = in_layer(x)
+            if g is not None:
+                cond_offset = i * 2 * self.hidden_channels
+                g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
+            else:
+                g_l = torch.zeros_like(x_in)
+
+            acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
+            acts = self.drop(acts)
+
+            res_skip_acts = res_skip_layer(acts)
+            if i < self.n_layers - 1:
+                res_acts = res_skip_acts[:, : self.hidden_channels, :]
+                x = (x + res_acts) * x_mask
+                output = output + res_skip_acts[:, self.hidden_channels :, :]
+            else:
+                output = output + res_skip_acts
+        return output * x_mask
+
+    def remove_weight_norm(self):
+        if self.gin_channels != 0:
+            torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            torch.nn.utils.remove_weight_norm(l)
+
+    def __prepare_scriptable__(self):
+        if self.gin_channels != 0:
+            for hook in self.cond_layer._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        return self
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        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)
+        self.lrelu_slope = LRELU_SLOPE
+
+    def forward(self, x: torch.Tensor, x_mask: Optional[torch.Tensor] = None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, self.lrelu_slope)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, self.lrelu_slope)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c2(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        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)
+
+    def __prepare_scriptable__(self):
+        for l in self.convs1:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        for l in self.convs2:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        return self
+
+
+class ResBlock2(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        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)
+        self.lrelu_slope = LRELU_SLOPE
+
+    def forward(self, x, x_mask: Optional[torch.Tensor] = None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, self.lrelu_slope)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+    def __prepare_scriptable__(self):
+        for l in self.convs:
+            for hook in l._forward_pre_hooks.values():
+                if (
+                    hook.__module__ == "torch.nn.utils.weight_norm"
+                    and hook.__class__.__name__ == "WeightNorm"
+                ):
+                    torch.nn.utils.remove_weight_norm(l)
+        return self
+
+
+class Log(nn.Module):
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        if not reverse:
+            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
+            logdet = torch.sum(-y, [1, 2])
+            return y, logdet
+        else:
+            x = torch.exp(x) * x_mask
+            return x
+
+
+class Flip(nn.Module):
+    # torch.jit.script() Compiled functions \
+    # can't take variable number of arguments or \
+    # use keyword-only arguments with defaults
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+        x = torch.flip(x, [1])
+        if not reverse:
+            logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+            return x, logdet
+        else:
+            return x, torch.zeros([1], device=x.device)
+
+
+class ElementwiseAffine(nn.Module):
+    def __init__(self, channels):
+        super(ElementwiseAffine, self).__init__()
+        self.channels = channels
+        self.m = nn.Parameter(torch.zeros(channels, 1))
+        self.logs = nn.Parameter(torch.zeros(channels, 1))
+
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = self.m + torch.exp(self.logs) * x
+            y = y * x_mask
+            logdet = torch.sum(self.logs * x_mask, [1, 2])
+            return y, logdet
+        else:
+            x = (x - self.m) * torch.exp(-self.logs) * x_mask
+            return x
+
+
+class ResidualCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        p_dropout=0,
+        gin_channels=0,
+        mean_only=False,
+    ):
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super(ResidualCouplingLayer, self).__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = WN(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            p_dropout=float(p_dropout),
+            gin_channels=gin_channels,
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse: bool = False,
+    ):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x, torch.zeros([1])
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+    def __prepare_scriptable__(self):
+        for hook in self.enc._forward_pre_hooks.values():
+            if (
+                hook.__module__ == "torch.nn.utils.weight_norm"
+                and hook.__class__.__name__ == "WeightNorm"
+            ):
+                torch.nn.utils.remove_weight_norm(self.enc)
+        return self
+
+
+class ConvFlow(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        n_layers,
+        num_bins=10,
+        tail_bound=5.0,
+    ):
+        super(ConvFlow, self).__init__()
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.num_bins = num_bins
+        self.tail_bound = tail_bound
+        self.half_channels = in_channels // 2
+
+        self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
+        self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
+        self.proj = nn.Conv1d(
+            filter_channels, self.half_channels * (num_bins * 3 - 1), 1
+        )
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        g: Optional[torch.Tensor] = None,
+        reverse=False,
+    ):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0)
+        h = self.convs(h, x_mask, g=g)
+        h = self.proj(h) * x_mask
+
+        b, c, t = x0.shape
+        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)  # [b, cx?, t] -> [b, c, t, ?]
+
+        unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
+        unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(
+            self.filter_channels
+        )
+        unnormalized_derivatives = h[..., 2 * self.num_bins :]
+
+        x1, logabsdet = piecewise_rational_quadratic_transform(
+            x1,
+            unnormalized_widths,
+            unnormalized_heights,
+            unnormalized_derivatives,
+            inverse=reverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+
+        x = torch.cat([x0, x1], 1) * x_mask
+        logdet = torch.sum(logabsdet * x_mask, [1, 2])
+        if not reverse:
+            return x, logdet
+        else:
+            return x

infer/lib/infer_pack/modules/F0Predictor/DioF0Predictor.py

@@ -0,0 +1,91 @@
+import numpy as np
+import pyworld
+
+from infer.lib.infer_pack.modules.F0Predictor.F0Predictor import F0Predictor
+
+
+class DioF0Predictor(F0Predictor):
+    def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        self.sampling_rate = sampling_rate
+
+    def interpolate_f0(self, f0):
+        """
+        对F0进行插值处理
+        """
+
+        data = np.reshape(f0, (f0.size, 1))
+
+        vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
+        vuv_vector[data > 0.0] = 1.0
+        vuv_vector[data <= 0.0] = 0.0
+
+        ip_data = data
+
+        frame_number = data.size
+        last_value = 0.0
+        for i in range(frame_number):
+            if data[i] <= 0.0:
+                j = i + 1
+                for j in range(i + 1, frame_number):
+                    if data[j] > 0.0:
+                        break
+                if j < frame_number - 1:
+                    if last_value > 0.0:
+                        step = (data[j] - data[i - 1]) / float(j - i)
+                        for k in range(i, j):
+                            ip_data[k] = data[i - 1] + step * (k - i + 1)
+                    else:
+                        for k in range(i, j):
+                            ip_data[k] = data[j]
+                else:
+                    for k in range(i, frame_number):
+                        ip_data[k] = last_value
+            else:
+                ip_data[i] = data[i]  # 这里可能存在一个没有必要的拷贝
+                last_value = data[i]
+
+        return ip_data[:, 0], vuv_vector[:, 0]
+
+    def resize_f0(self, x, target_len):
+        source = np.array(x)
+        source[source < 0.001] = np.nan
+        target = np.interp(
+            np.arange(0, len(source) * target_len, len(source)) / target_len,
+            np.arange(0, len(source)),
+            source,
+        )
+        res = np.nan_to_num(target)
+        return res
+
+    def compute_f0(self, wav, p_len=None):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        f0, t = pyworld.dio(
+            wav.astype(np.double),
+            fs=self.sampling_rate,
+            f0_floor=self.f0_min,
+            f0_ceil=self.f0_max,
+            frame_period=1000 * self.hop_length / self.sampling_rate,
+        )
+        f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.sampling_rate)
+        for index, pitch in enumerate(f0):
+            f0[index] = round(pitch, 1)
+        return self.interpolate_f0(self.resize_f0(f0, p_len))[0]
+
+    def compute_f0_uv(self, wav, p_len=None):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        f0, t = pyworld.dio(
+            wav.astype(np.double),
+            fs=self.sampling_rate,
+            f0_floor=self.f0_min,
+            f0_ceil=self.f0_max,
+            frame_period=1000 * self.hop_length / self.sampling_rate,
+        )
+        f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.sampling_rate)
+        for index, pitch in enumerate(f0):
+            f0[index] = round(pitch, 1)
+        return self.interpolate_f0(self.resize_f0(f0, p_len))

infer/lib/infer_pack/modules/F0Predictor/F0Predictor.py

@@ -0,0 +1,16 @@
+class F0Predictor(object):
+    def compute_f0(self, wav, p_len):
+        """
+        input: wav:[signal_length]
+               p_len:int
+        output: f0:[signal_length//hop_length]
+        """
+        pass
+
+    def compute_f0_uv(self, wav, p_len):
+        """
+        input: wav:[signal_length]
+               p_len:int
+        output: f0:[signal_length//hop_length],uv:[signal_length//hop_length]
+        """
+        pass

infer/lib/infer_pack/modules/F0Predictor/HarvestF0Predictor.py

@@ -0,0 +1,87 @@
+import numpy as np
+import pyworld
+
+from infer.lib.infer_pack.modules.F0Predictor.F0Predictor import F0Predictor
+
+
+class HarvestF0Predictor(F0Predictor):
+    def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        self.sampling_rate = sampling_rate
+
+    def interpolate_f0(self, f0):
+        """
+        对F0进行插值处理
+        """
+
+        data = np.reshape(f0, (f0.size, 1))
+
+        vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
+        vuv_vector[data > 0.0] = 1.0
+        vuv_vector[data <= 0.0] = 0.0
+
+        ip_data = data
+
+        frame_number = data.size
+        last_value = 0.0
+        for i in range(frame_number):
+            if data[i] <= 0.0:
+                j = i + 1
+                for j in range(i + 1, frame_number):
+                    if data[j] > 0.0:
+                        break
+                if j < frame_number - 1:
+                    if last_value > 0.0:
+                        step = (data[j] - data[i - 1]) / float(j - i)
+                        for k in range(i, j):
+                            ip_data[k] = data[i - 1] + step * (k - i + 1)
+                    else:
+                        for k in range(i, j):
+                            ip_data[k] = data[j]
+                else:
+                    for k in range(i, frame_number):
+                        ip_data[k] = last_value
+            else:
+                ip_data[i] = data[i]  # 这里可能存在一个没有必要的拷贝
+                last_value = data[i]
+
+        return ip_data[:, 0], vuv_vector[:, 0]
+
+    def resize_f0(self, x, target_len):
+        source = np.array(x)
+        source[source < 0.001] = np.nan
+        target = np.interp(
+            np.arange(0, len(source) * target_len, len(source)) / target_len,
+            np.arange(0, len(source)),
+            source,
+        )
+        res = np.nan_to_num(target)
+        return res
+
+    def compute_f0(self, wav, p_len=None):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        f0, t = pyworld.harvest(
+            wav.astype(np.double),
+            fs=self.sampling_rate,
+            f0_ceil=self.f0_max,
+            f0_floor=self.f0_min,
+            frame_period=1000 * self.hop_length / self.sampling_rate,
+        )
+        f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.fs)
+        return self.interpolate_f0(self.resize_f0(f0, p_len))[0]
+
+    def compute_f0_uv(self, wav, p_len=None):
+        if p_len is None:
+            p_len = wav.shape[0] // self.hop_length
+        f0, t = pyworld.harvest(
+            wav.astype(np.double),
+            fs=self.sampling_rate,
+            f0_floor=self.f0_min,
+            f0_ceil=self.f0_max,
+            frame_period=1000 * self.hop_length / self.sampling_rate,
+        )
+        f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.sampling_rate)
+        return self.interpolate_f0(self.resize_f0(f0, p_len))

infer/lib/infer_pack/modules/F0Predictor/PMF0Predictor.py

@@ -0,0 +1,98 @@
+import numpy as np
+import parselmouth
+
+from infer.lib.infer_pack.modules.F0Predictor.F0Predictor import F0Predictor
+
+
+class PMF0Predictor(F0Predictor):
+    def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        self.sampling_rate = sampling_rate
+
+    def interpolate_f0(self, f0):
+        """
+        对F0进行插值处理
+        """
+
+        data = np.reshape(f0, (f0.size, 1))
+
+        vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
+        vuv_vector[data > 0.0] = 1.0
+        vuv_vector[data <= 0.0] = 0.0
+
+        ip_data = data
+
+        frame_number = data.size
+        last_value = 0.0
+        for i in range(frame_number):
+            if data[i] <= 0.0:
+                j = i + 1
+                for j in range(i + 1, frame_number):
+                    if data[j] > 0.0:
+                        break
+                if j < frame_number - 1:
+                    if last_value > 0.0:
+                        step = (data[j] - data[i - 1]) / float(j - i)
+                        for k in range(i, j):
+                            ip_data[k] = data[i - 1] + step * (k - i + 1)
+                    else:
+                        for k in range(i, j):
+                            ip_data[k] = data[j]
+                else:
+                    for k in range(i, frame_number):
+                        ip_data[k] = last_value
+            else:
+                ip_data[i] = data[i]  # 这里可能存在一个没有必要的拷贝
+                last_value = data[i]
+
+        return ip_data[:, 0], vuv_vector[:, 0]
+
+    def compute_f0(self, wav, p_len=None):
+        x = wav
+        if p_len is None:
+            p_len = x.shape[0] // self.hop_length
+        else:
+            assert abs(p_len - x.shape[0] // self.hop_length) < 4, "pad length error"
+        time_step = self.hop_length / self.sampling_rate * 1000
+        f0 = (
+            parselmouth.Sound(x, self.sampling_rate)
+            .to_pitch_ac(
+                time_step=time_step / 1000,
+                voicing_threshold=0.6,
+                pitch_floor=self.f0_min,
+                pitch_ceiling=self.f0_max,
+            )
+            .selected_array["frequency"]
+        )
+
+        pad_size = (p_len - len(f0) + 1) // 2
+        if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+            f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+        f0, uv = self.interpolate_f0(f0)
+        return f0
+
+    def compute_f0_uv(self, wav, p_len=None):
+        x = wav
+        if p_len is None:
+            p_len = x.shape[0] // self.hop_length
+        else:
+            assert abs(p_len - x.shape[0] // self.hop_length) < 4, "pad length error"
+        time_step = self.hop_length / self.sampling_rate * 1000
+        f0 = (
+            parselmouth.Sound(x, self.sampling_rate)
+            .to_pitch_ac(
+                time_step=time_step / 1000,
+                voicing_threshold=0.6,
+                pitch_floor=self.f0_min,
+                pitch_ceiling=self.f0_max,
+            )
+            .selected_array["frequency"]
+        )
+
+        pad_size = (p_len - len(f0) + 1) // 2
+        if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+            f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+        f0, uv = self.interpolate_f0(f0)
+        return f0, uv

infer/lib/infer_pack/onnx_inference.py

@@ -0,0 +1,149 @@
+import librosa
+import numpy as np
+import onnxruntime
+import soundfile
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class ContentVec:
+    def __init__(self, vec_path="pretrained/vec-768-layer-12.onnx", device=None):
+        logger.info("Load model(s) from {}".format(vec_path))
+        if device == "cpu" or device is None:
+            providers = ["CPUExecutionProvider"]
+        elif device == "cuda":
+            providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        elif device == "dml":
+            providers = ["DmlExecutionProvider"]
+        else:
+            raise RuntimeError("Unsportted Device")
+        self.model = onnxruntime.InferenceSession(vec_path, providers=providers)
+
+    def __call__(self, wav):
+        return self.forward(wav)
+
+    def forward(self, wav):
+        feats = wav
+        if feats.ndim == 2:  # double channels
+            feats = feats.mean(-1)
+        assert feats.ndim == 1, feats.ndim
+        feats = np.expand_dims(np.expand_dims(feats, 0), 0)
+        onnx_input = {self.model.get_inputs()[0].name: feats}
+        logits = self.model.run(None, onnx_input)[0]
+        return logits.transpose(0, 2, 1)
+
+
+def get_f0_predictor(f0_predictor, hop_length, sampling_rate, **kargs):
+    if f0_predictor == "pm":
+        from lib.infer_pack.modules.F0Predictor.PMF0Predictor import PMF0Predictor
+
+        f0_predictor_object = PMF0Predictor(
+            hop_length=hop_length, sampling_rate=sampling_rate
+        )
+    elif f0_predictor == "harvest":
+        from lib.infer_pack.modules.F0Predictor.HarvestF0Predictor import (
+            HarvestF0Predictor,
+        )
+
+        f0_predictor_object = HarvestF0Predictor(
+            hop_length=hop_length, sampling_rate=sampling_rate
+        )
+    elif f0_predictor == "dio":
+        from lib.infer_pack.modules.F0Predictor.DioF0Predictor import DioF0Predictor
+
+        f0_predictor_object = DioF0Predictor(
+            hop_length=hop_length, sampling_rate=sampling_rate
+        )
+    else:
+        raise Exception("Unknown f0 predictor")
+    return f0_predictor_object
+
+
+class OnnxRVC:
+    def __init__(
+        self,
+        model_path,
+        sr=40000,
+        hop_size=512,
+        vec_path="vec-768-layer-12",
+        device="cpu",
+    ):
+        vec_path = f"pretrained/{vec_path}.onnx"
+        self.vec_model = ContentVec(vec_path, device)
+        if device == "cpu" or device is None:
+            providers = ["CPUExecutionProvider"]
+        elif device == "cuda":
+            providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        elif device == "dml":
+            providers = ["DmlExecutionProvider"]
+        else:
+            raise RuntimeError("Unsportted Device")
+        self.model = onnxruntime.InferenceSession(model_path, providers=providers)
+        self.sampling_rate = sr
+        self.hop_size = hop_size
+
+    def forward(self, hubert, hubert_length, pitch, pitchf, ds, rnd):
+        onnx_input = {
+            self.model.get_inputs()[0].name: hubert,
+            self.model.get_inputs()[1].name: hubert_length,
+            self.model.get_inputs()[2].name: pitch,
+            self.model.get_inputs()[3].name: pitchf,
+            self.model.get_inputs()[4].name: ds,
+            self.model.get_inputs()[5].name: rnd,
+        }
+        return (self.model.run(None, onnx_input)[0] * 32767).astype(np.int16)
+
+    def inference(
+        self,
+        raw_path,
+        sid,
+        f0_method="dio",
+        f0_up_key=0,
+        pad_time=0.5,
+        cr_threshold=0.02,
+    ):
+        f0_min = 50
+        f0_max = 1100
+        f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+        f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+        f0_predictor = get_f0_predictor(
+            f0_method,
+            hop_length=self.hop_size,
+            sampling_rate=self.sampling_rate,
+            threshold=cr_threshold,
+        )
+        wav, sr = librosa.load(raw_path, sr=self.sampling_rate)
+        org_length = len(wav)
+        if org_length / sr > 50.0:
+            raise RuntimeError("Reached Max Length")
+
+        wav16k = librosa.resample(wav, orig_sr=self.sampling_rate, target_sr=16000)
+        wav16k = wav16k
+
+        hubert = self.vec_model(wav16k)
+        hubert = np.repeat(hubert, 2, axis=2).transpose(0, 2, 1).astype(np.float32)
+        hubert_length = hubert.shape[1]
+
+        pitchf = f0_predictor.compute_f0(wav, hubert_length)
+        pitchf = pitchf * 2 ** (f0_up_key / 12)
+        pitch = pitchf.copy()
+        f0_mel = 1127 * np.log(1 + pitch / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
+            f0_mel_max - f0_mel_min
+        ) + 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > 255] = 255
+        pitch = np.rint(f0_mel).astype(np.int64)
+
+        pitchf = pitchf.reshape(1, len(pitchf)).astype(np.float32)
+        pitch = pitch.reshape(1, len(pitch))
+        ds = np.array([sid]).astype(np.int64)
+
+        rnd = np.random.randn(1, 192, hubert_length).astype(np.float32)
+        hubert_length = np.array([hubert_length]).astype(np.int64)
+
+        out_wav = self.forward(hubert, hubert_length, pitch, pitchf, ds, rnd).squeeze()
+        out_wav = np.pad(out_wav, (0, 2 * self.hop_size), "constant")
+        return out_wav[0:org_length]

infer/lib/infer_pack/transforms.py

@@ -0,0 +1,207 @@
+import numpy as np
+import torch
+from torch.nn import functional as F
+
+DEFAULT_MIN_BIN_WIDTH = 1e-3
+DEFAULT_MIN_BIN_HEIGHT = 1e-3
+DEFAULT_MIN_DERIVATIVE = 1e-3
+
+
+def piecewise_rational_quadratic_transform(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails=None,
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if tails is None:
+        spline_fn = rational_quadratic_spline
+        spline_kwargs = {}
+    else:
+        spline_fn = unconstrained_rational_quadratic_spline
+        spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
+
+    outputs, logabsdet = spline_fn(
+        inputs=inputs,
+        unnormalized_widths=unnormalized_widths,
+        unnormalized_heights=unnormalized_heights,
+        unnormalized_derivatives=unnormalized_derivatives,
+        inverse=inverse,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+        **spline_kwargs
+    )
+    return outputs, logabsdet
+
+
+def searchsorted(bin_locations, inputs, eps=1e-6):
+    bin_locations[..., -1] += eps
+    return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
+
+
+def unconstrained_rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails="linear",
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
+    outside_interval_mask = ~inside_interval_mask
+
+    outputs = torch.zeros_like(inputs)
+    logabsdet = torch.zeros_like(inputs)
+
+    if tails == "linear":
+        unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
+        constant = np.log(np.exp(1 - min_derivative) - 1)
+        unnormalized_derivatives[..., 0] = constant
+        unnormalized_derivatives[..., -1] = constant
+
+        outputs[outside_interval_mask] = inputs[outside_interval_mask]
+        logabsdet[outside_interval_mask] = 0
+    else:
+        raise RuntimeError("{} tails are not implemented.".format(tails))
+
+    (
+        outputs[inside_interval_mask],
+        logabsdet[inside_interval_mask],
+    ) = rational_quadratic_spline(
+        inputs=inputs[inside_interval_mask],
+        unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
+        unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
+        unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
+        inverse=inverse,
+        left=-tail_bound,
+        right=tail_bound,
+        bottom=-tail_bound,
+        top=tail_bound,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+    )
+
+    return outputs, logabsdet
+
+
+def rational_quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    left=0.0,
+    right=1.0,
+    bottom=0.0,
+    top=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if torch.min(inputs) < left or torch.max(inputs) > right:
+        raise ValueError("Input to a transform is not within its domain")
+
+    num_bins = unnormalized_widths.shape[-1]
+
+    if min_bin_width * num_bins > 1.0:
+        raise ValueError("Minimal bin width too large for the number of bins")
+    if min_bin_height * num_bins > 1.0:
+        raise ValueError("Minimal bin height too large for the number of bins")
+
+    widths = F.softmax(unnormalized_widths, dim=-1)
+    widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
+    cumwidths = torch.cumsum(widths, dim=-1)
+    cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
+    cumwidths = (right - left) * cumwidths + left
+    cumwidths[..., 0] = left
+    cumwidths[..., -1] = right
+    widths = cumwidths[..., 1:] - cumwidths[..., :-1]
+
+    derivatives = min_derivative + F.softplus(unnormalized_derivatives)
+
+    heights = F.softmax(unnormalized_heights, dim=-1)
+    heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
+    cumheights = torch.cumsum(heights, dim=-1)
+    cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
+    cumheights = (top - bottom) * cumheights + bottom
+    cumheights[..., 0] = bottom
+    cumheights[..., -1] = top
+    heights = cumheights[..., 1:] - cumheights[..., :-1]
+
+    if inverse:
+        bin_idx = searchsorted(cumheights, inputs)[..., None]
+    else:
+        bin_idx = searchsorted(cumwidths, inputs)[..., None]
+
+    input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
+    input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
+
+    input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
+    delta = heights / widths
+    input_delta = delta.gather(-1, bin_idx)[..., 0]
+
+    input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
+    input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
+
+    input_heights = heights.gather(-1, bin_idx)[..., 0]
+
+    if inverse:
+        a = (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        ) + input_heights * (input_delta - input_derivatives)
+        b = input_heights * input_derivatives - (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        )
+        c = -input_delta * (inputs - input_cumheights)
+
+        discriminant = b.pow(2) - 4 * a * c
+        assert (discriminant >= 0).all()
+
+        root = (2 * c) / (-b - torch.sqrt(discriminant))
+        outputs = root * input_bin_widths + input_cumwidths
+
+        theta_one_minus_theta = root * (1 - root)
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * root.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - root).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, -logabsdet
+    else:
+        theta = (inputs - input_cumwidths) / input_bin_widths
+        theta_one_minus_theta = theta * (1 - theta)
+
+        numerator = input_heights * (
+            input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
+        )
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        outputs = input_cumheights + numerator / denominator
+
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * theta.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - theta).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, logabsdet

infer/lib/jit/__init__.py

@@ -0,0 +1,163 @@
+from io import BytesIO
+import pickle
+import time
+import torch
+from tqdm import tqdm
+from collections import OrderedDict
+
+
+def load_inputs(path, device, is_half=False):
+    parm = torch.load(path, map_location=torch.device("cpu"), weights_only=False)
+    for key in parm.keys():
+        parm[key] = parm[key].to(device)
+        if is_half and parm[key].dtype == torch.float32:
+            parm[key] = parm[key].half()
+        elif not is_half and parm[key].dtype == torch.float16:
+            parm[key] = parm[key].float()
+    return parm
+
+
+def benchmark(
+    model, inputs_path, device=torch.device("cpu"), epoch=1000, is_half=False
+):
+    parm = load_inputs(inputs_path, device, is_half)
+    total_ts = 0.0
+    bar = tqdm(range(epoch))
+    for i in bar:
+        start_time = time.perf_counter()
+        o = model(**parm)
+        total_ts += time.perf_counter() - start_time
+    print(f"num_epoch: {epoch} | avg time(ms): {(total_ts*1000)/epoch}")
+
+
+def jit_warm_up(model, inputs_path, device=torch.device("cpu"), epoch=5, is_half=False):
+    benchmark(model, inputs_path, device, epoch=epoch, is_half=is_half)
+
+
+def to_jit_model(
+    model_path,
+    model_type: str,
+    mode: str = "trace",
+    inputs_path: str = None,
+    device=torch.device("cpu"),
+    is_half=False,
+):
+    model = None
+    if model_type.lower() == "synthesizer":
+        from .get_synthesizer import get_synthesizer
+
+        model, _ = get_synthesizer(model_path, device)
+        model.forward = model.infer
+    elif model_type.lower() == "rmvpe":
+        from .get_rmvpe import get_rmvpe
+
+        model = get_rmvpe(model_path, device)
+    elif model_type.lower() == "hubert":
+        from .get_hubert import get_hubert_model
+
+        model = get_hubert_model(model_path, device)
+        model.forward = model.infer
+    else:
+        raise ValueError(f"No model type named {model_type}")
+    model = model.eval()
+    model = model.half() if is_half else model.float()
+    if mode == "trace":
+        assert not inputs_path
+        inputs = load_inputs(inputs_path, device, is_half)
+        model_jit = torch.jit.trace(model, example_kwarg_inputs=inputs)
+    elif mode == "script":
+        model_jit = torch.jit.script(model)
+    model_jit.to(device)
+    model_jit = model_jit.half() if is_half else model_jit.float()
+    # model = model.half() if is_half else model.float()
+    return (model, model_jit)
+
+
+def export(
+    model: torch.nn.Module,
+    mode: str = "trace",
+    inputs: dict = None,
+    device=torch.device("cpu"),
+    is_half: bool = False,
+) -> dict:
+    model = model.half() if is_half else model.float()
+    model.eval()
+    if mode == "trace":
+        assert inputs is not None
+        model_jit = torch.jit.trace(model, example_kwarg_inputs=inputs)
+    elif mode == "script":
+        model_jit = torch.jit.script(model)
+    model_jit.to(device)
+    model_jit = model_jit.half() if is_half else model_jit.float()
+    buffer = BytesIO()
+    # model_jit=model_jit.cpu()
+    torch.jit.save(model_jit, buffer)
+    del model_jit
+    cpt = OrderedDict()
+    cpt["model"] = buffer.getvalue()
+    cpt["is_half"] = is_half
+    return cpt
+
+
+def load(path: str):
+    with open(path, "rb") as f:
+        return pickle.load(f)
+
+
+def save(ckpt: dict, save_path: str):
+    with open(save_path, "wb") as f:
+        pickle.dump(ckpt, f)
+
+
+def rmvpe_jit_export(
+    model_path: str,
+    mode: str = "script",
+    inputs_path: str = None,
+    save_path: str = None,
+    device=torch.device("cpu"),
+    is_half=False,
+):
+    if not save_path:
+        save_path = model_path.rstrip(".pth")
+        save_path += ".half.jit" if is_half else ".jit"
+    if "cuda" in str(device) and ":" not in str(device):
+        device = torch.device("cuda:0")
+    from .get_rmvpe import get_rmvpe
+
+    model = get_rmvpe(model_path, device)
+    inputs = None
+    if mode == "trace":
+        inputs = load_inputs(inputs_path, device, is_half)
+    ckpt = export(model, mode, inputs, device, is_half)
+    ckpt["device"] = str(device)
+    save(ckpt, save_path)
+    return ckpt
+
+
+def synthesizer_jit_export(
+    model_path: str,
+    mode: str = "script",
+    inputs_path: str = None,
+    save_path: str = None,
+    device=torch.device("cpu"),
+    is_half=False,
+):
+    if not save_path:
+        save_path = model_path.rstrip(".pth")
+        save_path += ".half.jit" if is_half else ".jit"
+    if "cuda" in str(device) and ":" not in str(device):
+        device = torch.device("cuda:0")
+    from .get_synthesizer import get_synthesizer
+
+    model, cpt = get_synthesizer(model_path, device)
+    assert isinstance(cpt, dict)
+    model.forward = model.infer
+    inputs = None
+    if mode == "trace":
+        inputs = load_inputs(inputs_path, device, is_half)
+    ckpt = export(model, mode, inputs, device, is_half)
+    cpt.pop("weight")
+    cpt["model"] = ckpt["model"]
+    cpt["device"] = device
+    save(cpt, save_path)
+    return cpt

infer/lib/jit/get_hubert.py

@@ -0,0 +1,342 @@
+import math
+import random
+from typing import Optional, Tuple
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+# from fairseq.data.data_utils import compute_mask_indices
+from fairseq.utils import index_put
+
+
+# @torch.jit.script
+def pad_to_multiple(x, multiple, dim=-1, value=0):
+    # Inspired from https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py#L41
+    if x is None:
+        return None, 0
+    tsz = x.size(dim)
+    m = tsz / multiple
+    remainder = math.ceil(m) * multiple - tsz
+    if int(tsz % multiple) == 0:
+        return x, 0
+    pad_offset = (0,) * (-1 - dim) * 2
+
+    return F.pad(x, (*pad_offset, 0, remainder), value=value), remainder
+
+
+def extract_features(
+    self,
+    x,
+    padding_mask=None,
+    tgt_layer=None,
+    min_layer=0,
+):
+    if padding_mask is not None:
+        x = index_put(x, padding_mask, 0)
+
+    x_conv = self.pos_conv(x.transpose(1, 2))
+    x_conv = x_conv.transpose(1, 2)
+    x = x + x_conv
+
+    if not self.layer_norm_first:
+        x = self.layer_norm(x)
+
+    # pad to the sequence length dimension
+    x, pad_length = pad_to_multiple(x, self.required_seq_len_multiple, dim=-2, value=0)
+    if pad_length > 0 and padding_mask is None:
+        padding_mask = x.new_zeros((x.size(0), x.size(1)), dtype=torch.bool)
+        padding_mask[:, -pad_length:] = True
+    else:
+        padding_mask, _ = pad_to_multiple(
+            padding_mask, self.required_seq_len_multiple, dim=-1, value=True
+        )
+    x = F.dropout(x, p=self.dropout, training=self.training)
+
+    # B x T x C -> T x B x C
+    x = x.transpose(0, 1)
+
+    layer_results = []
+    r = None
+    for i, layer in enumerate(self.layers):
+        dropout_probability = np.random.random() if self.layerdrop > 0 else 1
+        if not self.training or (dropout_probability > self.layerdrop):
+            x, (z, lr) = layer(
+                x, self_attn_padding_mask=padding_mask, need_weights=False
+            )
+            if i >= min_layer:
+                layer_results.append((x, z, lr))
+        if i == tgt_layer:
+            r = x
+            break
+
+    if r is not None:
+        x = r
+
+    # T x B x C -> B x T x C
+    x = x.transpose(0, 1)
+
+    # undo paddding
+    if pad_length > 0:
+        x = x[:, :-pad_length]
+
+        def undo_pad(a, b, c):
+            return (
+                a[:-pad_length],
+                b[:-pad_length] if b is not None else b,
+                c[:-pad_length],
+            )
+
+        layer_results = [undo_pad(*u) for u in layer_results]
+
+    return x, layer_results
+
+
+def compute_mask_indices(
+    shape: Tuple[int, int],
+    padding_mask: Optional[torch.Tensor],
+    mask_prob: float,
+    mask_length: int,
+    mask_type: str = "static",
+    mask_other: float = 0.0,
+    min_masks: int = 0,
+    no_overlap: bool = False,
+    min_space: int = 0,
+    require_same_masks: bool = True,
+    mask_dropout: float = 0.0,
+) -> torch.Tensor:
+    """
+    Computes random mask spans for a given shape
+
+    Args:
+        shape: the the shape for which to compute masks.
+            should be of size 2 where first element is batch size and 2nd is timesteps
+        padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements
+        mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
+            number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
+            however due to overlaps, the actual number will be smaller (unless no_overlap is True)
+        mask_type: how to compute mask lengths
+            static = fixed size
+            uniform = sample from uniform distribution [mask_other, mask_length*2]
+            normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element
+            poisson = sample from possion distribution with lambda = mask length
+        min_masks: minimum number of masked spans
+        no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping
+        min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans
+        require_same_masks: if true, will randomly drop out masks until same amount of masks remains in each sample
+        mask_dropout: randomly dropout this percentage of masks in each example
+    """
+
+    bsz, all_sz = shape
+    mask = torch.full((bsz, all_sz), False)
+
+    all_num_mask = int(
+        # add a random number for probabilistic rounding
+        mask_prob * all_sz / float(mask_length)
+        + torch.rand([1]).item()
+    )
+
+    all_num_mask = max(min_masks, all_num_mask)
+
+    mask_idcs = []
+    for i in range(bsz):
+        if padding_mask is not None:
+            sz = all_sz - padding_mask[i].long().sum().item()
+            num_mask = int(mask_prob * sz / float(mask_length) + np.random.rand())
+            num_mask = max(min_masks, num_mask)
+        else:
+            sz = all_sz
+            num_mask = all_num_mask
+
+        if mask_type == "static":
+            lengths = torch.full([num_mask], mask_length)
+        elif mask_type == "uniform":
+            lengths = torch.randint(mask_other, mask_length * 2 + 1, size=[num_mask])
+        elif mask_type == "normal":
+            lengths = torch.normal(mask_length, mask_other, size=[num_mask])
+            lengths = [max(1, int(round(x))) for x in lengths]
+        else:
+            raise Exception("unknown mask selection " + mask_type)
+
+        if sum(lengths) == 0:
+            lengths[0] = min(mask_length, sz - 1)
+
+        if no_overlap:
+            mask_idc = []
+
+            def arrange(s, e, length, keep_length):
+                span_start = torch.randint(low=s, high=e - length, size=[1]).item()
+                mask_idc.extend(span_start + i for i in range(length))
+
+                new_parts = []
+                if span_start - s - min_space >= keep_length:
+                    new_parts.append((s, span_start - min_space + 1))
+                if e - span_start - length - min_space > keep_length:
+                    new_parts.append((span_start + length + min_space, e))
+                return new_parts
+
+            parts = [(0, sz)]
+            min_length = min(lengths)
+            for length in sorted(lengths, reverse=True):
+                t = [e - s if e - s >= length + min_space else 0 for s, e in parts]
+                lens = torch.asarray(t, dtype=torch.int)
+                l_sum = torch.sum(lens)
+                if l_sum == 0:
+                    break
+                probs = lens / torch.sum(lens)
+                c = torch.multinomial(probs.float(), len(parts)).item()
+                s, e = parts.pop(c)
+                parts.extend(arrange(s, e, length, min_length))
+            mask_idc = torch.asarray(mask_idc)
+        else:
+            min_len = min(lengths)
+            if sz - min_len <= num_mask:
+                min_len = sz - num_mask - 1
+            mask_idc = torch.asarray(
+                random.sample([i for i in range(sz - min_len)], num_mask)
+            )
+            mask_idc = torch.asarray(
+                [
+                    mask_idc[j] + offset
+                    for j in range(len(mask_idc))
+                    for offset in range(lengths[j])
+                ]
+            )
+
+        mask_idcs.append(torch.unique(mask_idc[mask_idc < sz]))
+
+    min_len = min([len(m) for m in mask_idcs])
+    for i, mask_idc in enumerate(mask_idcs):
+        if isinstance(mask_idc, torch.Tensor):
+            mask_idc = torch.asarray(mask_idc, dtype=torch.float)
+        if len(mask_idc) > min_len and require_same_masks:
+            mask_idc = torch.asarray(
+                random.sample([i for i in range(mask_idc)], min_len)
+            )
+        if mask_dropout > 0:
+            num_holes = int(round(len(mask_idc) * mask_dropout))
+            mask_idc = torch.asarray(
+                random.sample([i for i in range(mask_idc)], len(mask_idc) - num_holes)
+            )
+
+        mask[i, mask_idc.int()] = True
+
+    return mask
+
+
+def apply_mask(self, x, padding_mask, target_list):
+    B, T, C = x.shape
+    torch.zeros_like(x)
+    if self.mask_prob > 0:
+        mask_indices = compute_mask_indices(
+            (B, T),
+            padding_mask,
+            self.mask_prob,
+            self.mask_length,
+            self.mask_selection,
+            self.mask_other,
+            min_masks=2,
+            no_overlap=self.no_mask_overlap,
+            min_space=self.mask_min_space,
+        )
+        mask_indices = mask_indices.to(x.device)
+        x[mask_indices] = self.mask_emb
+    else:
+        mask_indices = None
+
+    if self.mask_channel_prob > 0:
+        mask_channel_indices = compute_mask_indices(
+            (B, C),
+            None,
+            self.mask_channel_prob,
+            self.mask_channel_length,
+            self.mask_channel_selection,
+            self.mask_channel_other,
+            no_overlap=self.no_mask_channel_overlap,
+            min_space=self.mask_channel_min_space,
+        )
+        mask_channel_indices = (
+            mask_channel_indices.to(x.device).unsqueeze(1).expand(-1, T, -1)
+        )
+        x[mask_channel_indices] = 0
+
+    return x, mask_indices
+
+
+def get_hubert_model(
+    model_path="assets/hubert/hubert_base.pt", device=torch.device("cpu")
+):
+    models, _, _ = load_model_ensemble_and_task(
+        [model_path],
+        suffix="",
+    )
+    hubert_model = models[0]
+    hubert_model = hubert_model.to(device)
+
+    def _apply_mask(x, padding_mask, target_list):
+        return apply_mask(hubert_model, x, padding_mask, target_list)
+
+    hubert_model.apply_mask = _apply_mask
+
+    def _extract_features(
+        x,
+        padding_mask=None,
+        tgt_layer=None,
+        min_layer=0,
+    ):
+        return extract_features(
+            hubert_model.encoder,
+            x,
+            padding_mask=padding_mask,
+            tgt_layer=tgt_layer,
+            min_layer=min_layer,
+        )
+
+    hubert_model.encoder.extract_features = _extract_features
+
+    hubert_model._forward = hubert_model.forward
+
+    def hubert_extract_features(
+        self,
+        source: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        mask: bool = False,
+        ret_conv: bool = False,
+        output_layer: Optional[int] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        res = self._forward(
+            source,
+            padding_mask=padding_mask,
+            mask=mask,
+            features_only=True,
+            output_layer=output_layer,
+        )
+        feature = res["features"] if ret_conv else res["x"]
+        return feature, res["padding_mask"]
+
+    def _hubert_extract_features(
+        source: torch.Tensor,
+        padding_mask: Optional[torch.Tensor] = None,
+        mask: bool = False,
+        ret_conv: bool = False,
+        output_layer: Optional[int] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        return hubert_extract_features(
+            hubert_model, source, padding_mask, mask, ret_conv, output_layer
+        )
+
+    hubert_model.extract_features = _hubert_extract_features
+
+    def infer(source, padding_mask, output_layer: torch.Tensor):
+        output_layer = output_layer.item()
+        logits = hubert_model.extract_features(
+            source=source, padding_mask=padding_mask, output_layer=output_layer
+        )
+        feats = hubert_model.final_proj(logits[0]) if output_layer == 9 else logits[0]
+        return feats
+
+    hubert_model.infer = infer
+    # hubert_model.forward=infer
+    # hubert_model.forward
+
+    return hubert_model

