add

app.py

@@ -21,7 +21,8 @@ def list_files_tree(directory, indent=""):
 
 from huggingface_hub import snapshot_download
 print("Models...")
-models_id = """None1145/So-VITS-SVC-Vulpisfoglia"""
+models_id = """None1145/So-VITS-SVC-Lappland
+None1145/So-VITS-SVC-Lappland-the-Decadenza"""
 for model_id in models_id.split("\n"):
     if model_id in ["", " "]:
         break
@@ -95,12 +96,12 @@ def vc_fn(input_audio, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scal
 app = gr.Blocks()
 with app:
     with gr.Tabs():
-        with gr.TabItem("Model"):
+        with gr.Tabs():
             speaker = gr.Dropdown(label="讲话人", choices=speakers, value=speakers[0])
             model_submit = gr.Button("加载模型", variant="primary")
             model_output1 = gr.Textbox(label="Output Message")
-        model_submit.click(load, [speaker], [model_output1])
-        with gr.TabItem("Basic"):
+            model_submit.click(load, [speaker], [model_output1])
+        with gr.Tabs():
             # sid = gr.Dropdown(label="音色", choices=speakers, value=speakers[0])
             vc_input3 = gr.Audio(label="上传音频")
             vc_transform = gr.Number(label="变调（整数，可以正负，半音数量，升高八度就是12）", value=0)
@@ -111,5 +112,5 @@ with app:
             vc_submit = gr.Button("转换", variant="primary")
             vc_output1 = gr.Textbox(label="Output Message")
             vc_output2 = gr.Audio(label="Output Audio")
-        vc_submit.click(vc_fn, [vc_input3, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale], [vc_output1, vc_output2])
+            vc_submit.click(vc_fn, [vc_input3, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale], [vc_output1, vc_output2])
     app.launch()

export_index_for_onnx.py

@@ -1,20 +0,0 @@
-import os
-import pickle
-
-import faiss
-
-path = "crs"
-indexs_file_path = f"checkpoints/{path}/feature_and_index.pkl"
-indexs_out_dir = f"checkpoints/{path}/"
-
-with open("feature_and_index.pkl",mode="rb") as f:
-    indexs = pickle.load(f)
-
-for k in indexs:
-    print(f"Save {k} index")
-    faiss.write_index(
-        indexs[k],
-        os.path.join(indexs_out_dir,f"Index-{k}.index")
-    )
-
-print("Saved all index")

flask_api.py

@@ -1,60 +0,0 @@
-import io
-import logging
-
-import soundfile
-import torch
-import torchaudio
-from flask import Flask, request, send_file
-from flask_cors import CORS
-
-from inference.infer_tool import RealTimeVC, Svc
-
-app = Flask(__name__)
-
-CORS(app)
-
-logging.getLogger('numba').setLevel(logging.WARNING)
-
-
-@app.route("/voiceChangeModel", methods=["POST"])
-def voice_change_model():
-    request_form = request.form
-    wave_file = request.files.get("sample", None)
-    # 变调信息
-    f_pitch_change = float(request_form.get("fPitchChange", 0))
-    # DAW所需的采样率
-    daw_sample = int(float(request_form.get("sampleRate", 0)))
-    speaker_id = int(float(request_form.get("sSpeakId", 0)))
-    # http获得wav文件并转换
-    input_wav_path = io.BytesIO(wave_file.read())
-
-    # 模型推理
-    if raw_infer:
-        # out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path)
-        out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0,
-                                            auto_predict_f0=False, noice_scale=0.4, f0_filter=False)
-        tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample)
-    else:
-        out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0,
-                                auto_predict_f0=False, noice_scale=0.4, f0_filter=False)
-        tar_audio = torchaudio.functional.resample(torch.from_numpy(out_audio), svc_model.target_sample, daw_sample)
-    # 返回音频
-    out_wav_path = io.BytesIO()
-    soundfile.write(out_wav_path, tar_audio.cpu().numpy(), daw_sample, format="wav")
-    out_wav_path.seek(0)
-    return send_file(out_wav_path, download_name="temp.wav", as_attachment=True)
-
-
-if __name__ == '__main__':
-    # 启用则为直接切片合成，False为交叉淡化方式
-    # vst插件调整0.3-0.5s切片时间可以降低延迟，直接切片方法会有连接处爆音、交叉淡化会有轻微重叠声音
-    # 自行选择能接受的方法，或将vst最大切片时间调整为1s，此处设为Ture，延迟大音质稳定一些
-    raw_infer = True
-    # 每个模型和config是唯一对应的
-    model_name = "logs/32k/G_174000-Copy1.pth"
-    config_name = "configs/config.json"
-    cluster_model_path = "logs/44k/kmeans_10000.pt"
-    svc_model = Svc(model_name, config_name, cluster_model_path=cluster_model_path)
-    svc = RealTimeVC()
-    # 此处与vst插件对应，不建议更改
-    app.run(port=6842, host="0.0.0.0", debug=False, threaded=False)

flask_api_full_song.py

@@ -1,55 +0,0 @@
-import io
-
-import numpy as np
-import soundfile
-from flask import Flask, request, send_file
-
-from inference import infer_tool, slicer
-
-app = Flask(__name__)
-
-
-@app.route("/wav2wav", methods=["POST"])
-def wav2wav():
-    request_form = request.form
-    audio_path = request_form.get("audio_path", None)  # wav文件地址
-    tran = int(float(request_form.get("tran", 0)))  # 音调
-    spk = request_form.get("spk", 0)  # 说话人(id或者name都可以,具体看你的config)
-    wav_format = request_form.get("wav_format", 'wav')  # 范围文件格式
-    infer_tool.format_wav(audio_path)
-    chunks = slicer.cut(audio_path, db_thresh=-40)
-    audio_data, audio_sr = slicer.chunks2audio(audio_path, chunks)
-
-    audio = []
-    for (slice_tag, data) in audio_data:
-        print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======')
-
-        length = int(np.ceil(len(data) / audio_sr * svc_model.target_sample))
-        if slice_tag:
-            print('jump empty segment')
-            _audio = np.zeros(length)
-        else:
-            # padd
-            pad_len = int(audio_sr * 0.5)
-            data = np.concatenate([np.zeros([pad_len]), data, np.zeros([pad_len])])
-            raw_path = io.BytesIO()
-            soundfile.write(raw_path, data, audio_sr, format="wav")
-            raw_path.seek(0)
-            out_audio, out_sr = svc_model.infer(spk, tran, raw_path)
-            svc_model.clear_empty()
-            _audio = out_audio.cpu().numpy()
-            pad_len = int(svc_model.target_sample * 0.5)
-            _audio = _audio[pad_len:-pad_len]
-
-        audio.extend(list(infer_tool.pad_array(_audio, length)))
-    out_wav_path = io.BytesIO()
-    soundfile.write(out_wav_path, audio, svc_model.target_sample, format=wav_format)
-    out_wav_path.seek(0)
-    return send_file(out_wav_path, download_name=f"temp.{wav_format}", as_attachment=True)
-
-
-if __name__ == '__main__':
-    model_name = "logs/44k/G_60000.pth"  # 模型地址
-    config_name = "configs/config.json"  # config地址
-    svc_model = infer_tool.Svc(model_name, config_name)
-    app.run(port=1145, host="0.0.0.0", debug=False, threaded=False)

