Merge 5867122df2d08eacbdb6ffc64691403fa00e54bb into 0be59c8043a12112934d474ff4cc65658d848e8f

* 修复gpt的loss计算问题

* fallback tts config

GPT_SoVITS/AR/models/t2s_model.py

@@ -356,7 +356,7 @@ class Text2SemanticDecoder(nn.Module):
         x = self.ar_text_embedding(x)
         x = x + self.bert_proj(bert_feature.transpose(1, 2))
         x = self.ar_text_position(x)
-        x_mask = make_pad_mask(x_lens)
+        x_mask = make_pad_mask_left(x_lens)
 
         y_mask = make_pad_mask(y_lens)
         y_mask_int = y_mask.type(torch.int64)
@@ -420,7 +420,7 @@ class Text2SemanticDecoder(nn.Module):
             mask=xy_attn_mask,
         )
         x_len = x_lens.max()
-        logits = self.ar_predict_layer(xy_dec[:, x_len:])
+        logits = self.ar_predict_layer(xy_dec[:, x_len-1:])
 
         ###### DPO #############
         reject_xy_pos, reject_xy_attn_mask, reject_targets = self.make_input_data(
@@ -432,7 +432,7 @@ class Text2SemanticDecoder(nn.Module):
             mask=reject_xy_attn_mask,
         )
         x_len = x_lens.max()
-        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len:])
+        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len-1:])
 
         # loss
         # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
@@ -455,7 +455,7 @@ class Text2SemanticDecoder(nn.Module):
         x = self.ar_text_embedding(x)
         x = x + self.bert_proj(bert_feature.transpose(1, 2))
         x = self.ar_text_position(x)
-        x_mask = make_pad_mask(x_lens)
+        x_mask = make_pad_mask_left(x_lens)
 
         y_mask = make_pad_mask(y_lens)
         y_mask_int = y_mask.type(torch.int64)
@@ -502,7 +502,7 @@ class Text2SemanticDecoder(nn.Module):
             (xy_pos, None),
             mask=xy_attn_mask,
         )
-        logits = self.ar_predict_layer(xy_dec[:, x_len:]).permute(0, 2, 1)
+        logits = self.ar_predict_layer(xy_dec[:, x_len-1:]).permute(0, 2, 1)
         # loss
         # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
         loss = F.cross_entropy(logits, targets, reduction="sum")
@@ -578,7 +578,7 @@ class Text2SemanticDecoder(nn.Module):
     def pad_y_eos(self, y, y_mask_int, eos_id):
         targets = F.pad(y, (0, 1), value=0) + eos_id * F.pad(y_mask_int, (0, 1), value=1)
         # 错位
-        return targets[:, :-1], targets[:, 1:]
+        return targets[:, :-1], targets
 
     def infer_panel_batch_infer(
         self,

GPT_SoVITS/TTS_infer_pack/TTS.py

@@ -304,10 +304,10 @@ class TTS_Config:
             configs: dict = self._load_configs(self.configs_path)
 
         assert isinstance(configs, dict)
-        version = configs.get("version", "v2").lower()
-        assert version in ["v1", "v2", "v3", "v4", "v2pro", "v2proplus"]
-        self.default_configs[version] = configs.get(version, self.default_configs[version])
-        self.configs: dict = configs.get("custom", deepcopy(self.default_configs[version]))
+        configs_ = deepcopy(self.default_configs)
+        configs_.update(configs)
+        self.configs: dict = configs_.get("custom", configs_["v2"])
+        self.default_configs = deepcopy(configs_)
 
         self.device = self.configs.get("device", torch.device("cpu"))
         if "cuda" in str(self.device) and not torch.cuda.is_available():
@@ -315,11 +315,13 @@ class TTS_Config:
             self.device = torch.device("cpu")
 
         self.is_half = self.configs.get("is_half", False)
-        # if str(self.device) == "cpu" and self.is_half:
-        #     print(f"Warning: Half precision is not supported on CPU, set is_half to False.")
-        #     self.is_half = False
+        if str(self.device) == "cpu" and self.is_half:
+            print(f"Warning: Half precision is not supported on CPU, set is_half to False.")
+            self.is_half = False
 
+        version = self.configs.get("version", None)
         self.version = version
+        assert self.version in ["v1", "v2", "v3", "v4", "v2Pro", "v2ProPlus"], "Invalid version!"
         self.t2s_weights_path = self.configs.get("t2s_weights_path", None)
         self.vits_weights_path = self.configs.get("vits_weights_path", None)
         self.bert_base_path = self.configs.get("bert_base_path", None)
@@ -576,6 +578,10 @@ class TTS:
         if self.configs.is_half and str(self.configs.device) != "cpu":
             self.vits_model = self.vits_model.half()
 