infer/lib/jit/get_rmvpe.py

@@ -0,0 +1,12 @@
+import torch
+
+
+def get_rmvpe(model_path="assets/rmvpe/rmvpe.pt", device=torch.device("cpu")):
+    from infer.lib.rmvpe import E2E
+
+    model = E2E(4, 1, (2, 2))
+    ckpt = torch.load(model_path, map_location=device, weights_only=False)
+    model.load_state_dict(ckpt)
+    model.eval()
+    model = model.to(device)
+    return model

infer/lib/jit/get_synthesizer.py

@@ -0,0 +1,38 @@
+import torch
+
+
+def get_synthesizer(pth_path, device=torch.device("cpu")):
+    from infer.lib.infer_pack.models import (
+        SynthesizerTrnMs256NSFsid,
+        SynthesizerTrnMs256NSFsid_nono,
+        SynthesizerTrnMs768NSFsid,
+        SynthesizerTrnMs768NSFsid_nono,
+    )
+
+    cpt = torch.load(pth_path, map_location=torch.device("cpu"), weights_only=False)
+    # tgt_sr = cpt["config"][-1]
+    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+    if_f0 = cpt.get("f0", 1)
+    version = cpt.get("version", "v1")
+    if version == "v1":
+        if if_f0 == 1:
+            net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=False)
+        else:
+            net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
+    elif version == "v2":
+        if if_f0 == 1:
+            net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=False)
+        else:
+            net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
+    del net_g.enc_q
+    # net_g.forward = net_g.infer
+    # ckpt = {}
+    # ckpt["config"] = cpt["config"]
+    # ckpt["f0"] = if_f0
+    # ckpt["version"] = version
+    # ckpt["info"] = cpt.get("info", "0epoch")
+    net_g.load_state_dict(cpt["weight"], strict=False)
+    net_g = net_g.float()
+    net_g.eval().to(device)
+    net_g.remove_weight_norm()
+    return net_g, cpt

infer/lib/rmvpe.py

@@ -0,0 +1,670 @@
+from io import BytesIO
+import os
+from typing import List, Optional, Tuple
+import numpy as np
+import torch
+
+from infer.lib import jit
+
+try:
+    # Fix "Torch not compiled with CUDA enabled"
+    import intel_extension_for_pytorch as ipex  # pylint: disable=import-error, unused-import
+
+    if torch.xpu.is_available():
+        from infer.modules.ipex import ipex_init
+
+        ipex_init()
+except Exception:  # pylint: disable=broad-exception-caught
+    pass
+import torch.nn as nn
+import torch.nn.functional as F
+from librosa.util import normalize, pad_center, tiny
+from scipy.signal import get_window
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class STFT(torch.nn.Module):
+    def __init__(
+        self, filter_length=1024, hop_length=512, win_length=None, window="hann"
+    ):
+        """
+        This module implements an STFT using 1D convolution and 1D transpose convolutions.
+        This is a bit tricky so there are some cases that probably won't work as working
+        out the same sizes before and after in all overlap add setups is tough. Right now,
+        this code should work with hop lengths that are half the filter length (50% overlap
+        between frames).
+
+        Keyword Arguments:
+            filter_length {int} -- Length of filters used (default: {1024})
+            hop_length {int} -- Hop length of STFT (restrict to 50% overlap between frames) (default: {512})
+            win_length {[type]} -- Length of the window function applied to each frame (if not specified, it
+                equals the filter length). (default: {None})
+            window {str} -- Type of window to use (options are bartlett, hann, hamming, blackman, blackmanharris)
+                (default: {'hann'})
+        """
+        super(STFT, self).__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length if win_length else filter_length
+        self.window = window
+        self.forward_transform = None
+        self.pad_amount = int(self.filter_length / 2)
+        fourier_basis = np.fft.fft(np.eye(self.filter_length))
+
+        cutoff = int((self.filter_length / 2 + 1))
+        fourier_basis = np.vstack(
+            [np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])]
+        )
+        forward_basis = torch.FloatTensor(fourier_basis)
+        inverse_basis = torch.FloatTensor(np.linalg.pinv(fourier_basis))
+
+        assert filter_length >= self.win_length
+        # get window and zero center pad it to filter_length
+        fft_window = get_window(window, self.win_length, fftbins=True)
+        fft_window = pad_center(fft_window, size=filter_length)
+        fft_window = torch.from_numpy(fft_window).float()
+
+        # window the bases
+        forward_basis *= fft_window
+        inverse_basis = (inverse_basis.T * fft_window).T
+
+        self.register_buffer("forward_basis", forward_basis.float())
+        self.register_buffer("inverse_basis", inverse_basis.float())
+        self.register_buffer("fft_window", fft_window.float())
+
+    def transform(self, input_data, return_phase=False):
+        """Take input data (audio) to STFT domain.
+
+        Arguments:
+            input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
+
+        Returns:
+            magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
+                num_frequencies, num_frames)
+            phase {tensor} -- Phase of STFT with shape (num_batch,
+                num_frequencies, num_frames)
+        """
+        input_data = F.pad(
+            input_data,
+            (self.pad_amount, self.pad_amount),
+            mode="reflect",
+        )
+        forward_transform = input_data.unfold(
+            1, self.filter_length, self.hop_length
+        ).permute(0, 2, 1)
+        forward_transform = torch.matmul(self.forward_basis, forward_transform)
+        cutoff = int((self.filter_length / 2) + 1)
+        real_part = forward_transform[:, :cutoff, :]
+        imag_part = forward_transform[:, cutoff:, :]
+        magnitude = torch.sqrt(real_part**2 + imag_part**2)
+        if return_phase:
+            phase = torch.atan2(imag_part.data, real_part.data)
+            return magnitude, phase
+        else:
+            return magnitude
+
+    def inverse(self, magnitude, phase):
+        """Call the inverse STFT (iSTFT), given magnitude and phase tensors produced
+        by the ```transform``` function.
+
+        Arguments:
+            magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
+                num_frequencies, num_frames)
+            phase {tensor} -- Phase of STFT with shape (num_batch,
+                num_frequencies, num_frames)
+
+        Returns:
+            inverse_transform {tensor} -- Reconstructed audio given magnitude and phase. Of
+                shape (num_batch, num_samples)
+        """
+        cat = torch.cat(
+            [magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
+        )
+        fold = torch.nn.Fold(
+            output_size=(1, (cat.size(-1) - 1) * self.hop_length + self.filter_length),
+            kernel_size=(1, self.filter_length),
+            stride=(1, self.hop_length),
+        )
+        inverse_transform = torch.matmul(self.inverse_basis, cat)
+        inverse_transform = fold(inverse_transform)[
+            :, 0, 0, self.pad_amount : -self.pad_amount
+        ]
+        window_square_sum = (
+            self.fft_window.pow(2).repeat(cat.size(-1), 1).T.unsqueeze(0)
+        )
+        window_square_sum = fold(window_square_sum)[
+            :, 0, 0, self.pad_amount : -self.pad_amount
+        ]
+        inverse_transform /= window_square_sum
+        return inverse_transform
+
+    def forward(self, input_data):
+        """Take input data (audio) to STFT domain and then back to audio.
+
+        Arguments:
+            input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
+
+        Returns:
+            reconstruction {tensor} -- Reconstructed audio given magnitude and phase. Of
+                shape (num_batch, num_samples)
+        """
+        self.magnitude, self.phase = self.transform(input_data, return_phase=True)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
+
+
+from time import time as ttime
+
+
+class BiGRU(nn.Module):
+    def __init__(self, input_features, hidden_features, num_layers):
+        super(BiGRU, self).__init__()
+        self.gru = nn.GRU(
+            input_features,
+            hidden_features,
+            num_layers=num_layers,
+            batch_first=True,
+            bidirectional=True,
+        )
+
+    def forward(self, x):
+        return self.gru(x)[0]
+
+
+class ConvBlockRes(nn.Module):
+    def __init__(self, in_channels, out_channels, momentum=0.01):
+        super(ConvBlockRes, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=(1, 1),
+                padding=(1, 1),
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+            nn.Conv2d(
+                in_channels=out_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=(1, 1),
+                padding=(1, 1),
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+        )
+        # self.shortcut:Optional[nn.Module] = None
+        if in_channels != out_channels:
+            self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
+
+    def forward(self, x: torch.Tensor):
+        if not hasattr(self, "shortcut"):
+            return self.conv(x) + x
+        else:
+            return self.conv(x) + self.shortcut(x)
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        in_size,
+        n_encoders,
+        kernel_size,
+        n_blocks,
+        out_channels=16,
+        momentum=0.01,
+    ):
+        super(Encoder, self).__init__()
+        self.n_encoders = n_encoders
+        self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
+        self.layers = nn.ModuleList()
+        self.latent_channels = []
+        for i in range(self.n_encoders):
+            self.layers.append(
+                ResEncoderBlock(
+                    in_channels, out_channels, kernel_size, n_blocks, momentum=momentum
+                )
+            )
+            self.latent_channels.append([out_channels, in_size])
+            in_channels = out_channels
+            out_channels *= 2
+            in_size //= 2
+        self.out_size = in_size
+        self.out_channel = out_channels
+
+    def forward(self, x: torch.Tensor):
+        concat_tensors: List[torch.Tensor] = []
+        x = self.bn(x)
+        for i, layer in enumerate(self.layers):
+            t, x = layer(x)
+            concat_tensors.append(t)
+        return x, concat_tensors
+
+
+class ResEncoderBlock(nn.Module):
+    def __init__(
+        self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01
+    ):
+        super(ResEncoderBlock, self).__init__()
+        self.n_blocks = n_blocks
+        self.conv = nn.ModuleList()
+        self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
+        for i in range(n_blocks - 1):
+            self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
+        self.kernel_size = kernel_size
+        if self.kernel_size is not None:
+            self.pool = nn.AvgPool2d(kernel_size=kernel_size)
+
+    def forward(self, x):
+        for i, conv in enumerate(self.conv):
+            x = conv(x)
+        if self.kernel_size is not None:
+            return x, self.pool(x)
+        else:
+            return x
+
+
+class Intermediate(nn.Module):  #
+    def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
+        super(Intermediate, self).__init__()
+        self.n_inters = n_inters
+        self.layers = nn.ModuleList()
+        self.layers.append(
+            ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum)
+        )
+        for i in range(self.n_inters - 1):
+            self.layers.append(
+                ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum)
+            )
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = layer(x)
+        return x
+
+
+class ResDecoderBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
+        super(ResDecoderBlock, self).__init__()
+        out_padding = (0, 1) if stride == (1, 2) else (1, 1)
+        self.n_blocks = n_blocks
+        self.conv1 = nn.Sequential(
+            nn.ConvTranspose2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=(3, 3),
+                stride=stride,
+                padding=(1, 1),
+                output_padding=out_padding,
+                bias=False,
+            ),
+            nn.BatchNorm2d(out_channels, momentum=momentum),
+            nn.ReLU(),
+        )
+        self.conv2 = nn.ModuleList()
+        self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
+        for i in range(n_blocks - 1):
+            self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
+
+    def forward(self, x, concat_tensor):
+        x = self.conv1(x)
+        x = torch.cat((x, concat_tensor), dim=1)
+        for i, conv2 in enumerate(self.conv2):
+            x = conv2(x)
+        return x
+
+
+class Decoder(nn.Module):
+    def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
+        super(Decoder, self).__init__()
+        self.layers = nn.ModuleList()
+        self.n_decoders = n_decoders
+        for i in range(self.n_decoders):
+            out_channels = in_channels // 2
+            self.layers.append(
+                ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum)
+            )
+            in_channels = out_channels
+
+    def forward(self, x: torch.Tensor, concat_tensors: List[torch.Tensor]):
+        for i, layer in enumerate(self.layers):
+            x = layer(x, concat_tensors[-1 - i])
+        return x
+
+
+class DeepUnet(nn.Module):
+    def __init__(
+        self,
+        kernel_size,
+        n_blocks,
+        en_de_layers=5,
+        inter_layers=4,
+        in_channels=1,
+        en_out_channels=16,
+    ):
+        super(DeepUnet, self).__init__()
+        self.encoder = Encoder(
+            in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels
+        )
+        self.intermediate = Intermediate(
+            self.encoder.out_channel // 2,
+            self.encoder.out_channel,
+            inter_layers,
+            n_blocks,
+        )
+        self.decoder = Decoder(
+            self.encoder.out_channel, en_de_layers, kernel_size, n_blocks
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x, concat_tensors = self.encoder(x)
+        x = self.intermediate(x)
+        x = self.decoder(x, concat_tensors)
+        return x
+
+
+class E2E(nn.Module):
+    def __init__(
+        self,
+        n_blocks,
+        n_gru,
+        kernel_size,
+        en_de_layers=5,
+        inter_layers=4,
+        in_channels=1,
+        en_out_channels=16,
+    ):
+        super(E2E, self).__init__()
+        self.unet = DeepUnet(
+            kernel_size,
+            n_blocks,
+            en_de_layers,
+            inter_layers,
+            in_channels,
+            en_out_channels,
+        )
+        self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
+        if n_gru:
+            self.fc = nn.Sequential(
+                BiGRU(3 * 128, 256, n_gru),
+                nn.Linear(512, 360),
+                nn.Dropout(0.25),
+                nn.Sigmoid(),
+            )
+        else:
+            self.fc = nn.Sequential(
+                nn.Linear(3 * nn.N_MELS, nn.N_CLASS), nn.Dropout(0.25), nn.Sigmoid()
+            )
+
+    def forward(self, mel):
+        # print(mel.shape)
+        mel = mel.transpose(-1, -2).unsqueeze(1)
+        x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
+        x = self.fc(x)
+        # print(x.shape)
+        return x
+
+
+from librosa.filters import mel
+
+
+class MelSpectrogram(torch.nn.Module):
+    def __init__(
+        self,
+        is_half,
+        n_mel_channels,
+        sampling_rate,
+        win_length,
+        hop_length,
+        n_fft=None,
+        mel_fmin=0,
+        mel_fmax=None,
+        clamp=1e-5,
+    ):
+        super().__init__()
+        n_fft = win_length if n_fft is None else n_fft
+        self.hann_window = {}
+        mel_basis = mel(
+            sr=sampling_rate,
+            n_fft=n_fft,
+            n_mels=n_mel_channels,
+            fmin=mel_fmin,
+            fmax=mel_fmax,
+            htk=True,
+        )
+        mel_basis = torch.from_numpy(mel_basis).float()
+        self.register_buffer("mel_basis", mel_basis)
+        self.n_fft = win_length if n_fft is None else n_fft
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.sampling_rate = sampling_rate
+        self.n_mel_channels = n_mel_channels
+        self.clamp = clamp
+        self.is_half = is_half
+
+    def forward(self, audio, keyshift=0, speed=1, center=True):
+        factor = 2 ** (keyshift / 12)
+        n_fft_new = int(np.round(self.n_fft * factor))
+        win_length_new = int(np.round(self.win_length * factor))
+        hop_length_new = int(np.round(self.hop_length * speed))
+        keyshift_key = str(keyshift) + "_" + str(audio.device)
+        if keyshift_key not in self.hann_window:
+            self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(
+                audio.device
+            )
+        if "privateuseone" in str(audio.device):
+            if not hasattr(self, "stft"):
+                self.stft = STFT(
+                    filter_length=n_fft_new,
+                    hop_length=hop_length_new,
+                    win_length=win_length_new,
+                    window="hann",
+                ).to(audio.device)
+            magnitude = self.stft.transform(audio)
+        else:
+            fft = torch.stft(
+                audio,
+                n_fft=n_fft_new,
+                hop_length=hop_length_new,
+                win_length=win_length_new,
+                window=self.hann_window[keyshift_key],
+                center=center,
+                return_complex=True,
+            )
+            magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
+        if keyshift != 0:
+            size = self.n_fft // 2 + 1
+            resize = magnitude.size(1)
+            if resize < size:
+                magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
+            magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
+        mel_output = torch.matmul(self.mel_basis, magnitude)
+        if self.is_half == True:
+            mel_output = mel_output.half()
+        log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
+        return log_mel_spec
+
+
+class RMVPE:
+    def __init__(self, model_path: str, is_half, device=None, use_jit=False):
+        self.resample_kernel = {}
+        self.resample_kernel = {}
+        self.is_half = is_half
+        if device is None:
+            device = "cuda:0" if torch.cuda.is_available() else "cpu"
+        self.device = device
+        self.mel_extractor = MelSpectrogram(
+            is_half, 128, 16000, 1024, 160, None, 30, 8000
+        ).to(device)
+        if "privateuseone" in str(device):
+            import onnxruntime as ort
+
+            ort_session = ort.InferenceSession(
+                "%s/rmvpe.onnx" % os.environ["rmvpe_root"],
+                providers=["DmlExecutionProvider"],
+            )
+            self.model = ort_session
+        else:
+            if str(self.device) == "cuda":
+                self.device = torch.device("cuda:0")
+
+            def get_jit_model():
+                jit_model_path = model_path.rstrip(".pth")
+                jit_model_path += ".half.jit" if is_half else ".jit"
+                reload = False
+                if os.path.exists(jit_model_path):
+                    ckpt = jit.load(jit_model_path)
+                    model_device = ckpt["device"]
+                    if model_device != str(self.device):
+                        reload = True
+                else:
+                    reload = True
+
+                if reload:
+                    ckpt = jit.rmvpe_jit_export(
+                        model_path=model_path,
+                        mode="script",
+                        inputs_path=None,
+                        save_path=jit_model_path,
+                        device=device,
+                        is_half=is_half,
+                    )
+                model = torch.jit.load(BytesIO(ckpt["model"]), map_location=device)
+                return model
+
+            def get_default_model():
+                model = E2E(4, 1, (2, 2))
+                ckpt = torch.load(model_path, map_location="cpu", weights_only=False)
+                model.load_state_dict(ckpt)
+                model.eval()
+                if is_half:
+                    model = model.half()
+                else:
+                    model = model.float()
+                return model
+
+            if use_jit:
+                if is_half and "cpu" in str(self.device):
+                    logger.warning(
+                        "Use default rmvpe model. \
+                                 Jit is not supported on the CPU for half floating point"
+                    )
+                    self.model = get_default_model()
+                else:
+                    self.model = get_jit_model()
+            else:
+                self.model = get_default_model()
+
+            self.model = self.model.to(device)
+        cents_mapping = 20 * np.arange(360) + 1997.3794084376191
+        self.cents_mapping = np.pad(cents_mapping, (4, 4))  # 368
+
+    def mel2hidden(self, mel):
+        with torch.no_grad():
+            n_frames = mel.shape[-1]
+            n_pad = 32 * ((n_frames - 1) // 32 + 1) - n_frames
+            if n_pad > 0:
+                mel = F.pad(mel, (0, n_pad), mode="constant")
+            if "privateuseone" in str(self.device):
+                onnx_input_name = self.model.get_inputs()[0].name
+                onnx_outputs_names = self.model.get_outputs()[0].name
+                hidden = self.model.run(
+                    [onnx_outputs_names],
+                    input_feed={onnx_input_name: mel.cpu().numpy()},
+                )[0]
+            else:
+                mel = mel.half() if self.is_half else mel.float()
+                hidden = self.model(mel)
+            return hidden[:, :n_frames]
+
+    def decode(self, hidden, thred=0.03):
+        cents_pred = self.to_local_average_cents(hidden, thred=thred)
+        f0 = 10 * (2 ** (cents_pred / 1200))
+        f0[f0 == 10] = 0
+        # f0 = np.array([10 * (2 ** (cent_pred / 1200)) if cent_pred else 0 for cent_pred in cents_pred])
+        return f0
+
+    def infer_from_audio(self, audio, thred=0.03):
+        # torch.cuda.synchronize()
+        # t0 = ttime()
+        if not torch.is_tensor(audio):
+            audio = torch.from_numpy(audio)
+        mel = self.mel_extractor(
+            audio.float().to(self.device).unsqueeze(0), center=True
+        )
+        # print(123123123,mel.device.type)
+        # torch.cuda.synchronize()
+        # t1 = ttime()
+        hidden = self.mel2hidden(mel)
+        # torch.cuda.synchronize()
+        # t2 = ttime()
+        # print(234234,hidden.device.type)
+        if "privateuseone" not in str(self.device):
+            hidden = hidden.squeeze(0).cpu().numpy()
+        else:
+            hidden = hidden[0]
+        if self.is_half == True:
+            hidden = hidden.astype("float32")
+
+        f0 = self.decode(hidden, thred=thred)
+        # torch.cuda.synchronize()
+        # t3 = ttime()
+        # print("hmvpe:%s\t%s\t%s\t%s"%(t1-t0,t2-t1,t3-t2,t3-t0))
+        return f0
+
+    def to_local_average_cents(self, salience, thred=0.05):
+        # t0 = ttime()
+        center = np.argmax(salience, axis=1)  # 帧长#index
+        salience = np.pad(salience, ((0, 0), (4, 4)))  # 帧长,368
+        # t1 = ttime()
+        center += 4
+        todo_salience = []
+        todo_cents_mapping = []
+        starts = center - 4
+        ends = center + 5
+        for idx in range(salience.shape[0]):
+            todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
+            todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
+        # t2 = ttime()
+        todo_salience = np.array(todo_salience)  # 帧长，9
+        todo_cents_mapping = np.array(todo_cents_mapping)  # 帧长，9
+        product_sum = np.sum(todo_salience * todo_cents_mapping, 1)
+        weight_sum = np.sum(todo_salience, 1)  # 帧长
+        devided = product_sum / weight_sum  # 帧长
+        # t3 = ttime()
+        maxx = np.max(salience, axis=1)  # 帧长
+        devided[maxx <= thred] = 0
+        # t4 = ttime()
+        # print("decode:%s\t%s\t%s\t%s" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
+        return devided
+
+
+if __name__ == "__main__":
+    import librosa
+    import soundfile as sf
+
+    audio, sampling_rate = sf.read(r"C:\Users\liujing04\Desktop\Z\冬之花clip1.wav")
+    if len(audio.shape) > 1:
+        audio = librosa.to_mono(audio.transpose(1, 0))
+    audio_bak = audio.copy()
+    if sampling_rate != 16000:
+        audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
+    model_path = r"D:\BaiduNetdiskDownload\RVC-beta-v2-0727AMD_realtime\rmvpe.pt"
+    thred = 0.03  # 0.01
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    rmvpe = RMVPE(model_path, is_half=False, device=device)
+    t0 = ttime()
+    f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    # f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    # f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    # f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    # f0 = rmvpe.infer_from_audio(audio, thred=thred)
+    t1 = ttime()
+    logger.info("%s %.2f", f0.shape, t1 - t0)

infer/lib/rtrvc.py

@@ -0,0 +1,461 @@
+from io import BytesIO
+import os
+import sys
+import traceback
+from infer.lib import jit
+from infer.lib.jit.get_synthesizer import get_synthesizer
+from time import time as ttime
+import fairseq
+import faiss
+import numpy as np
+import parselmouth
+import pyworld
+import scipy.signal as signal
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchcrepe
+from torchaudio.transforms import Resample
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+from multiprocessing import Manager as M
+
+from configs.config import Config
+
+# config = Config()
+
+mm = M()
+
+
+def printt(strr, *args):
+    if len(args) == 0:
+        print(strr)
+    else:
+        print(strr % args)
+
+
+# config.device=torch.device("cpu")########强制cpu测试
+# config.is_half=False########强制cpu测试
+class RVC:
+    def __init__(
+        self,
+        key,
+        formant,
+        pth_path,
+        index_path,
+        index_rate,
+        n_cpu,
+        inp_q,
+        opt_q,
+        config: Config,
+        last_rvc=None,
+    ) -> None:
+        """
+        初始化
+        """
+        try:
+            if config.dml == True:
+
+                def forward_dml(ctx, x, scale):
+                    ctx.scale = scale
+                    res = x.clone().detach()
+                    return res
+
+                fairseq.modules.grad_multiply.GradMultiply.forward = forward_dml
+            # global config
+            self.config = config
+            self.inp_q = inp_q
+            self.opt_q = opt_q
+            # device="cpu"########强制cpu测试
+            self.device = config.device
+            self.f0_up_key = key
+            self.formant_shift = formant
+            self.f0_min = 50
+            self.f0_max = 1100
+            self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+            self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+            self.n_cpu = n_cpu
+            self.use_jit = self.config.use_jit
+            self.is_half = config.is_half
+
+            if index_rate != 0:
+                self.index = faiss.read_index(index_path)
+                self.big_npy = self.index.reconstruct_n(0, self.index.ntotal)
+                printt("Index search enabled")
+            self.pth_path: str = pth_path
+            self.index_path = index_path
+            self.index_rate = index_rate
+            self.cache_pitch: torch.Tensor = torch.zeros(
+                1024, device=self.device, dtype=torch.long
+            )
+            self.cache_pitchf = torch.zeros(
+                1024, device=self.device, dtype=torch.float32
+            )
+
+            self.resample_kernel = {}
+
+            if last_rvc is None:
+                models, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+                    ["assets/hubert/hubert_base.pt"],
+                    suffix="",
+                )
+                hubert_model = models[0]
+                hubert_model = hubert_model.to(self.device)
+                if self.is_half:
+                    hubert_model = hubert_model.half()
+                else:
+                    hubert_model = hubert_model.float()
+                hubert_model.eval()
+                self.model = hubert_model
+            else:
+                self.model = last_rvc.model
+
+            self.net_g: nn.Module = None
+
+            def set_default_model():
+                self.net_g, cpt = get_synthesizer(self.pth_path, self.device)
+                self.tgt_sr = cpt["config"][-1]
+                cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+                self.if_f0 = cpt.get("f0", 1)
+                self.version = cpt.get("version", "v1")
+                if self.is_half:
+                    self.net_g = self.net_g.half()
+                else:
+                    self.net_g = self.net_g.float()
+
+            def set_jit_model():
+                jit_pth_path = self.pth_path.rstrip(".pth")
+                jit_pth_path += ".half.jit" if self.is_half else ".jit"
+                reload = False
+                if str(self.device) == "cuda":
+                    self.device = torch.device("cuda:0")
+                if os.path.exists(jit_pth_path):
+                    cpt = jit.load(jit_pth_path)
+                    model_device = cpt["device"]
+                    if model_device != str(self.device):
+                        reload = True
+                else:
+                    reload = True
+
+                if reload:
+                    cpt = jit.synthesizer_jit_export(
+                        self.pth_path,
+                        "script",
+                        None,
+                        device=self.device,
+                        is_half=self.is_half,
+                    )
+
+                self.tgt_sr = cpt["config"][-1]
+                self.if_f0 = cpt.get("f0", 1)
+                self.version = cpt.get("version", "v1")
+                self.net_g = torch.jit.load(
+                    BytesIO(cpt["model"]), map_location=self.device
+                )
+                self.net_g.infer = self.net_g.forward
+                self.net_g.eval().to(self.device)
+
+            def set_synthesizer():
+                if self.use_jit and not config.dml:
+                    if self.is_half and "cpu" in str(self.device):
+                        printt(
+                            "Use default Synthesizer model. \
+                                    Jit is not supported on the CPU for half floating point"
+                        )
+                        set_default_model()
+                    else:
+                        set_jit_model()
+                else:
+                    set_default_model()
+
+            if last_rvc is None or last_rvc.pth_path != self.pth_path:
+                set_synthesizer()
+            else:
+                self.tgt_sr = last_rvc.tgt_sr
+                self.if_f0 = last_rvc.if_f0
+                self.version = last_rvc.version
+                self.is_half = last_rvc.is_half
+                if last_rvc.use_jit != self.use_jit:
+                    set_synthesizer()
+                else:
+                    self.net_g = last_rvc.net_g
+
+            if last_rvc is not None and hasattr(last_rvc, "model_rmvpe"):
+                self.model_rmvpe = last_rvc.model_rmvpe
+            if last_rvc is not None and hasattr(last_rvc, "model_fcpe"):
+                self.device_fcpe = last_rvc.device_fcpe
+                self.model_fcpe = last_rvc.model_fcpe
+        except:
+            printt(traceback.format_exc())
+
+    def change_key(self, new_key):
+        self.f0_up_key = new_key
+
+    def change_formant(self, new_formant):
+        self.formant_shift = new_formant
+
+    def change_index_rate(self, new_index_rate):
+        if new_index_rate != 0 and self.index_rate == 0:
+            self.index = faiss.read_index(self.index_path)
+            self.big_npy = self.index.reconstruct_n(0, self.index.ntotal)
+            printt("Index search enabled")
+        self.index_rate = new_index_rate
+
+    def get_f0_post(self, f0):
+        if not torch.is_tensor(f0):
+            f0 = torch.from_numpy(f0)
+        f0 = f0.float().to(self.device).squeeze()
+        f0_mel = 1127 * torch.log(1 + f0 / 700)
+        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * 254 / (
+            self.f0_mel_max - self.f0_mel_min
+        ) + 1
+        f0_mel[f0_mel <= 1] = 1
+        f0_mel[f0_mel > 255] = 255
+        f0_coarse = torch.round(f0_mel).long()
+        return f0_coarse, f0
+
+    def get_f0(self, x, f0_up_key, n_cpu, method="harvest"):
+        n_cpu = int(n_cpu)
+        if method == "crepe":
+            return self.get_f0_crepe(x, f0_up_key)
+        if method == "rmvpe":
+            return self.get_f0_rmvpe(x, f0_up_key)
+        if method == "fcpe":
+            return self.get_f0_fcpe(x, f0_up_key)
+        x = x.cpu().numpy()
+        if method == "pm":
+            p_len = x.shape[0] // 160 + 1
+            f0_min = 65
+            l_pad = int(np.ceil(1.5 / f0_min * 16000))
+            r_pad = l_pad + 1
+            s = parselmouth.Sound(np.pad(x, (l_pad, r_pad)), 16000).to_pitch_ac(
+                time_step=0.01,
+                voicing_threshold=0.6,
+                pitch_floor=f0_min,
+                pitch_ceiling=1100,
+            )
+            assert np.abs(s.t1 - 1.5 / f0_min) < 0.001
+            f0 = s.selected_array["frequency"]
+            if len(f0) < p_len:
+                f0 = np.pad(f0, (0, p_len - len(f0)))
+            f0 = f0[:p_len]
+            f0 *= pow(2, f0_up_key / 12)
+            return self.get_f0_post(f0)
+        if n_cpu == 1:
+            f0, t = pyworld.harvest(
+                x.astype(np.double),
+                fs=16000,
+                f0_ceil=1100,
+                f0_floor=50,
+                frame_period=10,
+            )
+            f0 = signal.medfilt(f0, 3)
+            f0 *= pow(2, f0_up_key / 12)
+            return self.get_f0_post(f0)
+        f0bak = np.zeros(x.shape[0] // 160 + 1, dtype=np.float64)
+        length = len(x)
+        part_length = 160 * ((length // 160 - 1) // n_cpu + 1)
+        n_cpu = (length // 160 - 1) // (part_length // 160) + 1
+        ts = ttime()
+        res_f0 = mm.dict()
+        for idx in range(n_cpu):
+            tail = part_length * (idx + 1) + 320
+            if idx == 0:
+                self.inp_q.put((idx, x[:tail], res_f0, n_cpu, ts))
+            else:
+                self.inp_q.put(
+                    (idx, x[part_length * idx - 320 : tail], res_f0, n_cpu, ts)
+                )
+        while 1:
+            res_ts = self.opt_q.get()
+            if res_ts == ts:
+                break
+        f0s = [i[1] for i in sorted(res_f0.items(), key=lambda x: x[0])]
+        for idx, f0 in enumerate(f0s):
+            if idx == 0:
+                f0 = f0[:-3]
+            elif idx != n_cpu - 1:
+                f0 = f0[2:-3]
+            else:
+                f0 = f0[2:]
+            f0bak[part_length * idx // 160 : part_length * idx // 160 + f0.shape[0]] = (
+                f0
+            )
+        f0bak = signal.medfilt(f0bak, 3)
+        f0bak *= pow(2, f0_up_key / 12)
+        return self.get_f0_post(f0bak)
+
+    def get_f0_crepe(self, x, f0_up_key):
+        if "privateuseone" in str(
+            self.device
+        ):  ###不支持dml，cpu又太慢用不成，拿fcpe顶替
+            return self.get_f0(x, f0_up_key, 1, "fcpe")
+        # printt("using crepe,device:%s"%self.device)
+        f0, pd = torchcrepe.predict(
+            x.unsqueeze(0).float(),
+            16000,
+            160,
+            self.f0_min,
+            self.f0_max,
+            "full",
+            batch_size=512,
+            # device=self.device if self.device.type!="privateuseone" else "cpu",###crepe不用半精度全部是全精度所以不愁###cpu延迟高到没法用
+            device=self.device,
+            return_periodicity=True,
+        )
+        pd = torchcrepe.filter.median(pd, 3)
+        f0 = torchcrepe.filter.mean(f0, 3)
+        f0[pd < 0.1] = 0
+        f0 *= pow(2, f0_up_key / 12)
+        return self.get_f0_post(f0)
+
+    def get_f0_rmvpe(self, x, f0_up_key):
+        if hasattr(self, "model_rmvpe") == False:
+            from infer.lib.rmvpe import RMVPE
+
+            printt("Loading rmvpe model")
+            self.model_rmvpe = RMVPE(
+                "assets/rmvpe/rmvpe.pt",
+                is_half=self.is_half,
+                device=self.device,
+                use_jit=self.config.use_jit,
+            )
+        f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+        f0 *= pow(2, f0_up_key / 12)
+        return self.get_f0_post(f0)
+
+    def get_f0_fcpe(self, x, f0_up_key):
+        if hasattr(self, "model_fcpe") == False:
+            from torchfcpe import spawn_bundled_infer_model
+
+            printt("Loading fcpe model")
+            if "privateuseone" in str(self.device):
+                self.device_fcpe = "cpu"
+            else:
+                self.device_fcpe = self.device
+            self.model_fcpe = spawn_bundled_infer_model(self.device_fcpe)
+        f0 = self.model_fcpe.infer(
+            x.to(self.device_fcpe).unsqueeze(0).float(),
+            sr=16000,
+            decoder_mode="local_argmax",
+            threshold=0.006,
+        )
+        f0 *= pow(2, f0_up_key / 12)
+        return self.get_f0_post(f0)
+
+    def infer(
+        self,
+        input_wav: torch.Tensor,
+        block_frame_16k,
+        skip_head,
+        return_length,
+        f0method,
+    ) -> np.ndarray:
+        t1 = ttime()
+        with torch.no_grad():
+            if self.config.is_half:
+                feats = input_wav.half().view(1, -1)
+            else:
+                feats = input_wav.float().view(1, -1)
+            padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
+            inputs = {
+                "source": feats,
+                "padding_mask": padding_mask,
+                "output_layer": 9 if self.version == "v1" else 12,
+            }
+            logits = self.model.extract_features(**inputs)
+            feats = (
+                self.model.final_proj(logits[0]) if self.version == "v1" else logits[0]
+            )
+            feats = torch.cat((feats, feats[:, -1:, :]), 1)
+        t2 = ttime()
+        try:
+            if hasattr(self, "index") and self.index_rate != 0:
+                npy = feats[0][skip_head // 2 :].cpu().numpy().astype("float32")
+                score, ix = self.index.search(npy, k=8)
+                if (ix >= 0).all():
+                    weight = np.square(1 / score)
+                    weight /= weight.sum(axis=1, keepdims=True)
+                    npy = np.sum(
+                        self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1
+                    )
+                    if self.config.is_half:
+                        npy = npy.astype("float16")
+                    feats[0][skip_head // 2 :] = (
+                        torch.from_numpy(npy).unsqueeze(0).to(self.device)
+                        * self.index_rate
+                        + (1 - self.index_rate) * feats[0][skip_head // 2 :]
+                    )
+                else:
+                    printt(
+                        "Invalid index. You MUST use added_xxxx.index but not trained_xxxx.index!"
+                    )
+            else:
+                printt("Index search FAILED or disabled")
+        except:
+            traceback.print_exc()
+            printt("Index search FAILED")
+        t3 = ttime()
+        p_len = input_wav.shape[0] // 160
+        factor = pow(2, self.formant_shift / 12)
+        return_length2 = int(np.ceil(return_length * factor))
+        if self.if_f0 == 1:
+            f0_extractor_frame = block_frame_16k + 800
+            if f0method == "rmvpe":
+                f0_extractor_frame = 5120 * ((f0_extractor_frame - 1) // 5120 + 1) - 160
+            pitch, pitchf = self.get_f0(
+                input_wav[-f0_extractor_frame:], self.f0_up_key - self.formant_shift, self.n_cpu, f0method
+            )
+            shift = block_frame_16k // 160
+            self.cache_pitch[:-shift] = self.cache_pitch[shift:].clone()
+            self.cache_pitchf[:-shift] = self.cache_pitchf[shift:].clone()
+            self.cache_pitch[4 - pitch.shape[0] :] = pitch[3:-1]
+            self.cache_pitchf[4 - pitch.shape[0] :] = pitchf[3:-1]
+            cache_pitch = self.cache_pitch[None, -p_len:]
+            cache_pitchf = self.cache_pitchf[None, -p_len:] * return_length2 / return_length
+        t4 = ttime()
+        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+        feats = feats[:, :p_len, :]
+        p_len = torch.LongTensor([p_len]).to(self.device)
+        sid = torch.LongTensor([0]).to(self.device)
+        skip_head = torch.LongTensor([skip_head])
+        return_length2 = torch.LongTensor([return_length2])
+        return_length = torch.LongTensor([return_length])
+        with torch.no_grad():
+            if self.if_f0 == 1:
+                infered_audio, _, _ = self.net_g.infer(
+                    feats,
+                    p_len,
+                    cache_pitch,
+                    cache_pitchf,
+                    sid,
+                    skip_head,
+                    return_length,
+                    return_length2,
+                )
+            else:
+                infered_audio, _, _ = self.net_g.infer(
+                    feats, p_len, sid, skip_head, return_length, return_length2
+                )
+        infered_audio = infered_audio.squeeze(1).float()
+        upp_res = int(np.floor(factor * self.tgt_sr // 100))
+        if upp_res != self.tgt_sr // 100:
+            if upp_res not in self.resample_kernel:
+                self.resample_kernel[upp_res] = Resample(
+                    orig_freq=upp_res,
+                    new_freq=self.tgt_sr // 100,
+                    dtype=torch.float32,
+                ).to(self.device)
+            infered_audio = self.resample_kernel[upp_res](
+                infered_audio[:, : return_length * upp_res]
+            )
+        t5 = ttime()
+        printt(
+            "Spent time: fea = %.3fs, index = %.3fs, f0 = %.3fs, model = %.3fs",
+            t2 - t1,
+            t3 - t2,
+            t4 - t3,
+            t5 - t4,
+        )
+        return infered_audio.squeeze()