onnx_export.py

@@ -1,144 +0,0 @@
-import argparse
-import json
-
-import torch
-
-import utils
-from onnxexport.model_onnx_speaker_mix import SynthesizerTrn
-
-parser = argparse.ArgumentParser(description='SoVitsSvc OnnxExport')
-
-def OnnxExport(path=None):
-    device = torch.device("cpu")
-    hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
-    SVCVITS = SynthesizerTrn(
-        hps.data.filter_length // 2 + 1,
-        hps.train.segment_size // hps.data.hop_length,
-        **hps.model)
-    _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", SVCVITS, None)
-    _ = SVCVITS.eval().to(device)
-    for i in SVCVITS.parameters():
-        i.requires_grad = False
-    
-    num_frames = 200
-
-    test_hidden_unit = torch.rand(1, num_frames, SVCVITS.gin_channels)
-    test_pitch = torch.rand(1, num_frames)
-    test_vol = torch.rand(1, num_frames)
-    test_mel2ph = torch.LongTensor(torch.arange(0, num_frames)).unsqueeze(0)
-    test_uv = torch.ones(1, num_frames, dtype=torch.float32)
-    test_noise = torch.randn(1, 192, num_frames)
-    test_sid = torch.LongTensor([0])
-    export_mix = True
-    if len(hps.spk) < 2:
-        export_mix = False
-    
-    if export_mix:
-        spk_mix = []
-        n_spk = len(hps.spk)
-        for i in range(n_spk):
-            spk_mix.append(1.0/float(n_spk))
-        test_sid = torch.tensor(spk_mix)
-        SVCVITS.export_chara_mix(hps.spk)
-        test_sid = test_sid.unsqueeze(0)
-        test_sid = test_sid.repeat(num_frames, 1)
-    
-    SVCVITS.eval()
-
-    if export_mix:
-        daxes = {
-            "c": [0, 1],
-            "f0": [1],
-            "mel2ph": [1],
-            "uv": [1],
-            "noise": [2],
-            "sid":[0]
-        }
-    else:
-        daxes = {
-            "c": [0, 1],
-            "f0": [1],
-            "mel2ph": [1],
-            "uv": [1],
-            "noise": [2]
-        }
-    
-    input_names = ["c", "f0", "mel2ph", "uv", "noise", "sid"]
-    output_names = ["audio", ]
-
-    if SVCVITS.vol_embedding:
-        input_names.append("vol")
-        vol_dadict = {"vol" : [1]}
-        daxes.update(vol_dadict)
-        test_inputs = (
-            test_hidden_unit.to(device),
-            test_pitch.to(device),
-            test_mel2ph.to(device),
-            test_uv.to(device),
-            test_noise.to(device),
-            test_sid.to(device),
-            test_vol.to(device)
-        )
-    else:
-        test_inputs = (
-            test_hidden_unit.to(device),
-            test_pitch.to(device),
-            test_mel2ph.to(device),
-            test_uv.to(device),
-            test_noise.to(device),
-            test_sid.to(device)
-        )
-
-    # SVCVITS = torch.jit.script(SVCVITS)
-    SVCVITS(test_hidden_unit.to(device),
-            test_pitch.to(device),
-            test_mel2ph.to(device),
-            test_uv.to(device),
-            test_noise.to(device),
-            test_sid.to(device),
-            test_vol.to(device))
-
-    SVCVITS.dec.OnnxExport()
-
-    torch.onnx.export(
-        SVCVITS,
-        test_inputs,
-        f"checkpoints/{path}/{path}_SoVits.onnx",
-        dynamic_axes=daxes,
-        do_constant_folding=False,
-        opset_version=16,
-        verbose=False,
-        input_names=input_names,
-        output_names=output_names
-    )
-
-    vec_lay = "layer-12" if SVCVITS.gin_channels == 768 else "layer-9"
-    spklist = []
-    for key in hps.spk.keys():
-        spklist.append(key)
-
-    MoeVSConf = {
-        "Folder" : f"{path}",
-        "Name" : f"{path}",
-        "Type" : "SoVits",
-        "Rate" : hps.data.sampling_rate,
-        "Hop" : hps.data.hop_length,
-        "Hubert": f"vec-{SVCVITS.gin_channels}-{vec_lay}",
-        "SoVits4": True,
-        "SoVits3": False,
-        "CharaMix": export_mix,
-        "Volume": SVCVITS.vol_embedding,
-        "HiddenSize": SVCVITS.gin_channels,
-        "Characters": spklist,
-        "Cluster": ""
-    }
-
-    with open(f"checkpoints/{path}.json", 'w') as MoeVsConfFile:
-        json.dump(MoeVSConf, MoeVsConfFile, indent = 4)
-
-
-if __name__ == '__main__':
-    parser.add_argument('-n', '--model_name', type=str, default="TransformerFlow", help='模型文件夹名（根目录下新建ckeckpoints文件夹，在此文件夹下建立一个新的文件夹，放置模型，该文件夹名即为此项）')
-    args = parser.parse_args()
-    path = args.model_name
-    OnnxExport(path)

onnx_export_old.py

@@ -1,56 +0,0 @@
-import torch
-
-import utils
-from onnxexport.model_onnx import SynthesizerTrn
-
-
-def main(NetExport):
-    path = "SoVits4.0"
-    if NetExport:
-        device = torch.device("cpu")
-        hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
-        SVCVITS = SynthesizerTrn(
-            hps.data.filter_length // 2 + 1,
-            hps.train.segment_size // hps.data.hop_length,
-            **hps.model)
-        _ = utils.load_checkpoint(f"checkpoints/{path}/model.pth", SVCVITS, None)
-        _ = SVCVITS.eval().to(device)
-        for i in SVCVITS.parameters():
-            i.requires_grad = False
-        
-        n_frame = 10
-        test_hidden_unit = torch.rand(1, n_frame, 256)
-        test_pitch = torch.rand(1, n_frame)
-        test_mel2ph = torch.arange(0, n_frame, dtype=torch.int64)[None] # torch.LongTensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]).unsqueeze(0)
-        test_uv = torch.ones(1, n_frame, dtype=torch.float32)
-        test_noise = torch.randn(1, 192, n_frame)
-        test_sid = torch.LongTensor([0])
-        input_names = ["c", "f0", "mel2ph", "uv", "noise", "sid"]
-        output_names = ["audio", ]
-        
-        torch.onnx.export(SVCVITS,
-                          (
-                              test_hidden_unit.to(device),
-                              test_pitch.to(device),
-                              test_mel2ph.to(device),
-                              test_uv.to(device),
-                              test_noise.to(device),
-                              test_sid.to(device)
-                          ),
-                          f"checkpoints/{path}/model.onnx",
-                          dynamic_axes={
-                              "c": [0, 1],
-                              "f0": [1],
-                              "mel2ph": [1],
-                              "uv": [1],
-                              "noise": [2],
-                          },
-                          do_constant_folding=False,
-                          opset_version=16,
-                          verbose=False,
-                          input_names=input_names,
-                          output_names=output_names)
-
-
-if __name__ == '__main__':
-    main(True)

onnxexport/model_onnx.py

@@ -1,333 +0,0 @@
-import torch
-from torch import nn
-from torch.nn import Conv1d, Conv2d
-from torch.nn import functional as F
-from torch.nn.utils import spectral_norm, weight_norm
-
-import modules.attentions as attentions
-import modules.commons as commons
-import modules.modules as modules
-import utils
-from modules.commons import get_padding
-from utils import f0_to_coarse
-from vdecoder.hifigan.models import Generator
-
-
-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
-
-
-class Encoder(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):
-        # print(x.shape,x_lengths.shape)
-        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
-
-
-class TextEncoder(nn.Module):
-    def __init__(self,
-                 out_channels,
-                 hidden_channels,
-                 kernel_size,
-                 n_layers,
-                 gin_channels=0,
-                 filter_channels=None,
-                 n_heads=None,
-                 p_dropout=None):
-        super().__init__()
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.n_layers = n_layers
-        self.gin_channels = gin_channels
-        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
-        self.f0_emb = nn.Embedding(256, hidden_channels)
-
-        self.enc_ = attentions.Encoder(
-            hidden_channels,
-            filter_channels,
-            n_heads,
-            n_layers,
-            kernel_size,
-            p_dropout)
-
-    def forward(self, x, x_mask, f0=None, z=None):
-        x = x + self.f0_emb(f0).transpose(1, 2)
-        x = self.enc_(x * x_mask, x_mask)
-        stats = self.proj(x) * x_mask
-        m, logs = torch.split(stats, self.out_channels, dim=1)
-        z = (m + z * torch.exp(logs)) * x_mask
-        return z, m, logs, x_mask
-
-
-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 is False else spectral_norm
-        self.convs = nn.ModuleList([
-            norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
-            norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
-            norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
-            norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
-            norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
-        ])
-        self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
-
-    def forward(self, x):
-        fmap = []
-
-        # 1d to 2d
-        b, c, t = x.shape
-        if t % self.period != 0:  # pad first
-            n_pad = self.period - (t % self.period)
-            x = F.pad(x, (0, n_pad), "reflect")
-            t = t + n_pad
-        x = x.view(b, c, t // self.period, self.period)
-
-        for l in self.convs:
-            x = l(x)
-            x = F.leaky_relu(x, modules.LRELU_SLOPE)
-            fmap.append(x)
-        x = self.conv_post(x)
-        fmap.append(x)
-        x = torch.flatten(x, 1, -1)
-
-        return x, fmap
-
-
-class DiscriminatorS(torch.nn.Module):
-    def __init__(self, use_spectral_norm=False):
-        super(DiscriminatorS, self).__init__()
-        norm_f = weight_norm if use_spectral_norm is 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 F0Decoder(nn.Module):
-    def __init__(self,
-                 out_channels,
-                 hidden_channels,
-                 filter_channels,
-                 n_heads,
-                 n_layers,
-                 kernel_size,
-                 p_dropout,
-                 spk_channels=0):
-        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.spk_channels = spk_channels
-
-        self.prenet = nn.Conv1d(hidden_channels, hidden_channels, 3, padding=1)
-        self.decoder = attentions.FFT(
-            hidden_channels,
-            filter_channels,
-            n_heads,
-            n_layers,
-            kernel_size,
-            p_dropout)
-        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
-        self.f0_prenet = nn.Conv1d(1, hidden_channels, 3, padding=1)
-        self.cond = nn.Conv1d(spk_channels, hidden_channels, 1)
-
-    def forward(self, x, norm_f0, x_mask, spk_emb=None):
-        x = torch.detach(x)
-        if spk_emb is not None:
-            x = x + self.cond(spk_emb)
-        x += self.f0_prenet(norm_f0)
-        x = self.prenet(x) * x_mask
-        x = self.decoder(x * x_mask, x_mask)
-        x = self.proj(x) * x_mask
-        return x
-
-
-class SynthesizerTrn(nn.Module):
-    """
-  Synthesizer for Training
-  """
-
-    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,
-                 gin_channels,
-                 ssl_dim,
-                 n_speakers,
-                 sampling_rate=44100,
-                 **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.ssl_dim = ssl_dim
-        self.emb_g = nn.Embedding(n_speakers, gin_channels)
-
-        self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)
-
-        self.enc_p = TextEncoder(
-            inter_channels,
-            hidden_channels,
-            filter_channels=filter_channels,
-            n_heads=n_heads,
-            n_layers=n_layers,
-            kernel_size=kernel_size,
-            p_dropout=p_dropout
-        )
-        hps = {
-            "sampling_rate": sampling_rate,
-            "inter_channels": inter_channels,
-            "resblock": resblock,
-            "resblock_kernel_sizes": resblock_kernel_sizes,
-            "resblock_dilation_sizes": resblock_dilation_sizes,
-            "upsample_rates": upsample_rates,
-            "upsample_initial_channel": upsample_initial_channel,
-            "upsample_kernel_sizes": upsample_kernel_sizes,
-            "gin_channels": gin_channels,
-        }
-        self.dec = Generator(h=hps)
-        self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
-        self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
-        self.f0_decoder = F0Decoder(
-            1,
-            hidden_channels,
-            filter_channels,
-            n_heads,
-            n_layers,
-            kernel_size,
-            p_dropout,
-            spk_channels=gin_channels
-        )
-        self.emb_uv = nn.Embedding(2, hidden_channels)
-        self.predict_f0 = False
-
-    def forward(self, c, f0, mel2ph, uv, noise=None, g=None):
-
-        decoder_inp = F.pad(c, [0, 0, 1, 0])
-        mel2ph_ = mel2ph.unsqueeze(2).repeat([1, 1, c.shape[-1]])
-        c = torch.gather(decoder_inp, 1, mel2ph_).transpose(1, 2)  # [B, T, H]
-
-        c_lengths = (torch.ones(c.size(0)) * c.size(-1)).to(c.device)
-        g = g.unsqueeze(0)
-        g = self.emb_g(g).transpose(1, 2)
-        x_mask = torch.unsqueeze(commons.sequence_mask(c_lengths, c.size(2)), 1).to(c.dtype)
-        x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2)
-
-        if self.predict_f0:
-            lf0 = 2595. * torch.log10(1. + f0.unsqueeze(1) / 700.) / 500
-            norm_lf0 = utils.normalize_f0(lf0, x_mask, uv, random_scale=False)
-            pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
-            f0 = (700 * (torch.pow(10, pred_lf0 * 500 / 2595) - 1)).squeeze(1)
-
-        z_p, m_p, logs_p, c_mask = self.enc_p(x, x_mask, f0=f0_to_coarse(f0), z=noise)
-        z = self.flow(z_p, c_mask, g=g, reverse=True)
-        o = self.dec(z * c_mask, g=g, f0=f0)
-        return o

onnxexport/model_onnx_speaker_mix.py

@@ -1,366 +0,0 @@
-import torch
-from torch import nn
-from torch.nn import functional as F
-
-import modules.attentions as attentions
-import modules.commons as commons
-import modules.modules as modules
-import utils
-from utils import f0_to_coarse
-
-
-class ResidualCouplingBlock(nn.Module):
-    def __init__(self,
-                 channels,
-                 hidden_channels,
-                 kernel_size,
-                 dilation_rate,
-                 n_layers,
-                 n_flows=4,
-                 gin_channels=0,
-                 share_parameter=False
-                 ):
-        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()
-
-        self.wn = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=0, gin_channels=gin_channels) if share_parameter else None
-
-        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, wn_sharing_parameter=self.wn))
-            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
-
-class TransformerCouplingBlock(nn.Module):
-    def __init__(self,
-                 channels,
-                 hidden_channels,
-                 filter_channels,
-                 n_heads,
-                 n_layers,
-                 kernel_size,
-                 p_dropout,
-                 n_flows=4,
-                 gin_channels=0,
-                 share_parameter=False
-                 ):
-            
-        super().__init__()
-        self.channels = channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.n_layers = n_layers
-        self.n_flows = n_flows
-        self.gin_channels = gin_channels
-
-        self.flows = nn.ModuleList()
-
-        self.wn = attentions.FFT(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, isflow = True, gin_channels = self.gin_channels) if share_parameter else None
-
-        for i in range(n_flows):
-            self.flows.append(
-                modules.TransformerCouplingLayer(channels, hidden_channels, kernel_size, n_layers, n_heads, p_dropout, filter_channels, mean_only=True, wn_sharing_parameter=self.wn, gin_channels = self.gin_channels))
-            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
-
-
-class Encoder(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):
-        # print(x.shape,x_lengths.shape)
-        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
-
-
-class TextEncoder(nn.Module):
-    def __init__(self,
-                 out_channels,
-                 hidden_channels,
-                 kernel_size,
-                 n_layers,
-                 gin_channels=0,
-                 filter_channels=None,
-                 n_heads=None,
-                 p_dropout=None):
-        super().__init__()
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.n_layers = n_layers
-        self.gin_channels = gin_channels
-        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
-        self.f0_emb = nn.Embedding(256, hidden_channels)
-
-        self.enc_ = attentions.Encoder(
-            hidden_channels,
-            filter_channels,
-            n_heads,
-            n_layers,
-            kernel_size,
-            p_dropout)
-
-    def forward(self, x, x_mask, f0=None, z=None):
-        x = x + self.f0_emb(f0).transpose(1, 2)
-        x = self.enc_(x * x_mask, x_mask)
-        stats = self.proj(x) * x_mask
-        m, logs = torch.split(stats, self.out_channels, dim=1)
-        z = (m + z * torch.exp(logs)) * x_mask
-
-        return z, m, logs, x_mask
-
-
-class F0Decoder(nn.Module):
-    def __init__(self,
-                 out_channels,
-                 hidden_channels,
-                 filter_channels,
-                 n_heads,
-                 n_layers,
-                 kernel_size,
-                 p_dropout,
-                 spk_channels=0):
-        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.spk_channels = spk_channels
-
-        self.prenet = nn.Conv1d(hidden_channels, hidden_channels, 3, padding=1)
-        self.decoder = attentions.FFT(
-            hidden_channels,
-            filter_channels,
-            n_heads,
-            n_layers,
-            kernel_size,
-            p_dropout)
-        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
-        self.f0_prenet = nn.Conv1d(1, hidden_channels, 3, padding=1)
-        self.cond = nn.Conv1d(spk_channels, hidden_channels, 1)
-
-    def forward(self, x, norm_f0, x_mask, spk_emb=None):
-        x = torch.detach(x)
-        if (spk_emb is not None):
-            x = x + self.cond(spk_emb)
-        x += self.f0_prenet(norm_f0)
-        x = self.prenet(x) * x_mask
-        x = self.decoder(x * x_mask, x_mask)
-        x = self.proj(x) * x_mask
-        return x
-
-
-class SynthesizerTrn(nn.Module):
-    """
-    Synthesizer for Training
-    """
-
-    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,
-                 gin_channels,
-                 ssl_dim,
-                 n_speakers,
-                 sampling_rate=44100,
-                 vol_embedding=False,
-                 vocoder_name = "nsf-hifigan",
-                 use_depthwise_conv = False,
-                 use_automatic_f0_prediction = True,
-                 flow_share_parameter = False,
-                 n_flow_layer = 4,
-                 n_layers_trans_flow = 3,
-                 use_transformer_flow = False,
-                 **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.ssl_dim = ssl_dim
-        self.vol_embedding = vol_embedding
-        self.emb_g = nn.Embedding(n_speakers, gin_channels)
-        self.use_depthwise_conv = use_depthwise_conv
-        self.use_automatic_f0_prediction = use_automatic_f0_prediction
-        self.n_layers_trans_flow = n_layers_trans_flow
-        if vol_embedding:
-           self.emb_vol = nn.Linear(1, hidden_channels)
-
-        self.pre = nn.Conv1d(ssl_dim, hidden_channels, kernel_size=5, padding=2)
-
-        self.enc_p = TextEncoder(
-            inter_channels,
-            hidden_channels,
-            filter_channels=filter_channels,
-            n_heads=n_heads,
-            n_layers=n_layers,
-            kernel_size=kernel_size,
-            p_dropout=p_dropout
-        )
-        hps = {
-            "sampling_rate": sampling_rate,
-            "inter_channels": inter_channels,
-            "resblock": resblock,
-            "resblock_kernel_sizes": resblock_kernel_sizes,
-            "resblock_dilation_sizes": resblock_dilation_sizes,
-            "upsample_rates": upsample_rates,
-            "upsample_initial_channel": upsample_initial_channel,
-            "upsample_kernel_sizes": upsample_kernel_sizes,
-            "gin_channels": gin_channels,
-            "use_depthwise_conv":use_depthwise_conv
-        }
-        
-        modules.set_Conv1dModel(self.use_depthwise_conv)
-
-        if vocoder_name == "nsf-hifigan":
-            from vdecoder.hifigan.models import Generator
-            self.dec = Generator(h=hps)
-        elif vocoder_name == "nsf-snake-hifigan":
-            from vdecoder.hifiganwithsnake.models import Generator
-            self.dec = Generator(h=hps)
-        else:
-            print("[?] Unkown vocoder: use default(nsf-hifigan)")
-            from vdecoder.hifigan.models import Generator
-            self.dec = Generator(h=hps)
-
-        self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
-        if use_transformer_flow:
-            self.flow = TransformerCouplingBlock(inter_channels, hidden_channels, filter_channels, n_heads, n_layers_trans_flow, 5, p_dropout, n_flow_layer, gin_channels=gin_channels, share_parameter=flow_share_parameter)
-        else:
-            self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, n_flow_layer, gin_channels=gin_channels, share_parameter=flow_share_parameter)
-        if self.use_automatic_f0_prediction:
-            self.f0_decoder = F0Decoder(
-                1,
-                hidden_channels,
-                filter_channels,
-                n_heads,
-                n_layers,
-                kernel_size,
-                p_dropout,
-                spk_channels=gin_channels
-            )
-        self.emb_uv = nn.Embedding(2, hidden_channels)
-        self.predict_f0 = False
-        self.speaker_map = []
-        self.export_mix = False
-
-    def export_chara_mix(self, speakers_mix):
-        self.speaker_map = torch.zeros((len(speakers_mix), 1, 1, self.gin_channels))
-        i = 0
-        for key in speakers_mix.keys():
-            spkidx = speakers_mix[key]
-            self.speaker_map[i] = self.emb_g(torch.LongTensor([[spkidx]]))
-            i = i + 1
-        self.speaker_map = self.speaker_map.unsqueeze(0)
-        self.export_mix = True
-
-    def forward(self, c, f0, mel2ph, uv, noise=None, g=None, vol = None):
-        decoder_inp = F.pad(c, [0, 0, 1, 0])
-        mel2ph_ = mel2ph.unsqueeze(2).repeat([1, 1, c.shape[-1]])
-        c = torch.gather(decoder_inp, 1, mel2ph_).transpose(1, 2)  # [B, T, H]
-
-        if self.export_mix:   # [N, S]  *  [S, B, 1, H]
-            g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1))  # [N, S, B, 1, 1]
-            g = g * self.speaker_map  # [N, S, B, 1, H]
-            g = torch.sum(g, dim=1) # [N, 1, B, 1, H]
-            g = g.transpose(0, -1).transpose(0, -2).squeeze(0) # [B, H, N]
-        else:
-            if g.dim() == 1:
-                g = g.unsqueeze(0)
-            g = self.emb_g(g).transpose(1, 2)
-        
-        x_mask = torch.unsqueeze(torch.ones_like(f0), 1).to(c.dtype)
-        # vol proj
-
-        vol = self.emb_vol(vol[:,:,None]).transpose(1,2) if vol is not None and self.vol_embedding else 0
-
-        x = self.pre(c) * x_mask + self.emb_uv(uv.long()).transpose(1, 2) + vol
-
-        if self.use_automatic_f0_prediction and self.predict_f0:
-            lf0 = 2595. * torch.log10(1. + f0.unsqueeze(1) / 700.) / 500
-            norm_lf0 = utils.normalize_f0(lf0, x_mask, uv, random_scale=False)
-            pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
-            f0 = (700 * (torch.pow(10, pred_lf0 * 500 / 2595) - 1)).squeeze(1)
-        
-        z_p, m_p, logs_p, c_mask = self.enc_p(x, x_mask, f0=f0_to_coarse(f0), z=noise)
-        z = self.flow(z_p, c_mask, g=g, reverse=True)
-        o = self.dec(z * c_mask, g=g, f0=f0)
-        return o
-

