Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+stftpitchshift filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,34 @@
----
-license: mit
----
+# **Harmonify (RVC No UI Colab)**
+https://colab.research.google.com/drive/1X8YR4Ruv7zzY8YAMPTfC7hkxqT_d4Q5d
+
+## **Credits**
+[Eempostor](https://discordapp.com/users/818050831034613771) - Made everything work together, Notebook creator
+
+[Applio](https://github.com/IAHispano/Applio-RVC-Fork) by [IAHispano](https://github.com/IAHispano) - The repo this colab is based on
+
+[CNChTu](https://github.com/CNChTu) - [FCPE](https://github.com/CNChTu/FCPE) F0 method
+
+[So Vits SVC](https://github.com/svc-develop-team/so-vits-svc) - [FCPE](https://github.com/CNChTu/FCPE) F0 method script
+
+[ChatGPT](https://chat.openai.com/) - Helper
+
+[Phind](https://www.phind.com/) - Helper
+
+If you have any suggestions or problems on this colab, dm [me](https://discordapp.com/users/818050831034613771) on discord.
+
+## **Changelogs**
+9/3/2024 | Huge changes
+- Pitch extraction `fcpe` now uses the `torchfcpe` library. You can still use the previous version with `fcpe_legacy`
+- `MIN_PITCH` and `MAX_PITCH` now accepts pitch notations
+- Fixed error when running inference without a GPU (GPU is still recommended as its way faster and more stable)
+- Fixed error when running `hybrid` with `pm`, `dio`, and `harvest`
+- Overhaulled the directory structure and some code
+
+14/1/2024 | Fixes
+- Fixed `Cannot retrieve the public link of the file` issue while downloading models from google drive
+
+10/1/2024 | Fixes
+- Fixed `Cannot retrieve the public link of the file` issue on installation
+
+12/12/2023 | Small adjustments
+- Moved `DOWNLOAD` and `SAVE_TO_DRIVE` option on inference cell into a new cell

assets/fcpe/fcpe.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3a8dd2dbd51baf19ed295006f2ac25dba6dd60adc7ec578ae5fbd94970951da
+size 69005189

assets/hubert/hubert_base.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f54b40fd2802423a5643779c4861af1e9ee9c1564dc9d32f54f20b5ffba7db96
+size 189507909

assets/rmvpe/rmvpe.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5ed4719f59085d1affc5d81354c70828c740584f2d24e782523345a6a278962
+size 181189687

lib/infer.py

@@ -0,0 +1,221 @@
+import os
+import shutil
+import gc
+import torch
+from multiprocessing import cpu_count
+from lib.modules import VC
+from lib.split_audio import split_silence_nonsilent, adjust_audio_lengths, combine_silence_nonsilent
+
+class Configs:
+    def __init__(self, device, is_half):
+        self.device = device
+        self.is_half = is_half
+        self.n_cpu = 0
+        self.gpu_name = None
+        self.gpu_mem = None
+        self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
+
+    def device_config(self) -> tuple:
+        if torch.cuda.is_available():
+            i_device = int(self.device.split(":")[-1])
+            self.gpu_name = torch.cuda.get_device_name(i_device)
+            #if (
+#                    ("16" in self.gpu_name and "V100" not in self.gpu_name.upper())
+#                    or "P40" in self.gpu_name.upper()
+#                    or "1060" in self.gpu_name
+#                    or "1070" in self.gpu_name
+#                    or "1080" in self.gpu_name
+#            ):
+#                print("16 series/10 series P40 forced single precision")
+#                self.is_half = False
+#                for config_file in ["32k.json", "40k.json", "48k.json"]:
+#                    with open(BASE_DIR / "src" / "configs" / config_file, "r") as f:
+#                        strr = f.read().replace("true", "false")
+#                    with open(BASE_DIR / "src" / "configs" / config_file, "w") as f:
+#                        f.write(strr)
+#                with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "r") as f:
+#                    strr = f.read().replace("3.7", "3.0")
+#                with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "w") as f:
+#                    f.write(strr)
+#            else:
+#                self.gpu_name = None
+#            self.gpu_mem = int(
+#                torch.cuda.get_device_properties(i_device).total_memory
+#                / 1024
+#                / 1024
+#                / 1024
+#                + 0.4
+#            )
+#            if self.gpu_mem <= 4:
+#                with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "r") as f:
+#                    strr = f.read().replace("3.7", "3.0")
+#                with open(BASE_DIR / "src" / "trainset_preprocess_pipeline_print.py", "w") as f:
+#                    f.write(strr)
+        elif torch.backends.mps.is_available():
+            print("No supported N-card found, use MPS for inference")
+            self.device = "mps"
+        else:
+            print("No supported N-card found, use CPU for inference")
+            self.device = "cpu"
+
+        if self.n_cpu == 0:
+            self.n_cpu = cpu_count()
+
+        if self.is_half:
+            # 6G memory config
+            x_pad = 3
+            x_query = 10
+            x_center = 60
+            x_max = 65
+        else:
+            # 5G memory config
+            x_pad = 1
+            x_query = 6
+            x_center = 38
+            x_max = 41
+
+        if self.gpu_mem != None and self.gpu_mem <= 4:
+            x_pad = 1
+            x_query = 5
+            x_center = 30
+            x_max = 32
+
+        return x_pad, x_query, x_center, x_max
+
+def get_model(voice_model):
+    model_dir = os.path.join(os.getcwd(), "models", voice_model)
+    model_filename, index_filename = None, None
+    for file in os.listdir(model_dir):
+        ext = os.path.splitext(file)[1]
+        if ext == '.pth':
+            model_filename = file
+        if ext == '.index':
+            index_filename = file
+
+    if model_filename is None:
+        print(f'No model file exists in {models_dir}.')
+        return None, None
+
+    return os.path.join(model_dir, model_filename), os.path.join(model_dir, index_filename) if index_filename else ''
+
+def infer_audio(
+    model_name,
+    audio_path,
+    f0_change=0,
+    f0_method="rmvpe+",
+    min_pitch="50",
+    max_pitch="1100",
+    crepe_hop_length=128,
+    index_rate=0.75,
+    filter_radius=3,
+    rms_mix_rate=0.25,
+    protect=0.33,
+    split_infer=False,
+    min_silence=500,
+    silence_threshold=-50,
+    seek_step=1,
+    keep_silence=100,
+    do_formant=False,
+    quefrency=0,
+    timbre=1,
+    f0_autotune=False,
+    audio_format="wav",
+    resample_sr=0,
+    hubert_model_path="assets/hubert/hubert_base.pt",
+    rmvpe_model_path="assets/rmvpe/rmvpe.pt",
+    fcpe_model_path="assets/fcpe/fcpe.pt"
+    ):
+    os.environ["rmvpe_model_path"] = rmvpe_model_path
+    os.environ["fcpe_model_path"] = fcpe_model_path
+    configs = Configs('cuda:0', True)
+    vc = VC(configs)
+    pth_path, index_path = get_model(model_name)
+    vc_data = vc.get_vc(pth_path, protect, 0.5)
+    
+    if split_infer:
+        inferred_files = []
+        temp_dir = os.path.join(os.getcwd(), "seperate", "temp")
+        os.makedirs(temp_dir, exist_ok=True)
+        print("Splitting audio to silence and nonsilent segments.")
+        silence_files, nonsilent_files = split_silence_nonsilent(audio_path, min_silence, silence_threshold, seek_step, keep_silence)
+        print(f"Total silence segments: {len(silence_files)}.\nTotal nonsilent segments: {len(nonsilent_files)}.")
+        for i, nonsilent_file in enumerate(nonsilent_files):
+            print(f"Inferring nonsilent audio {i+1}")
+            inference_info, audio_data, output_path = vc.vc_single(
+            0,
+            nonsilent_file,
+            f0_change,
+            f0_method,
+            index_path,
+            index_path,
+            index_rate,
+            filter_radius,
+            resample_sr,
+            rms_mix_rate,
+            protect,
+            audio_format,
+            crepe_hop_length,
+            do_formant,
+            quefrency,
+            timbre,
+            min_pitch,
+            max_pitch,
+            f0_autotune,
+            hubert_model_path
+            )
+            if inference_info[0] == "Success.":
+                print("Inference ran successfully.")
+                print(inference_info[1])
+                print("Times:\nnpy: %.2fs f0: %.2fs infer: %.2fs\nTotal time: %.2fs" % (*inference_info[2],))
+            else:
+                print(f"An error occurred while processing.\n{inference_info[0]}")
+                return None
+            inferred_files.append(output_path)
+        print("Adjusting inferred audio lengths.")
+        adjusted_inferred_files = adjust_audio_lengths(nonsilent_files, inferred_files)
+        print("Combining silence and inferred audios.")
+        output_count = 1
+        while True:
+            output_path = os.path.join(os.getcwd(), "output", f"{os.path.splitext(os.path.basename(audio_path))[0]}{model_name}{f0_method.capitalize()}_{output_count}.{audio_format}")
+            if not os.path.exists(output_path):
+                break
+            output_count += 1
+        output_path = combine_silence_nonsilent(silence_files, adjusted_inferred_files, keep_silence, output_path)
+        [shutil.move(inferred_file, temp_dir) for inferred_file in inferred_files]
+        shutil.rmtree(temp_dir)
+    else:
+        inference_info, audio_data, output_path = vc.vc_single(
+            0,
+            audio_path,
+            f0_change,
+            f0_method,
+            index_path,
+            index_path,
+            index_rate,
+            filter_radius,
+            resample_sr,
+            rms_mix_rate,
+            protect,
+            audio_format,
+            crepe_hop_length,
+            do_formant,
+            quefrency,
+            timbre,
+            min_pitch,
+            max_pitch,
+            f0_autotune,
+            hubert_model_path
+        )
+        if inference_info[0] == "Success.":
+            print("Inference ran successfully.")
+            print(inference_info[1])
+            print("Times:\nnpy: %.2fs f0: %.2fs infer: %.2fs\nTotal time: %.2fs" % (*inference_info[2],))
+        else:
+            print(f"An error occurred while processing.\n{inference_info[0]}")
+            del configs, vc
+            gc.collect()
+            return inference_info[0]
+    
+    del configs, vc
+    gc.collect()
+    return output_path

lib/infer_libs/audio.py

@@ -0,0 +1,87 @@
+import numpy as np
+import av
+import ffmpeg
+import os
+import traceback
+import sys
+import subprocess
+
+platform_stft_mapping = {
+    'linux': os.path.join(os.getcwd(), 'stftpitchshift'),
+    'darwin': os.path.join(os.getcwd(), 'stftpitchshift'),
+    'win32': os.path.join(os.getcwd(), 'stftpitchshift.exe'),
+}
+
+stft = platform_stft_mapping.get(sys.platform)
+
+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, DoFormant=False, Quefrency=1.0, Timbre=1.0):
+    formanted = False
+    file = file.strip(' \n"')
+    if not os.path.exists(file):
+        raise RuntimeError(
+            "Wrong audio path, that does not exist."
+        )
+
+    try:
+        if DoFormant:
+            print("Starting formant shift. Please wait as this process takes a while.")
+            formanted_file = f"{os.path.splitext(os.path.basename(file))[0]}_formanted{os.path.splitext(os.path.basename(file))[1]}"
+            command = (
+                f'{stft} -i "{file}" -q "{Quefrency}" '
+                f'-t "{Timbre}" -o "{formanted_file}"'
+            )
+            subprocess.run(command, shell=True)
+            file = formanted_file
+            print(f"Formanted {file}\n")
+
+        # 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 = (
+            file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        )  # Prevent small white copy path head and tail with spaces and " and return
+        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)
+        )
+
+        return np.frombuffer(out, np.float32).flatten()
+
+    except Exception as e:
+        raise RuntimeError(f"Failed to load audio: {e}")
+
+def check_audio_duration(file):
+    try:
+        file = file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+
+        probe = ffmpeg.probe(file)
+
+        duration = float(probe['streams'][0]['duration'])
+
+        if duration < 0.76:
+            print(
+                f"Audio file, {file.split('/')[-1]}, under ~0.76s detected - file is too short. Target at least 1-2s for best results."
+            )
+            return False
+
+        return True
+    except Exception as e:
+        raise RuntimeError(f"Failed to check audio duration: {e}")

lib/infer_libs/fcpe.py

@@ -0,0 +1,873 @@
+from typing import Union
+
+import torch.nn.functional as F
+import numpy as np
+import torch
+import torch.nn as nn
+from torch.nn.utils import weight_norm
+from torchaudio.transforms import Resample
+import os
+import librosa
+import soundfile as sf
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+import math
+from functools import partial
+
+from einops import rearrange, repeat
+from local_attention import LocalAttention
+from torch import nn
+
+os.environ["LRU_CACHE_CAPACITY"] = "3"
+
+def load_wav_to_torch(full_path, target_sr=None, return_empty_on_exception=False):
+    sampling_rate = None
+    try:
+        data, sampling_rate = sf.read(full_path, always_2d=True)# than soundfile.
+    except Exception as ex:
+        print(f"'{full_path}' failed to load.\nException:")
+        print(ex)
+        if return_empty_on_exception:
+            return [], sampling_rate or target_sr or 48000
+        else:
+            raise Exception(ex)
+    
+    if len(data.shape) > 1:
+        data = data[:, 0]
+        assert len(data) > 2# check duration of audio file is > 2 samples (because otherwise the slice operation was on the wrong dimension)
+    
+    if np.issubdtype(data.dtype, np.integer): # if audio data is type int
+        max_mag = -np.iinfo(data.dtype).min # maximum magnitude = min possible value of intXX
+    else: # if audio data is type fp32
+        max_mag = max(np.amax(data), -np.amin(data))
+        max_mag = (2**31)+1 if max_mag > (2**15) else ((2**15)+1 if max_mag > 1.01 else 1.0) # data should be either 16-bit INT, 32-bit INT or [-1 to 1] float32
+    
+    data = torch.FloatTensor(data.astype(np.float32))/max_mag
+    
+    if (torch.isinf(data) | torch.isnan(data)).any() and return_empty_on_exception:# resample will crash with inf/NaN inputs. return_empty_on_exception will return empty arr instead of except
+        return [], sampling_rate or target_sr or 48000
+    if target_sr is not None and sampling_rate != target_sr:
+        data = torch.from_numpy(librosa.core.resample(data.numpy(), orig_sr=sampling_rate, target_sr=target_sr))
+        sampling_rate = target_sr
+    
+    return data, sampling_rate
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
+
+def dynamic_range_decompression(x, C=1):
+    return np.exp(x) / C
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+class STFT():
+    def __init__(self, sr=22050, n_mels=80, n_fft=1024, win_size=1024, hop_length=256, fmin=20, fmax=11025, clip_val=1e-5):
+        self.target_sr = sr
+        
+        self.n_mels     = n_mels
+        self.n_fft      = n_fft
+        self.win_size   = win_size
+        self.hop_length = hop_length
+        self.fmin     = fmin
+        self.fmax     = fmax
+        self.clip_val = clip_val
+        self.mel_basis = {}
+        self.hann_window = {}
+    
+    def get_mel(self, y, keyshift=0, speed=1, center=False, train=False):
+        sampling_rate = self.target_sr
+        n_mels     = self.n_mels
+        n_fft      = self.n_fft
+        win_size   = self.win_size
+        hop_length = self.hop_length
+        fmin       = self.fmin
+        fmax       = self.fmax
+        clip_val   = self.clip_val
+        
+        factor = 2 ** (keyshift / 12)       
+        n_fft_new = int(np.round(n_fft * factor))
+        win_size_new = int(np.round(win_size * factor))
+        hop_length_new = int(np.round(hop_length * speed))
+        if not train:
+            mel_basis = self.mel_basis
+            hann_window = self.hann_window
+        else:
+            mel_basis = {}
+            hann_window = {}
+        
+        if torch.min(y) < -1.:
+            print('min value is ', torch.min(y))
+        if torch.max(y) > 1.:
+            print('max value is ', torch.max(y))
+        
+        mel_basis_key = str(fmax)+'_'+str(y.device)
+        if mel_basis_key not in mel_basis:
+            mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=n_mels, fmin=fmin, fmax=fmax)
+            mel_basis[mel_basis_key] = torch.from_numpy(mel).float().to(y.device)
+        
+        keyshift_key = str(keyshift)+'_'+str(y.device)
+        if keyshift_key not in hann_window:
+            hann_window[keyshift_key] = torch.hann_window(win_size_new).to(y.device)
+        
+        pad_left = (win_size_new - hop_length_new) //2
+        pad_right = max((win_size_new- hop_length_new + 1) //2, win_size_new - y.size(-1) - pad_left)
+        if pad_right < y.size(-1):
+            mode = 'reflect'
+        else:
+            mode = 'constant'
+        y = torch.nn.functional.pad(y.unsqueeze(1), (pad_left, pad_right), mode = mode)
+        y = y.squeeze(1)
+        
+        spec = torch.stft(y, n_fft_new, hop_length=hop_length_new, win_length=win_size_new, window=hann_window[keyshift_key],
+                          center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=True)                          
+        spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + (1e-9))
+        if keyshift != 0:
+            size = n_fft // 2 + 1
+            resize = spec.size(1)
+            if resize < size:
+                spec = F.pad(spec, (0, 0, 0, size-resize))
+            spec = spec[:, :size, :] * win_size / win_size_new   
+        spec = torch.matmul(mel_basis[mel_basis_key], spec)
+        spec = dynamic_range_compression_torch(spec, clip_val=clip_val)
+        return spec
+    
+    def __call__(self, audiopath):
+        audio, sr = load_wav_to_torch(audiopath, target_sr=self.target_sr)
+        spect = self.get_mel(audio.unsqueeze(0)).squeeze(0)
+        return spect
+
+stft = STFT()
+
+#import fast_transformers.causal_product.causal_product_cuda
+
+def softmax_kernel(data, *, projection_matrix, is_query, normalize_data=True, eps=1e-4, device = None):
+    b, h, *_ = data.shape
+    # (batch size, head, length, model_dim)
+
+    # normalize model dim
+    data_normalizer = (data.shape[-1] ** -0.25) if normalize_data else 1.
+
+    # what is ration?, projection_matrix.shape[0] --> 266
+    
+    ratio = (projection_matrix.shape[0] ** -0.5)
+
+    projection = repeat(projection_matrix, 'j d -> b h j d', b = b, h = h)
+    projection = projection.type_as(data)
+
+    #data_dash = w^T x
+    data_dash = torch.einsum('...id,...jd->...ij', (data_normalizer * data), projection)
+
+    
+    # diag_data = D**2 
+    diag_data = data ** 2
+    diag_data = torch.sum(diag_data, dim=-1)
+    diag_data = (diag_data / 2.0) * (data_normalizer ** 2)
+    diag_data = diag_data.unsqueeze(dim=-1)
+    
+    #print ()
+    if is_query:
+        data_dash = ratio * (
+            torch.exp(data_dash - diag_data -
+                    torch.max(data_dash, dim=-1, keepdim=True).values) + eps)
+    else:
+        data_dash = ratio * (
+            torch.exp(data_dash - diag_data + eps))#- torch.max(data_dash)) + eps)
+
+    return data_dash.type_as(data)
+
+def orthogonal_matrix_chunk(cols, qr_uniform_q = False, device = None):
+    unstructured_block = torch.randn((cols, cols), device = device)
+    q, r = torch.linalg.qr(unstructured_block.cpu(), mode='reduced')
+    q, r = map(lambda t: t.to(device), (q, r))
+
+    # proposed by @Parskatt
+    # to make sure Q is uniform https://arxiv.org/pdf/math-ph/0609050.pdf
+    if qr_uniform_q:
+        d = torch.diag(r, 0)
+        q *= d.sign()
+    return q.t()
+def exists(val):
+    return val is not None
+
+def empty(tensor):
+    return tensor.numel() == 0
+
+def default(val, d):
+    return val if exists(val) else d
+
+def cast_tuple(val):
+    return (val,) if not isinstance(val, tuple) else val
+
+class PCmer(nn.Module):
+    """The encoder that is used in the Transformer model."""
+    
+    def __init__(self, 
+                num_layers,
+                num_heads,
+                dim_model,
+                dim_keys,
+                dim_values,
+                residual_dropout,
+                attention_dropout):
+        super().__init__()
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.dim_model = dim_model
+        self.dim_values = dim_values
+        self.dim_keys = dim_keys
+        self.residual_dropout = residual_dropout
+        self.attention_dropout = attention_dropout
+
+        self._layers = nn.ModuleList([_EncoderLayer(self) for _ in range(num_layers)])
+        
+    #  METHODS  ########################################################################################################
+    
+    def forward(self, phone, mask=None):
+        
+        # apply all layers to the input
+        for (i, layer) in enumerate(self._layers):
+            phone = layer(phone, mask)
+        # provide the final sequence
+        return phone
+
+
+# ==================================================================================================================== #
+#  CLASS  _ E N C O D E R  L A Y E R                                                                                   #
+# ==================================================================================================================== #
+
+
+class _EncoderLayer(nn.Module):
+    """One layer of the encoder.
+    
+    Attributes:
+        attn: (:class:`mha.MultiHeadAttention`): The attention mechanism that is used to read the input sequence.
+        feed_forward (:class:`ffl.FeedForwardLayer`): The feed-forward layer on top of the attention mechanism.
+    """
+    
+    def __init__(self, parent: PCmer):
+        """Creates a new instance of ``_EncoderLayer``.
+        
+        Args:
+            parent (Encoder): The encoder that the layers is created for.
+        """
+        super().__init__()
+        
+        
+        self.conformer = ConformerConvModule(parent.dim_model)
+        self.norm = nn.LayerNorm(parent.dim_model)
+        self.dropout = nn.Dropout(parent.residual_dropout)
+        
+        # selfatt -> fastatt: performer!
+        self.attn = SelfAttention(dim = parent.dim_model,
+                                  heads = parent.num_heads,
+                                  causal = False)
+        
+    #  METHODS  ########################################################################################################
+
+    def forward(self, phone, mask=None):
+        
+        # compute attention sub-layer
+        phone = phone + (self.attn(self.norm(phone), mask=mask))
+        
+        phone = phone + (self.conformer(phone))
+        
+        return phone 
+
+def calc_same_padding(kernel_size):
+    pad = kernel_size // 2
+    return (pad, pad - (kernel_size + 1) % 2)
+
+# helper classes
+
+class Swish(nn.Module):
+    def forward(self, x):
+        return x * x.sigmoid()
+
+class Transpose(nn.Module):
+    def __init__(self, dims):
+        super().__init__()
+        assert len(dims) == 2, 'dims must be a tuple of two dimensions'
+        self.dims = dims
+
+    def forward(self, x):
+        return x.transpose(*self.dims)
+
+class GLU(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x):
+        out, gate = x.chunk(2, dim=self.dim)
+        return out * gate.sigmoid()
+
+class DepthWiseConv1d(nn.Module):
+    def __init__(self, chan_in, chan_out, kernel_size, padding):
+        super().__init__()
+        self.padding = padding
+        self.conv = nn.Conv1d(chan_in, chan_out, kernel_size, groups = chan_in)
+
+    def forward(self, x):
+        x = F.pad(x, self.padding)
+        return self.conv(x)
+
+class ConformerConvModule(nn.Module):
+    def __init__(
+        self,
+        dim,
+        causal = False,
+        expansion_factor = 2,
+        kernel_size = 31,
+        dropout = 0.):
+        super().__init__()
+
+        inner_dim = dim * expansion_factor
+        padding = calc_same_padding(kernel_size) if not causal else (kernel_size - 1, 0)
+
+        self.net = nn.Sequential(
+            nn.LayerNorm(dim),
+            Transpose((1, 2)),
+            nn.Conv1d(dim, inner_dim * 2, 1),
+            GLU(dim=1),
+            DepthWiseConv1d(inner_dim, inner_dim, kernel_size = kernel_size, padding = padding),
+            #nn.BatchNorm1d(inner_dim) if not causal else nn.Identity(),
+            Swish(),
+            nn.Conv1d(inner_dim, dim, 1),
+            Transpose((1, 2)),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+def linear_attention(q, k, v):
+    if v is None:
+        #print (k.size(), q.size())
+        out = torch.einsum('...ed,...nd->...ne', k, q)
+        return out
+
+    else:
+        k_cumsum = k.sum(dim = -2) 
+        #k_cumsum = k.sum(dim = -2)
+        D_inv = 1. / (torch.einsum('...nd,...d->...n', q, k_cumsum.type_as(q)) + 1e-8)
+
+        context = torch.einsum('...nd,...ne->...de', k, v)
+        #print ("TRUEEE: ", context.size(), q.size(), D_inv.size())
+        out = torch.einsum('...de,...nd,...n->...ne', context, q, D_inv)
+        return out
+
+def gaussian_orthogonal_random_matrix(nb_rows, nb_columns, scaling = 0, qr_uniform_q = False, device = None):
+    nb_full_blocks = int(nb_rows / nb_columns)
+    #print (nb_full_blocks)
+    block_list = []
+
+    for _ in range(nb_full_blocks):
+        q = orthogonal_matrix_chunk(nb_columns, qr_uniform_q = qr_uniform_q, device = device)
+        block_list.append(q)
+    # block_list[n] is a orthogonal matrix ... (model_dim * model_dim)
+    #print (block_list[0].size(), torch.einsum('...nd,...nd->...n', block_list[0], torch.roll(block_list[0],1,1)))
+    #print (nb_rows, nb_full_blocks, nb_columns)
+    remaining_rows = nb_rows - nb_full_blocks * nb_columns
+    #print (remaining_rows)
+    if remaining_rows > 0:
+        q = orthogonal_matrix_chunk(nb_columns, qr_uniform_q = qr_uniform_q, device = device)
+        #print (q[:remaining_rows].size())
+        block_list.append(q[:remaining_rows])
+
+    final_matrix = torch.cat(block_list)
+    
+    if scaling == 0:
+        multiplier = torch.randn((nb_rows, nb_columns), device = device).norm(dim = 1)
+    elif scaling == 1:
+        multiplier = math.sqrt((float(nb_columns))) * torch.ones((nb_rows,), device = device)
+    else:
+        raise ValueError(f'Invalid scaling {scaling}')
+
+    return torch.diag(multiplier) @ final_matrix
+
+class FastAttention(nn.Module):
+    def __init__(self, dim_heads, nb_features = None, ortho_scaling = 0, causal = False, generalized_attention = False, kernel_fn = nn.ReLU(), qr_uniform_q = False, no_projection = False):
+        super().__init__()
+        nb_features = default(nb_features, int(dim_heads * math.log(dim_heads)))
+
+        self.dim_heads = dim_heads
+        self.nb_features = nb_features
+        self.ortho_scaling = ortho_scaling
+
+        self.create_projection = partial(gaussian_orthogonal_random_matrix, nb_rows = self.nb_features, nb_columns = dim_heads, scaling = ortho_scaling, qr_uniform_q = qr_uniform_q)
+        projection_matrix = self.create_projection()
+        self.register_buffer('projection_matrix', projection_matrix)
+
+        self.generalized_attention = generalized_attention
+        self.kernel_fn = kernel_fn
+
+        # if this is turned on, no projection will be used
+        # queries and keys will be softmax-ed as in the original efficient attention paper
+        self.no_projection = no_projection
+
+        self.causal = causal
+
+    @torch.no_grad()
+    def redraw_projection_matrix(self):
+        projections = self.create_projection()
+        self.projection_matrix.copy_(projections)
+        del projections
+
+    def forward(self, q, k, v):
+        device = q.device
+
+        if self.no_projection:
+            q = q.softmax(dim = -1)
+            k = torch.exp(k) if self.causal else k.softmax(dim = -2)
+        else:
+            create_kernel = partial(softmax_kernel, projection_matrix = self.projection_matrix, device = device)
+            
+            q = create_kernel(q, is_query = True)
+            k = create_kernel(k, is_query = False)
+
+        attn_fn = linear_attention if not self.causal else self.causal_linear_fn
+        if v is None:
+            out = attn_fn(q, k, None)
+            return out
+        else:
+            out = attn_fn(q, k, v)
+            return out
+class SelfAttention(nn.Module):
+    def __init__(self, dim, causal = False, heads = 8, dim_head = 64, local_heads = 0, local_window_size = 256, nb_features = None, feature_redraw_interval = 1000, generalized_attention = False, kernel_fn = nn.ReLU(), qr_uniform_q = False, dropout = 0., no_projection = False):
+        super().__init__()
+        assert dim % heads == 0, 'dimension must be divisible by number of heads'
+        dim_head = default(dim_head, dim // heads)
+        inner_dim = dim_head * heads