+        self.configs.save_configs()
+
+
+
     def init_t2s_weights(self, weights_path: str):
         print(f"Loading Text2Semantic weights from {weights_path}")
         self.configs.t2s_weights_path = weights_path

GPT_SoVITS/configs/tts_infer.yaml

@@ -1,4 +1,3 @@
-version: v2ProPlus
 custom:
   bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
   cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base

GPT_SoVITS/inference_webui_fast.py

@@ -125,7 +125,8 @@ is_exist_s2gv4 = os.path.exists(path_sovits_v4)
 tts_config = TTS_Config("GPT_SoVITS/configs/tts_infer.yaml")
 tts_config.device = device
 tts_config.is_half = is_half
-tts_config.version = version
+# tts_config.version = version
+tts_config.update_version(version)
 if gpt_path is not None:
     if "！" in gpt_path or "!" in gpt_path:
         gpt_path = name2gpt_path[gpt_path]

api_model_manager.py

@@ -0,0 +1,212 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+import os
+import json
+import glob
+import re
+from typing import Dict, List, Tuple, Optional
+import logging
+
+logger = logging.getLogger("gpt-sovits-api")
+
+class ModelManager:
+    """
+    GPT-SoVITS模型管理器
+    用于管理GPT和SoVITS模型的映射关系
+    """
+    def __init__(self):
+        self.gpt_weights_dir = "GPT_weights"
+        self.sovits_weights_dir = "SoVITS_weights"
+        
+        # 扫描多个版本的模型目录
+        self.gpt_dirs = [
+            "GPT_weights", 
+            "GPT_weights_v2", 
+            "GPT_weights_v3", 
+            "GPT_weights_v4"
+        ]
+        
+        self.sovits_dirs = [
+            "SoVITS_weights", 
+            "SoVITS_weights_v2", 
+            "SoVITS_weights_v3", 
+            "SoVITS_weights_v4"
+        ]
+        
+        # 模型映射缓存
+        self.model_mapping = {}
+        self.voice_info = {}
+        
+        # 加载模型映射
+        self.load_model_mapping()
+    
+    def _extract_model_info(self, filename: str) -> Dict:
+        """
+        从模型文件名中提取信息
+        支持多种命名格式:
+        1. 模型名_e迭代次数_s批次.pth 
+        2. 模型名-e迭代次数.ckpt
+        
+        Args:
+            filename: 模型文件名
+        
+        Returns:
+            Dict: 包含模型名称、迭代次数和批次的字典
+        """
+        basename = os.path.basename(filename)
+        name_parts = basename.split('.')
+        base_name = name_parts[0]
+        
+        # 尝试匹配迭代次数 (e参数)，支持连字符(-)和下划线(_)
+        e_match = re.search(r"[-_]e(\d+)", base_name)
+        
+        # 尝试匹配批次 (s参数)，主要在SoVITS模型中使用
+        s_match = re.search(r"[-_]s(\d+)", base_name)
+        
+        # 提取模型名称（去掉e和s参数部分）
+        model_name = base_name
+        
+        # 如果找到了e参数
+        if e_match:
+            # 获取e参数之前的部分作为模型名称
+            e_pos = base_name.find(e_match.group(0))
+            if e_pos > 0:
+                separator = base_name[e_pos]  # 获取分隔符 (- 或 _)
+                model_name = base_name.split(f"{separator}e")[0]
+        
+        # 提取扩展名
+        ext = os.path.splitext(basename)[1].lower()
+        
+        iteration = int(e_match.group(1)) if e_match else 0
+        batch = int(s_match.group(1)) if s_match else 0
+        
+        logger.debug(f"解析模型: {basename} -> 名称={model_name}, 迭代={iteration}, 批次={batch}")
+        
+        return {
+            "name": model_name,
+            "iteration": iteration,
+            "batch": batch,
+            "filename": filename
+        }
+    
+    def load_model_mapping(self):
+        """
+        扫描模型目录，创建模型映射关系
+        将相同名称的GPT和SoVITS模型进行匹配
+        """
+        # 扫描GPT模型
+        gpt_models = {}
+        for dir_path in self.gpt_dirs:
+            if not os.path.exists(dir_path):
+                continue
+            
+            for file_path in glob.glob(f"{dir_path}/*.ckpt"):
+                model_info = self._extract_model_info(file_path)
+                model_name = model_info["name"]
+                
+                # 使用更高迭代次数和批次的模型
+                if model_name not in gpt_models or \
+                   (model_info["iteration"] > gpt_models[model_name]["iteration"] or \
+                   (model_info["iteration"] == gpt_models[model_name]["iteration"] and \
+                    model_info["batch"] > gpt_models[model_name]["batch"])):
+                    gpt_models[model_name] = model_info
+        
+        # 扫描SoVITS模型
+        sovits_models = {}