infer/lib/slicer2.py

@@ -0,0 +1,260 @@
+import numpy as np
+
+
+# This function is obtained from librosa.
+def get_rms(
+    y,
+    frame_length=2048,
+    hop_length=512,
+    pad_mode="constant",
+):
+    padding = (int(frame_length // 2), int(frame_length // 2))
+    y = np.pad(y, padding, mode=pad_mode)
+
+    axis = -1
+    # put our new within-frame axis at the end for now
+    out_strides = y.strides + tuple([y.strides[axis]])
+    # Reduce the shape on the framing axis
+    x_shape_trimmed = list(y.shape)
+    x_shape_trimmed[axis] -= frame_length - 1
+    out_shape = tuple(x_shape_trimmed) + tuple([frame_length])
+    xw = np.lib.stride_tricks.as_strided(y, shape=out_shape, strides=out_strides)
+    if axis < 0:
+        target_axis = axis - 1
+    else:
+        target_axis = axis + 1
+    xw = np.moveaxis(xw, -1, target_axis)
+    # Downsample along the target axis
+    slices = [slice(None)] * xw.ndim
+    slices[axis] = slice(0, None, hop_length)
+    x = xw[tuple(slices)]
+
+    # Calculate power
+    power = np.mean(np.abs(x) ** 2, axis=-2, keepdims=True)
+
+    return np.sqrt(power)
+
+
+class Slicer:
+    def __init__(
+        self,
+        sr: int,
+        threshold: float = -40.0,
+        min_length: int = 5000,
+        min_interval: int = 300,
+        hop_size: int = 20,
+        max_sil_kept: int = 5000,
+    ):
+        if not min_length >= min_interval >= hop_size:
+            raise ValueError(
+                "The following condition must be satisfied: min_length >= min_interval >= hop_size"
+            )
+        if not max_sil_kept >= hop_size:
+            raise ValueError(
+                "The following condition must be satisfied: max_sil_kept >= hop_size"
+            )
+        min_interval = sr * min_interval / 1000
+        self.threshold = 10 ** (threshold / 20.0)
+        self.hop_size = round(sr * hop_size / 1000)
+        self.win_size = min(round(min_interval), 4 * self.hop_size)
+        self.min_length = round(sr * min_length / 1000 / self.hop_size)
+        self.min_interval = round(min_interval / self.hop_size)
+        self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
+
+    def _apply_slice(self, waveform, begin, end):
+        if len(waveform.shape) > 1:
+            return waveform[
+                :, begin * self.hop_size : min(waveform.shape[1], end * self.hop_size)
+            ]
+        else:
+            return waveform[
+                begin * self.hop_size : min(waveform.shape[0], end * self.hop_size)
+            ]
+
+    # @timeit
+    def slice(self, waveform):
+        if len(waveform.shape) > 1:
+            samples = waveform.mean(axis=0)
+        else:
+            samples = waveform
+        if samples.shape[0] <= self.min_length:
+            return [waveform]
+        rms_list = get_rms(
+            y=samples, frame_length=self.win_size, hop_length=self.hop_size
+        ).squeeze(0)
+        sil_tags = []
+        silence_start = None
+        clip_start = 0
+        for i, rms in enumerate(rms_list):
+            # Keep looping while frame is silent.
+            if rms < self.threshold:
+                # Record start of silent frames.
+                if silence_start is None:
+                    silence_start = i
+                continue
+            # Keep looping while frame is not silent and silence start has not been recorded.
+            if silence_start is None:
+                continue
+            # Clear recorded silence start if interval is not enough or clip is too short
+            is_leading_silence = silence_start == 0 and i > self.max_sil_kept
+            need_slice_middle = (
+                i - silence_start >= self.min_interval
+                and i - clip_start >= self.min_length
+            )
+            if not is_leading_silence and not need_slice_middle:
+                silence_start = None
+                continue
+            # Need slicing. Record the range of silent frames to be removed.
+            if i - silence_start <= self.max_sil_kept:
+                pos = rms_list[silence_start : i + 1].argmin() + silence_start
+                if silence_start == 0:
+                    sil_tags.append((0, pos))
+                else:
+                    sil_tags.append((pos, pos))
+                clip_start = pos
+            elif i - silence_start <= self.max_sil_kept * 2:
+                pos = rms_list[
+                    i - self.max_sil_kept : silence_start + self.max_sil_kept + 1
+                ].argmin()
+                pos += i - self.max_sil_kept
+                pos_l = (
+                    rms_list[
+                        silence_start : silence_start + self.max_sil_kept + 1
+                    ].argmin()
+                    + silence_start
+                )
+                pos_r = (
+                    rms_list[i - self.max_sil_kept : i + 1].argmin()
+                    + i
+                    - self.max_sil_kept
+                )
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                    clip_start = pos_r
+                else:
+                    sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
+                    clip_start = max(pos_r, pos)
+            else:
+                pos_l = (
+                    rms_list[
+                        silence_start : silence_start + self.max_sil_kept + 1
+                    ].argmin()
+                    + silence_start
+                )
+                pos_r = (
+                    rms_list[i - self.max_sil_kept : i + 1].argmin()
+                    + i
+                    - self.max_sil_kept
+                )
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                else:
+                    sil_tags.append((pos_l, pos_r))
+                clip_start = pos_r
+            silence_start = None
+        # Deal with trailing silence.
+        total_frames = rms_list.shape[0]
+        if (
+            silence_start is not None
+            and total_frames - silence_start >= self.min_interval
+        ):
+            silence_end = min(total_frames, silence_start + self.max_sil_kept)
+            pos = rms_list[silence_start : silence_end + 1].argmin() + silence_start
+            sil_tags.append((pos, total_frames + 1))
+        # Apply and return slices.
+        if len(sil_tags) == 0:
+            return [waveform]
+        else:
+            chunks = []
+            if sil_tags[0][0] > 0:
+                chunks.append(self._apply_slice(waveform, 0, sil_tags[0][0]))
+            for i in range(len(sil_tags) - 1):
+                chunks.append(
+                    self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0])
+                )
+            if sil_tags[-1][1] < total_frames:
+                chunks.append(
+                    self._apply_slice(waveform, sil_tags[-1][1], total_frames)
+                )
+            return chunks
+
+
+def main():
+    import os.path
+    from argparse import ArgumentParser
+
+    import librosa
+    import soundfile
+
+    parser = ArgumentParser()
+    parser.add_argument("audio", type=str, help="The audio to be sliced")
+    parser.add_argument(
+        "--out", type=str, help="Output directory of the sliced audio clips"
+    )
+    parser.add_argument(
+        "--db_thresh",
+        type=float,
+        required=False,
+        default=-40,
+        help="The dB threshold for silence detection",
+    )
+    parser.add_argument(
+        "--min_length",
+        type=int,
+        required=False,
+        default=5000,
+        help="The minimum milliseconds required for each sliced audio clip",
+    )
+    parser.add_argument(
+        "--min_interval",
+        type=int,
+        required=False,
+        default=300,
+        help="The minimum milliseconds for a silence part to be sliced",
+    )
+    parser.add_argument(
+        "--hop_size",
+        type=int,
+        required=False,
+        default=10,
+        help="Frame length in milliseconds",
+    )
+    parser.add_argument(
+        "--max_sil_kept",
+        type=int,
+        required=False,
+        default=500,
+        help="The maximum silence length kept around the sliced clip, presented in milliseconds",
+    )
+    args = parser.parse_args()
+    out = args.out
+    if out is None:
+        out = os.path.dirname(os.path.abspath(args.audio))
+    audio, sr = librosa.load(args.audio, sr=None, mono=False)
+    slicer = Slicer(
+        sr=sr,
+        threshold=args.db_thresh,
+        min_length=args.min_length,
+        min_interval=args.min_interval,
+        hop_size=args.hop_size,
+        max_sil_kept=args.max_sil_kept,
+    )
+    chunks = slicer.slice(audio)
+    if not os.path.exists(out):
+        os.makedirs(out)
+    for i, chunk in enumerate(chunks):
+        if len(chunk.shape) > 1:
+            chunk = chunk.T
+        soundfile.write(
+            os.path.join(
+                out,
+                f"%s_%d.wav"
+                % (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),
+            ),
+            chunk,
+            sr,
+        )
+
+
+if __name__ == "__main__":
+    main()