train.py

@@ -1,329 +0,0 @@
-import logging
-import multiprocessing
-import os
-import time
-
-import torch
-import torch.distributed as dist
-import torch.multiprocessing as mp
-from torch.cuda.amp import GradScaler, autocast
-from torch.nn import functional as F
-from torch.nn.parallel import DistributedDataParallel as DDP
-from torch.utils.data import DataLoader
-from torch.utils.tensorboard import SummaryWriter
-
-import modules.commons as commons
-import utils
-from data_utils import TextAudioCollate, TextAudioSpeakerLoader
-from models import (
-    MultiPeriodDiscriminator,
-    SynthesizerTrn,
-)
-from modules.losses import discriminator_loss, feature_loss, generator_loss, kl_loss
-from modules.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
-
-logging.getLogger('matplotlib').setLevel(logging.WARNING)
-logging.getLogger('numba').setLevel(logging.WARNING)
-
-torch.backends.cudnn.benchmark = True
-global_step = 0
-start_time = time.time()
-
-# os.environ['TORCH_DISTRIBUTED_DEBUG'] = 'INFO'
-
-
-def main():
-    """Assume Single Node Multi GPUs Training Only"""
-    assert torch.cuda.is_available(), "CPU training is not allowed."
-    hps = utils.get_hparams()
-
-    n_gpus = torch.cuda.device_count()
-    os.environ['MASTER_ADDR'] = 'localhost'
-    os.environ['MASTER_PORT'] = hps.train.port
-
-    mp.spawn(run, nprocs=n_gpus, args=(n_gpus, hps,))
-
-
-def run(rank, n_gpus, hps):
-    global global_step
-    if rank == 0:
-        logger = utils.get_logger(hps.model_dir)
-        logger.info(hps)
-        utils.check_git_hash(hps.model_dir)
-        writer = SummaryWriter(log_dir=hps.model_dir)
-        writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
-    
-    # for pytorch on win, backend use gloo    
-    dist.init_process_group(backend=  'gloo' if os.name == 'nt' else 'nccl', init_method='env://', world_size=n_gpus, rank=rank)
-    torch.manual_seed(hps.train.seed)
-    torch.cuda.set_device(rank)
-    collate_fn = TextAudioCollate()
-    all_in_mem = hps.train.all_in_mem   # If you have enough memory, turn on this option to avoid disk IO and speed up training.
-    train_dataset = TextAudioSpeakerLoader(hps.data.training_files, hps, all_in_mem=all_in_mem)
-    num_workers = 5 if multiprocessing.cpu_count() > 4 else multiprocessing.cpu_count()
-    if all_in_mem:
-        num_workers = 0
-    train_loader = DataLoader(train_dataset, num_workers=num_workers, shuffle=False, pin_memory=True,
-                              batch_size=hps.train.batch_size, collate_fn=collate_fn)
-    if rank == 0:
-        eval_dataset = TextAudioSpeakerLoader(hps.data.validation_files, hps, all_in_mem=all_in_mem,vol_aug = False)
-        eval_loader = DataLoader(eval_dataset, num_workers=1, shuffle=False,
-                                 batch_size=1, pin_memory=False,
-                                 drop_last=False, collate_fn=collate_fn)
-
-    net_g = SynthesizerTrn(
-        hps.data.filter_length // 2 + 1,
-        hps.train.segment_size // hps.data.hop_length,
-        **hps.model).cuda(rank)
-    net_d = MultiPeriodDiscriminator(hps.model.use_spectral_norm).cuda(rank)
-    optim_g = torch.optim.AdamW(
-        net_g.parameters(),
-        hps.train.learning_rate,
-        betas=hps.train.betas,
-        eps=hps.train.eps)
-    optim_d = torch.optim.AdamW(
-        net_d.parameters(),
-        hps.train.learning_rate,
-        betas=hps.train.betas,
-        eps=hps.train.eps)
-    net_g = DDP(net_g, device_ids=[rank])  # , find_unused_parameters=True)
-    net_d = DDP(net_d, device_ids=[rank])
-
-    skip_optimizer = False
-    try:
-        _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g,
-                                                   optim_g, skip_optimizer)
-        _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d,
-                                                   optim_d, skip_optimizer)
-        epoch_str = max(epoch_str, 1)
-        name=utils.latest_checkpoint_path(hps.model_dir, "D_*.pth")
-        global_step=int(name[name.rfind("_")+1:name.rfind(".")])+1
-        #global_step = (epoch_str - 1) * len(train_loader)
-    except Exception:
-        print("load old checkpoint failed...")
-        epoch_str = 1
-        global_step = 0
-    if skip_optimizer:
-        epoch_str = 1
-        global_step = 0
-
-    warmup_epoch = hps.train.warmup_epochs
-    scheduler_g = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2)
-    scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2)
-
-    scaler = GradScaler(enabled=hps.train.fp16_run)
-
-    for epoch in range(epoch_str, hps.train.epochs + 1):
-        # set up warm-up learning rate
-        if epoch <= warmup_epoch:
-            for param_group in optim_g.param_groups:
-                param_group['lr'] = hps.train.learning_rate / warmup_epoch * epoch
-            for param_group in optim_d.param_groups:
-                param_group['lr'] = hps.train.learning_rate / warmup_epoch * epoch
-        # training
-        if rank == 0:
-            train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler,
-                               [train_loader, eval_loader], logger, [writer, writer_eval])
-        else:
-            train_and_evaluate(rank, epoch, hps, [net_g, net_d], [optim_g, optim_d], [scheduler_g, scheduler_d], scaler,
-                               [train_loader, None], None, None)
-        # update learning rate
-        scheduler_g.step()
-        scheduler_d.step()
-
-
-def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers):
-    net_g, net_d = nets
-    optim_g, optim_d = optims
-    scheduler_g, scheduler_d = schedulers
-    train_loader, eval_loader = loaders
-    if writers is not None:
-        writer, writer_eval = writers
-    
-    half_type = torch.bfloat16 if hps.train.half_type=="bf16" else torch.float16
-
-    # train_loader.batch_sampler.set_epoch(epoch)
-    global global_step
-
-    net_g.train()
-    net_d.train()
-    for batch_idx, items in enumerate(train_loader):
-        c, f0, spec, y, spk, lengths, uv,volume = items
-        g = spk.cuda(rank, non_blocking=True)
-        spec, y = spec.cuda(rank, non_blocking=True), y.cuda(rank, non_blocking=True)
-        c = c.cuda(rank, non_blocking=True)
-        f0 = f0.cuda(rank, non_blocking=True)
-        uv = uv.cuda(rank, non_blocking=True)
-        lengths = lengths.cuda(rank, non_blocking=True)
-        mel = spec_to_mel_torch(
-            spec,
-            hps.data.filter_length,
-            hps.data.n_mel_channels,
-            hps.data.sampling_rate,
-            hps.data.mel_fmin,
-            hps.data.mel_fmax)
-        
-        with autocast(enabled=hps.train.fp16_run, dtype=half_type):
-            y_hat, ids_slice, z_mask, \
-            (z, z_p, m_p, logs_p, m_q, logs_q), pred_lf0, norm_lf0, lf0 = net_g(c, f0, uv, spec, g=g, c_lengths=lengths,
-                                                                                spec_lengths=lengths,vol = volume)
-
-            y_mel = commons.slice_segments(mel, ids_slice, hps.train.segment_size // hps.data.hop_length)
-            y_hat_mel = mel_spectrogram_torch(
-                y_hat.squeeze(1),
-                hps.data.filter_length,
-                hps.data.n_mel_channels,
-                hps.data.sampling_rate,
-                hps.data.hop_length,
-                hps.data.win_length,
-                hps.data.mel_fmin,
-                hps.data.mel_fmax
-            )
-            y = commons.slice_segments(y, ids_slice * hps.data.hop_length, hps.train.segment_size)  # slice
-
-            # Discriminator
-            y_d_hat_r, y_d_hat_g, _, _ = net_d(y, y_hat.detach())
-
-            with autocast(enabled=False, dtype=half_type):
-                loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
-                loss_disc_all = loss_disc
-        
-        optim_d.zero_grad()
-        scaler.scale(loss_disc_all).backward()
-        scaler.unscale_(optim_d)
-        grad_norm_d = commons.clip_grad_value_(net_d.parameters(), None)
-        scaler.step(optim_d)
-        
-
-        with autocast(enabled=hps.train.fp16_run, dtype=half_type):
-            # Generator
-            y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(y, y_hat)
-            with autocast(enabled=False, dtype=half_type):
-                loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
-                loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
-                loss_fm = feature_loss(fmap_r, fmap_g)
-                loss_gen, losses_gen = generator_loss(y_d_hat_g)
-                loss_lf0 = F.mse_loss(pred_lf0, lf0) if net_g.module.use_automatic_f0_prediction else 0
-                loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl + loss_lf0
-        optim_g.zero_grad()
-        scaler.scale(loss_gen_all).backward()
-        scaler.unscale_(optim_g)
-        grad_norm_g = commons.clip_grad_value_(net_g.parameters(), None)
-        scaler.step(optim_g)
-        scaler.update()
-
-        if rank == 0:
-            if global_step % hps.train.log_interval == 0:
-                lr = optim_g.param_groups[0]['lr']
-                losses = [loss_disc, loss_gen, loss_fm, loss_mel, loss_kl]
-                reference_loss=0
-                for i in losses:
-                    reference_loss += i
-                logger.info('Train Epoch: {} [{:.0f}%]'.format(
-                    epoch,
-                    100. * batch_idx / len(train_loader)))
-                logger.info(f"Losses: {[x.item() for x in losses]}, step: {global_step}, lr: {lr}, reference_loss: {reference_loss}")
-
-                scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc_all, "learning_rate": lr,
-                               "grad_norm_d": grad_norm_d, "grad_norm_g": grad_norm_g}
-                scalar_dict.update({"loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/kl": loss_kl,
-                                    "loss/g/lf0": loss_lf0})
-
-                # scalar_dict.update({"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)})
-                # scalar_dict.update({"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)})
-                # scalar_dict.update({"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)})
-                image_dict = {
-                    "slice/mel_org": utils.plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()),
-                    "slice/mel_gen": utils.plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()),
-                    "all/mel": utils.plot_spectrogram_to_numpy(mel[0].data.cpu().numpy())
-                }
-
-                if net_g.module.use_automatic_f0_prediction:
-                    image_dict.update({
-                        "all/lf0": utils.plot_data_to_numpy(lf0[0, 0, :].cpu().numpy(),
-                                                              pred_lf0[0, 0, :].detach().cpu().numpy()),
-                        "all/norm_lf0": utils.plot_data_to_numpy(lf0[0, 0, :].cpu().numpy(),
-                                                                   norm_lf0[0, 0, :].detach().cpu().numpy())
-                    })
-
-                utils.summarize(
-                    writer=writer,
-                    global_step=global_step,
-                    images=image_dict,
-                    scalars=scalar_dict
-                )
-
-            if global_step % hps.train.eval_interval == 0:
-                evaluate(hps, net_g, eval_loader, writer_eval)
-                utils.save_checkpoint(net_g, optim_g, hps.train.learning_rate, epoch,
-                                      os.path.join(hps.model_dir, "G_{}.pth".format(global_step)))
-                utils.save_checkpoint(net_d, optim_d, hps.train.learning_rate, epoch,
-                                      os.path.join(hps.model_dir, "D_{}.pth".format(global_step)))
-                keep_ckpts = getattr(hps.train, 'keep_ckpts', 0)
-                if keep_ckpts > 0:
-                    utils.clean_checkpoints(path_to_models=hps.model_dir, n_ckpts_to_keep=keep_ckpts, sort_by_time=True)
-
-        global_step += 1
-
-    if rank == 0:
-        global start_time
-        now = time.time()
-        durtaion = format(now - start_time, '.2f')
-        logger.info(f'====> Epoch: {epoch}, cost {durtaion} s')
-        start_time = now
-
-
-def evaluate(hps, generator, eval_loader, writer_eval):
-    generator.eval()
-    image_dict = {}
-    audio_dict = {}
-    with torch.no_grad():
-        for batch_idx, items in enumerate(eval_loader):
-            c, f0, spec, y, spk, _, uv,volume = items
-            g = spk[:1].cuda(0)
-            spec, y = spec[:1].cuda(0), y[:1].cuda(0)
-            c = c[:1].cuda(0)
-            f0 = f0[:1].cuda(0)
-            uv= uv[:1].cuda(0)
-            if volume is not None:
-                volume = volume[:1].cuda(0)
-            mel = spec_to_mel_torch(
-                spec,
-                hps.data.filter_length,
-                hps.data.n_mel_channels,
-                hps.data.sampling_rate,
-                hps.data.mel_fmin,
-                hps.data.mel_fmax)
-            y_hat,_ = generator.module.infer(c, f0, uv, g=g,vol = volume)
-
-            y_hat_mel = mel_spectrogram_torch(
-                y_hat.squeeze(1).float(),
-                hps.data.filter_length,
-                hps.data.n_mel_channels,
-                hps.data.sampling_rate,
-                hps.data.hop_length,
-                hps.data.win_length,
-                hps.data.mel_fmin,
-                hps.data.mel_fmax
-            )
-
-            audio_dict.update({
-                f"gen/audio_{batch_idx}": y_hat[0],
-                f"gt/audio_{batch_idx}": y[0]
-            })
-        image_dict.update({
-            "gen/mel": utils.plot_spectrogram_to_numpy(y_hat_mel[0].cpu().numpy()),
-            "gt/mel": utils.plot_spectrogram_to_numpy(mel[0].cpu().numpy())
-        })
-    utils.summarize(
-        writer=writer_eval,
-        global_step=global_step,
-        images=image_dict,
-        audios=audio_dict,
-        audio_sampling_rate=hps.data.sampling_rate
-    )
-    generator.train()
-
-
-if __name__ == "__main__":
-    main()