+        for dir_path in self.sovits_dirs:
+            if not os.path.exists(dir_path):
+                continue
+            
+            for file_path in glob.glob(f"{dir_path}/*.pth"):
+                model_info = self._extract_model_info(file_path)
+                model_name = model_info["name"]
+                
+                # 使用更高迭代次数和批次的模型
+                if model_name not in sovits_models or \
+                   (model_info["iteration"] > sovits_models[model_name]["iteration"] or \
+                   (model_info["iteration"] == sovits_models[model_name]["iteration"] and \
+                    model_info["batch"] > sovits_models[model_name]["batch"])):
+                    sovits_models[model_name] = model_info
+        
+        # 创建映射关系
+        for name in set(list(gpt_models.keys()) + list(sovits_models.keys())):
+            gpt_model = gpt_models.get(name)
+            sovits_model = sovits_models.get(name)
+            
+            if gpt_model and sovits_model:
+                self.model_mapping[name] = {
+                    "gpt_path": gpt_model["filename"],
+                    "sovits_path": sovits_model["filename"],
+                    "iteration": min(gpt_model["iteration"], sovits_model["iteration"]),
+                    "batch": min(gpt_model["batch"], sovits_model["batch"])
+                }
+                
+                self.voice_info[name] = {
+                    "id": name,
+                    "name": name,
+                    "iteration": min(gpt_model["iteration"], sovits_model["iteration"]),
+                    "batch": min(gpt_model["batch"], sovits_model["batch"])
+                }
+        
+        logger.info(f"已加载 {len(self.model_mapping)} 个模型映射")
+    
+    def get_model_paths(self, voice_name: str) -> Tuple[Optional[str], Optional[str]]:
+        """
+        获取指定voice对应的GPT和SoVITS模型路径
+        
+        Args:
+            voice_name: 声音名称
+        
+        Returns:
+            Tuple[str, str]: (GPT模型路径, SoVITS模型路径)
+        """
+        if voice_name in self.model_mapping:
+            return (
+                self.model_mapping[voice_name]["gpt_path"],
+                self.model_mapping[voice_name]["sovits_path"]
+            )
+        return None, None
+    
+    def get_all_voices(self) -> List[Dict]:
+        """
+        获取所有可用的声音列表
+        
+        Returns:
+            List[Dict]: 声音信息列表
+        """
+        return [self.voice_info[name] for name in self.voice_info]
+    
+    def get_voice_details(self, voice_name: str) -> Optional[Dict]:
+        """
+        获取指定声音的详细信息
+        
+        Args:
+            voice_name: 声音名称
+        
+        Returns:
+            Dict: 声音详细信息
+        """
+        if voice_name in self.voice_info:
+            info = self.voice_info[voice_name].copy()
+            info.update({
+                "gpt_path": self.model_mapping[voice_name]["gpt_path"],
+                "sovits_path": self.model_mapping[voice_name]["sovits_path"]
+            })
+            return info
+        return None
+
+# 单例模式
+model_manager = ModelManager()
+
+if __name__ == "__main__":
+    # 测试代码
+    logging.basicConfig(level=logging.INFO)
+    manager = ModelManager()
+    voices = manager.get_all_voices()
+    print(f"发现 {len(voices)} 个声音模型:")
+    for voice in voices:
+        print(f"- {voice['name']}, 迭代次数: {voice['iteration']}, 批次: {voice['batch']}")
+        gpt_path, sovits_path = manager.get_model_paths(voice['name'])
+        print(f"  GPT: {gpt_path}")
+        print(f"  SoVITS: {sovits_path}") 

tools/assets.py

@@ -59,7 +59,7 @@ top_html = """
         <a href="https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e" target="_blank">
             <img src="https://img.shields.io/badge/简体中文-阅读文档-blue?style=for-the-badge&logo=googledocs&logoColor=white" style="width: auto; height: 30px;">
         </a>
-        <a href="https://github.com/RVC-Boss/GPT-SoVITS" target="_blank">
+        <a href="https://lj1995-gpt-sovits-proplus.hf.space/" target="_blank">
             <img src="https://img.shields.io/badge/免费在线体验-free_online_demo-yellow.svg?style=for-the-badge&logo=huggingface" style="width: auto; height: 30px;">
         </a>
         <a href="https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e" target="_blank">