+        self.fast_attention = FastAttention(dim_head, nb_features, causal = causal, generalized_attention = generalized_attention, kernel_fn = kernel_fn, qr_uniform_q = qr_uniform_q, no_projection = no_projection)
+
+        self.heads = heads
+        self.global_heads = heads - local_heads
+        self.local_attn = LocalAttention(window_size = local_window_size, causal = causal, autopad = True, dropout = dropout, look_forward = int(not causal), rel_pos_emb_config = (dim_head, local_heads)) if local_heads > 0 else None
+
+        #print (heads, nb_features, dim_head)
+        #name_embedding = torch.zeros(110, heads, dim_head, dim_head)
+        #self.name_embedding = nn.Parameter(name_embedding, requires_grad=True)
+        
+
+        self.to_q = nn.Linear(dim, inner_dim)
+        self.to_k = nn.Linear(dim, inner_dim)
+        self.to_v = nn.Linear(dim, inner_dim)
+        self.to_out = nn.Linear(inner_dim, dim)
+        self.dropout = nn.Dropout(dropout)
+
+    @torch.no_grad()
+    def redraw_projection_matrix(self):
+        self.fast_attention.redraw_projection_matrix()
+        #torch.nn.init.zeros_(self.name_embedding)
+        #print (torch.sum(self.name_embedding))
+    def forward(self, x, context = None, mask = None, context_mask = None, name=None, inference=False, **kwargs):
+        _, _, _, h, gh = *x.shape, self.heads, self.global_heads
+        
+        cross_attend = exists(context)
+
+        context = default(context, x)
+        context_mask = default(context_mask, mask) if not cross_attend else context_mask
+        #print (torch.sum(self.name_embedding))
+        q, k, v = self.to_q(x), self.to_k(context), self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), (q, k, v))
+        (q, lq), (k, lk), (v, lv) = map(lambda t: (t[:, :gh], t[:, gh:]), (q, k, v))
+
+        attn_outs = []
+        #print (name)
+        #print (self.name_embedding[name].size())
+        if not empty(q):
+            if exists(context_mask):
+                global_mask = context_mask[:, None, :, None]
+                v.masked_fill_(~global_mask, 0.)
+            if cross_attend:
+                pass
+                #print (torch.sum(self.name_embedding))
+                #out = self.fast_attention(q,self.name_embedding[name],None)
+                #print (torch.sum(self.name_embedding[...,-1:]))
+            else:
+                out = self.fast_attention(q, k, v)
+            attn_outs.append(out)
+
+        if not empty(lq):
+            assert not cross_attend, 'local attention is not compatible with cross attention'
+            out = self.local_attn(lq, lk, lv, input_mask = mask)
+            attn_outs.append(out)
+
+        out = torch.cat(attn_outs, dim = 1)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        out =  self.to_out(out)
+        return self.dropout(out)
+
+def l2_regularization(model, l2_alpha):
+    l2_loss = []
+    for module in model.modules():
+        if type(module) is nn.Conv2d:
+            l2_loss.append((module.weight ** 2).sum() / 2.0)
+    return l2_alpha * sum(l2_loss)
+
+
+class FCPEModel(nn.Module):
+    def __init__(
+            self,
+            input_channel=128,
+            out_dims=360,
+            n_layers=12,
+            n_chans=512,
+            use_siren=False,
+            use_full=False,
+            loss_mse_scale=10,
+            loss_l2_regularization=False,
+            loss_l2_regularization_scale=1,
+            loss_grad1_mse=False,
+            loss_grad1_mse_scale=1,
+            f0_max=1975.5,
+            f0_min=32.70,
+            confidence=False,
+            threshold=0.05,
+            use_input_conv=True
+    ):
+        super().__init__()
+        if use_siren is True:
+            raise ValueError("Siren is not supported yet.")
+        if use_full is True:
+            raise ValueError("Full model is not supported yet.")
+
+        self.loss_mse_scale = loss_mse_scale if (loss_mse_scale is not None) else 10
+        self.loss_l2_regularization = loss_l2_regularization if (loss_l2_regularization is not None) else False
+        self.loss_l2_regularization_scale = loss_l2_regularization_scale if (loss_l2_regularization_scale
+                                                                             is not None) else 1
+        self.loss_grad1_mse = loss_grad1_mse if (loss_grad1_mse is not None) else False
+        self.loss_grad1_mse_scale = loss_grad1_mse_scale if (loss_grad1_mse_scale is not None) else 1
+        self.f0_max = f0_max if (f0_max is not None) else 1975.5
+        self.f0_min = f0_min if (f0_min is not None) else 32.70
+        self.confidence = confidence if (confidence is not None) else False
+        self.threshold = threshold if (threshold is not None) else 0.05
+        self.use_input_conv = use_input_conv if (use_input_conv is not None) else True
+
+        self.cent_table_b = torch.Tensor(
+            np.linspace(self.f0_to_cent(torch.Tensor([f0_min]))[0], self.f0_to_cent(torch.Tensor([f0_max]))[0],
+                        out_dims))
+        self.register_buffer("cent_table", self.cent_table_b)
+
+        # conv in stack
+        _leaky = nn.LeakyReLU()
+        self.stack = nn.Sequential(
+            nn.Conv1d(input_channel, n_chans, 3, 1, 1),
+            nn.GroupNorm(4, n_chans),
+            _leaky,
+            nn.Conv1d(n_chans, n_chans, 3, 1, 1))
+
+        # transformer
+        self.decoder = PCmer(
+            num_layers=n_layers,
+            num_heads=8,
+            dim_model=n_chans,
+            dim_keys=n_chans,
+            dim_values=n_chans,
+            residual_dropout=0.1,
+            attention_dropout=0.1)
+        self.norm = nn.LayerNorm(n_chans)
+
+        # out
+        self.n_out = out_dims
+        self.dense_out = weight_norm(
+            nn.Linear(n_chans, self.n_out))
+
+    def forward(self, mel, infer=True, gt_f0=None, return_hz_f0=False, cdecoder = "local_argmax"):
+        """
+        input:
+            B x n_frames x n_unit
+        return:
+            dict of B x n_frames x feat
+        """
+        if cdecoder == "argmax":
+            self.cdecoder = self.cents_decoder
+        elif cdecoder == "local_argmax":
+            self.cdecoder = self.cents_local_decoder
+        if self.use_input_conv:
+            x = self.stack(mel.transpose(1, 2)).transpose(1, 2)
+        else:
+            x = mel
+        x = self.decoder(x)
+        x = self.norm(x)
+        x = self.dense_out(x)  # [B,N,D]
+        x = torch.sigmoid(x)
+        if not infer:
+            gt_cent_f0 = self.f0_to_cent(gt_f0)  # mel f0  #[B,N,1]
+            gt_cent_f0 = self.gaussian_blurred_cent(gt_cent_f0)  # #[B,N,out_dim]
+            loss_all = self.loss_mse_scale * F.binary_cross_entropy(x, gt_cent_f0)  # bce loss
+            # l2 regularization
+            if self.loss_l2_regularization:
+                loss_all = loss_all + l2_regularization(model=self, l2_alpha=self.loss_l2_regularization_scale)
+            x = loss_all
+        if infer:
+            x = self.cdecoder(x)
+            x = self.cent_to_f0(x)
+            if not return_hz_f0:
+                x = (1 + x / 700).log()
+        return x
+
+    def cents_decoder(self, y, mask=True):
+        B, N, _ = y.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        rtn = torch.sum(ci * y, dim=-1, keepdim=True) / torch.sum(y, dim=-1, keepdim=True)  # cents: [B,N,1]
+        if mask:
+            confident = torch.max(y, dim=-1, keepdim=True)[0]
+            confident_mask = torch.ones_like(confident)
+            confident_mask[confident <= self.threshold] = float("-INF")
+            rtn = rtn * confident_mask
+        if self.confidence:
+            return rtn, confident
+        else:
+            return rtn
+        
+    def cents_local_decoder(self, y, mask=True):
+        B, N, _ = y.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        confident, max_index = torch.max(y, dim=-1, keepdim=True)
+        local_argmax_index = torch.arange(0,9).to(max_index.device) + (max_index - 4)
+        local_argmax_index[local_argmax_index<0] = 0
+        local_argmax_index[local_argmax_index>=self.n_out] = self.n_out - 1
+        ci_l = torch.gather(ci,-1,local_argmax_index)
+        y_l = torch.gather(y,-1,local_argmax_index)
+        rtn = torch.sum(ci_l * y_l, dim=-1, keepdim=True) / torch.sum(y_l, dim=-1, keepdim=True)  # cents: [B,N,1]
+        if mask:
+            confident_mask = torch.ones_like(confident)
+            confident_mask[confident <= self.threshold] = float("-INF")
+            rtn = rtn * confident_mask
+        if self.confidence:
+            return rtn, confident
+        else:
+            return rtn
+
+    def cent_to_f0(self, cent):
+        return 10. * 2 ** (cent / 1200.)
+
+    def f0_to_cent(self, f0):
+        return 1200. * torch.log2(f0 / 10.)
+
+    def gaussian_blurred_cent(self, cents):  # cents: [B,N,1]
+        mask = (cents > 0.1) & (cents < (1200. * np.log2(self.f0_max / 10.)))
+        B, N, _ = cents.size()
+        ci = self.cent_table[None, None, :].expand(B, N, -1)
+        return torch.exp(-torch.square(ci - cents) / 1250) * mask.float()
+
+
+class FCPEInfer:
+    def __init__(self, model_path, device=None, dtype=torch.float32):
+        if device is None:
+            device = 'cuda' if torch.cuda.is_available() else 'cpu'
+        self.device = device
+        ckpt = torch.load(model_path, map_location=torch.device(self.device))
+        self.args = DotDict(ckpt["config"])
+        self.dtype = dtype
+        model = FCPEModel(
+            input_channel=self.args.model.input_channel,
+            out_dims=self.args.model.out_dims,
+            n_layers=self.args.model.n_layers,
+            n_chans=self.args.model.n_chans,
+            use_siren=self.args.model.use_siren,
+            use_full=self.args.model.use_full,
+            loss_mse_scale=self.args.loss.loss_mse_scale,
+            loss_l2_regularization=self.args.loss.loss_l2_regularization,
+            loss_l2_regularization_scale=self.args.loss.loss_l2_regularization_scale,
+            loss_grad1_mse=self.args.loss.loss_grad1_mse,
+            loss_grad1_mse_scale=self.args.loss.loss_grad1_mse_scale,
+            f0_max=self.args.model.f0_max,
+            f0_min=self.args.model.f0_min,
+            confidence=self.args.model.confidence,
+        )
+        model.to(self.device).to(self.dtype)
+        model.load_state_dict(ckpt['model'])
+        model.eval()
+        self.model = model
+        self.wav2mel = Wav2Mel(self.args, dtype=self.dtype, device=self.device)
+
+    @torch.no_grad()
+    def __call__(self, audio, sr, threshold=0.05):
+        self.model.threshold = threshold
+        audio = audio[None,:]
+        mel = self.wav2mel(audio=audio, sample_rate=sr).to(self.dtype)
+        f0 = self.model(mel=mel, infer=True, return_hz_f0=True)
+        return f0
+
+
+class Wav2Mel:
+
+    def __init__(self, args, device=None, dtype=torch.float32):
+        # self.args = args
+        self.sampling_rate = args.mel.sampling_rate
+        self.hop_size = args.mel.hop_size
+        if device is None:
+            device = 'cuda' if torch.cuda.is_available() else 'cpu'
+        self.device = device
+        self.dtype = dtype
+        self.stft = STFT(
+            args.mel.sampling_rate,
+            args.mel.num_mels,
+            args.mel.n_fft,
+            args.mel.win_size,
+            args.mel.hop_size,
+            args.mel.fmin,
+            args.mel.fmax
+        )
+        self.resample_kernel = {}
+
+    def extract_nvstft(self, audio, keyshift=0, train=False):
+        mel = self.stft.get_mel(audio, keyshift=keyshift, train=train).transpose(1, 2)  # B, n_frames, bins
+        return mel
+
+    def extract_mel(self, audio, sample_rate, keyshift=0, train=False):
+        audio = audio.to(self.dtype).to(self.device)
+        # resample
+        if sample_rate == self.sampling_rate:
+            audio_res = audio
+        else:
+            key_str = str(sample_rate)
+            if key_str not in self.resample_kernel:
+                self.resample_kernel[key_str] = Resample(sample_rate, self.sampling_rate, lowpass_filter_width=128)
+            self.resample_kernel[key_str] = self.resample_kernel[key_str].to(self.dtype).to(self.device)
+            audio_res = self.resample_kernel[key_str](audio)
+
+        # extract
+        mel = self.extract_nvstft(audio_res, keyshift=keyshift, train=train)  # B, n_frames, bins
+        n_frames = int(audio.shape[1] // self.hop_size) + 1
+        if n_frames > int(mel.shape[1]):
+            mel = torch.cat((mel, mel[:, -1:, :]), 1)
+        if n_frames < int(mel.shape[1]):
+            mel = mel[:, :n_frames, :]
+        return mel
+
+    def __call__(self, audio, sample_rate, keyshift=0, train=False):
+        return self.extract_mel(audio, sample_rate, keyshift=keyshift, train=train)
+
+
+class DotDict(dict):
+    def __getattr__(*args):
+        val = dict.get(*args)
+        return DotDict(val) if type(val) is dict else val
+
+    __setattr__ = dict.__setitem__
+    __delattr__ = dict.__delitem__
+
+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
+
+class FCPE(F0Predictor):
+    def __init__(self, model_path, hop_length=512, f0_min=50, f0_max=1100, dtype=torch.float32, device=None, sampling_rate=44100,
+                 threshold=0.05):
+        self.fcpe = FCPEInfer(model_path, device=device, dtype=dtype)
+        self.hop_length = hop_length
+        self.f0_min = f0_min
+        self.f0_max = f0_max
+        if device is None:
+            self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
+        else:
+            self.device = device
+        self.threshold = threshold
+        self.sampling_rate = sampling_rate
+        self.dtype = dtype
+        self.name = "fcpe"
+
+    def repeat_expand(
+            self, content: Union[torch.Tensor, np.ndarray], target_len: int, mode: str = "nearest"
+    ):
+        ndim = content.ndim
+
+        if content.ndim == 1:
+            content = content[None, None]
+        elif content.ndim == 2:
+            content = content[None]
+
+        assert content.ndim == 3
+
+        is_np = isinstance(content, np.ndarray)
+        if is_np:
+            content = torch.from_numpy(content)
+
+        results = torch.nn.functional.interpolate(content, size=target_len, mode=mode)
+
+        if is_np:
+            results = results.numpy()
+
+        if ndim == 1:
+            return results[0, 0]
+        elif ndim == 2:
+            return results[0]
+
+    def post_process(self, x, sampling_rate, f0, pad_to):
+        if isinstance(f0, np.ndarray):
+            f0 = torch.from_numpy(f0).float().to(x.device)
+
+        if pad_to is None:
+            return f0
+
+        f0 = self.repeat_expand(f0, pad_to)
+
+        vuv_vector = torch.zeros_like(f0)
+        vuv_vector[f0 > 0.0] = 1.0
+        vuv_vector[f0 <= 0.0] = 0.0
+
+        # 去掉0频率, 并线性插值
+        nzindex = torch.nonzero(f0).squeeze()
+        f0 = torch.index_select(f0, dim=0, index=nzindex).cpu().numpy()
+        time_org = self.hop_length / sampling_rate * nzindex.cpu().numpy()
+        time_frame = np.arange(pad_to) * self.hop_length / sampling_rate
+
+        vuv_vector = F.interpolate(vuv_vector[None, None, :], size=pad_to)[0][0]
+
+        if f0.shape[0] <= 0:
+            return torch.zeros(pad_to, dtype=torch.float, device=x.device).cpu().numpy(), vuv_vector.cpu().numpy()
+        if f0.shape[0] == 1:
+            return (torch.ones(pad_to, dtype=torch.float, device=x.device) * f0[
+                0]).cpu().numpy(), vuv_vector.cpu().numpy()
+
+        # 大概可以用 torch 重写?
+        f0 = np.interp(time_frame, time_org, f0, left=f0[0], right=f0[-1])
+        # vuv_vector = np.ceil(scipy.ndimage.zoom(vuv_vector,pad_to/len(vuv_vector),order = 0))
+
+        return f0, vuv_vector.cpu().numpy()
+
+    def compute_f0(self, wav, p_len=None):
+        x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
+        if p_len is None:
+            print("fcpe 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"
+        f0 = self.fcpe(x, sr=self.sampling_rate, threshold=self.threshold)[0,:,0]
+        if torch.all(f0 == 0):
+            rtn = f0.cpu().numpy() if p_len is None else np.zeros(p_len)
+            return rtn, rtn
+        return self.post_process(x, self.sampling_rate, f0, p_len)[0]
+
+    def compute_f0_uv(self, wav, p_len=None):
+        x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
+        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"
+        f0 = self.fcpe(x, sr=self.sampling_rate, threshold=self.threshold)[0,:,0]
+        if torch.all(f0 == 0):
+            rtn = f0.cpu().numpy() if p_len is None else np.zeros(p_len)
+            return rtn, rtn
+        return self.post_process(x, self.sampling_rate, f0, p_len)