train_diff.py

@@ -1,77 +0,0 @@
-import argparse
-
-import torch
-from loguru import logger
-from torch.optim import lr_scheduler
-
-from diffusion.data_loaders import get_data_loaders
-from diffusion.logger import utils
-from diffusion.solver import train
-from diffusion.unit2mel import Unit2Mel
-from diffusion.vocoder import Vocoder
-
-
-def parse_args(args=None, namespace=None):
-    """Parse command-line arguments."""
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "-c",
-        "--config",
-        type=str,
-        required=True,
-        help="path to the config file")
-    return parser.parse_args(args=args, namespace=namespace)
-
-
-if __name__ == '__main__':
-    # parse commands
-    cmd = parse_args()
-    
-    # load config
-    args = utils.load_config(cmd.config)
-    logger.info(' > config:'+ cmd.config)
-    logger.info(' > exp:'+ args.env.expdir)
-    
-    # load vocoder
-    vocoder = Vocoder(args.vocoder.type, args.vocoder.ckpt, device=args.device)
-    
-    # load model
-    model = Unit2Mel(
-                args.data.encoder_out_channels, 
-                args.model.n_spk,
-                args.model.use_pitch_aug,
-                vocoder.dimension,
-                args.model.n_layers,
-                args.model.n_chans,
-                args.model.n_hidden,
-                args.model.timesteps,
-                args.model.k_step_max
-                )
-    
-    logger.info(f' > Now model timesteps is {model.timesteps}, and k_step_max is {model.k_step_max}')
-    
-    # load parameters
-    optimizer = torch.optim.AdamW(model.parameters())
-    initial_global_step, model, optimizer = utils.load_model(args.env.expdir, model, optimizer, device=args.device)
-    for param_group in optimizer.param_groups:
-        param_group['initial_lr'] = args.train.lr
-        param_group['lr'] = args.train.lr * (args.train.gamma ** max(((initial_global_step-2)//args.train.decay_step),0) )
-        param_group['weight_decay'] = args.train.weight_decay
-    scheduler = lr_scheduler.StepLR(optimizer, step_size=args.train.decay_step, gamma=args.train.gamma,last_epoch=initial_global_step-2)
-    
-    # device
-    if args.device == 'cuda':
-        torch.cuda.set_device(args.env.gpu_id)
-    model.to(args.device)
-    
-    for state in optimizer.state.values():
-        for k, v in state.items():
-            if torch.is_tensor(v):
-                state[k] = v.to(args.device)
-                    
-    # datas
-    loader_train, loader_valid = get_data_loaders(args, whole_audio=False)
-    
-    # run
-    train(args, initial_global_step, model, optimizer, scheduler, vocoder, loader_train, loader_valid)
-    

train_index.py

@@ -1,30 +0,0 @@
-import argparse
-import os
-import pickle
-
-import utils
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--root_dir", type=str, default="dataset/44k", help="path to root dir"
-    )
-    parser.add_argument('-c', '--config', type=str, default="./configs/config.json",
-                    help='JSON file for configuration')
-    parser.add_argument(
-        "--output_dir", type=str, default="logs/44k", help="path to output dir"
-    )
-
-    args = parser.parse_args()
-
-    hps = utils.get_hparams_from_file(args.config)
-    spk_dic = hps.spk
-    result = {}
-    
-    for k,v in spk_dic.items():
-        print(f"now, index {k} feature...")
-        index = utils.train_index(k,args.root_dir)
-        result[v] = index
-
-    with open(os.path.join(args.output_dir,"feature_and_index.pkl"),"wb") as f:
-        pickle.dump(result,f)

wav_upload.py

@@ -1,25 +0,0 @@
-import argparse
-import os
-import shutil
-
-from google.colab import files
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--type", type=str, required=True, help="type of file to upload")
-    args = parser.parse_args()
-    file_type = args.type
-
-    basepath = os.getcwd()
-    uploaded = files.upload() # 上传文件
-    assert(file_type in ['zip', 'audio'])
-    if file_type == "zip":
-        upload_path = "./upload/"
-        for filename in uploaded.keys():
-            #将上传的文件移动到指定的位置上
-            shutil.move(os.path.join(basepath, filename), os.path.join(upload_path, "userzip.zip"))
-    elif file_type == "audio":
-        upload_path = "./raw/"
-        for filename in uploaded.keys():
-            #将上传的文件移动到指定的位置上
-            shutil.move(os.path.join(basepath, filename), os.path.join(upload_path, filename))