infer/lib/train/data_utils.py

@@ -0,0 +1,517 @@
+import os
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import torch
+import torch.utils.data
+
+from infer.lib.train.mel_processing import spectrogram_torch
+from infer.lib.train.utils import load_filepaths_and_text, load_wav_to_torch
+
+
+class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
+    """
+    1) loads audio, text pairs
+    2) normalizes text and converts them to sequences of integers
+    3) computes spectrograms from audio files.
+    """
+
+    def __init__(self, audiopaths_and_text, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.filter_length = hparams.filter_length
+        self.hop_length = hparams.hop_length
+        self.win_length = hparams.win_length
+        self.sampling_rate = hparams.sampling_rate
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 5000)
+        self._filter()
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+        audiopaths_and_text_new = []
+        lengths = []
+        for audiopath, text, pitch, pitchf, dv in self.audiopaths_and_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_and_text_new.append([audiopath, text, pitch, pitchf, dv])
+                lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
+        self.audiopaths_and_text = audiopaths_and_text_new
+        self.lengths = lengths
+
+    def get_sid(self, sid):
+        sid = torch.LongTensor([int(sid)])
+        return sid
+
+    def get_audio_text_pair(self, audiopath_and_text):
+        # separate filename and text
+        file = audiopath_and_text[0]
+        phone = audiopath_and_text[1]
+        pitch = audiopath_and_text[2]
+        pitchf = audiopath_and_text[3]
+        dv = audiopath_and_text[4]
+
+        phone, pitch, pitchf = self.get_labels(phone, pitch, pitchf)
+        spec, wav = self.get_audio(file)
+        dv = self.get_sid(dv)
+
+        len_phone = phone.size()[0]
+        len_spec = spec.size()[-1]
+        # print(123,phone.shape,pitch.shape,spec.shape)
+        if len_phone != len_spec:
+            len_min = min(len_phone, len_spec)
+            # amor
+            len_wav = len_min * self.hop_length
+
+            spec = spec[:, :len_min]
+            wav = wav[:, :len_wav]
+
+            phone = phone[:len_min, :]
+            pitch = pitch[:len_min]
+            pitchf = pitchf[:len_min]
+
+        return (spec, wav, phone, pitch, pitchf, dv)
+
+    def get_labels(self, phone, pitch, pitchf):
+        phone = np.load(phone)
+        phone = np.repeat(phone, 2, axis=0)
+        pitch = np.load(pitch)
+        pitchf = np.load(pitchf)
+        n_num = min(phone.shape[0], 900)  # DistributedBucketSampler
+        # print(234,phone.shape,pitch.shape)
+        phone = phone[:n_num, :]
+        pitch = pitch[:n_num]
+        pitchf = pitchf[:n_num]
+        phone = torch.FloatTensor(phone)
+        pitch = torch.LongTensor(pitch)
+        pitchf = torch.FloatTensor(pitchf)
+        return phone, pitch, pitchf
+
+    def get_audio(self, filename):
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError(
+                "{} SR doesn't match target {} SR".format(
+                    sampling_rate, self.sampling_rate
+                )
+            )
+        audio_norm = audio
+        #        audio_norm = audio / self.max_wav_value
+        #        audio_norm = audio / np.abs(audio).max()
+
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if os.path.exists(spec_filename):
+            try:
+                spec = torch.load(spec_filename, weights_only=False)
+            except:
+                logger.warning("%s %s", spec_filename, traceback.format_exc())
+                spec = spectrogram_torch(
+                    audio_norm,
+                    self.filter_length,
+                    self.sampling_rate,
+                    self.hop_length,
+                    self.win_length,
+                    center=False,
+                )
+                spec = torch.squeeze(spec, 0)
+                torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        else:
+            spec = spectrogram_torch(
+                audio_norm,
+                self.filter_length,
+                self.sampling_rate,
+                self.hop_length,
+                self.win_length,
+                center=False,
+            )
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        return spec, audio_norm
+
+    def __getitem__(self, index):
+        return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextAudioCollateMultiNSFsid:
+    """Zero-pads model inputs and targets"""
+
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text and aduio
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True
+        )
+
+        max_spec_len = max([x[0].size(1) for x in batch])
+        max_wave_len = max([x[1].size(1) for x in batch])
+        spec_lengths = torch.LongTensor(len(batch))
+        wave_lengths = torch.LongTensor(len(batch))
+        spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
+        wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
+        spec_padded.zero_()
+        wave_padded.zero_()
+
+        max_phone_len = max([x[2].size(0) for x in batch])
+        phone_lengths = torch.LongTensor(len(batch))
+        phone_padded = torch.FloatTensor(
+            len(batch), max_phone_len, batch[0][2].shape[1]
+        )  # (spec, wav, phone, pitch)
+        pitch_padded = torch.LongTensor(len(batch), max_phone_len)
+        pitchf_padded = torch.FloatTensor(len(batch), max_phone_len)
+        phone_padded.zero_()
+        pitch_padded.zero_()
+        pitchf_padded.zero_()
+        # dv = torch.FloatTensor(len(batch), 256)#gin=256
+        sid = torch.LongTensor(len(batch))
+
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            spec = row[0]
+            spec_padded[i, :, : spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wave = row[1]
+            wave_padded[i, :, : wave.size(1)] = wave
+            wave_lengths[i] = wave.size(1)
+
+            phone = row[2]
+            phone_padded[i, : phone.size(0), :] = phone
+            phone_lengths[i] = phone.size(0)
+
+            pitch = row[3]
+            pitch_padded[i, : pitch.size(0)] = pitch
+            pitchf = row[4]
+            pitchf_padded[i, : pitchf.size(0)] = pitchf
+
+            # dv[i] = row[5]
+            sid[i] = row[5]
+
+        return (
+            phone_padded,
+            phone_lengths,
+            pitch_padded,
+            pitchf_padded,
+            spec_padded,
+            spec_lengths,
+            wave_padded,
+            wave_lengths,
+            # dv
+            sid,
+        )
+
+
+class TextAudioLoader(torch.utils.data.Dataset):
+    """
+    1) loads audio, text pairs
+    2) normalizes text and converts them to sequences of integers
+    3) computes spectrograms from audio files.
+    """
+
+    def __init__(self, audiopaths_and_text, hparams):
+        self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+        self.max_wav_value = hparams.max_wav_value
+        self.sampling_rate = hparams.sampling_rate
+        self.filter_length = hparams.filter_length
+        self.hop_length = hparams.hop_length
+        self.win_length = hparams.win_length
+        self.sampling_rate = hparams.sampling_rate
+        self.min_text_len = getattr(hparams, "min_text_len", 1)
+        self.max_text_len = getattr(hparams, "max_text_len", 5000)
+        self._filter()
+
+    def _filter(self):
+        """
+        Filter text & store spec lengths
+        """
+        # Store spectrogram lengths for Bucketing
+        # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+        # spec_length = wav_length // hop_length
+        audiopaths_and_text_new = []
+        lengths = []
+        for audiopath, text, dv in self.audiopaths_and_text:
+            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+                audiopaths_and_text_new.append([audiopath, text, dv])
+                lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
+        self.audiopaths_and_text = audiopaths_and_text_new
+        self.lengths = lengths
+
+    def get_sid(self, sid):
+        sid = torch.LongTensor([int(sid)])
+        return sid
+
+    def get_audio_text_pair(self, audiopath_and_text):
+        # separate filename and text
+        file = audiopath_and_text[0]
+        phone = audiopath_and_text[1]
+        dv = audiopath_and_text[2]
+
+        phone = self.get_labels(phone)
+        spec, wav = self.get_audio(file)
+        dv = self.get_sid(dv)
+
+        len_phone = phone.size()[0]
+        len_spec = spec.size()[-1]
+        if len_phone != len_spec:
+            len_min = min(len_phone, len_spec)
+            len_wav = len_min * self.hop_length
+            spec = spec[:, :len_min]
+            wav = wav[:, :len_wav]
+            phone = phone[:len_min, :]
+        return (spec, wav, phone, dv)
+
+    def get_labels(self, phone):
+        phone = np.load(phone)
+        phone = np.repeat(phone, 2, axis=0)
+        n_num = min(phone.shape[0], 900)  # DistributedBucketSampler
+        phone = phone[:n_num, :]
+        phone = torch.FloatTensor(phone)
+        return phone
+
+    def get_audio(self, filename):
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError(
+                "{} SR doesn't match target {} SR".format(
+                    sampling_rate, self.sampling_rate
+                )
+            )
+        audio_norm = audio
+        #        audio_norm = audio / self.max_wav_value
+        #        audio_norm = audio / np.abs(audio).max()
+
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pt")
+        if os.path.exists(spec_filename):
+            try:
+                spec = torch.load(spec_filename, weights_only=False)
+            except:
+                logger.warning("%s %s", spec_filename, traceback.format_exc())
+                spec = spectrogram_torch(
+                    audio_norm,
+                    self.filter_length,
+                    self.sampling_rate,
+                    self.hop_length,
+                    self.win_length,
+                    center=False,
+                )
+                spec = torch.squeeze(spec, 0)
+                torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        else:
+            spec = spectrogram_torch(
+                audio_norm,
+                self.filter_length,
+                self.sampling_rate,
+                self.hop_length,
+                self.win_length,
+                center=False,
+            )
+            spec = torch.squeeze(spec, 0)
+            torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+        return spec, audio_norm
+
+    def __getitem__(self, index):
+        return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+    def __len__(self):
+        return len(self.audiopaths_and_text)
+
+
+class TextAudioCollate:
+    """Zero-pads model inputs and targets"""
+
+    def __init__(self, return_ids=False):
+        self.return_ids = return_ids
+
+    def __call__(self, batch):
+        """Collate's training batch from normalized text and aduio
+        PARAMS
+        ------
+        batch: [text_normalized, spec_normalized, wav_normalized]
+        """
+        # Right zero-pad all one-hot text sequences to max input length
+        _, ids_sorted_decreasing = torch.sort(
+            torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True
+        )
+
+        max_spec_len = max([x[0].size(1) for x in batch])
+        max_wave_len = max([x[1].size(1) for x in batch])
+        spec_lengths = torch.LongTensor(len(batch))
+        wave_lengths = torch.LongTensor(len(batch))
+        spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
+        wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
+        spec_padded.zero_()
+        wave_padded.zero_()
+
+        max_phone_len = max([x[2].size(0) for x in batch])
+        phone_lengths = torch.LongTensor(len(batch))
+        phone_padded = torch.FloatTensor(
+            len(batch), max_phone_len, batch[0][2].shape[1]
+        )
+        phone_padded.zero_()
+        sid = torch.LongTensor(len(batch))
+
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            spec = row[0]
+            spec_padded[i, :, : spec.size(1)] = spec
+            spec_lengths[i] = spec.size(1)
+
+            wave = row[1]
+            wave_padded[i, :, : wave.size(1)] = wave
+            wave_lengths[i] = wave.size(1)
+
+            phone = row[2]
+            phone_padded[i, : phone.size(0), :] = phone
+            phone_lengths[i] = phone.size(0)
+
+            sid[i] = row[3]
+
+        return (
+            phone_padded,
+            phone_lengths,
+            spec_padded,
+            spec_lengths,
+            wave_padded,
+            wave_lengths,
+            sid,
+        )
+
+
+class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
+    """
+    Maintain similar input lengths in a batch.
+    Length groups are specified by boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
+
+    It removes samples which are not included in the boundaries.
+    Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
+    """
+
+    def __init__(
+        self,
+        dataset,
+        batch_size,
+        boundaries,
+        num_replicas=None,
+        rank=None,
+        shuffle=True,
+    ):
+        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
+        self.lengths = dataset.lengths
+        self.batch_size = batch_size
+        self.boundaries = boundaries
+
+        self.buckets, self.num_samples_per_bucket = self._create_buckets()
+        self.total_size = sum(self.num_samples_per_bucket)
+        self.num_samples = self.total_size // self.num_replicas
+
+    def _create_buckets(self):
+        buckets = [[] for _ in range(len(self.boundaries) - 1)]
+        for i in range(len(self.lengths)):
+            length = self.lengths[i]
+            idx_bucket = self._bisect(length)
+            if idx_bucket != -1:
+                buckets[idx_bucket].append(i)
+
+        for i in range(len(buckets) - 1, -1, -1):  #
+            if len(buckets[i]) == 0:
+                buckets.pop(i)
+                self.boundaries.pop(i + 1)
+
+        num_samples_per_bucket = []
+        for i in range(len(buckets)):
+            len_bucket = len(buckets[i])
+            total_batch_size = self.num_replicas * self.batch_size
+            rem = (
+                total_batch_size - (len_bucket % total_batch_size)
+            ) % total_batch_size
+            num_samples_per_bucket.append(len_bucket + rem)
+        return buckets, num_samples_per_bucket
+
+    def __iter__(self):
+        # deterministically shuffle based on epoch
+        g = torch.Generator()
+        g.manual_seed(self.epoch)
+
+        indices = []
+        if self.shuffle:
+            for bucket in self.buckets:
+                indices.append(torch.randperm(len(bucket), generator=g).tolist())
+        else:
+            for bucket in self.buckets:
+                indices.append(list(range(len(bucket))))
+
+        batches = []
+        for i in range(len(self.buckets)):
+            bucket = self.buckets[i]
+            len_bucket = len(bucket)
+            ids_bucket = indices[i]
+            num_samples_bucket = self.num_samples_per_bucket[i]
+
+            # add extra samples to make it evenly divisible
+            rem = num_samples_bucket - len_bucket
+            ids_bucket = (
+                ids_bucket
+                + ids_bucket * (rem // len_bucket)
+                + ids_bucket[: (rem % len_bucket)]
+            )
+
+            # subsample
+            ids_bucket = ids_bucket[self.rank :: self.num_replicas]
+
+            # batching
+            for j in range(len(ids_bucket) // self.batch_size):
+                batch = [
+                    bucket[idx]
+                    for idx in ids_bucket[
+                        j * self.batch_size : (j + 1) * self.batch_size
+                    ]
+                ]
+                batches.append(batch)
+
+        if self.shuffle:
+            batch_ids = torch.randperm(len(batches), generator=g).tolist()
+            batches = [batches[i] for i in batch_ids]
+        self.batches = batches
+
+        assert len(self.batches) * self.batch_size == self.num_samples
+        return iter(self.batches)
+
+    def _bisect(self, x, lo=0, hi=None):
+        if hi is None:
+            hi = len(self.boundaries) - 1
+
+        if hi > lo:
+            mid = (hi + lo) // 2
+            if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]:
+                return mid
+            elif x <= self.boundaries[mid]:
+                return self._bisect(x, lo, mid)
+            else:
+                return self._bisect(x, mid + 1, hi)
+        else:
+            return -1
+
+    def __len__(self):
+        return self.num_samples // self.batch_size

infer/lib/train/losses.py

@@ -0,0 +1,58 @@
+import torch
+
+
+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):
+            rl = rl.float().detach()
+            gl = gl.float()
+            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):
+        dr = dr.float()
+        dg = dg.float()
+        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:
+        dg = dg.float()
+        l = torch.mean((1 - dg) ** 2)
+        gen_losses.append(l)
+        loss += l
+
+    return loss, gen_losses
+
+
+def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
+    """
+    z_p, logs_q: [b, h, t_t]
+    m_p, logs_p: [b, h, t_t]
+    """
+    z_p = z_p.float()
+    logs_q = logs_q.float()
+    m_p = m_p.float()
+    logs_p = logs_p.float()
+    z_mask = z_mask.float()
+
+    kl = logs_p - logs_q - 0.5
+    kl += 0.5 * ((z_p - m_p) ** 2) * torch.exp(-2.0 * logs_p)
+    kl = torch.sum(kl * z_mask)
+    l = kl / torch.sum(z_mask)
+    return l

infer/lib/train/mel_processing.py

@@ -0,0 +1,127 @@
+import torch
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+import logging
+
+logger = logging.getLogger(__name__)
+
+MAX_WAV_VALUE = 32768.0
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    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)
+
+
+# Reusable banks
+mel_basis = {}
+hann_window = {}
+
+
+def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
+    """Convert waveform into Linear-frequency Linear-amplitude spectrogram.
+
+    Args:
+        y             :: (B, T) - Audio waveforms
+        n_fft
+        sampling_rate
+        hop_size
+        win_size
+        center
+    Returns:
+        :: (B, Freq, Frame) - Linear-frequency Linear-amplitude spectrogram
+    """
+
+    # Window - Cache if needed
+    global hann_window
+    dtype_device = str(y.dtype) + "_" + str(y.device)
+    wnsize_dtype_device = str(win_size) + "_" + dtype_device
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(
+            dtype=y.dtype, device=y.device
+        )
+
+    # Padding
+    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)
+
+    # Complex Spectrogram :: (B, T) -> (B, Freq, Frame, RealComplex=2)
+    spec = torch.stft(
+        y,
+        n_fft,
+        hop_length=hop_size,
+        win_length=win_size,
+        window=hann_window[wnsize_dtype_device],
+        center=center,
+        pad_mode="reflect",
+        normalized=False,
+        onesided=True,
+        return_complex=True,
+    )
+
+    # Linear-frequency Linear-amplitude spectrogram :: (B, Freq, Frame, RealComplex=2) -> (B, Freq, Frame)
+    spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + 1e-6)
+    return spec
+
+
+def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
+    # MelBasis - Cache if needed
+    global mel_basis
+    dtype_device = str(spec.dtype) + "_" + str(spec.device)
+    fmax_dtype_device = str(fmax) + "_" + dtype_device
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(
+            sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax
+        )
+        mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(
+            dtype=spec.dtype, device=spec.device
+        )
+
+    # Mel-frequency Log-amplitude spectrogram :: (B, Freq=num_mels, Frame)
+    melspec = torch.matmul(mel_basis[fmax_dtype_device], spec)
+    melspec = spectral_normalize_torch(melspec)
+    return melspec
+
+
+def mel_spectrogram_torch(
+    y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False
+):
+    """Convert waveform into Mel-frequency Log-amplitude spectrogram.
+
+    Args:
+        y       :: (B, T)           - Waveforms
+    Returns:
+        melspec :: (B, Freq, Frame) - Mel-frequency Log-amplitude spectrogram
+    """
+    # Linear-frequency Linear-amplitude spectrogram :: (B, T) -> (B, Freq, Frame)
+    spec = spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center)
+
+    # Mel-frequency Log-amplitude spectrogram :: (B, Freq, Frame) -> (B, Freq=num_mels, Frame)
+    melspec = spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax)
+
+    return melspec