lib/infer_libs/infer_pack/attentions.py

@@ -0,0 +1,414 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from lib.infer_libs.infer_pack import commons
+from lib.infer_libs.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().__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.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
+        for i in range(self.n_layers):
+            y = self.attn_layers[i](x, x, 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 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().__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().__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, c, attn_mask=None):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(self, query, key, value, mask=None):
+        # reshape [b, d, t] -> [b, n_h, t, d_k]
+        b, d, t_s, t_t = (*key.size(), 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):
+        max_relative_position = 2 * self.window_size + 1
+        # Pad first before slice to avoid using cond ops.
+        pad_length = 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]]),
+            )
+        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]]))
+
+        # 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, length - 1]])
+        )
+
+        # 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, length - 1]])
+        )
+        x_flat = x.view([batch, heads, length**2 + 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], [length, 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=None,
+        causal=False,
+    ):
+        super().__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
+
+        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 forward(self, x, x_mask):
+        x = self.conv_1(self.padding(x * x_mask))
+        if self.activation == "gelu":
+            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))
+        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))
+        return x

lib/infer_libs/infer_pack/commons.py

@@ -0,0 +1,164 @@
+import math
+import torch
+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 shift_1d(x):
+    x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+    return x
+
+
+def sequence_mask(length, max_length=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

lib/infer_libs/infer_pack/models.py

@@ -0,0 +1,1174 @@
+import math
+import logging
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import Conv1d, Conv2d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, spectral_norm, weight_norm
+
+from lib.infer_libs.infer_pack import attentions, commons, modules
+from lib.infer_libs.infer_pack.commons import get_padding, init_weights
+has_xpu = bool(hasattr(torch, "xpu") and torch.xpu.is_available())
+
+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 forward(self, f0, upp):
+        """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[:, None].transpose(1, 2)
+            f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
+            # fundamental component
+            f0_buf[:, :, 0] = f0[:, :, 0]
+            for idx in np.arange(self.harmonic_num):
+                f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (
+                    idx + 2
+                )  # idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
+            rad_values = (f0_buf / self.sampling_rate) % 1  ###%1意味着n_har的乘积无法后处理优化
+            rand_ini = torch.rand(
+                f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device
+            )
+            rand_ini[:, 0] = 0
+            rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+            tmp_over_one = torch.cumsum(rad_values, 1)  # % 1  #####%1意味着后面的cumsum无法再优化
+            tmp_over_one *= upp
+            tmp_over_one = F.interpolate(
+                tmp_over_one.transpose(2, 1),
+                scale_factor=upp,
+                mode="linear",
+                align_corners=True,
+            ).transpose(2, 1)
+            rad_values = F.interpolate(
+                rad_values.transpose(2, 1), scale_factor=upp, mode="nearest"
+            ).transpose(
+                2, 1
+            )  #######
+            tmp_over_one %= 1
+            tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
+            cumsum_shift = torch.zeros_like(rad_values)
+            cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+            sine_waves = torch.sin(
+                torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi
+            )
+            sine_waves = sine_waves * self.sine_amp
+            uv = self._f02uv(f0)
+            uv = F.interpolate(
+                uv.transpose(2, 1), scale_factor=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):
+        if hasattr(self, "ddtype") == False:
+            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.ddtype:
+            sine_wavs = sine_wavs.to(self.ddtype)
+        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 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().__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
+        self.enc_p = TextEncoder256(
+            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)
+        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()
+        self.enc_q.remove_weight_norm()
+
+    def forward(
+        self, phone, phone_lengths, pitch, pitchf, y, y_lengths, ds
+    ):  # 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)
+
+    def infer(self, phone, phone_lengths, pitch, nsff0, sid, rate=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        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
+        if rate:
+            head = int(z_p.shape[2] * rate)
+            z_p = z_p[:, :, -head:]
+            x_mask = x_mask[:, :, -head:]
+            nsff0 = nsff0[:, -head:]
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec(z * x_mask, nsff0, g=g)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+class SynthesizerTrnMs768NSFsid(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().__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
+        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)
+        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()
+        self.enc_q.remove_weight_norm()
+
+    def forward(
+        self, phone, phone_lengths, pitch, pitchf, y, y_lengths, ds
+    ):  # 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)
+
+    def infer(self, phone, phone_lengths, pitch, nsff0, sid, rate=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        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
+        if rate:
+            head = int(z_p.shape[2] * rate)
+            z_p = z_p[:, :, -head:]
+            x_mask = x_mask[:, :, -head:]
+            nsff0 = nsff0[:, -head:]
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec(z * x_mask, nsff0, g=g)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+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().__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 = 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 = TextEncoder256(
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            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()
+        self.enc_q.remove_weight_norm()
+
+    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)
+
+    def infer(self, phone, phone_lengths, sid, rate=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        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
+        if rate:
+            head = int(z_p.shape[2] * rate)
+            z_p = z_p[:, :, -head:]
+            x_mask = x_mask[:, :, -head:]
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec(z * x_mask, g=g)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+class SynthesizerTrnMs768NSFsid_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().__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 = 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 = TextEncoder768(
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            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()
+        self.enc_q.remove_weight_norm()
+
+    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)
+
+    def infer(self, phone, phone_lengths, sid, rate=None):
+        g = self.emb_g(sid).unsqueeze(-1)
+        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
+        if rate:
+            head = int(z_p.shape[2] * rate)
+            z_p = z_p[:, :, -head:]
+            x_mask = x_mask[:, :, -head:]
+        z = self.flow(z_p, x_mask, g=g, reverse=True)
+        o = self.dec(z * x_mask, g=g)
+        return o, x_mask, (z, z_p, m_p, logs_p)
+
+
+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

lib/infer_libs/infer_pack/modules.py

@@ -0,0 +1,517 @@
+import math
+import torch
+from torch import nn
+from torch.nn import Conv1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, weight_norm
+
+from lib.infer_libs.infer_pack import commons
+from lib.infer_libs.infer_pack.commons import get_padding, init_weights
+from lib.infer_libs.infer_pack.transforms import piecewise_rational_quadratic_transform
+
+LRELU_SLOPE = 0.1
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__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().__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 = 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(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().__init__()
+        self.channels = channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+
+        self.drop = nn.Dropout(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=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 = p_dropout
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = nn.Dropout(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, x_mask, g=None, **kwargs):
+        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 range(self.n_layers):
+            x_in = self.in_layers[i](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 = self.res_skip_layers[i](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)
+
+
+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)
+
+    def forward(self, x, x_mask=None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, 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)
+
+
+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)
+
+    def forward(self, x, x_mask=None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, 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)
+
+
+class Log(nn.Module):
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        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):
+    def forward(self, x, *args, reverse=False, **kwargs):
+        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
+
+
+class ElementwiseAffine(nn.Module):
+    def __init__(self, channels):
+        super().__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().__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=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, x_mask, g=None, reverse=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
+
+    def remove_weight_norm(self):
+        self.enc.remove_weight_norm()
+
+
+class ConvFlow(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        n_layers,
+        num_bins=10,
+        tail_bound=5.0,
+    ):
+        super().__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, x_mask, g=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