webUI.py

@@ -1,430 +0,0 @@
-import glob
-import json
-import logging
-import os
-import re
-import subprocess
-import sys
-import time
-import traceback
-from itertools import chain
-from pathlib import Path
-
-# os.system("wget -P cvec/ https://huggingface.co/spaces/innnky/nanami/resolve/main/checkpoint_best_legacy_500.pt")
-import gradio as gr
-import librosa
-import numpy as np
-import soundfile
-import torch
-
-from compress_model import removeOptimizer
-from edgetts.tts_voices import SUPPORTED_LANGUAGES
-from inference.infer_tool import Svc
-from utils import mix_model
-
-logging.getLogger('numba').setLevel(logging.WARNING)
-logging.getLogger('markdown_it').setLevel(logging.WARNING)
-logging.getLogger('urllib3').setLevel(logging.WARNING)
-logging.getLogger('matplotlib').setLevel(logging.WARNING)
-logging.getLogger('multipart').setLevel(logging.WARNING)
-
-model = None
-spk = None
-debug = False
-
-local_model_root = './trained'
-
-cuda = {}
-if torch.cuda.is_available():
-    for i in range(torch.cuda.device_count()):
-        device_name = torch.cuda.get_device_properties(i).name
-        cuda[f"CUDA:{i} {device_name}"] = f"cuda:{i}"
-
-def upload_mix_append_file(files,sfiles):
-    try:
-        if(sfiles is None):
-            file_paths = [file.name for file in files]
-        else:
-            file_paths = [file.name for file in chain(files,sfiles)]
-        p = {file:100 for file in file_paths}
-        return file_paths,mix_model_output1.update(value=json.dumps(p,indent=2))
-    except Exception as e:
-        if debug:
-            traceback.print_exc()
-        raise gr.Error(e)
-
-def mix_submit_click(js,mode):
-    try:
-        assert js.lstrip()!=""
-        modes = {"凸组合":0, "线性组合":1}
-        mode = modes[mode]
-        data = json.loads(js)
-        data = list(data.items())
-        model_path,mix_rate = zip(*data)
-        path = mix_model(model_path,mix_rate,mode)
-        return f"成功，文件被保存在了{path}"
-    except Exception as e:
-        if debug:
-            traceback.print_exc()
-        raise gr.Error(e)
-
-def updata_mix_info(files):
-    try:
-        if files is None :
-            return mix_model_output1.update(value="")
-        p = {file.name:100 for file in files}
-        return mix_model_output1.update(value=json.dumps(p,indent=2))
-    except Exception as e:
-        if debug:
-            traceback.print_exc()
-        raise gr.Error(e)
-
-def modelAnalysis(model_path,config_path,cluster_model_path,device,enhance,diff_model_path,diff_config_path,only_diffusion,use_spk_mix,local_model_enabled,local_model_selection):
-    global model
-    try:
-        device = cuda[device] if "CUDA" in device else device
-        cluster_filepath = os.path.split(cluster_model_path.name) if cluster_model_path is not None else "no_cluster"
-        # get model and config path
-        if (local_model_enabled):
-            # local path
-            model_path = glob.glob(os.path.join(local_model_selection, '*.pth'))[0]
-            config_path = glob.glob(os.path.join(local_model_selection, '*.json'))[0]
-        else:
-            # upload from webpage
-            model_path = model_path.name
-            config_path = config_path.name
-        fr = ".pkl" in cluster_filepath[1]
-        model = Svc(model_path,
-                config_path,
-                device=device if device != "Auto" else None,
-                cluster_model_path = cluster_model_path.name if cluster_model_path is not None else "",
-                nsf_hifigan_enhance=enhance,
-                diffusion_model_path = diff_model_path.name if diff_model_path is not None else "",
-                diffusion_config_path = diff_config_path.name if diff_config_path is not None else "",
-                shallow_diffusion = True if diff_model_path is not None else False,
-                only_diffusion = only_diffusion,
-                spk_mix_enable = use_spk_mix,
-                feature_retrieval = fr
-                )
-        spks = list(model.spk2id.keys())
-        device_name = torch.cuda.get_device_properties(model.dev).name if "cuda" in str(model.dev) else str(model.dev)
-        msg = f"成功加载模型到设备{device_name}上\n"
-        if cluster_model_path is None:
-            msg += "未加载聚类模型或特征检索模型\n"
-        elif fr:
-            msg += f"特征检索模型{cluster_filepath[1]}加载成功\n"
-        else:
-            msg += f"聚类模型{cluster_filepath[1]}加载成功\n"
-        if diff_model_path is None:
-            msg += "未加载扩散模型\n"
-        else:
-            msg += f"扩散模型{diff_model_path.name}加载成功\n"
-        msg += "当前模型的可用音色：\n"
-        for i in spks:
-            msg += i + " "
-        return sid.update(choices = spks,value=spks[0]), msg
-    except Exception as e:
-        if debug:
-            traceback.print_exc()
-        raise gr.Error(e)
-
-    
-def modelUnload():
-    global model
-    if model is None:
-        return sid.update(choices = [],value=""),"没有模型需要卸载!"
-    else:
-        model.unload_model()
-        model = None
-        torch.cuda.empty_cache()
-        return sid.update(choices = [],value=""),"模型卸载完毕!"
-    
-def vc_infer(output_format, sid, audio_path, truncated_basename, vc_transform, auto_f0, cluster_ratio, slice_db, noise_scale, pad_seconds, cl_num, lg_num, lgr_num, f0_predictor, enhancer_adaptive_key, cr_threshold, k_step, use_spk_mix, second_encoding, loudness_envelope_adjustment):
-    global model
-    _audio = model.slice_inference(
-        audio_path,
-        sid,
-        vc_transform,
-        slice_db,
-        cluster_ratio,
-        auto_f0,
-        noise_scale,
-        pad_seconds,
-        cl_num,
-        lg_num,
-        lgr_num,
-        f0_predictor,
-        enhancer_adaptive_key,
-        cr_threshold,
-        k_step,
-        use_spk_mix,
-        second_encoding,
-        loudness_envelope_adjustment
-    )  
-    model.clear_empty()
-    #构建保存文件的路径，并保存到results文件夹内
-    str(int(time.time()))
-    if not os.path.exists("results"):
-        os.makedirs("results")
-    key = "auto" if auto_f0 else f"{int(vc_transform)}key"
-    cluster = "_" if cluster_ratio == 0 else f"_{cluster_ratio}_"
-    isdiffusion = "sovits"
-    if model.shallow_diffusion:
-        isdiffusion = "sovdiff"
-
-    if model.only_diffusion:
-        isdiffusion = "diff"
-    
-    output_file_name = 'result_'+truncated_basename+f'_{sid}_{key}{cluster}{isdiffusion}.{output_format}'
-    output_file = os.path.join("results", output_file_name)
-    soundfile.write(output_file, _audio, model.target_sample, format=output_format)
-    return output_file
-
-def vc_fn(sid, input_audio, output_format, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,f0_predictor,enhancer_adaptive_key,cr_threshold,k_step,use_spk_mix,second_encoding,loudness_envelope_adjustment):
-    global model
-    try:
-        if input_audio is None:
-            return "You need to upload an audio", None
-        if model is None:
-            return "You need to upload an model", None
-        if getattr(model, 'cluster_model', None) is None and model.feature_retrieval is False:
-            if cluster_ratio != 0:
-                return "You need to upload an cluster model or feature retrieval model before assigning cluster ratio!", None
-        #print(input_audio)    