infer/lib/train/process_ckpt.py

@@ -0,0 +1,261 @@
+import os
+import sys
+import traceback
+from collections import OrderedDict
+
+import torch
+
+from i18n.i18n import I18nAuto
+
+i18n = I18nAuto()
+
+
+def savee(ckpt, sr, if_f0, name, epoch, version, hps):
+    try:
+        opt = OrderedDict()
+        opt["weight"] = {}
+        for key in ckpt.keys():
+            if "enc_q" in key:
+                continue
+            opt["weight"][key] = ckpt[key].half()
+        opt["config"] = [
+            hps.data.filter_length // 2 + 1,
+            32,
+            hps.model.inter_channels,
+            hps.model.hidden_channels,
+            hps.model.filter_channels,
+            hps.model.n_heads,
+            hps.model.n_layers,
+            hps.model.kernel_size,
+            hps.model.p_dropout,
+            hps.model.resblock,
+            hps.model.resblock_kernel_sizes,
+            hps.model.resblock_dilation_sizes,
+            hps.model.upsample_rates,
+            hps.model.upsample_initial_channel,
+            hps.model.upsample_kernel_sizes,
+            hps.model.spk_embed_dim,
+            hps.model.gin_channels,
+            hps.data.sampling_rate,
+        ]
+        opt["info"] = "%sepoch" % epoch
+        opt["sr"] = sr
+        opt["f0"] = if_f0
+        opt["version"] = version
+        torch.save(opt, "assets/weights/%s.pth" % name)
+        return "Success."
+    except:
+        return traceback.format_exc()
+
+
+def show_info(path):
+    try:
+        a = torch.load(path, map_location="cpu", weights_only=False)
+        return "模型信息:%s\n采样率:%s\n模型是否输入音高引导:%s\n版本:%s" % (
+            a.get("info", "None"),
+            a.get("sr", "None"),
+            a.get("f0", "None"),
+            a.get("version", "None"),
+        )
+    except:
+        return traceback.format_exc()
+
+
+def extract_small_model(path, name, sr, if_f0, info, version):
+    try:
+        ckpt = torch.load(path, map_location="cpu", weights_only=False)
+        if "model" in ckpt:
+            ckpt = ckpt["model"]
+        opt = OrderedDict()
+        opt["weight"] = {}
+        for key in ckpt.keys():
+            if "enc_q" in key:
+                continue
+            opt["weight"][key] = ckpt[key].half()
+        if sr == "40k":
+            opt["config"] = [
+                1025,
+                32,
+                192,
+                192,
+                768,
+                2,
+                6,
+                3,
+                0,
+                "1",
+                [3, 7, 11],
+                [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                [10, 10, 2, 2],
+                512,
+                [16, 16, 4, 4],
+                109,
+                256,
+                40000,
+            ]
+        elif sr == "48k":
+            if version == "v1":
+                opt["config"] = [
+                    1025,
+                    32,
+                    192,
+                    192,
+                    768,
+                    2,
+                    6,
+                    3,
+                    0,
+                    "1",
+                    [3, 7, 11],
+                    [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    [10, 6, 2, 2, 2],
+                    512,
+                    [16, 16, 4, 4, 4],
+                    109,
+                    256,
+                    48000,
+                ]
+            else:
+                opt["config"] = [
+                    1025,
+                    32,
+                    192,
+                    192,
+                    768,
+                    2,
+                    6,
+                    3,
+                    0,
+                    "1",
+                    [3, 7, 11],
+                    [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    [12, 10, 2, 2],
+                    512,
+                    [24, 20, 4, 4],
+                    109,
+                    256,
+                    48000,
+                ]
+        elif sr == "32k":
+            if version == "v1":
+                opt["config"] = [
+                    513,
+                    32,
+                    192,
+                    192,
+                    768,
+                    2,
+                    6,
+                    3,
+                    0,
+                    "1",
+                    [3, 7, 11],
+                    [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    [10, 4, 2, 2, 2],
+                    512,
+                    [16, 16, 4, 4, 4],
+                    109,
+                    256,
+                    32000,
+                ]
+            else:
+                opt["config"] = [
+                    513,
+                    32,
+                    192,
+                    192,
+                    768,
+                    2,
+                    6,
+                    3,
+                    0,
+                    "1",
+                    [3, 7, 11],
+                    [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+                    [10, 8, 2, 2],
+                    512,
+                    [20, 16, 4, 4],
+                    109,
+                    256,
+                    32000,
+                ]
+        if info == "":
+            info = "Extracted model."
+        opt["info"] = info
+        opt["version"] = version
+        opt["sr"] = sr
+        opt["f0"] = int(if_f0)
+        torch.save(opt, "assets/weights/%s.pth" % name)
+        return "Success."
+    except:
+        return traceback.format_exc()
+
+
+def change_info(path, info, name):
+    try:
+        ckpt = torch.load(path, map_location="cpu", weights_only=False)
+        ckpt["info"] = info
+        if name == "":
+            name = os.path.basename(path)
+        torch.save(ckpt, "assets/weights/%s" % name)
+        return "Success."
+    except:
+        return traceback.format_exc()
+
+
+def merge(path1, path2, alpha1, sr, f0, info, name, version):
+    try:
+
+        def extract(ckpt):
+            a = ckpt["model"]
+            opt = OrderedDict()
+            opt["weight"] = {}
+            for key in a.keys():
+                if "enc_q" in key:
+                    continue
+                opt["weight"][key] = a[key]
+            return opt
+
+        ckpt1 = torch.load(path1, map_location="cpu", weights_only=False)
+        ckpt2 = torch.load(path2, map_location="cpu", weights_only=False)
+        cfg = ckpt1["config"]
+        if "model" in ckpt1:
+            ckpt1 = extract(ckpt1)
+        else:
+            ckpt1 = ckpt1["weight"]
+        if "model" in ckpt2:
+            ckpt2 = extract(ckpt2)
+        else:
+            ckpt2 = ckpt2["weight"]
+        if sorted(list(ckpt1.keys())) != sorted(list(ckpt2.keys())):
+            return "Fail to merge the models. The model architectures are not the same."
+        opt = OrderedDict()
+        opt["weight"] = {}
+        for key in ckpt1.keys():
+            # try:
+            if key == "emb_g.weight" and ckpt1[key].shape != ckpt2[key].shape:
+                min_shape0 = min(ckpt1[key].shape[0], ckpt2[key].shape[0])
+                opt["weight"][key] = (
+                    alpha1 * (ckpt1[key][:min_shape0].float())
+                    + (1 - alpha1) * (ckpt2[key][:min_shape0].float())
+                ).half()
+            else:
+                opt["weight"][key] = (
+                    alpha1 * (ckpt1[key].float()) + (1 - alpha1) * (ckpt2[key].float())
+                ).half()
+        # except:
+        #     pdb.set_trace()
+        opt["config"] = cfg
+        """
+        if(sr=="40k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 10, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 40000]
+        elif(sr=="48k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10,6,2,2,2], 512, [16, 16, 4, 4], 109, 256, 48000]
+        elif(sr=="32k"):opt["config"] = [513, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 4, 2, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 32000]
+        """
+        opt["sr"] = sr
+        opt["f0"] = 1 if f0 == i18n("是") else 0
+        opt["version"] = version
+        opt["info"] = info
+        torch.save(opt, "assets/weights/%s.pth" % name)
+        return "Success."
+    except:
+        return traceback.format_exc()

infer/lib/train/utils.py

@@ -0,0 +1,483 @@
+import argparse
+import glob
+import json
+import logging
+import os
+import subprocess
+import sys
+import shutil
+
+import numpy as np
+import torch
+from scipy.io.wavfile import read
+
+MATPLOTLIB_FLAG = False
+
+logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
+logger = logging
+
+
+def load_checkpoint_d(checkpoint_path, combd, sbd, optimizer=None, load_opt=1):
+    assert os.path.isfile(checkpoint_path)
+    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
+
+    ##################
+    def go(model, bkey):
+        saved_state_dict = checkpoint_dict[bkey]
+        if hasattr(model, "module"):
+            state_dict = model.module.state_dict()
+        else:
+            state_dict = model.state_dict()
+        new_state_dict = {}
+        for k, v in state_dict.items():  # 模型需要的shape
+            try:
+                new_state_dict[k] = saved_state_dict[k]
+                if saved_state_dict[k].shape != state_dict[k].shape:
+                    logger.warning(
+                        "shape-%s-mismatch. need: %s, get: %s",
+                        k,
+                        state_dict[k].shape,
+                        saved_state_dict[k].shape,
+                    )  #
+                    raise KeyError
+            except:
+                # logger.info(traceback.format_exc())
+                logger.info("%s is not in the checkpoint", k)  # pretrain缺失的
+                new_state_dict[k] = v  # 模型自带的随机值
+        if hasattr(model, "module"):
+            model.module.load_state_dict(new_state_dict, strict=False)
+        else:
+            model.load_state_dict(new_state_dict, strict=False)
+        return model
+
+    go(combd, "combd")
+    model = go(sbd, "sbd")
+    #############
+    logger.info("Loaded model weights")
+
+    iteration = checkpoint_dict["iteration"]
+    learning_rate = checkpoint_dict["learning_rate"]
+    if (
+        optimizer is not None and load_opt == 1
+    ):  ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
+        #   try:
+        optimizer.load_state_dict(checkpoint_dict["optimizer"])
+    #   except:
+    #     traceback.print_exc()
+    logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
+    return model, optimizer, learning_rate, iteration
+
+
+# def load_checkpoint(checkpoint_path, model, optimizer=None):
+#   assert os.path.isfile(checkpoint_path)
+#   checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
+#   iteration = checkpoint_dict['iteration']
+#   learning_rate = checkpoint_dict['learning_rate']
+#   if optimizer is not None:
+#     optimizer.load_state_dict(checkpoint_dict['optimizer'])
+#   # print(1111)
+#   saved_state_dict = checkpoint_dict['model']
+#   # print(1111)
+#
+#   if hasattr(model, 'module'):
+#     state_dict = model.module.state_dict()
+#   else:
+#     state_dict = model.state_dict()
+#   new_state_dict= {}
+#   for k, v in state_dict.items():
+#     try:
+#       new_state_dict[k] = saved_state_dict[k]
+#     except:
+#       logger.info("%s is not in the checkpoint" % k)
+#       new_state_dict[k] = v
+#   if hasattr(model, 'module'):
+#     model.module.load_state_dict(new_state_dict)
+#   else:
+#     model.load_state_dict(new_state_dict)
+#   logger.info("Loaded checkpoint '{}' (epoch {})" .format(
+#     checkpoint_path, iteration))
+#   return model, optimizer, learning_rate, iteration
+def load_checkpoint(checkpoint_path, model, optimizer=None, load_opt=1):
+    assert os.path.isfile(checkpoint_path)
+    checkpoint_dict = torch.load(checkpoint_path, map_location="cpu", weights_only=False)
+
+    saved_state_dict = checkpoint_dict["model"]
+    if hasattr(model, "module"):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+    new_state_dict = {}
+    for k, v in state_dict.items():  # 模型需要的shape
+        try:
+            new_state_dict[k] = saved_state_dict[k]
+            if saved_state_dict[k].shape != state_dict[k].shape:
+                logger.warning(
+                    "shape-%s-mismatch|need-%s|get-%s",
+                    k,
+                    state_dict[k].shape,
+                    saved_state_dict[k].shape,
+                )  #
+                raise KeyError
+        except:
+            # logger.info(traceback.format_exc())
+            logger.info("%s is not in the checkpoint", k)  # pretrain缺失的
+            new_state_dict[k] = v  # 模型自带的随机值
+    if hasattr(model, "module"):
+        model.module.load_state_dict(new_state_dict, strict=False)
+    else:
+        model.load_state_dict(new_state_dict, strict=False)
+    logger.info("Loaded model weights")
+
+    iteration = checkpoint_dict["iteration"]
+    learning_rate = checkpoint_dict["learning_rate"]
+    if (
+        optimizer is not None and load_opt == 1
+    ):  ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
+        #   try:
+        optimizer.load_state_dict(checkpoint_dict["optimizer"])
+    #   except:
+    #     traceback.print_exc()
+    logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
+    return model, optimizer, learning_rate, iteration
+
+
+def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
+    logger.info(
+        "Saving model and optimizer state at epoch {} to {}".format(
+            iteration, checkpoint_path
+        )
+    )
+    if hasattr(model, "module"):
+        state_dict = model.module.state_dict()
+    else:
+        state_dict = model.state_dict()
+    torch.save(
+        {
+            "model": state_dict,
+            "iteration": iteration,
+            "optimizer": optimizer.state_dict(),
+            "learning_rate": learning_rate,
+        },
+        checkpoint_path,
+    )
+
+
+def save_checkpoint_d(combd, sbd, optimizer, learning_rate, iteration, checkpoint_path):
+    logger.info(
+        "Saving model and optimizer state at epoch {} to {}".format(
+            iteration, checkpoint_path
+        )
+    )
+    if hasattr(combd, "module"):
+        state_dict_combd = combd.module.state_dict()
+    else:
+        state_dict_combd = combd.state_dict()
+    if hasattr(sbd, "module"):
+        state_dict_sbd = sbd.module.state_dict()
+    else:
+        state_dict_sbd = sbd.state_dict()
+    torch.save(
+        {
+            "combd": state_dict_combd,
+            "sbd": state_dict_sbd,
+            "iteration": iteration,
+            "optimizer": optimizer.state_dict(),
+            "learning_rate": learning_rate,
+        },
+        checkpoint_path,
+    )
+
+
+def summarize(
+    writer,
+    global_step,
+    scalars={},
+    histograms={},
+    images={},
+    audios={},
+    audio_sampling_rate=22050,
+):
+    for k, v in scalars.items():
+        writer.add_scalar(k, v, global_step)
+    for k, v in histograms.items():
+        writer.add_histogram(k, v, global_step)
+    for k, v in images.items():
+        writer.add_image(k, v, global_step, dataformats="HWC")
+    for k, v in audios.items():
+        writer.add_audio(k, v, global_step, audio_sampling_rate)
+
+
+def latest_checkpoint_path(dir_path, regex="G_*.pth"):
+    f_list = glob.glob(os.path.join(dir_path, regex))
+    f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
+    x = f_list[-1]
+    logger.debug(x)
+    return x
+
+
+def plot_spectrogram_to_numpy(spectrogram):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        MATPLOTLIB_FLAG = True
+        mpl_logger = logging.getLogger("matplotlib")
+        mpl_logger.setLevel(logging.WARNING)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(10, 2))
+    im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+    plt.colorbar(im, ax=ax)
+    plt.xlabel("Frames")
+    plt.ylabel("Channels")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close()
+    return data
+
+
+def plot_alignment_to_numpy(alignment, info=None):
+    global MATPLOTLIB_FLAG
+    if not MATPLOTLIB_FLAG:
+        import matplotlib
+
+        matplotlib.use("Agg")
+        MATPLOTLIB_FLAG = True
+        mpl_logger = logging.getLogger("matplotlib")
+        mpl_logger.setLevel(logging.WARNING)
+    import matplotlib.pylab as plt
+    import numpy as np
+
+    fig, ax = plt.subplots(figsize=(6, 4))
+    im = ax.imshow(
+        alignment.transpose(), aspect="auto", origin="lower", interpolation="none"
+    )
+    fig.colorbar(im, ax=ax)
+    xlabel = "Decoder timestep"
+    if info is not None:
+        xlabel += "\n\n" + info
+    plt.xlabel(xlabel)
+    plt.ylabel("Encoder timestep")
+    plt.tight_layout()
+
+    fig.canvas.draw()
+    data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+    data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+    plt.close()
+    return data
+
+
+def load_wav_to_torch(full_path):
+    sampling_rate, data = read(full_path)
+    return torch.FloatTensor(data.astype(np.float32)), sampling_rate
+
+
+def load_filepaths_and_text(filename, split="|"):
+    try:
+        with open(filename, encoding="utf-8") as f:
+            filepaths_and_text = [line.strip().split(split) for line in f]
+    except UnicodeDecodeError:
+        with open(filename) as f:
+            filepaths_and_text = [line.strip().split(split) for line in f]
+    
+    return filepaths_and_text
+
+
+def get_hparams(init=True):
+    """
+    todo:
+      结尾七人���：
+        保存频率、总epoch                     done
+        bs                                    done
+        pretrainG、pretrainD                  done
+        卡号：os.en["CUDA_VISIBLE_DEVICES"]   done
+        if_latest                             done
+      模型：if_f0                             done
+      采样率：自动选择config                  done
+      是否缓存数据集进GPU:if_cache_data_in_gpu done
+
+      -m:
+        自动决定training_files路径,改掉train_nsf_load_pretrain.py里的hps.data.training_files    done
+      -c不要了
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-se",
+        "--save_every_epoch",
+        type=int,
+        required=True,
+        help="checkpoint save frequency (epoch)",
+    )
+    parser.add_argument(
+        "-te", "--total_epoch", type=int, required=True, help="total_epoch"
+    )
+    parser.add_argument(
+        "-pg", "--pretrainG", type=str, default="", help="Pretrained Generator path"
+    )
+    parser.add_argument(
+        "-pd", "--pretrainD", type=str, default="", help="Pretrained Discriminator path"
+    )
+    parser.add_argument("-g", "--gpus", type=str, default="0", help="split by -")
+    parser.add_argument(
+        "-bs", "--batch_size", type=int, required=True, help="batch size"
+    )
+    parser.add_argument(
+        "-e", "--experiment_dir", type=str, required=True, help="experiment dir"
+    )  # -m
+    parser.add_argument(
+        "-sr", "--sample_rate", type=str, required=True, help="sample rate, 32k/40k/48k"
+    )
+    parser.add_argument(
+        "-sw",
+        "--save_every_weights",
+        type=str,
+        default="0",
+        help="save the extracted model in weights directory when saving checkpoints",
+    )
+    parser.add_argument(
+        "-v", "--version", type=str, required=True, help="model version"
+    )
+    parser.add_argument(
+        "-f0",
+        "--if_f0",
+        type=int,
+        required=True,
+        help="use f0 as one of the inputs of the model, 1 or 0",
+    )
+    parser.add_argument(
+        "-l",
+        "--if_latest",
+        type=int,
+        required=True,
+        help="if only save the latest G/D pth file, 1 or 0",
+    )
+    parser.add_argument(
+        "-c",
+        "--if_cache_data_in_gpu",
+        type=int,
+        required=True,
+        help="if caching the dataset in GPU memory, 1 or 0",
+    )
+
+    args = parser.parse_args()
+    name = args.experiment_dir
+    experiment_dir = os.path.join("./logs", args.experiment_dir)
+
+    config_save_path = os.path.join(experiment_dir, "config.json")
+    with open(config_save_path, "r") as f:
+        config = json.load(f)
+
+    hparams = HParams(**config)
+    hparams.model_dir = hparams.experiment_dir = experiment_dir
+    hparams.save_every_epoch = args.save_every_epoch
+    hparams.name = name
+    hparams.total_epoch = args.total_epoch
+    hparams.pretrainG = args.pretrainG
+    hparams.pretrainD = args.pretrainD
+    hparams.version = args.version
+    hparams.gpus = args.gpus
+    hparams.train.batch_size = args.batch_size
+    hparams.sample_rate = args.sample_rate
+    hparams.if_f0 = args.if_f0
+    hparams.if_latest = args.if_latest
+    hparams.save_every_weights = args.save_every_weights
+    hparams.if_cache_data_in_gpu = args.if_cache_data_in_gpu
+    hparams.data.training_files = "%s/filelist.txt" % experiment_dir
+    return hparams
+
+
+def get_hparams_from_dir(model_dir):
+    config_save_path = os.path.join(model_dir, "config.json")
+    with open(config_save_path, "r") as f:
+        data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    hparams.model_dir = model_dir
+    return hparams
+
+
+def get_hparams_from_file(config_path):
+    with open(config_path, "r") as f:
+        data = f.read()
+    config = json.loads(data)
+
+    hparams = HParams(**config)
+    return hparams
+
+
+def check_git_hash(model_dir):
+    source_dir = os.path.dirname(os.path.realpath(__file__))
+    if not os.path.exists(os.path.join(source_dir, ".git")):
+        logger.warning(
+            "{} is not a git repository, therefore hash value comparison will be ignored.".format(
+                source_dir
+            )
+        )
+        return
+
+    cur_hash = subprocess.getoutput("git rev-parse HEAD")
+
+    path = os.path.join(model_dir, "githash")
+    if os.path.exists(path):
+        saved_hash = open(path).read()
+        if saved_hash != cur_hash:
+            logger.warning(
+                "git hash values are different. {}(saved) != {}(current)".format(
+                    saved_hash[:8], cur_hash[:8]
+                )
+            )
+    else:
+        open(path, "w").write(cur_hash)
+
+
+def get_logger(model_dir, filename="train.log"):
+    global logger
+    logger = logging.getLogger(os.path.basename(model_dir))
+    logger.setLevel(logging.DEBUG)
+
+    formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
+    if not os.path.exists(model_dir):
+        os.makedirs(model_dir)
+    h = logging.FileHandler(os.path.join(model_dir, filename))
+    h.setLevel(logging.DEBUG)
+    h.setFormatter(formatter)
+    logger.addHandler(h)
+    return logger
+
+
+class HParams:
+    def __init__(self, **kwargs):
+        for k, v in kwargs.items():
+            if type(v) == dict:
+                v = HParams(**v)
+            self[k] = v
+
+    def keys(self):
+        return self.__dict__.keys()
+
+    def items(self):
+        return self.__dict__.items()
+
+    def values(self):
+        return self.__dict__.values()
+
+    def __len__(self):
+        return len(self.__dict__)
+
+    def __getitem__(self, key):
+        return getattr(self, key)
+
+    def __setitem__(self, key, value):
+        return setattr(self, key, value)
+
+    def __contains__(self, key):
+        return key in self.__dict__
+
+    def __repr__(self):
+        return self.__dict__.__repr__()