lib/infer_libs/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

lib/infer_libs/rmvpe.py

@@ -0,0 +1,705 @@
+import os
+
+import numpy as np
+import torch
+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 lib.infer.modules.ipex import ipex_init
+        ipex_init()
+except Exception:
+    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__)
+
+
+###stft codes from https://github.com/pseeth/torch-stft/blob/master/torch_stft/util.py
+def window_sumsquare(
+    window,
+    n_frames,
+    hop_length=200,
+    win_length=800,
+    n_fft=800,
+    dtype=np.float32,
+    norm=None,
+):
+    """
+    # from librosa 0.6
+    Compute the sum-square envelope of a window function at a given hop length.
+    This is used to estimate modulation effects induced by windowing
+    observations in short-time fourier transforms.
+    Parameters
+    ----------
+    window : string, tuple, number, callable, or list-like
+        Window specification, as in `get_window`
+    n_frames : int > 0
+        The number of analysis frames
+    hop_length : int > 0
+        The number of samples to advance between frames
+    win_length : [optional]
+        The length of the window function.  By default, this matches `n_fft`.
+    n_fft : int > 0
+        The length of each analysis frame.
+    dtype : np.dtype
+        The data type of the output
+    Returns
+    -------
+    wss : np.ndarray, shape=`(n_fft + hop_length * (n_frames - 1))`
+        The sum-squared envelope of the window function
+    """
+    if win_length is None:
+        win_length = n_fft
+
+    n = n_fft + hop_length * (n_frames - 1)
+    x = np.zeros(n, dtype=dtype)
+
+    # Compute the squared window at the desired length
+    win_sq = get_window(window, win_length, fftbins=True)
+    win_sq = normalize(win_sq, norm=norm) ** 2
+    win_sq = pad_center(win_sq, n_fft)
+
+    # Fill the envelope
+    for i in range(n_frames):
+        sample = i * hop_length
+        x[sample : min(n, sample + n_fft)] += win_sq[: max(0, min(n_fft, n - sample))]
+    return x
+
+
+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)
+        #scale = self.filter_length / self.hop_length
+        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)
+        """
+        # num_batches = input_data.shape[0]
+        # num_samples = input_data.shape[-1]
+
+        # self.num_samples = num_samples
+
+        # similar to librosa, reflect-pad the input
+        # input_data = input_data.view(num_batches, 1, num_samples)
+        # print(1234,input_data.shape)
+        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)
+        # phase = torch.atan2(imag_part.data, real_part.data)
+
+        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(),
+        )
+        if in_channels != out_channels:
+            self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
+            self.is_shortcut = True
+        else:
+            self.is_shortcut = False
+
+    def forward(self, x):
+        if self.is_shortcut:
+            return self.conv(x) + self.shortcut(x)
+        else:
+            return self.conv(x) + 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):
+        concat_tensors = []
+        x = self.bn(x)
+        for i in range(self.n_encoders):
+            _, x = self.layers[i](x)
+            concat_tensors.append(_)
+        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 in range(self.n_blocks):
+            x = self.conv[i](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 in range(self.n_inters):
+            x = self.layers[i](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 in range(self.n_blocks):
+            x = self.conv2[i](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, concat_tensors):
+        for i in range(self.n_decoders):
+            x = self.layers[i](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):
+        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(
+                # "cpu"if(audio.device.type=="privateuseone") else audio.device
+                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 (audio.device.type == "privateuseone"):
+        #     magnitude=magnitude.to(audio.device)
+        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))
+        # print(log_mel_spec.device.type)
+        return log_mel_spec
+
+
+class RMVPE:
+    def __init__(self, model_path, is_half, device=None):
+        self.resample_kernel = {}
+        self.resample_kernel = {}
+        self.is_half = is_half
+        if device is None:
+            device = "cuda" 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:
+            model = E2E(4, 1, (2, 2))
+            ckpt = torch.load(model_path, map_location="cpu")
+            model.load_state_dict(ckpt)
+            model.eval()
+            if is_half == True:
+                model = model.half()
+            self.model = 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:
+                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()
+        mel = self.mel_extractor(
+            torch.from_numpy(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 infer_from_audio_with_pitch(self, audio, thred=0.03, f0_min=50, f0_max=1100):
+        t0 = ttime()
+        audio = torch.from_numpy(audio).float().to(self.device).unsqueeze(0)
+        mel = self.mel_extractor(audio, center=True)
+        t1 = ttime()
+        hidden = self.mel2hidden(mel)
+        t2 = ttime()
+        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)
+        f0[(f0 < f0_min) | (f0 > f0_max)] = 0  
+        t3 = ttime()
+        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)

lib/modules.py

@@ -0,0 +1,559 @@
+import os, sys
+import traceback
+import logging
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+logger = logging.getLogger(__name__)
+import numpy as np
+import soundfile as sf
+import torch
+from io import BytesIO
+from lib.infer_libs.audio import load_audio
+from lib.infer_libs.audio import wav2
+from lib.infer_libs.infer_pack.models import (
+    SynthesizerTrnMs256NSFsid,
+    SynthesizerTrnMs256NSFsid_nono,
+    SynthesizerTrnMs768NSFsid,
+    SynthesizerTrnMs768NSFsid_nono,
+)
+from lib.pipeline import Pipeline
+import time
+import glob
+from shutil import move
+from fairseq import checkpoint_utils
+
+sup_audioext = {
+    "wav",
+    "mp3",
+    "flac",
+    "ogg",
+    "opus",
+    "m4a",
+    "mp4",
+    "aac",
+    "alac",
+    "wma",
+    "aiff",
+    "webm",
+    "ac3",
+}
+
+def note_to_hz(note_name):
+    try:
+        SEMITONES = {'C': -9, 'C#': -8, 'D': -7, 'D#': -6, 'E': -5, 'F': -4, 'F#': -3, 'G': -2, 'G#': -1, 'A': 0, 'A#': 1, 'B': 2}
+        pitch_class, octave = note_name[:-1], int(note_name[-1])
+        semitone = SEMITONES[pitch_class]
+        note_number = 12 * (octave - 4) + semitone
+        frequency = 440.0 * (2.0 ** (1.0/12)) ** note_number
+        return frequency
+    except:
+        return None
+
+def load_hubert(hubert_model_path, config):
+    models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
+        [hubert_model_path],
+        suffix="",
+    )
+    hubert_model = models[0]
+    hubert_model = hubert_model.to(config.device)
+    if config.is_half:
+        hubert_model = hubert_model.half()
+    else:
+        hubert_model = hubert_model.float()
+    return hubert_model.eval()
+
+class VC:
+    def __init__(self, config):
+        self.n_spk = None
+        self.tgt_sr = None
+        self.net_g = None
+        self.pipeline = None
+        self.cpt = None
+        self.version = None
+        self.if_f0 = None
+        self.version = None
+        self.hubert_model = None
+
+        self.config = config
+
+    def get_vc(self, sid, *to_return_protect):
+        logger.info("Get sid: " + sid)
+
+        to_return_protect0 = {
+            "visible": self.if_f0 != 0,
+            "value": to_return_protect[0]
+            if self.if_f0 != 0 and to_return_protect
+            else 0.5,
+            "__type__": "update",
+        }
+        to_return_protect1 = {
+            "visible": self.if_f0 != 0,
+            "value": to_return_protect[1]
+            if self.if_f0 != 0 and to_return_protect
+            else 0.33,
+            "__type__": "update",
+        }
+
+        if sid == "" or sid == []:
+            if self.hubert_model is not None:  # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的
+                logger.info("Clean model cache")
+                del (
+                    self.net_g,
+                    self.n_spk,
+                    self.vc,
+                    self.hubert_model,
+                    self.tgt_sr,
+                )  # ,cpt
+                self.hubert_model = (
+                    self.net_g
+                ) = self.n_spk = self.vc = self.hubert_model = self.tgt_sr = None
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+                ###楼下不这么折腾清理不干净
+                self.if_f0 = self.cpt.get("f0", 1)
+                self.version = self.cpt.get("version", "v1")
+                if self.version == "v1":
+                    if self.if_f0 == 1:
+                        self.net_g = SynthesizerTrnMs256NSFsid(
+                            *self.cpt["config"], is_half=self.config.is_half
+                        )
+                    else:
+                        self.net_g = SynthesizerTrnMs256NSFsid_nono(*self.cpt["config"])
+                elif self.version == "v2":
+                    if self.if_f0 == 1:
+                        self.net_g = SynthesizerTrnMs768NSFsid(
+                            *self.cpt["config"], is_half=self.config.is_half
+                        )
+                    else:
+                        self.net_g = SynthesizerTrnMs768NSFsid_nono(*self.cpt["config"])
+                del self.net_g, self.cpt
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+            return (
+                {"visible": False, "__type__": "update"},
+                {
+                    "visible": True,
+                    "value": to_return_protect0,
+                    "__type__": "update",
+                },
+                {
+                    "visible": True,
+                    "value": to_return_protect1,
+                    "__type__": "update",
+                },
+                "",
+                "",
+            )
+        #person = f'{os.getenv("weight_root")}/{sid}'
+        person = f'{sid}'
+        #logger.info(f"Loading: {person}")
+        logger.info(f"Loading...")
+        self.cpt = torch.load(person, map_location="cpu")
+        self.tgt_sr = self.cpt["config"][-1]
+        self.cpt["config"][-3] = self.cpt["weight"]["emb_g.weight"].shape[0]  # n_spk
+        self.if_f0 = self.cpt.get("f0", 1)
+        self.version = self.cpt.get("version", "v1")
+
+        synthesizer_class = {
+            ("v1", 1): SynthesizerTrnMs256NSFsid,
+            ("v1", 0): SynthesizerTrnMs256NSFsid_nono,
+            ("v2", 1): SynthesizerTrnMs768NSFsid,
+            ("v2", 0): SynthesizerTrnMs768NSFsid_nono,
+        }
+
+        self.net_g = synthesizer_class.get(
+            (self.version, self.if_f0), SynthesizerTrnMs256NSFsid
+        )(*self.cpt["config"], is_half=self.config.is_half)
+
+        del self.net_g.enc_q
+
+        self.net_g.load_state_dict(self.cpt["weight"], strict=False)
+        self.net_g.eval().to(self.config.device)
+        if self.config.is_half:
+            self.net_g = self.net_g.half()
+        else:
+            self.net_g = self.net_g.float()
+
+        self.pipeline = Pipeline(self.tgt_sr, self.config)
+        n_spk = self.cpt["config"][-3]
+        #index = {"value": get_index_path_from_model(sid), "__type__": "update"}
+        #logger.info("Select index: " + index["value"])
+
+        return (
+            (
+                {"visible": False, "maximum": n_spk, "__type__": "update"},
+                to_return_protect0,
+                to_return_protect1
+            )
+            if to_return_protect
+            else {"visible": False, "maximum": n_spk, "__type__": "update"}
+        )
+
+    def vc_single_dont_save(
+        self,
+        sid,
+        input_audio_path1,
+        f0_up_key,
+        f0_method,
+        file_index,
+        file_index2,
+        index_rate,
+        filter_radius,
+        resample_sr,
+        rms_mix_rate,
+        protect,
+        crepe_hop_length,
+        do_formant,
+        quefrency,
+        timbre,
+        f0_min,
+        f0_max,
+        f0_autotune,
+        hubert_model_path = "assets/hubert/hubert_base.pt"
+    ):
+        """
+        Performs inference without saving
+    
+        Parameters:
+        - sid (int)
+        - input_audio_path1 (str)
+        - f0_up_key (int)
+        - f0_method (str)
+        - file_index (str)
+        - file_index2 (str)
+        - index_rate (float)
+        - filter_radius (int)
+        - resample_sr (int)
+        - rms_mix_rate (float)
+        - protect (float)
+        - crepe_hop_length (int)
+        - do_formant (bool)
+        - quefrency (float)
+        - timbre (float)
+        - f0_min (str)
+        - f0_max (str)
+        - f0_autotune (bool)
+        - hubert_model_path (str)
+
+        Returns:
+        Tuple(Tuple(status, index_info, times), Tuple(sr, data)):
+            - Tuple(status, index_info, times):
+                - status (str): either "Success." or an error
+                - index_info (str): index path if used
+                - times (list): [npy_time, f0_time, infer_time, total_time]
+            - Tuple(sr, data): Audio data results.
+        """
+        global total_time
+        total_time = 0
+        start_time = time.time()
+        
+        if not input_audio_path1:
+            return "You need to upload an audio", None
+        
+        if not os.path.exists(input_audio_path1):
+            return "Audio was not properly selected or doesn't exist", None
+        
+        f0_up_key = int(f0_up_key)
+        if not f0_min.isdigit():
+            f0_min = note_to_hz(f0_min)
+            if f0_min:
+                print(f"Converted Min pitch: freq - {f0_min}")
+            else:
+                f0_min = 50
+                print("Invalid minimum pitch note. Defaulting to 50hz.")
+        else:
+            f0_min = float(f0_min)
+        if not f0_max.isdigit():
+            f0_max = note_to_hz(f0_max)
+            if f0_max:
+                print(f"Converted Max pitch: freq - {f0_max}")
+            else:
+                f0_max = 1100
+                print("Invalid maximum pitch note. Defaulting to 1100hz.")
+        else:
+            f0_max = float(f0_max)
+        
+        try:
+            print(f"Attempting to load {input_audio_path1}....")
+            audio = load_audio(file=input_audio_path1,
+                               sr=16000,
+                               DoFormant=do_formant,
+                               Quefrency=quefrency,
+                               Timbre=timbre)
+            
+            audio_max = np.abs(audio).max() / 0.95
+            if audio_max > 1:
+                audio /= audio_max
+            times = [0, 0, 0]
+
+            if self.hubert_model is None:
+                self.hubert_model = load_hubert(hubert_model_path, self.config)
+
+            try:
+                self.if_f0 = self.cpt.get("f0", 1)
+            except NameError:
+                message = "Model was not properly selected"
+                print(message)
+                return message, None
+            
+            if file_index and not file_index == "" and isinstance(file_index, str):
+                file_index = file_index.strip(" ") \
+                .strip('"') \
+                .strip("\n") \
+                .strip('"') \
+                .strip(" ") \
+                .replace("trained", "added")
+            elif file_index2:
+                file_index = file_index2
+            else:
+                file_index = ""  
+
+            audio_opt = self.pipeline.pipeline(
+                self.hubert_model,
+                self.net_g,
+                sid,
+                audio,
+                input_audio_path1,
+                times,
+                f0_up_key,
+                f0_method,
+                file_index,
+                index_rate,
+                self.if_f0,
+                filter_radius,
+                self.tgt_sr,
+                resample_sr,
+                rms_mix_rate,
+                self.version,
+                protect,
+                crepe_hop_length,
+                f0_autotune,
+                f0_min=f0_min,
+                f0_max=f0_max                 
+            )
+
+            if self.tgt_sr != resample_sr >= 16000:
+                tgt_sr = resample_sr
+            else:
+                tgt_sr = self.tgt_sr
+            index_info = (
+                "Index: %s." % file_index
+                if isinstance(file_index, str) and os.path.exists(file_index)
+                else "Index not used."
+            )
+            end_time = time.time()
+            total_time = end_time - start_time
+            times.append(total_time)
+            return (
+                ("Success.", index_info, times),
+                (tgt_sr, audio_opt),
+            )
+        except:
+            info = traceback.format_exc()
+            logger.warn(info)
+            return (
+                (info, None, [None, None, None, None]),
+                (None, None)
+            )
+
+    def vc_single(
+        self,
+        sid,
+        input_audio_path1,
+        f0_up_key,
+        f0_method,
+        file_index,
+        file_index2,
+        index_rate,
+        filter_radius,
+        resample_sr,
+        rms_mix_rate,
+        protect,
+        format1,
+        crepe_hop_length,
+        do_formant,
+        quefrency,
+        timbre,
+        f0_min,
+        f0_max,
+        f0_autotune,
+        hubert_model_path = "assets/hubert/hubert_base.pt"
+    ):
+        """
+        Performs inference with saving
+    
+        Parameters:
+        - sid (int)
+        - input_audio_path1 (str)
+        - f0_up_key (int)
+        - f0_method (str)
+        - file_index (str)
+        - file_index2 (str)
+        - index_rate (float)
+        - filter_radius (int)
+        - resample_sr (int)
+        - rms_mix_rate (float)
+        - protect (float)
+        - format1 (str)
+        - crepe_hop_length (int)
+        - do_formant (bool)
+        - quefrency (float)
+        - timbre (float)
+        - f0_min (str)
+        - f0_max (str)
+        - f0_autotune (bool)
+        - hubert_model_path (str)
+
+        Returns:
+        Tuple(Tuple(status, index_info, times), Tuple(sr, data), output_path):
+            - Tuple(status, index_info, times):
+                - status (str): either "Success." or an error
+                - index_info (str): index path if used
+                - times (list): [npy_time, f0_time, infer_time, total_time]
+            - Tuple(sr, data): Audio data results.
+            - output_path (str): Audio results path
+        """
+        global total_time
+        total_time = 0
+        start_time = time.time()
+        
+        if not input_audio_path1:
+            return "You need to upload an audio", None, None
+        
+        if not os.path.exists(input_audio_path1):
+            return "Audio was not properly selected or doesn't exist", None, None
+
+        f0_up_key = int(f0_up_key)
+        if not f0_min.isdigit():
+            f0_min = note_to_hz(f0_min)
+            if f0_min:
+                print(f"Converted Min pitch: freq - {f0_min}")
+            else:
+                f0_min = 50
+                print("Invalid minimum pitch note. Defaulting to 50hz.")
+        else:
+            f0_min = float(f0_min)
+        if not f0_max.isdigit():
+            f0_max = note_to_hz(f0_max)
+            if f0_max:
+                print(f"Converted Max pitch: freq - {f0_max}")
+            else:
+                f0_max = 1100
+                print("Invalid maximum pitch note. Defaulting to 1100hz.")
+        else:
+            f0_max = float(f0_max)
+
+        try:
+            print(f"Attempting to load {input_audio_path1}...")
+            audio = load_audio(file=input_audio_path1,
+                               sr=16000,
+                               DoFormant=do_formant,
+                               Quefrency=quefrency,
+                               Timbre=timbre)
+            
+            audio_max = np.abs(audio).max() / 0.95
+            if audio_max > 1:
+                audio /= audio_max
+            times = [0, 0, 0]
+
+            if self.hubert_model is None:
+                self.hubert_model = load_hubert(hubert_model_path, self.config)
+
+            try:
+                self.if_f0 = self.cpt.get("f0", 1)
+            except NameError:
+                message = "Model was not properly selected"
+                print(message)
+                return message, None
+            if file_index and not file_index == "" and isinstance(file_index, str):
+                file_index = file_index.strip(" ") \
+                .strip('"') \
+                .strip("\n") \
+                .strip('"') \
+                .strip(" ") \
+                .replace("trained", "added")
+            elif file_index2:
+                file_index = file_index2
+            else:
+                file_index = ""
+
+            audio_opt = self.pipeline.pipeline(
+                self.hubert_model,
+                self.net_g,
+                sid,
+                audio,
+                input_audio_path1,
+                times,
+                f0_up_key,
+                f0_method,
+                file_index,
+                index_rate,
+                self.if_f0,
+                filter_radius,
+                self.tgt_sr,
+                resample_sr,
+                rms_mix_rate,
+                self.version,
+                protect,
+                crepe_hop_length,
+                f0_autotune,
+                f0_min=f0_min,
+                f0_max=f0_max                 
+            )
+
+            if self.tgt_sr != resample_sr >= 16000:
+                tgt_sr = resample_sr
+            else:
+                tgt_sr = self.tgt_sr
+            index_info = (
+                "Index: %s." % file_index
+                if isinstance(file_index, str) and os.path.exists(file_index)
+                else "Index not used."
+            )
+            
+            opt_root = os.path.join(os.getcwd(), "output")
+            os.makedirs(opt_root, exist_ok=True)
+            output_count = 1
+            
+            while True:
+                opt_filename = f"{os.path.splitext(os.path.basename(input_audio_path1))[0]}{os.path.basename(os.path.dirname(file_index))}{f0_method.capitalize()}_{output_count}.{format1}"
+                current_output_path = os.path.join(opt_root, opt_filename)
+                if not os.path.exists(current_output_path):
+                    break
+                output_count += 1
+            try:
+                if format1 in ["wav", "flac"]:
+                    sf.write(
+                        current_output_path,
+                        audio_opt,
+                        self.tgt_sr,
+                    )
+                else:
+                    with BytesIO() as wavf:
+                        sf.write(
+                            wavf,
+                            audio_opt,
+                            self.tgt_sr,
+                            format="wav"
+                        )
+                        wavf.seek(0, 0)
+                        with open(current_output_path, "wb") as outf:
+                                wav2(wavf, outf, format1)
+            except:
+                info = traceback.format_exc()
+            end_time = time.time()
+            total_time = end_time - start_time
+            times.append(total_time)
+            return (
+                ("Success.", index_info, times),
+                (tgt_sr, audio_opt),
+                current_output_path
+            )
+        except:
+            info = traceback.format_exc()
+            logger.warn(info)
+            return (
+                (info, None, [None, None, None, None]),
+                (None, None),
+                None
+            )