-        audio, sampling_rate = soundfile.read(input_audio)
-        #print(audio.shape,sampling_rate)
-        if np.issubdtype(audio.dtype, np.integer):
-            audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
-        #print(audio.dtype)
-        if len(audio.shape) > 1:
-            audio = librosa.to_mono(audio.transpose(1, 0))
-        # 未知原因Gradio上传的filepath会有一个奇怪的固定后缀，这里去掉
-        truncated_basename = Path(input_audio).stem[:-6]
-        processed_audio = os.path.join("raw", f"{truncated_basename}.wav")
-        soundfile.write(processed_audio, audio, sampling_rate, format="wav")
-        output_file = vc_infer(output_format, sid, processed_audio, truncated_basename, vc_transform, auto_f0, cluster_ratio, slice_db, noise_scale, pad_seconds, cl_num, lg_num, lgr_num, f0_predictor, enhancer_adaptive_key, cr_threshold, k_step, use_spk_mix, second_encoding, loudness_envelope_adjustment)
-
-        return "Success", output_file
-    except Exception as e:
-        if debug:
-            traceback.print_exc()
-        raise gr.Error(e)
-
-def text_clear(text):
-    return re.sub(r"[\n\,\(\) ]", "", text)
-
-def vc_fn2(_text, _lang, _gender, _rate, _volume, sid, output_format, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,f0_predictor,enhancer_adaptive_key,cr_threshold, k_step,use_spk_mix,second_encoding,loudness_envelope_adjustment):
-    global model
-    try:
-        if model is None:
-            return "You need to upload an model", None
-        if getattr(model, 'cluster_model', None) is None and model.feature_retrieval is False:
-            if cluster_ratio != 0:
-                return "You need to upload an cluster model or feature retrieval model before assigning cluster ratio!", None
-        _rate = f"+{int(_rate*100)}%" if _rate >= 0 else f"{int(_rate*100)}%"
-        _volume = f"+{int(_volume*100)}%" if _volume >= 0 else f"{int(_volume*100)}%"
-        if _lang == "Auto":
-            _gender = "Male" if _gender == "男" else "Female"
-            subprocess.run([sys.executable, "edgetts/tts.py", _text, _lang, _rate, _volume, _gender])
-        else:
-            subprocess.run([sys.executable, "edgetts/tts.py", _text, _lang, _rate, _volume])
-        target_sr = 44100
-        y, sr = librosa.load("tts.wav")
-        resampled_y = librosa.resample(y, orig_sr=sr, target_sr=target_sr)
-        soundfile.write("tts.wav", resampled_y, target_sr, subtype = "PCM_16")
-        input_audio = "tts.wav"
-        #audio, _ = soundfile.read(input_audio)
-        output_file_path = vc_infer(output_format, sid, input_audio, "tts", vc_transform, auto_f0, cluster_ratio, slice_db, noise_scale, pad_seconds, cl_num, lg_num, lgr_num, f0_predictor, enhancer_adaptive_key, cr_threshold, k_step, use_spk_mix, second_encoding, loudness_envelope_adjustment)
-        os.remove("tts.wav")
-        return "Success", output_file_path
-    except Exception as e:
-        if debug: traceback.print_exc()  # noqa: E701
-        raise gr.Error(e)
-
-def model_compression(_model):
-    if _model == "":
-        return "请先选择要压缩的模型"
-    else:
-        model_path = os.path.split(_model.name)
-        filename, extension = os.path.splitext(model_path[1])
-        output_model_name = f"{filename}_compressed{extension}"
-        output_path = os.path.join(os.getcwd(), output_model_name)
-        removeOptimizer(_model.name, output_path)
-        return f"模型已成功被保存在了{output_path}"
-
-def scan_local_models():
-    res = []
-    candidates = glob.glob(os.path.join(local_model_root, '**', '*.json'), recursive=True)
-    candidates = set([os.path.dirname(c) for c in candidates])
-    for candidate in candidates:
-        jsons = glob.glob(os.path.join(candidate, '*.json'))
-        pths = glob.glob(os.path.join(candidate, '*.pth'))
-        if (len(jsons) == 1 and len(pths) == 1):
-            # must contain exactly one json and one pth file
-            res.append(candidate)
-    return res
-
-def local_model_refresh_fn():
-    choices = scan_local_models()
-    return gr.Dropdown.update(choices=choices)
-
-def debug_change():
-    global debug
-    debug = debug_button.value
-
-with gr.Blocks(
-    theme=gr.themes.Base(
-        primary_hue = gr.themes.colors.green,
-        font=["Source Sans Pro", "Arial", "sans-serif"],
-        font_mono=['JetBrains mono', "Consolas", 'Courier New']
-    ),
-) as app:
-    with gr.Tabs():
-        with gr.TabItem("推理"):
-            gr.Markdown(value="""
-                So-vits-svc 4.0 推理 webui
-                """)
-            with gr.Row(variant="panel"):
-                with gr.Column():
-                    gr.Markdown(value="""
-                        <font size=2> 模型设置</font>
-                        """)
-                    with gr.Tabs():
-                        # invisible checkbox that tracks tab status
-                        local_model_enabled = gr.Checkbox(value=False, visible=False)
-                        with gr.TabItem('上传') as local_model_tab_upload:
-                            with gr.Row():
-                                model_path = gr.File(label="选择模型文件")
-                                config_path = gr.File(label="选择配置文件")
-                        with gr.TabItem('本地') as local_model_tab_local:
-                            gr.Markdown(f'模型应当放置于{local_model_root}文件夹下')
-                            local_model_refresh_btn = gr.Button('刷新本地模型列表')
-                            local_model_selection = gr.Dropdown(label='选择模型文件夹', choices=[], interactive=True)
-                    with gr.Row():
-                        diff_model_path = gr.File(label="选择扩散模型文件")
-                        diff_config_path = gr.File(label="选择扩散模型配置文件")
-                    cluster_model_path = gr.File(label="选择聚类模型或特征检索文件（没有可以不选）")
-                    device = gr.Dropdown(label="推理设备，默认为自动选择CPU和GPU", choices=["Auto",*cuda.keys(),"cpu"], value="Auto")
-                    enhance = gr.Checkbox(label="是否使用NSF_HIFIGAN增强,该选项对部分训练集少的模型有一定的音质增强效果，但是对训练好的模型有反面效果，默认关闭", value=False)
-                    only_diffusion = gr.Checkbox(label="是否使用全扩散推理，开启后将不使用So-VITS模型，仅使用扩散模型进行完整扩散推理，默认关闭", value=False)
-                with gr.Column():
-                    gr.Markdown(value="""
-                        <font size=3>左侧文件全部选择完毕后(全部文件模块显示download)，点击“加载模型”进行解析：</font>
-                        """)
-                    model_load_button = gr.Button(value="加载模型", variant="primary")
-                    model_unload_button = gr.Button(value="卸载模型", variant="primary")
-                    sid = gr.Dropdown(label="音色（说话人）")
-                    sid_output = gr.Textbox(label="Output Message")
-
-
-            with gr.Row(variant="panel"):
-                with gr.Column():
-                    gr.Markdown(value="""
-                        <font size=2> 推理设置</font>
-                        """)
-                    auto_f0 = gr.Checkbox(label="自动f0预测，配合聚类模型f0预测效果更好,会导致变调功能失效（仅限转换语音，歌声勾选此项会究极跑调）", value=False)
-                    f0_predictor = gr.Dropdown(label="选择F0预测器,可选择crepe,pm,dio,harvest,rmvpe,默认为pm(注意：crepe为原F0使用均值滤波器)", choices=["pm","dio","harvest","crepe","rmvpe"], value="pm")
-                    vc_transform = gr.Number(label="变调（整数，可以正负，半音数量，升高八度就是12）", value=0)
