Upload 3 files

lib/infer_libs/audio.py

@@ -0,0 +1,142 @@
+import librosa
+import numpy as np
+import av
+from io import BytesIO
+import ffmpeg
+import os
+import traceback
+import sys
+import random
+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 audio2(i, o, format, sr):
+    inp = av.open(i, 'rb')
+    out = av.open(o, 'wb', format=format)
+    if format == "ogg": format = "libvorbis"
+    if format == "f32le": format = "pcm_f32le"
+
+    ostream = out.add_stream(format, channels=1)
+    ostream.sample_rate = sr
+
+    for frame in inp.decode(audio=0):
+        for p in ostream.encode(frame): out.mux(p)
+
+    out.close()
+    inp.close()
+
+def load_audion(file, sr):
+    try:
+        file = (
+            file.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+        )  # 防止小白拷路径头尾带了空格和"和回车
+        with open(file, "rb") as f:
+            with BytesIO() as out:
+                audio2(f, out, "f32le", sr)
+                return np.frombuffer(out.getvalue(), np.float32).flatten()
+
+    except AttributeError:
+        audio = file[1] / 32768.0
+        if len(audio.shape) == 2:
+            audio = np.mean(audio, -1)
+        return librosa.resample(audio, orig_sr=file[0], target_sr=16000)
+
+    except Exception as e:
+        raise RuntimeError(f"Failed to load audio: {e}")
+
+def load_audio(file, sr, DoFormant=False, Quefrency=1.0, Timbre=1.0):
+    converted = False
+    formanted = False
+    file = file.strip(' \n"')
+    if not os.path.exists(file):
+        raise RuntimeError(
+            "Wrong audio path, that does not exist."
+        )
+
+    try:
+        if not file.endswith(".wav"):
+            converted = True
+            formatted_file = f"{os.path.splitext(os.path.basename(file))[0]}.wav"
+            subprocess.run(
+                ["ffmpeg", "-nostdin", "-i", file, formatted_file],
+                capture_output=True,
+                text=True,
+            )
+            file = formatted_file
+            print(f"File formatted to wav format: {file}\n")
+
+        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")
+
+        with open(file, "rb") as f:
+            with BytesIO() as out:
+                audio2(f, out, "f32le", sr)
+                audio_data = np.frombuffer(out.getvalue(), np.float32).flatten()
+
+        if converted:
+            try:
+                os.remove(formatted_file)
+            except Exception as error:
+                print(f"Couldn't remove converted type of file due to {error}")
+                error = None
+            converted = False
+
+        return audio_data
+
+    except AttributeError:
+        audio = file[1] / 32768.0
+        if len(audio.shape) == 2:
+            audio = np.mean(audio, -1)
+        return librosa.resample(audio, orig_sr=file[0], target_sr=16000)
+    except Exception:
+        raise RuntimeError(traceback.format_exc())
+
+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 FCPE(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 = FCPE(
+            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/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)