lib/pipeline.py

@@ -0,0 +1,773 @@
+import os
+import sys
+import gc
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+from functools import lru_cache
+from time import time as ttime
+from torch import Tensor
+import faiss
+import librosa
+import numpy as np
+import parselmouth
+import pyworld
+import torch.nn.functional as F
+from scipy import signal
+from tqdm import tqdm
+
+import random
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import re
+from functools import partial
+bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
+
+input_audio_path2wav = {}
+import torchcrepe  # Fork Feature. Crepe algo for training and preprocess
+from torchfcpe import spawn_bundled_infer_model
+import torch
+from lib.infer_libs.rmvpe import RMVPE
+from lib.infer_libs.fcpe import FCPE
+
+@lru_cache
+def cache_harvest_f0(input_audio_path, fs, f0max, f0min, frame_period):
+    audio = input_audio_path2wav[input_audio_path]
+    f0, t = pyworld.harvest(
+        audio,
+        fs=fs,
+        f0_ceil=f0max,
+        f0_floor=f0min,
+        frame_period=frame_period,
+    )
+    f0 = pyworld.stonemask(audio, f0, t, fs)
+    return f0
+
+
+def change_rms(data1, sr1, data2, sr2, rate):  # 1是输入音频，2是输出音频,rate是2的占比
+    # print(data1.max(),data2.max())
+    rms1 = librosa.feature.rms(
+        y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2
+    )  # 每半秒一个点
+    rms2 = librosa.feature.rms(y=data2, frame_length=sr2 // 2 * 2, hop_length=sr2 // 2)
+    rms1 = torch.from_numpy(rms1)
+    rms1 = F.interpolate(
+        rms1.unsqueeze(0), size=data2.shape[0], mode="linear"
+    ).squeeze()
+    rms2 = torch.from_numpy(rms2)
+    rms2 = F.interpolate(
+        rms2.unsqueeze(0), size=data2.shape[0], mode="linear"
+    ).squeeze()
+    rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-6)
+    data2 *= (
+        torch.pow(rms1, torch.tensor(1 - rate))
+        * torch.pow(rms2, torch.tensor(rate - 1))
+    ).numpy()
+    return data2
+
+
+class Pipeline(object):
+    def __init__(self, tgt_sr, config):
+        self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
+            config.x_pad,
+            config.x_query,
+            config.x_center,
+            config.x_max,
+            config.is_half,
+        )
+        self.sr = 16000  # hubert输入采样率
+        self.window = 160  # 每帧点数
+        self.t_pad = self.sr * self.x_pad  # 每条前后pad时间
+        self.t_pad_tgt = tgt_sr * self.x_pad
+        self.t_pad2 = self.t_pad * 2
+        self.t_query = self.sr * self.x_query  # 查询切点前后查询时间
+        self.t_center = self.sr * self.x_center  # 查询切点位置
+        self.t_max = self.sr * self.x_max  # 免查询时长阈值
+        self.device = config.device
+        self.model_rmvpe = RMVPE(os.environ["rmvpe_model_path"], is_half=self.is_half, device=self.device)
+
+        self.note_dict = [
+            65.41, 69.30, 73.42, 77.78, 82.41, 87.31,
+            92.50, 98.00, 103.83, 110.00, 116.54, 123.47,
+            130.81, 138.59, 146.83, 155.56, 164.81, 174.61,
+            185.00, 196.00, 207.65, 220.00, 233.08, 246.94,
+            261.63, 277.18, 293.66, 311.13, 329.63, 349.23,
+            369.99, 392.00, 415.30, 440.00, 466.16, 493.88,
+            523.25, 554.37, 587.33, 622.25, 659.25, 698.46,
+            739.99, 783.99, 830.61, 880.00, 932.33, 987.77,
+            1046.50, 1108.73, 1174.66, 1244.51, 1318.51, 1396.91,
+            1479.98, 1567.98, 1661.22, 1760.00, 1864.66, 1975.53,
+            2093.00, 2217.46, 2349.32, 2489.02, 2637.02, 2793.83,
+            2959.96, 3135.96, 3322.44, 3520.00, 3729.31, 3951.07
+        ]
+
+    # Fork Feature: Get the best torch device to use for f0 algorithms that require a torch device. Will return the type (torch.device)
+    def get_optimal_torch_device(self, index: int = 0) -> torch.device:
+        if torch.cuda.is_available():
+            return torch.device(
+                f"cuda:{index % torch.cuda.device_count()}"
+            )  # Very fast
+        elif torch.backends.mps.is_available():
+            return torch.device("mps")
+        return torch.device("cpu")
+
+    # Fork Feature: Compute f0 with the crepe method
+    def get_f0_crepe_computation(
+        self,
+        x,
+        f0_min,
+        f0_max,
+        p_len,
+        *args,  # 512 before. Hop length changes the speed that the voice jumps to a different dramatic pitch. Lower hop lengths means more pitch accuracy but longer inference time.
+        **kwargs,  # Either use crepe-tiny "tiny" or crepe "full". Default is full
+    ):
+        x = x.astype(
+            np.float32
+        )  # fixes the F.conv2D exception. We needed to convert double to float.
+        x /= np.quantile(np.abs(x), 0.999)
+        torch_device = self.get_optimal_torch_device()
+        audio = torch.from_numpy(x).to(torch_device, copy=True)
+        audio = torch.unsqueeze(audio, dim=0)
+        if audio.ndim == 2 and audio.shape[0] > 1:
+            audio = torch.mean(audio, dim=0, keepdim=True).detach()
+        audio = audio.detach()
+        hop_length = kwargs.get('crepe_hop_length', 160)
+        model = kwargs.get('model', 'full') 
+        print("Initiating prediction with a crepe_hop_length of: " + str(hop_length))
+        pitch: Tensor = torchcrepe.predict(
+            audio,
+            self.sr,
+            hop_length,
+            f0_min,
+            f0_max,
+            model,
+            batch_size=hop_length * 2,
+            device=torch_device,
+            pad=True,
+        )
+        p_len = p_len or x.shape[0] // hop_length
+        # Resize the pitch for final f0
+        source = np.array(pitch.squeeze(0).cpu().float().numpy())
+        source[source < 0.001] = np.nan
+        target = np.interp(
+            np.arange(0, len(source) * p_len, len(source)) / p_len,
+            np.arange(0, len(source)),
+            source,
+        )
+        f0 = np.nan_to_num(target)
+        return f0  # Resized f0
+    
+    def get_f0_official_crepe_computation(
+        self,
+        x,
+        f0_min,
+        f0_max,
+        *args,
+        **kwargs
+    ):
+        # Pick a batch size that doesn't cause memory errors on your gpu
+        batch_size = 512
+        # Compute pitch using first gpu
+        audio = torch.tensor(np.copy(x))[None].float()
+        model = kwargs.get('model', 'full')
+        f0, pd = torchcrepe.predict(
+            audio,
+            self.sr,
+            self.window,
+            f0_min,
+            f0_max,
+            model,
+            batch_size=batch_size,
+            device=self.device,
+            return_periodicity=True,
+        )
+        pd = torchcrepe.filter.median(pd, 3)
+        f0 = torchcrepe.filter.mean(f0, 3)
+        f0[pd < 0.1] = 0
+        f0 = f0[0].cpu().numpy()
+        return f0
+
+    # Fork Feature: Compute pYIN f0 method
+    def get_f0_pyin_computation(self, x, f0_min, f0_max):
+        y, sr = librosa.load(x, sr=self.sr, mono=True)
+        f0, _, _ = librosa.pyin(y, fmin=f0_min, fmax=f0_max, sr=self.sr)
+        f0 = f0[1:]  # Get rid of extra first frame
+        return f0
+
+    def get_rmvpe(self, x, *args, **kwargs):
+        if not hasattr(self, "model_rmvpe"):
+            from lib.infer.infer_libs.rmvpe import RMVPE
+            
+            logger.info(
+                f"Loading rmvpe model, {os.environ['rmvpe_model_path']}"
+            )
+            self.model_rmvpe = RMVPE(
+                os.environ["rmvpe_model_path"],
+                is_half=self.is_half,
+                device=self.device,
+            )
+        f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+
+        if "privateuseone" in str(self.device):  # clean ortruntime memory
+                del self.model_rmvpe.model
+                del self.model_rmvpe
+                logger.info("Cleaning ortruntime memory")
+            
+        return f0
+    
+
+    def get_pitch_dependant_rmvpe(self, x, f0_min=1, f0_max=40000, *args, **kwargs):
+        if not hasattr(self, "model_rmvpe"):
+            from lib.infer.infer_libs.rmvpe import RMVPE
+            
+            logger.info(
+                f"Loading rmvpe model, {os.environ['rmvpe_model_path']}"
+            )
+            self.model_rmvpe = RMVPE(
+                os.environ["rmvpe_model_path"],
+                is_half=self.is_half,
+                device=self.device,
+            )
+        f0 = self.model_rmvpe.infer_from_audio_with_pitch(x, thred=0.03, f0_min=f0_min, f0_max=f0_max)   
+        if "privateuseone" in str(self.device):  # clean ortruntime memory
+                del self.model_rmvpe.model
+                del self.model_rmvpe
+                logger.info("Cleaning ortruntime memory")
+            
+        return f0
+        
+    def get_fcpe(self, x, f0_min, f0_max, p_len, *args, **kwargs):
+        self.model_fcpe = FCPE(os.environ["fcpe_model_path"], f0_min=f0_min, f0_max=f0_max, dtype=torch.float32, device=self.device, sampling_rate=self.sr, threshold=0.03)
+        f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
+        del self.model_fcpe
+        gc.collect()
+        return f0
+
+    def get_torchfcpe(self, x, sr, f0_min, f0_max, p_len, *args, **kwargs):
+        self.model_torchfcpe = spawn_bundled_infer_model(device=self.device)
+        f0 = self.model_torchfcpe.infer(
+            torch.from_numpy(x).float().unsqueeze(0).unsqueeze(-1).to(self.device),
+            sr=sr,
+            decoder_mode="local_argmax",
+            threshold=0.006,
+            f0_min=f0_min,
+            f0_max=f0_max,
+            output_interp_target_length=p_len
+        )
+        return f0.squeeze().cpu().numpy()
+
+    def autotune_f0(self, f0):
+        autotuned_f0 = []
+        for freq in f0:
+            closest_notes = [x for x in self.note_dict if abs(x - freq) == min(abs(n - freq) for n in self.note_dict)]
+            autotuned_f0.append(random.choice(closest_notes))
+        return np.array(autotuned_f0, np.float64)
+
+
+    # Fork Feature: Acquire median hybrid f0 estimation calculation
+    def get_f0_hybrid_computation(
+        self,
+        methods_str,
+        input_audio_path,
+        x,
+        f0_min,
+        f0_max,
+        p_len,
+        filter_radius,
+        crepe_hop_length,
+        time_step,
+    ):
+        # Get various f0 methods from input to use in the computation stack
+        methods_str = re.search('hybrid\[(.+)\]', methods_str)
+        if methods_str:  # Ensure a match was found
+            methods = [method.strip() for method in methods_str.group(1).split('+')]
+        f0_computation_stack = []
+
+        print("Calculating f0 pitch estimations for methods: %s" % str(methods))
+        x = x.astype(np.float32)
+        x /= np.quantile(np.abs(x), 0.999)
+        # Get f0 calculations for all methods specified
+        for method in methods:
+            f0 = None
+            if method == "pm":
+                f0 = (
+                    parselmouth.Sound(x, self.sr)
+                    .to_pitch_ac(
+                        time_step=time_step / 1000,
+                        voicing_threshold=0.6,
+                        pitch_floor=f0_min,
+                        pitch_ceiling=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"
+                    )
+            elif method == "crepe":
+                f0 = self.get_f0_official_crepe_computation(x, f0_min, f0_max, model="full")
+                f0 = f0[1:]
+            elif method == "crepe-tiny":
+                f0 = self.get_f0_official_crepe_computation(x, f0_min, f0_max, model="tiny")
+                f0 = f0[1:]  # Get rid of extra first frame
+            elif method == "mangio-crepe":
+                f0 = self.get_f0_crepe_computation(
+                    x, f0_min, f0_max, p_len, crepe_hop_length=crepe_hop_length
+                )
+            elif method == "mangio-crepe-tiny":
+                f0 = self.get_f0_crepe_computation(
+                    x, f0_min, f0_max, p_len, crepe_hop_length=crepe_hop_length, model="tiny"
+                )
+            elif method == "harvest":
+                input_audio_path2wav[input_audio_path] = x.astype(np.double)
+                f0 = cache_harvest_f0(input_audio_path, self.sr, f0_max, f0_min, 10)
+                if filter_radius > 2:
+                    f0 = signal.medfilt(f0, 3)
+            elif method == "dio":
+                f0, t = pyworld.dio(
+                    x.astype(np.double),
+                    fs=self.sr,
+                    f0_ceil=f0_max,
+                    f0_floor=f0_min,
+                    frame_period=10,
+                )
+                f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sr)
+                f0 = signal.medfilt(f0, 3)
+                f0 = f0[1:]
+            elif method == "rmvpe":
+                f0 = self.get_rmvpe(x)
+                f0 = f0[1:]
+            elif method == "fcpe_legacy":
+                f0 = self.get_fcpe(x, f0_min=f0_min, f0_max=f0_max, p_len=p_len)
+            elif method == "fcpe":
+                f0 = self.get_torchfcpe(x, self.sr, f0_min, f0_max, p_len)
+            elif method == "pyin":
+                f0 = self.get_f0_pyin_computation(input_audio_path, f0_min, f0_max)
+            # Push method to the stack
+            f0_computation_stack.append(f0)
+
+        for fc in f0_computation_stack:
+            print(len(fc))
+
+        print("Calculating hybrid median f0 from the stack of: %s" % str(methods))
+        f0_median_hybrid = None
+        if len(f0_computation_stack) == 1:
+            f0_median_hybrid = f0_computation_stack[0]
+        else:
+            f0_median_hybrid = np.nanmedian(f0_computation_stack, axis=0)
+        return f0_median_hybrid
+    
+    def get_f0(
+        self,
+        input_audio_path,
+        x,
+        p_len,
+        f0_up_key,
+        f0_method,
+        filter_radius,
+        crepe_hop_length,
+        f0_autotune,
+        inp_f0=None,
+        f0_min=50,
+        f0_max=1100,
+    ):
+        global input_audio_path2wav
+        time_step = self.window / self.sr * 1000
+        f0_min = f0_min
+        f0_max = f0_max
+        f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+        f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+
+        if f0_method == "pm":
+            f0 = (
+                parselmouth.Sound(x, self.sr)
+                .to_pitch_ac(
+                    time_step=time_step / 1000,
+                    voicing_threshold=0.6,