-                    cluster_ratio = gr.Number(label="聚类模型/特征检索混合比例，0-1之间，0即不启用聚类/特征检索。使用聚类/特征检索能提升音色相似度，但会导致咬字下降（如果使用建议0.5左右）", value=0)
-                    slice_db = gr.Number(label="切片阈值", value=-40)
-                    output_format = gr.Radio(label="音频输出格式", choices=["wav", "flac", "mp3"], value = "wav")
-                    noise_scale = gr.Number(label="noise_scale 建议不要动，会影响音质，玄学参数", value=0.4)
-                    k_step = gr.Slider(label="浅扩散步数，只有使用了扩散模型才有效，步数越大越接近扩散模型的结果", value=100, minimum = 1, maximum = 1000)
-                with gr.Column():
-                    pad_seconds = gr.Number(label="推理音频pad秒数，由于未知原因开头结尾会有异响，pad一小段静音段后就不会出现", value=0.5)
-                    cl_num = gr.Number(label="音频自动切片，0为不切片，单位为秒(s)", value=0)
-                    lg_num = gr.Number(label="两端音频切片的交叉淡入长度，如果自动切片后出现人声不连贯可调整该数值，如果连贯建议采用默认值0，注意，该设置会影响推理速度，单位为秒/s", value=0)
-                    lgr_num = gr.Number(label="自动音频切片后，需要舍弃每段切片的头尾。该参数设置交叉长度保留的比例，范围0-1,左开右闭", value=0.75)
-                    enhancer_adaptive_key = gr.Number(label="使增强器适应更高的音域(单位为半音数)|默认为0", value=0)
-                    cr_threshold = gr.Number(label="F0过滤阈值，只有启动crepe时有效. 数值范围从0-1. 降低该值可减少跑调概率，但会增加哑音", value=0.05)
-                    loudness_envelope_adjustment = gr.Number(label="输入源响度包络替换输出响度包络融合比例，越靠近1越使用输出响度包络", value = 0)
-                    second_encoding = gr.Checkbox(label = "二次编码，浅扩散前会对原始音频进行二次编码，玄学选项，效果时好时差，默认关闭", value=False)
-                    use_spk_mix = gr.Checkbox(label = "动态声线融合", value = False, interactive = False)
-            with gr.Tabs():
-                with gr.TabItem("音频转音频"):
-                    vc_input3 = gr.Audio(label="选择音频", type="filepath")
-                    vc_submit = gr.Button("音频转换", variant="primary")
-                with gr.TabItem("文字转音频"):
-                    text2tts=gr.Textbox(label="在此输入要转译的文字。注意，使用该功能建议打开F0预测，不然会很怪")
-                    with gr.Row():
-                        tts_gender = gr.Radio(label = "说话人性别", choices = ["男","女"], value = "男")
-                        tts_lang = gr.Dropdown(label = "选择语言，Auto为根据输入文字自动识别", choices=SUPPORTED_LANGUAGES, value = "Auto")
-                        tts_rate = gr.Slider(label = "TTS语音变速（倍速相对值）", minimum = -1, maximum = 3, value = 0, step = 0.1)
-                        tts_volume = gr.Slider(label = "TTS语音音量（相对值）", minimum = -1, maximum = 1.5, value = 0, step = 0.1)
-                    vc_submit2 = gr.Button("文字转换", variant="primary")
-            with gr.Row():
-                with gr.Column():
-                    vc_output1 = gr.Textbox(label="Output Message")
-                with gr.Column():
-                    vc_output2 = gr.Audio(label="Output Audio", interactive=False)
-
-        with gr.TabItem("小工具/实验室特性"):
-            gr.Markdown(value="""
-                        <font size=2> So-vits-svc 4.0 小工具/实验室特性</font>
-                        """)
-            with gr.Tabs():
-                with gr.TabItem("静态声线融合"):
-                    gr.Markdown(value="""
-                        <font size=2> 介绍:该功能可以将多个声音模型合成为一个声音模型(多个模型参数的凸组合或线性组合)，从而制造出现实中不存在的声线 
-                                          注意：
-                                          1.该功能仅支持单说话人的模型
-                                          2.如果强行使用多说话人模型，需要保证多个模型的说话人数量相同，这样可以混合同一个SpaekerID下的声音
-                                          3.保证所有待混合模型的config.json中的model字段是相同的
-                                          4.输出的混合模型可以使用待合成模型的任意一个config.json，但聚类模型将不能使用
-                                          5.批量上传模型��时候最好把模型放到一个文件夹选中后一起上传
-                                          6.混合比例调整建议大小在0-100之间，也可以调为其他数字，但在线性组合模式下会出现未知的效果
-                                          7.混合完毕后，文件将会保存在项目根目录中，文件名为output.pth
-                                          8.凸组合模式会将混合比例执行Softmax使混合比例相加为1，而线性组合模式不会
-                        </font>
-                        """)
-                    mix_model_path = gr.Files(label="选择需要混合模型文件")
-                    mix_model_upload_button = gr.UploadButton("选择/追加需要混合模型文件", file_count="multiple")
-                    mix_model_output1 = gr.Textbox(
-                                            label="混合比例调整，单位/%",
-                                            interactive = True
-                                         )
-                    mix_mode = gr.Radio(choices=["凸组合", "线性组合"], label="融合模式",value="凸组合",interactive = True)
-                    mix_submit = gr.Button("声线融合启动", variant="primary")
-                    mix_model_output2 = gr.Textbox(
-                                            label="Output Message"
-                                         )
-                    mix_model_path.change(updata_mix_info,[mix_model_path],[mix_model_output1])
-                    mix_model_upload_button.upload(upload_mix_append_file, [mix_model_upload_button,mix_model_path], [mix_model_path,mix_model_output1])
-                    mix_submit.click(mix_submit_click, [mix_model_output1,mix_mode], [mix_model_output2])
-                
-                with gr.TabItem("模型压缩工具"):
-                    gr.Markdown(value="""
-                        该工具可以实现对模型的体积压缩，在**不影响模型推理功能**的情况下，将原本约600M的So-VITS模型压缩至约200M, 大大减少了硬盘的压力。
-                        **注意：压缩后的模型将无法继续训练，请在确认封炉后再压缩。**
-                    """)
-                    model_to_compress = gr.File(label="模型上传")
-                    compress_model_btn = gr.Button("压缩模型", variant="primary")
-                    compress_model_output = gr.Textbox(label="输出信息", value="")
-
-                    compress_model_btn.click(model_compression, [model_to_compress], [compress_model_output])
-                    
-                    
-    with gr.Tabs():
-        with gr.Row(variant="panel"):
-            with gr.Column():
-                gr.Markdown(value="""
-                    <font size=2> WebUI设置</font>
-                    """)
-                debug_button = gr.Checkbox(label="Debug模式，如果向社区反馈BUG需要打开，打开后控制台可以显示具体错误提示", value=debug)
-        # refresh local model list
-        local_model_refresh_btn.click(local_model_refresh_fn, outputs=local_model_selection)
-        # set local enabled/disabled on tab switch
-        local_model_tab_upload.select(lambda: False, outputs=local_model_enabled)
-        local_model_tab_local.select(lambda: True, outputs=local_model_enabled)
-        
-        vc_submit.click(vc_fn, [sid, vc_input3, output_format, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,f0_predictor,enhancer_adaptive_key,cr_threshold,k_step,use_spk_mix,second_encoding,loudness_envelope_adjustment], [vc_output1, vc_output2])
-        vc_submit2.click(vc_fn2, [text2tts, tts_lang, tts_gender, tts_rate, tts_volume, sid, output_format, vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale,pad_seconds,cl_num,lg_num,lgr_num,f0_predictor,enhancer_adaptive_key,cr_threshold,k_step,use_spk_mix,second_encoding,loudness_envelope_adjustment], [vc_output1, vc_output2])
-
-        debug_button.change(debug_change,[],[])
-        model_load_button.click(modelAnalysis,[model_path,config_path,cluster_model_path,device,enhance,diff_model_path,diff_config_path,only_diffusion,use_spk_mix,local_model_enabled,local_model_selection],[sid,sid_output])
-        model_unload_button.click(modelUnload,[],[sid,sid_output])
-    os.system("start http://127.0.0.1:7860")
-    app.launch()
-
-
-