remove choice model

vits/tts_inferencer.py

@@ -33,44 +33,33 @@ def get_text(text, hps):
 
 class TTSInferencer:
     def __init__(self, hps_path, device="cpu"):
+        print("init")
         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)
+        self.model_paths = {}
+        self.models = {}
+        for mode in mode_dict:
+            self.model_paths[mode] = self.get_latest_model_path_by_mode(mode)
+        self.load_models()
 
-    def select_mode(self, mode):
-        self.mode = mode
-        self.model_dir_path = os.path.join(ROOT_PATH, "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 get_latest_model_path_by_mode(self, mode):
+        model_dir_path = os.path.join(ROOT_PATH, "models", mode_dict[mode])
+        return utils.latest_checkpoint_path(model_dir_path, "G_*.pth")
 
-    def infer(self, text, noise_scale=.667, noise_scale_w=0.8, length_scale=1):
+    def infer(self, text, mode, 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()
+            audio = self.models[mode].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）")
@@ -80,7 +69,6 @@ class TTSInferencer:
             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>
@@ -88,22 +76,14 @@ class TTSInferencer:
                 <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")
-
+        tts_submit.click(self.infer, [tts_input, choice_mode, 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(
+    def load_models(self):
+        for key, model_path in self.model_paths.items():
+            self.models[key] = 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)
+            _ = self.models[key].eval()
+            _ = utils.load_checkpoint(model_path, self.models[key], None)