+                    pitch_floor=f0_min,
+                    pitch_ceiling=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"
+                )
+        elif f0_method == "harvest":
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            f0 = cache_harvest_f0(input_audio_path, self.sr, f0_max, f0_min, 10)
+            if filter_radius > 2:
+                f0 = signal.medfilt(f0, 3)
+        elif f0_method == "dio":  # Potentially Buggy?
+            f0, t = pyworld.dio(
+                x.astype(np.double),
+                fs=self.sr,
+                f0_ceil=f0_max,
+                f0_floor=f0_min,
+                frame_period=10,
+            )
+            f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sr)
+            f0 = signal.medfilt(f0, 3)
+        elif f0_method == "crepe":
+            model = "full"
+            # Pick a batch size that doesn't cause memory errors on your gpu
+            batch_size = 512
+            # Compute pitch using first gpu
+            audio = torch.tensor(np.copy(x))[None].float()
+            f0, pd = torchcrepe.predict(
+                audio,
+                self.sr,
+                self.window,
+                f0_min,
+                f0_max,
+                model,
+                batch_size=batch_size,
+                device=self.device,
+                return_periodicity=True,
+            )
+            pd = torchcrepe.filter.median(pd, 3)
+            f0 = torchcrepe.filter.mean(f0, 3)
+            f0[pd < 0.1] = 0
+            f0 = f0[0].cpu().numpy()
+        elif f0_method == "crepe-tiny":
+            f0 = self.get_f0_official_crepe_computation(x, f0_min, f0_max, model="tiny")
+        elif f0_method == "mangio-crepe":
+            f0 = self.get_f0_crepe_computation(
+                x, f0_min, f0_max, p_len, crepe_hop_length=crepe_hop_length
+            )
+        elif f0_method == "mangio-crepe-tiny":
+            f0 = self.get_f0_crepe_computation(
+                x, f0_min, f0_max, p_len, crepe_hop_length=crepe_hop_length, model="tiny"
+            )
+        elif f0_method == "rmvpe":
+            if not hasattr(self, "model_rmvpe"):
+                from lib.infer.infer_libs.rmvpe import RMVPE
+
+                logger.info(
+                    f"Loading rmvpe model, {os.environ['rmvpe_model_path']}"
+                )
+                self.model_rmvpe = RMVPE(
+                    os.environ["rmvpe_model_path"],
+                    is_half=self.is_half,
+                    device=self.device,
+                )
+            f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
+
+            if "privateuseone" in str(self.device):  # clean ortruntime memory
+                del self.model_rmvpe.model
+                del self.model_rmvpe
+                logger.info("Cleaning ortruntime memory")
+        elif f0_method == "rmvpe+":
+            params = {'x': x, 'p_len': p_len, 'f0_up_key': f0_up_key, 'f0_min': f0_min, 
+                      'f0_max': f0_max, 'time_step': time_step, 'filter_radius': filter_radius, 
+                      'crepe_hop_length': crepe_hop_length, 'model': "full"
+                      }
+            f0 = self.get_pitch_dependant_rmvpe(**params)
+        elif f0_method == "pyin":
+            f0 = self.get_f0_pyin_computation(input_audio_path, f0_min, f0_max)
+        elif f0_method == "fcpe_legacy":
+            f0 = self.get_fcpe(x, f0_min=f0_min, f0_max=f0_max, p_len=p_len)
+        elif f0_method == "fcpe":
+            f0 = self.get_torchfcpe(x, self.sr, f0_min, f0_max, p_len)
+        elif "hybrid" in f0_method:
+            # Perform hybrid median pitch estimation
+            input_audio_path2wav[input_audio_path] = x.astype(np.double)
+            f0 = self.get_f0_hybrid_computation(
+                f0_method,
+                input_audio_path,
+                x,
+                f0_min,
+                f0_max,
+                p_len,
+                filter_radius,
+                crepe_hop_length,
+                time_step,
+            )
+        #print("Autotune:", f0_autotune)
+        if f0_autotune == True:
+            print("Autotune:", f0_autotune)
+            f0 = self.autotune_f0(f0)
+
+        f0 *= pow(2, f0_up_key / 12)
+        # with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
+        tf0 = self.sr // self.window  # 每秒f0点数
+        if inp_f0 is not None:
+            delta_t = np.round(
+                (inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
+            ).astype("int16")
+            replace_f0 = np.interp(
+                list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
+            )
+            shape = f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)].shape[0]
+            f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[
+                :shape
+            ]
+        # with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
+        f0bak = f0.copy()
+        f0_mel = 1127 * np.log(1 + f0 / 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
+        f0_coarse = np.rint(f0_mel).astype(np.int32)
+        return f0_coarse, f0bak  # 1-0
+
+    def vc(
+        self,
+        model,
+        net_g,
+        sid,
+        audio0,
+        pitch,
+        pitchf,
+        times,
+        index,
+        big_npy,
+        index_rate,
+        version,
+        protect,
+    ):  # ,file_index,file_big_npy
+        feats = torch.from_numpy(audio0)
+        if self.is_half:
+            feats = feats.half()
+        else:
+            feats = feats.float()
+        if feats.dim() == 2:  # double channels
+            feats = feats.mean(-1)
+        assert feats.dim() == 1, feats.dim()
+        feats = feats.view(1, -1)
+        padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
+
+        inputs = {
+            "source": feats.to(self.device),
+            "padding_mask": padding_mask,
+            "output_layer": 9 if version == "v1" else 12,
+        }
+        t0 = ttime()
+        with torch.no_grad():
+            logits = model.extract_features(**inputs)
+            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+        if protect < 0.5 and pitch is not None and pitchf is not None:
+            feats0 = feats.clone()
+        if (
+            not isinstance(index, type(None))
+            and not isinstance(big_npy, type(None))
+            and index_rate != 0
+        ):
+            npy = feats[0].cpu().numpy()
+            if self.is_half:
+                npy = npy.astype("float32")
+
+            # _, I = index.search(npy, 1)
+            # npy = big_npy[I.squeeze()]
+
+            score, ix = index.search(npy, k=8)
+            weight = np.square(1 / score)
+            weight /= weight.sum(axis=1, keepdims=True)
+            npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
+
+            if self.is_half:
+                npy = npy.astype("float16")
+            feats = (
+                torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
+                + (1 - index_rate) * feats
+            )
+
+        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+        if protect < 0.5 and pitch is not None and pitchf is not None:
+            feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(
+                0, 2, 1
+            )
+        t1 = ttime()
+        p_len = audio0.shape[0] // self.window
+        if feats.shape[1] < p_len:
+            p_len = feats.shape[1]
+            if pitch is not None and pitchf is not None:
+                pitch = pitch[:, :p_len]
+                pitchf = pitchf[:, :p_len]
+
+        if protect < 0.5 and pitch is not None and pitchf is not None:
+            pitchff = pitchf.clone()
+            pitchff[pitchf > 0] = 1
+            pitchff[pitchf < 1] = protect
+            pitchff = pitchff.unsqueeze(-1)
+            feats = feats * pitchff + feats0 * (1 - pitchff)
+            feats = feats.to(feats0.dtype)
+        p_len = torch.tensor([p_len], device=self.device).long()
+        with torch.no_grad():
+            hasp = pitch is not None and pitchf is not None
+            arg = (feats, p_len, pitch, pitchf, sid) if hasp else (feats, p_len, sid)
+            audio1 = (net_g.infer(*arg)[0][0, 0]).data.cpu().float().numpy()
+            del hasp, arg
+        del feats, p_len, padding_mask
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        t2 = ttime()
+        times[0] += t1 - t0
+        times[2] += t2 - t1
+        return audio1
+    def process_t(self, t, s, window, audio_pad, pitch, pitchf, times, index, big_npy, index_rate, version, protect, t_pad_tgt, if_f0, sid, model, net_g):
+        t = t // window * window
+        if if_f0 == 1:
+            return self.vc(
+                model,
+                net_g,
+                sid,
+                audio_pad[s : t + t_pad_tgt + window],
+                pitch[:, s // window : (t + t_pad_tgt) // window],
+                pitchf[:, s // window : (t + t_pad_tgt) // window],
+                times,
+                index,
+                big_npy,
+                index_rate,
+                version,
+                protect,
+            )[t_pad_tgt : -t_pad_tgt]
+        else:
+            return self.vc(
+                model,
+                net_g,
+                sid,
+                audio_pad[s : t + t_pad_tgt + window],
+                None,
+                None,
+                times,
+                index,
+                big_npy,
+                index_rate,
+                version,
+                protect,
+            )[t_pad_tgt : -t_pad_tgt]
+
+
+    def pipeline(
+        self,
+        model,
+        net_g,
+        sid,
+        audio,
+        input_audio_path,
+        times,
+        f0_up_key,
+        f0_method,
+        file_index,
+        index_rate,
+        if_f0,
+        filter_radius,
+        tgt_sr,
+        resample_sr,
+        rms_mix_rate,
+        version,
+        protect,
+        crepe_hop_length, 
+        f0_autotune, 
+        f0_min=50, 
+        f0_max=1100
+    ):
+        if (
+            file_index != ""
+            and isinstance(file_index, str)
+            # and file_big_npy != ""
+            # and os.path.exists(file_big_npy) == True
+            and os.path.exists(file_index)
+            and index_rate != 0
+        ):
+            try:
+                index = faiss.read_index(file_index)
+                # big_npy = np.load(file_big_npy)
+                big_npy = index.reconstruct_n(0, index.ntotal)
+            except:
+                traceback.print_exc()
+                index = big_npy = None
+        else:
+            index = big_npy = None
+        audio = signal.filtfilt(bh, ah, audio)
+        audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
+        opt_ts = []
+        if audio_pad.shape[0] > self.t_max:
+            audio_sum = np.zeros_like(audio)
+            for i in range(self.window):
+                audio_sum += audio_pad[i : i - self.window]
+            for t in range(self.t_center, audio.shape[0], self.t_center):
+                opt_ts.append(
+                    t
+                    - self.t_query
+                    + np.where(
+                        np.abs(audio_sum[t - self.t_query : t + self.t_query])
+                        == np.abs(audio_sum[t - self.t_query : t + self.t_query]).min()
+                    )[0][0]
+                )
+        s = 0
+        audio_opt = []
+        t = None
+        t1 = ttime()
+        audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
+        p_len = audio_pad.shape[0] // self.window
+        inp_f0 = None
+        
+        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
+        pitch, pitchf = None, None
+        if if_f0:
+            pitch, pitchf = self.get_f0(
+                input_audio_path,
+                audio_pad,
+                p_len,
+                f0_up_key,
+                f0_method,
+                filter_radius, 
+                crepe_hop_length, 
+                f0_autotune,
+                inp_f0, 
+                f0_min, 
+                f0_max
+            )
+            pitch = pitch[:p_len]
+            pitchf = pitchf[:p_len]
+            if "mps" not in str(self.device) or "xpu" not in str(self.device):
+                pitchf = pitchf.astype(np.float32)
+            pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
+            pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
+        t2 = ttime()
+        times[1] += t2 - t1
+
+        with tqdm(total=len(opt_ts), desc="Processing", unit="window") as pbar:
+            for i, t in enumerate(opt_ts):
+                t = t // self.window * self.window
+                start = s
+                end = t + self.t_pad2 + self.window
+                audio_slice = audio_pad[start:end]
+                pitch_slice = pitch[:, start // self.window:end // self.window] if if_f0 else None
+                pitchf_slice = pitchf[:, start // self.window:end // self.window] if if_f0 else None
+                audio_opt.append(self.vc(model, net_g, sid, audio_slice, pitch_slice, pitchf_slice, times, index, big_npy, index_rate, version, protect)[self.t_pad_tgt : -self.t_pad_tgt])
+                s = t
+                pbar.update(1)
+                pbar.refresh()
+
+        audio_slice = audio_pad[t:]
+        pitch_slice = pitch[:, t // self.window:] if if_f0 and t is not None else pitch
+        pitchf_slice = pitchf[:, t // self.window:] if if_f0 and t is not None else pitchf
+        audio_opt.append(self.vc(model, net_g, sid, audio_slice, pitch_slice, pitchf_slice, times, index, big_npy, index_rate, version, protect)[self.t_pad_tgt : -self.t_pad_tgt])
+        
+        audio_opt = np.concatenate(audio_opt)
+        if rms_mix_rate != 1:
+            audio_opt = change_rms(audio, 16000, audio_opt, tgt_sr, rms_mix_rate)
+        if tgt_sr != resample_sr >= 16000:
+            audio_opt = librosa.resample(
+                audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
+            )
+        audio_max = np.abs(audio_opt).max() / 0.99
+        max_int16 = 32768
+        if audio_max > 1:
+            max_int16 /= audio_max
+        audio_opt = (audio_opt * max_int16).astype(np.int16)
+        del pitch, pitchf, sid
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+
+        print("Returning completed audio...")
+        return audio_opt

lib/split_audio.py

@@ -0,0 +1,91 @@
+import os
+from pydub import AudioSegment
+from pydub.silence import detect_silence, detect_nonsilent
+
+SEPERATE_DIR = os.path.join(os.getcwd(), "seperate")
+TEMP_DIR = os.path.join(SEPERATE_DIR, "temp")
+cache = {}
+
+os.makedirs(SEPERATE_DIR, exist_ok=True)
+os.makedirs(TEMP_DIR, exist_ok=True)
+
+def cache_result(func):
+    def wrapper(*args, **kwargs):
+        key = (args, frozenset(kwargs.items()))
+        if key in cache:
+            return cache[key]
+        else:
+            result = func(*args, **kwargs)
+            cache[key] = result
+            return result
+    return wrapper   
+
+def get_non_silent(audio_name, audio, min_silence, silence_thresh, seek_step, keep_silence):
+    """
+    Function to get non-silent parts of the audio.
+    """
+    nonsilent_ranges = detect_nonsilent(audio, min_silence_len=min_silence, silence_thresh=silence_thresh, seek_step=seek_step)
+    nonsilent_files = []
+    for index, range in enumerate(nonsilent_ranges):
+        nonsilent_name = os.path.join(SEPERATE_DIR, f"{audio_name}_min{min_silence}_t{silence_thresh}_ss{seek_step}_ks{keep_silence}", f"nonsilent{index}-{audio_name}.wav")
+        start, end = range[0] - keep_silence, range[1] + keep_silence
+        audio[start:end].export(nonsilent_name, format="wav")
+        nonsilent_files.append(nonsilent_name)
+    return nonsilent_files
+
+def get_silence(audio_name, audio, min_silence, silence_thresh, seek_step, keep_silence):
+    """
+    Function to get silent parts of the audio.
+    """
+    silence_ranges = detect_silence(audio, min_silence_len=min_silence, silence_thresh=silence_thresh, seek_step=seek_step)
+    silence_files = []
+    for index, range in enumerate(silence_ranges):
+        silence_name = os.path.join(SEPERATE_DIR, f"{audio_name}_min{min_silence}_t{silence_thresh}_ss{seek_step}_ks{keep_silence}", f"silence{index}-{audio_name}.wav")
+        start, end = range[0] + keep_silence, range[1] - keep_silence
+        audio[start:end].export(silence_name, format="wav")
+        silence_files.append(silence_name)
+    return silence_files
+
+@cache_result
+def split_silence_nonsilent(input_path, min_silence=500, silence_thresh=-40, seek_step=1, keep_silence=100):
+    """
+    Function to split the audio into silent and non-silent parts.
+    """
+    audio_name = os.path.splitext(os.path.basename(input_path))[0]
+    os.makedirs(os.path.join(SEPERATE_DIR, f"{audio_name}_min{min_silence}_t{silence_thresh}_ss{seek_step}_ks{keep_silence}"), exist_ok=True)
+    audio = AudioSegment.silent(duration=1000) + AudioSegment.from_file(input_path) + AudioSegment.silent(duration=1000)
+    silence_files = get_silence(audio_name, audio, min_silence, silence_thresh, seek_step, keep_silence)
+    nonsilent_files = get_non_silent(audio_name, audio, min_silence, silence_thresh, seek_step, keep_silence)
+    return silence_files, nonsilent_files
+
+def adjust_audio_lengths(original_audios, inferred_audios):
+    """
+    Function to adjust the lengths of the inferred audio files list to match the original audio files length.
+    """
+    adjusted_audios = []
+    for original_audio, inferred_audio in zip(original_audios, inferred_audios):
+        audio_1 = AudioSegment.from_file(original_audio)
+        audio_2 = AudioSegment.from_file(inferred_audio)
+        
+        if len(audio_1) > len(audio_2):
+            audio_2 += AudioSegment.silent(duration=len(audio_1) - len(audio_2))
+        else:
+            audio_2 = audio_2[:len(audio_1)]
+        
+        adjusted_file = os.path.join(TEMP_DIR, f"adjusted-{os.path.basename(inferred_audio)}")
+        audio_2.export(adjusted_file, format="wav")
+        adjusted_audios.append(adjusted_file)
+    
+    return adjusted_audios
+
+def combine_silence_nonsilent(silence_files, nonsilent_files, keep_silence, output):
+    """
+    Function to combine the silent and non-silent parts of the audio.
+    """
+    combined = AudioSegment.empty()
+    for silence, nonsilent in zip(silence_files, nonsilent_files):
+        combined += AudioSegment.from_wav(silence) + AudioSegment.from_wav(nonsilent)
+    combined += AudioSegment.from_wav(silence_files[-1])
+    combined = AudioSegment.silent(duration=keep_silence) + combined[1000:-1000] + AudioSegment.silent(duration=keep_silence)
+    combined.export(output, format="wav")
+    return output

stftpitchshift

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb2f50ea8e5ca1a11a587f11f25ba9182f9b24e2367ac480f430b3f04062782e
+size 1822104