infer/lib/uvr5_pack/lib_v5/dataset.py

@@ -0,0 +1,183 @@
+import os
+import random
+
+import numpy as np
+import torch
+import torch.utils.data
+from tqdm import tqdm
+
+from . import spec_utils
+
+
+class VocalRemoverValidationSet(torch.utils.data.Dataset):
+    def __init__(self, patch_list):
+        self.patch_list = patch_list
+
+    def __len__(self):
+        return len(self.patch_list)
+
+    def __getitem__(self, idx):
+        path = self.patch_list[idx]
+        data = np.load(path)
+
+        X, y = data["X"], data["y"]
+
+        X_mag = np.abs(X)
+        y_mag = np.abs(y)
+
+        return X_mag, y_mag
+
+
+def make_pair(mix_dir, inst_dir):
+    input_exts = [".wav", ".m4a", ".mp3", ".mp4", ".flac"]
+
+    X_list = sorted(
+        [
+            os.path.join(mix_dir, fname)
+            for fname in os.listdir(mix_dir)
+            if os.path.splitext(fname)[1] in input_exts
+        ]
+    )
+    y_list = sorted(
+        [
+            os.path.join(inst_dir, fname)
+            for fname in os.listdir(inst_dir)
+            if os.path.splitext(fname)[1] in input_exts
+        ]
+    )
+
+    filelist = list(zip(X_list, y_list))
+
+    return filelist
+
+
+def train_val_split(dataset_dir, split_mode, val_rate, val_filelist):
+    if split_mode == "random":
+        filelist = make_pair(
+            os.path.join(dataset_dir, "mixtures"),
+            os.path.join(dataset_dir, "instruments"),
+        )
+
+        random.shuffle(filelist)
+
+        if len(val_filelist) == 0:
+            val_size = int(len(filelist) * val_rate)
+            train_filelist = filelist[:-val_size]
+            val_filelist = filelist[-val_size:]
+        else:
+            train_filelist = [
+                pair for pair in filelist if list(pair) not in val_filelist
+            ]
+    elif split_mode == "subdirs":
+        if len(val_filelist) != 0:
+            raise ValueError(
+                "The `val_filelist` option is not available in `subdirs` mode"
+            )
+
+        train_filelist = make_pair(
+            os.path.join(dataset_dir, "training/mixtures"),
+            os.path.join(dataset_dir, "training/instruments"),
+        )
+
+        val_filelist = make_pair(
+            os.path.join(dataset_dir, "validation/mixtures"),
+            os.path.join(dataset_dir, "validation/instruments"),
+        )
+
+    return train_filelist, val_filelist
+
+
+def augment(X, y, reduction_rate, reduction_mask, mixup_rate, mixup_alpha):
+    perm = np.random.permutation(len(X))
+    for i, idx in enumerate(tqdm(perm)):
+        if np.random.uniform() < reduction_rate:
+            y[idx] = spec_utils.reduce_vocal_aggressively(
+                X[idx], y[idx], reduction_mask
+            )
+
+        if np.random.uniform() < 0.5:
+            # swap channel
+            X[idx] = X[idx, ::-1]
+            y[idx] = y[idx, ::-1]
+        if np.random.uniform() < 0.02:
+            # mono
+            X[idx] = X[idx].mean(axis=0, keepdims=True)
+            y[idx] = y[idx].mean(axis=0, keepdims=True)
+        if np.random.uniform() < 0.02:
+            # inst
+            X[idx] = y[idx]
+
+        if np.random.uniform() < mixup_rate and i < len(perm) - 1:
+            lam = np.random.beta(mixup_alpha, mixup_alpha)
+            X[idx] = lam * X[idx] + (1 - lam) * X[perm[i + 1]]
+            y[idx] = lam * y[idx] + (1 - lam) * y[perm[i + 1]]
+
+    return X, y
+
+
+def make_padding(width, cropsize, offset):
+    left = offset
+    roi_size = cropsize - left * 2
+    if roi_size == 0:
+        roi_size = cropsize
+    right = roi_size - (width % roi_size) + left
+
+    return left, right, roi_size
+
+
+def make_training_set(filelist, cropsize, patches, sr, hop_length, n_fft, offset):
+    len_dataset = patches * len(filelist)
+
+    X_dataset = np.zeros((len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
+    y_dataset = np.zeros((len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
+
+    for i, (X_path, y_path) in enumerate(tqdm(filelist)):
+        X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
+        coef = np.max([np.abs(X).max(), np.abs(y).max()])
+        X, y = X / coef, y / coef
+
+        l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
+        X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode="constant")
+        y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode="constant")
+
+        starts = np.random.randint(0, X_pad.shape[2] - cropsize, patches)
+        ends = starts + cropsize
+        for j in range(patches):
+            idx = i * patches + j
+            X_dataset[idx] = X_pad[:, :, starts[j] : ends[j]]
+            y_dataset[idx] = y_pad[:, :, starts[j] : ends[j]]
+
+    return X_dataset, y_dataset
+
+
+def make_validation_set(filelist, cropsize, sr, hop_length, n_fft, offset):
+    patch_list = []
+    patch_dir = "cs{}_sr{}_hl{}_nf{}_of{}".format(
+        cropsize, sr, hop_length, n_fft, offset
+    )
+    os.makedirs(patch_dir, exist_ok=True)
+
+    for i, (X_path, y_path) in enumerate(tqdm(filelist)):
+        basename = os.path.splitext(os.path.basename(X_path))[0]
+
+        X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
+        coef = np.max([np.abs(X).max(), np.abs(y).max()])
+        X, y = X / coef, y / coef
+
+        l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
+        X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode="constant")
+        y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode="constant")
+
+        len_dataset = int(np.ceil(X.shape[2] / roi_size))
+        for j in range(len_dataset):
+            outpath = os.path.join(patch_dir, "{}_p{}.npz".format(basename, j))
+            start = j * roi_size
+            if not os.path.exists(outpath):
+                np.savez(
+                    outpath,
+                    X=X_pad[:, :, start : start + cropsize],
+                    y=y_pad[:, :, start : start + cropsize],
+                )
+            patch_list.append(outpath)
+
+    return VocalRemoverValidationSet(patch_list)

infer/lib/uvr5_pack/lib_v5/layers.py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_123812KB .py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_123821KB.py

@@ -0,0 +1,118 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_33966KB.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv6 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv7 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        feat6 = self.conv6(x)
+        feat7 = self.conv7(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_537227KB.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv6 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv7 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        feat6 = self.conv6(x)
+        feat7 = self.conv7(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_537238KB.py

@@ -0,0 +1,126 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(SeperableConv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nin,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                groups=nin,
+                bias=False,
+            ),
+            nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+    def __call__(self, x):
+        skip = self.conv1(x)
+        h = self.conv2(skip)
+
+        return h, skip
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+        h = self.conv(x)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+        self.conv3 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv6 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.conv7 = SeperableConv2DBNActiv(
+            nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = nn.Sequential(
+            Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+        )
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        feat6 = self.conv6(x)
+        feat7 = self.conv7(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+        bottle = self.bottleneck(out)
+        return bottle

infer/lib/uvr5_pack/lib_v5/layers_new.py

@@ -0,0 +1,125 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+        super(Conv2DBNActiv, self).__init__()
+        self.conv = nn.Sequential(
+            nn.Conv2d(
+                nin,
+                nout,
+                kernel_size=ksize,
+                stride=stride,
+                padding=pad,
+                dilation=dilation,
+                bias=False,
+            ),
+            nn.BatchNorm2d(nout),
+            activ(),
+        )
+
+    def __call__(self, x):
+        return self.conv(x)
+
+
+class Encoder(nn.Module):
+    def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+        super(Encoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, stride, pad, activ=activ)
+        self.conv2 = Conv2DBNActiv(nout, nout, ksize, 1, pad, activ=activ)
+
+    def __call__(self, x):
+        h = self.conv1(x)
+        h = self.conv2(h)
+
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+    ):
+        super(Decoder, self).__init__()
+        self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+        # self.conv2 = Conv2DBNActiv(nout, nout, ksize, 1, pad, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def __call__(self, x, skip=None):
+        x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+
+        if skip is not None:
+            skip = spec_utils.crop_center(skip, x)
+            x = torch.cat([x, skip], dim=1)
+
+        h = self.conv1(x)
+        # h = self.conv2(h)
+
+        if self.dropout is not None:
+            h = self.dropout(h)
+
+        return h
+
+
+class ASPPModule(nn.Module):
+    def __init__(self, nin, nout, dilations=(4, 8, 12), activ=nn.ReLU, dropout=False):
+        super(ASPPModule, self).__init__()
+        self.conv1 = nn.Sequential(
+            nn.AdaptiveAvgPool2d((1, None)),
+            Conv2DBNActiv(nin, nout, 1, 1, 0, activ=activ),
+        )
+        self.conv2 = Conv2DBNActiv(nin, nout, 1, 1, 0, activ=activ)
+        self.conv3 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[0], dilations[0], activ=activ
+        )
+        self.conv4 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[1], dilations[1], activ=activ
+        )
+        self.conv5 = Conv2DBNActiv(
+            nin, nout, 3, 1, dilations[2], dilations[2], activ=activ
+        )
+        self.bottleneck = Conv2DBNActiv(nout * 5, nout, 1, 1, 0, activ=activ)
+        self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+    def forward(self, x):
+        _, _, h, w = x.size()
+        feat1 = F.interpolate(
+            self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+        )
+        feat2 = self.conv2(x)
+        feat3 = self.conv3(x)
+        feat4 = self.conv4(x)
+        feat5 = self.conv5(x)
+        out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+        out = self.bottleneck(out)
+
+        if self.dropout is not None:
+            out = self.dropout(out)
+
+        return out
+
+
+class LSTMModule(nn.Module):
+    def __init__(self, nin_conv, nin_lstm, nout_lstm):
+        super(LSTMModule, self).__init__()
+        self.conv = Conv2DBNActiv(nin_conv, 1, 1, 1, 0)
+        self.lstm = nn.LSTM(
+            input_size=nin_lstm, hidden_size=nout_lstm // 2, bidirectional=True
+        )
+        self.dense = nn.Sequential(
+            nn.Linear(nout_lstm, nin_lstm), nn.BatchNorm1d(nin_lstm), nn.ReLU()
+        )
+
+    def forward(self, x):
+        N, _, nbins, nframes = x.size()
+        h = self.conv(x)[:, 0]  # N, nbins, nframes
+        h = h.permute(2, 0, 1)  # nframes, N, nbins
+        h, _ = self.lstm(h)
+        h = self.dense(h.reshape(-1, h.size()[-1]))  # nframes * N, nbins
+        h = h.reshape(nframes, N, 1, nbins)
+        h = h.permute(1, 2, 3, 0)
+
+        return h

infer/lib/uvr5_pack/lib_v5/model_param_init.py

@@ -0,0 +1,69 @@
+import json
+import os
+import pathlib
+
+default_param = {}
+default_param["bins"] = 768
+default_param["unstable_bins"] = 9  # training only
+default_param["reduction_bins"] = 762  # training only
+default_param["sr"] = 44100
+default_param["pre_filter_start"] = 757
+default_param["pre_filter_stop"] = 768
+default_param["band"] = {}
+
+
+default_param["band"][1] = {
+    "sr": 11025,
+    "hl": 128,
+    "n_fft": 960,
+    "crop_start": 0,
+    "crop_stop": 245,
+    "lpf_start": 61,  # inference only
+    "res_type": "polyphase",
+}
+
+default_param["band"][2] = {
+    "sr": 44100,
+    "hl": 512,
+    "n_fft": 1536,
+    "crop_start": 24,
+    "crop_stop": 547,
+    "hpf_start": 81,  # inference only
+    "res_type": "sinc_best",
+}
+
+
+def int_keys(d):
+    r = {}
+    for k, v in d:
+        if k.isdigit():
+            k = int(k)
+        r[k] = v
+    return r
+
+
+class ModelParameters(object):
+    def __init__(self, config_path=""):
+        if ".pth" == pathlib.Path(config_path).suffix:
+            import zipfile
+
+            with zipfile.ZipFile(config_path, "r") as zip:
+                self.param = json.loads(
+                    zip.read("param.json"), object_pairs_hook=int_keys
+                )
+        elif ".json" == pathlib.Path(config_path).suffix:
+            with open(config_path, "r") as f:
+                self.param = json.loads(f.read(), object_pairs_hook=int_keys)
+        else:
+            self.param = default_param
+
+        for k in [
+            "mid_side",
+            "mid_side_b",
+            "mid_side_b2",
+            "stereo_w",
+            "stereo_n",
+            "reverse",
+        ]:
+            if not k in self.param:
+                self.param[k] = False

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 16000,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 16000,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 32000,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "kaiser_fast"
+		}
+	},
+	"sr": 32000,
+	"pre_filter_start": 1000,
+	"pre_filter_stop": 1021
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 33075,
+			"hl": 384,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 33075,
+	"pre_filter_start": 1000,
+	"pre_filter_stop": 1021
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 1024,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json

@@ -0,0 +1,19 @@
+{
+	"bins": 256,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 256,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 256,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 256,
+	"pre_filter_stop": 256
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 1024,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 1024
+}

infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512_cut.json

@@ -0,0 +1,19 @@
+{
+	"bins": 1024,
+	"unstable_bins": 0,
+	"reduction_bins": 0,
+	"band": {
+		"1": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 2048,
+			"crop_start": 0,
+			"crop_stop": 700,
+			"hpf_start": -1,
+			"res_type": "sinc_best"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 1023,
+	"pre_filter_stop": 700
+}

infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json

@@ -0,0 +1,30 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 705,
+	"band": {
+		"1": {
+			"sr": 6000,
+			"hl": 66,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 240,
+			"lpf_start": 60,
+			"lpf_stop": 118,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 32000,
+			"hl": 352,
+			"n_fft": 1024,
+			"crop_start": 22,
+			"crop_stop": 505,
+			"hpf_start": 44,
+			"hpf_stop": 23,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 32000,
+	"pre_filter_start": 710,
+	"pre_filter_stop": 731
+}

infer/lib/uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json

@@ -0,0 +1,30 @@
+{
+	"bins": 512,
+	"unstable_bins": 7,
+	"reduction_bins": 510,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 160,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 192,
+			"lpf_start": 41,
+			"lpf_stop": 139,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 44100,
+			"hl": 640,
+			"n_fft": 1024,
+			"crop_start": 10,
+			"crop_stop": 320,
+			"hpf_start": 47,
+			"hpf_stop": 15,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 510,
+	"pre_filter_stop": 512
+}

infer/lib/uvr5_pack/lib_v5/modelparams/2band_48000.json

@@ -0,0 +1,30 @@
+{
+	"bins": 768,
+	"unstable_bins": 7,
+	"reduction_bins": 705,
+	"band": {
+		"1": {
+			"sr": 6000,
+			"hl": 66,
+			"n_fft": 512,
+			"crop_start": 0,
+			"crop_stop": 240,
+			"lpf_start": 60,
+			"lpf_stop": 240,
+			"res_type": "sinc_fastest"
+		},
+		"2": {
+			"sr": 48000,
+			"hl": 528,
+			"n_fft": 1536,
+			"crop_start": 22,
+			"crop_stop": 505,
+			"hpf_start": 82,
+			"hpf_stop": 22,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 48000,
+	"pre_filter_start": 710,
+	"pre_filter_stop": 731
+}

infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json

@@ -0,0 +1,42 @@
+{
+	"bins": 768,
+	"unstable_bins": 5,
+	"reduction_bins": 733,
+	"band": {
+		"1": {
+			"sr": 11025,
+			"hl": 128,
+			"n_fft": 768,
+			"crop_start": 0,
+			"crop_stop": 278,
+			"lpf_start": 28,
+			"lpf_stop": 140,
+			"res_type": "polyphase"
+		},
+		"2": {
+			"sr": 22050,
+			"hl": 256,
+			"n_fft": 768,
+			"crop_start": 14,
+			"crop_stop": 322,
+			"hpf_start": 70,
+			"hpf_stop": 14,
+			"lpf_start": 283,
+			"lpf_stop": 314,
+			"res_type": "polyphase"
+		},	
+		"3": {
+			"sr": 44100,
+			"hl": 512,
+			"n_fft": 768,
+			"crop_start": 131,
+			"crop_stop": 313,
+			"hpf_start": 154,
+			"hpf_stop": 141,
+			"res_type": "sinc_medium"
+		}
+	},
+	"sr": 44100,
+	"pre_filter_start": 757,
+	"pre_filter_stop": 768
+}