vits

app.py

@@ -1,7 +1,5 @@
 import gradio as gr
-from tts_inferencer import TTSInferencer
-
-tts_inferencer = TTSInferencer("./configs/hoshimi_base.json")
+from vits.tts_inferencer import TTSInferencer
 
 app = gr.Blocks()
 with app:
@@ -9,5 +7,6 @@ with app:
         gr.HTML(f.read())
     with gr.Tabs():
         with gr.TabItem("语音合成"):
+            tts_inferencer = TTSInferencer("vits/configs/hoshimi_base.json")
             tts_inferencer.render()
     app.launch()

tts_inferencer.py

@@ -1,108 +0,0 @@
-import os
-import json
-import math
-import torch
-from torch import nn
-from torch.nn import functional as F
-from torch.utils.data import DataLoader
-
-import commons
-import utils
-from models import SynthesizerTrn
-from text.symbols import symbols
-from text import text_to_sequence
-import gradio as gr
-
-mode_dict = {
-    "普通声线": "normal",
-    "营业声线": "formal"
-}
-
-default_mode = "普通声线"
-default_noise_scale = 0.667
-default_noise_scale_w = 0.8
-default_length_scale = 1
-
-def get_text(text, hps):
-    text_norm = text_to_sequence(text, hps.data.text_cleaners)
-    if hps.data.add_blank:
-        text_norm = commons.intersperse(text_norm, 0)
-    text_norm = torch.LongTensor(text_norm)
-    return text_norm
-
-class TTSInferencer:
-    def __init__(self, hps_path, device="cpu"):
-        self.device = torch.device(device)
-        self.hps = utils.get_hparams_from_file(hps_path)
-        self.select_mode(default_mode)
-        self.load_model(self.latest_model_path)
-
-    def select_mode(self, mode):
-        self.mode = mode
-        self.model_dir_path = os.path.join("models", mode_dict[mode])
-        self.models = []
-        for f in os.listdir(self.model_dir_path):
-            if (f.startswith("D_")):
-                continue
-            if (f.endswith(".pth")):
-                self.models.append(f)
-        self.latest_model_path = utils.latest_checkpoint_path(self.model_dir_path, "G_*.pth")
-
-    def infer(self, text, noise_scale=.667, noise_scale_w=0.8, length_scale=1):
-        stn_tst = get_text(text, self.hps)
-        with torch.no_grad():
-            x_tst = stn_tst.unsqueeze(0).to(self.device)
-            x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(self.device)
-            audio = self.net_g.infer(x_tst, x_tst_lengths, noise_scale=noise_scale, noise_scale_w=noise_scale_w, length_scale=length_scale)[0][0,0].data.float().numpy()
-        return (self.hps.data.sampling_rate, audio)
-
-    def change_mode(self, mode):
-        self.select_mode(mode)
-        return gr.update(choices=self.models, value=os.path.basename(self.latest_model_path))
-
-    def change_model(self, model_file_name):
-        self.load_model(os.path.join(self.model_dir_path, model_file_name))
-        return f"载入模型：{model_file_name}（{self.mode}）"
-    
-    def render(self):
-        choice_mode = gr.Radio(choices=["普通声线", "营业声线"], label="声线选择", value=default_mode)
-        choice_model = gr.Dropdown(choices=self.models, label=f"模型迭代版本选择", value=os.path.basename(self.pth_path))
-        # with gr.Row():
-        #     advanced = gr.Checkbox(label="显示高级设置（效果不可控）") 
-        #     default = gr.Button("恢复默认设置").style(full_width=False)
-        noise_scale = gr.Slider(minimum=0, maximum=3, value=default_noise_scale, step=0.001, label="noise_scale（效果不可控，谨慎修改）")
-        noise_scale_w = gr.Slider(minimum=0, maximum=3, value=default_noise_scale_w, step=0.001, label="noise_scale_w（效果不可控，谨慎修改）")
-        length_scale = gr.Slider(minimum=0, maximum=3, value=default_length_scale, step=0.001, label="length_scale（数值越大输出音频越长）")
-        
-        tts_input = gr.TextArea(
-            label="请输入文本（目前只支持汉字和单个英文字母，可以使用常用符号和空格来改变语调和停顿，请勿一次性输入过长文本）",
-            value="这里是爱喝奶茶，穿得也像奶茶魅力点是普通话二乙的星弥吼西咪，晚上齁。")
-        tts_submit = gr.Button("合成", variant="primary")
-        tts_output = gr.Audio(label="Output")
-        tts_model = gr.Markdown(f"载入模型：{os.path.basename(self.latest_model_path)}（{self.mode})")
-        gr.HTML('''
-            <div style="text-align:right;font-size:12px;color:#4D4D4D">
-                <div class="font-medium">版权声明</div>
-                <div>本项目数据集和模型版权属于星弥Hoshimi</div>
-                <div>仅供学习交流，不可用于任何商业和非法用途，否则后果自负</div>
-            </div>
-        ''')
-        # advanced.change(fn=lambda visible: gr.update(visible=visible), inputs=advanced, outputs=noise_scale)
-        # advanced.change(fn=lambda visible: gr.update(visible=visible), inputs=advanced, outputs=noise_scale_w)
-        # default.click(fn=lambda visible: gr.update(value=default_noise_scale), inputs=advanced, outputs=noise_scale)
-        # default.click(fn=lambda visible: gr.update(value=default_noise_scale_w), inputs=advanced, outputs=noise_scale_w)
-        # default.click(fn=lambda visible: gr.update(value=default_length_scale), inputs=advanced, outputs=length_scale)        
-        choice_mode.change(self.change_mode, inputs=choice_mode, outputs=choice_model)
-        choice_model.change(self.change_model, inputs=[choice_model], outputs=[tts_model])
-        tts_submit.click(self.infer, [tts_input, noise_scale, noise_scale_w, length_scale], [tts_output], api_name=f"infer")
-
-    
-    def load_model(self, model_path):
-        self.pth_path = model_path
-        self.net_g = SynthesizerTrn(
-            len(symbols),
-            self.hps.data.filter_length // 2 + 1,
-            self.hps.train.segment_size // self.hps.data.hop_length,
-            **self.hps.model).to(self.device)
-        _ = self.net_g.eval()
-        _ = utils.load_checkpoint(self.pth_path, self.net_g, None)

vits/__init__.py

@@ -0,0 +1,5 @@
+import os
+import sys
+
+ROOT_PATH = os.path.dirname(os.path.abspath(__file__))
+sys.path.append(ROOT_PATH)

models.py → vits/models.py

@@ -7,7 +7,6 @@ from torch.nn import functional as F
 import commons
 import modules
 import attentions
-import monotonic_align
 
 from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
 from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm