debug

app.py

@@ -9,19 +9,6 @@ import shutil
 from huggingface_hub import snapshot_download, hf_hub_download
 import requests
 import subprocess
-import json
-
-# 获取当前工作目录的绝对路径
-BASE_DIR = os.path.abspath(os.getcwd())
-
-# 修复相对路径为绝对路径的函数
-def fix_path(path):
-    if path is None:
-        return None
-    # 如果是相对路径（以./开头），转换为绝对路径
-    if isinstance(path, str) and path.startswith('./'):
-        return os.path.join(BASE_DIR, path[2:])
-    return path
 
 # 检查并安装必要的依赖
 def install_dependencies():
@@ -113,33 +100,6 @@ def download_amphion_code():
             if not os.path.exists(local_path):
                 with open(local_path, "w") as f:
                     f.write("# Placeholder file\n")
-    
-    # 下载特殊文件：hubert_large_l18_mean_std.npz
-    try:
-        # 确保目录存在
-        os.makedirs(os.path.join(os.getcwd(), "models/vc/vevo/config"), exist_ok=True)
-        
-        # 尝试从HuggingFace下载
-        try:
-            hf_hub_download(
-                repo_id=REPO_ID,
-                filename="tokenizer/vq8192/hubert_large_l18_mean_std.npz",
-                cache_dir=CACHE_DIR,
-                local_dir=os.path.join(os.getcwd(), "models/vc/vevo/config"),
-                local_dir_use_symlinks=False
-            )
-            print("成功下载: hubert_large_l18_mean_std.npz")
-        except Exception as e:
-            print(f"无法从HuggingFace下载hubert_large_l18_mean_std.npz: {str(e)}")
-            # 尝试从GitHub下载
-            hubert_url = "https://huggingface.co/amphion/Vevo/resolve/main/tokenizer/vq8192/hubert_large_l18_mean_std.npz"
-            response = requests.get(hubert_url)
-            if response.status_code == 200:
-                with open(os.path.join(os.getcwd(), "models/vc/vevo/config/hubert_large_l18_mean_std.npz"), "wb") as f:
-                    f.write(response.content)
-                print("成功从HuggingFace下载: hubert_large_l18_mean_std.npz")
-    except Exception as e:
-        print(f"下载hubert_large_l18_mean_std.npz时出错: {str(e)}")
 
 # 先下载必要的代码文件
 download_amphion_code()
@@ -199,10 +159,10 @@ class VevoGradioApp:
         self.pipelines = {}
         # 配置文件路径
         self.config_paths = {
-            "vq32tovq8192": fix_path("./models/vc/vevo/config/Vq32ToVq8192.json"),
-            "vq8192tomels": fix_path("./models/vc/vevo/config/Vq8192ToMels.json"),
-            "phonetovq8192": fix_path("./models/vc/vevo/config/PhoneToVq8192.json"),
-            "vocoder": fix_path("./models/vc/vevo/config/Vocoder.json")
+            "vq32tovq8192": "./models/vc/vevo/config/Vq32ToVq8192.json",
+            "vq8192tomels": "./models/vc/vevo/config/Vq8192ToMels.json",
+            "phonetovq8192": "./models/vc/vevo/config/PhoneToVq8192.json",
+            "vocoder": "./models/vc/vevo/config/Vocoder.json"
         }
         
         # 确保配置文件存在
@@ -210,8 +170,7 @@ class VevoGradioApp:
         
     def download_configs(self):
         """下载必要的配置文件"""
-        config_dir = fix_path("./models/vc/vevo/config")
-        os.makedirs(config_dir, exist_ok=True)
+        os.makedirs("./models/vc/vevo/config", exist_ok=True)
         config_files = {
             "Vq32ToVq8192.json": "https://raw.githubusercontent.com/open-mmlab/Amphion/main/models/vc/vevo/config/Vq32ToVq8192.json",
             "Vq8192ToMels.json": "https://raw.githubusercontent.com/open-mmlab/Amphion/main/models/vc/vevo/config/Vq8192ToMels.json",
@@ -219,8 +178,13 @@ class VevoGradioApp:
             "Vocoder.json": "https://raw.githubusercontent.com/open-mmlab/Amphion/main/models/vc/vevo/config/Vocoder.json"
         }
         
+        # 额外下载必要的统计文件
+        stat_files = {
+            "hubert_large_l18_mean_std.npz": "https://huggingface.co/amphion/Vevo/resolve/main/tokenizer/vq32/hubert_large_l18_mean_std.npz"
+        }
+        
         for filename, url in config_files.items():
-            target_path = os.path.join(config_dir, filename)
+            target_path = f"./models/vc/vevo/config/{filename}"
             if not os.path.exists(target_path):
                 try:
                     response = requests.get(url)
@@ -238,36 +202,81 @@ class VevoGradioApp:
                     with open(target_path, 'w') as f:
                         f.write('{}')
                     print(f"无法下载配置文件 {filename}，已创建占位符")
-                    
-        # 修改配置文件中的相对路径
-        self.update_config_paths()
+        
+        # 下载统计文件
+        for filename, url in stat_files.items():
+            # 同时支持两个位置：配置目录和标准位置
+            target_paths = [
+                f"./models/vc/vevo/config/{filename}",  # 配置文件夹中
+                f"./tokenizer/vq32/{filename}"  # HuggingFace仓库标准位置
+            ]
+            
+            # 确保目录存在
+            for target_path in target_paths:
+                os.makedirs(os.path.dirname(target_path), exist_ok=True)
+                
+                if not os.path.exists(target_path):
+                    try:
+                        response = requests.get(url)
+                        if response.status_code == 200:
+                            with open(target_path, "wb") as f:
+                                f.write(response.content)
+                            print(f"成功下载统计文件到: {target_path}")
+                        else:
+                            print(f"无法下载统计文件 {filename} 到 {target_path}, 状态码: {response.status_code}")
+                    except Exception as e:
+                        print(f"下载统计文件 {filename} 到 {target_path} 时出错: {str(e)}")
+        
+        # 修复配置文件中的路径
+        self.fix_config_paths()
     
-    def update_config_paths(self):
-        """修改配置文件中的相对路径"""
+    def fix_config_paths(self):
+        """修复配置文件中的相对路径"""
         try:
-            # 修改Vq8192ToMels.json文件中的路径引用
-            vq8192_path = self.config_paths["vq8192tomels"]
-            if os.path.exists(vq8192_path):
-                with open(vq8192_path, 'r') as f:
-                    config = json.load(f)
-                
-                # 修改hubert_large_l18_mean_std.npz路径
-                if "model" in config and "representation_stat_mean_var_path" in config["model"]:
-                    # 替换为绝对路径
-                    hubert_stat_path = fix_path("./models/vc/vevo/config/hubert_large_l18_mean_std.npz")
-                    config["model"]["representation_stat_mean_var_path"] = hubert_stat_path
+            for config_name, config_path in self.config_paths.items():
+                if os.path.exists(config_path):
+                    with open(config_path, 'r') as f:
+                        config_data = f.read()
+                    
+                    # 获取当前工作目录的绝对路径
+                    base_dir = os.path.abspath(os.getcwd())
                     
-                    # 保存修改后的配置
-                    with open(vq8192_path, 'w') as f:
-                        json.dump(config, f, indent=2)
-                    print("成功更新配置文件中的路径")
+                    # 替换配置中的相对路径
+                    if 'representation_stat_mean_var_path' in config_data:
+                        # 正确的统计文件路径
+                        stat_file_path = f"{base_dir}/models/vc/vevo/config/hubert_large_l18_mean_std.npz"
+                        
+                        # 替换所有可能的路径格式
+                        replacements = [
+                            ('"representation_stat_mean_var_path": "./models/vc/vevo/config/hubert_large_l18_mean_std.npz"', f'"representation_stat_mean_var_path": "{stat_file_path}"'),
+                            ('"representation_stat_mean_var_path": "models/vc/vevo/config/hubert_large_l18_mean_std.npz"', f'"representation_stat_mean_var_path": "{stat_file_path}"'),
+                            ('"representation_stat_mean_var_path": "./tokenizer/vq32/hubert_large_l18_mean_std.npz"', f'"representation_stat_mean_var_path": "{stat_file_path}"'),
+                            ('"representation_stat_mean_var_path": "tokenizer/vq32/hubert_large_l18_mean_std.npz"', f'"representation_stat_mean_var_path": "{stat_file_path}"'),
+                        ]
+                        
+                        for old, new in replacements:
+                            config_data = config_data.replace(old, new)
+                    
+                    # 保存修复后的配置
+                    with open(config_path, 'w') as f:
+                        f.write(config_data)
+                    
+                    print(f"已修复配置文件路径: {config_path}")
         except Exception as e:
-            print(f"更新配置文件路径时出错: {str(e)}")
+            print(f"修复配置文件路径时出错: {str(e)}")
     
     def init_voice_conversion_pipeline(self):
         """初始化语音转换管道"""
         if "voice" not in self.pipelines:
             try:
+                # 确保配置文件路径是绝对路径
+                absolute_config_paths = {}
+                for key, path in self.config_paths.items():
+                    if path and not os.path.isabs(path):
+                        absolute_config_paths[key] = os.path.abspath(path)
+                    else:
+                        absolute_config_paths[key] = path
+                
                 # 内容标记器
                 local_dir = snapshot_download(
                     repo_id=REPO_ID,
@@ -315,31 +324,43 @@ class VevoGradioApp:
                 )
                 vocoder_ckpt_path = os.path.join(local_dir, "acoustic_modeling/Vocoder")
                 
+                # 确保统计文件存在
+                possible_stat_file_paths = [
+                    os.path.join(os.getcwd(), "models/vc/vevo/config/hubert_large_l18_mean_std.npz"),
+                    os.path.join(os.getcwd(), "tokenizer/vq32/hubert_large_l18_mean_std.npz")
+                ]
+                
+                # 检查是否有任一路径存在
+                stat_file_exists = any(os.path.exists(path) for path in possible_stat_file_paths)
+                
+                if not stat_file_exists:
+                    print(f"警告: 找不到统计文件，将尝试创建空文件")
+                    try:
+                        import numpy as np
+                        # 在两个位置都创建一个简单的统计文件
+                        for stat_path in possible_stat_file_paths:
+                            os.makedirs(os.path.dirname(stat_path), exist_ok=True)
+                            np.savez(stat_path, mean=np.zeros(1024), std=np.ones(1024))
+                            print(f"已创建占位符统计文件: {stat_path}")
+                    except Exception as e:
+                        print(f"创建统计文件时出错: {str(e)}")
+                
                 # 创建推理管道
                 self.pipelines["voice"] = VevoInferencePipeline(
                     content_tokenizer_ckpt_path=content_tokenizer_ckpt_path,
                     content_style_tokenizer_ckpt_path=content_style_tokenizer_ckpt_path,
-                    ar_cfg_path=self.config_paths["vq32tovq8192"],
+                    ar_cfg_path=absolute_config_paths["vq32tovq8192"],
                     ar_ckpt_path=ar_ckpt_path,
-                    fmt_cfg_path=self.config_paths["vq8192tomels"],
+                    fmt_cfg_path=absolute_config_paths["vq8192tomels"],
                     fmt_ckpt_path=fmt_ckpt_path,
-                    vocoder_cfg_path=self.config_paths["vocoder"],
+                    vocoder_cfg_path=absolute_config_paths["vocoder"],
                     vocoder_ckpt_path=vocoder_ckpt_path,
                     device=self.device,
                 )
             except Exception as e:
                 print(f"初始化语音转换管道时出错: {str(e)}")
                 # 创建一个占位符管道
-                try:
-                    # 尝试提供必要的配置文件
-                    self.pipelines["voice"] = VevoInferencePipeline(
-                        device=self.device,
-                        fmt_cfg_path=self.config_paths["vq8192tomels"],
-                        vocoder_cfg_path=self.config_paths["vocoder"],
-                    )
-                except:
-                    # 如果还是失败，创建最简单的管道
-                    self.pipelines["voice"] = VevoInferencePipeline(device=self.device)
+                self.pipelines["voice"] = VevoInferencePipeline(device=self.device)
             
         return self.pipelines["voice"]
     
@@ -347,6 +368,14 @@ class VevoGradioApp:
         """初始化音色转换管道"""
         if "timbre" not in self.pipelines:
             try:
+                # 确保配置文件路径是绝对路径
+                absolute_config_paths = {}
+                for key, path in self.config_paths.items():
+                    if path and not os.path.isabs(path):
+                        absolute_config_paths[key] = os.path.abspath(path)
+                    else:
+                        absolute_config_paths[key] = path
+                
                 # 内容-风格标记器
                 local_dir = snapshot_download(
                     repo_id=REPO_ID,
@@ -377,25 +406,16 @@ class VevoGradioApp:
                 # 创建推理管道
                 self.pipelines["timbre"] = VevoInferencePipeline(
                     content_style_tokenizer_ckpt_path=tokenizer_ckpt_path,
-                    fmt_cfg_path=self.config_paths["vq8192tomels"],
+                    fmt_cfg_path=absolute_config_paths["vq8192tomels"],
                     fmt_ckpt_path=fmt_ckpt_path,
-                    vocoder_cfg_path=self.config_paths["vocoder"],
+                    vocoder_cfg_path=absolute_config_paths["vocoder"],
                     vocoder_ckpt_path=vocoder_ckpt_path,
                     device=self.device,
                 )
             except Exception as e:
                 print(f"初始化音色转换管道时出错: {str(e)}")
                 # 创建一个占位符管道
-                try:
-                    # 尝试提供必要的配置文件
-                    self.pipelines["timbre"] = VevoInferencePipeline(
-                        device=self.device,
-                        fmt_cfg_path=self.config_paths["vq8192tomels"],
-                        vocoder_cfg_path=self.config_paths["vocoder"],
-                    )
-                except:
-                    # 如果还是失败，创建最简单的管道
-                    self.pipelines["timbre"] = VevoInferencePipeline(device=self.device)
+                self.pipelines["timbre"] = VevoInferencePipeline(device=self.device)
             
         return self.pipelines["timbre"]
     
@@ -403,6 +423,14 @@ class VevoGradioApp:
         """初始化文本转语音管道"""
         if "tts" not in self.pipelines:
             try:
+                # 确保配置文件路径是绝对路径
+                absolute_config_paths = {}
+                for key, path in self.config_paths.items():
+                    if path and not os.path.isabs(path):
+                        absolute_config_paths[key] = os.path.abspath(path)
+                    else:
+                        absolute_config_paths[key] = path
+                
                 # 内容-风格标记器
                 local_dir = snapshot_download(
                     repo_id=REPO_ID,
@@ -442,28 +470,18 @@ class VevoGradioApp:
                 # 创建推理管道
                 self.pipelines["tts"] = VevoInferencePipeline(
                     content_style_tokenizer_ckpt_path=content_style_tokenizer_ckpt_path,
-                    ar_cfg_path=self.config_paths["phonetovq8192"],
+                    ar_cfg_path=absolute_config_paths["phonetovq8192"],
                     ar_ckpt_path=ar_ckpt_path,
-                    fmt_cfg_path=self.config_paths["vq8192tomels"],
+                    fmt_cfg_path=absolute_config_paths["vq8192tomels"],
                     fmt_ckpt_path=fmt_ckpt_path,
-                    vocoder_cfg_path=self.config_paths["vocoder"],
+                    vocoder_cfg_path=absolute_config_paths["vocoder"],
                     vocoder_ckpt_path=vocoder_ckpt_path,
                     device=self.device,
                 )
             except Exception as e:
                 print(f"初始化TTS管道时出错: {str(e)}")
                 # 创建一个占位符管道
-                try:
-                    # 尝试提供必要的配置文件
-                    self.pipelines["tts"] = VevoInferencePipeline(
-                        device=self.device,
-                        fmt_cfg_path=self.config_paths["vq8192tomels"],
-                        vocoder_cfg_path=self.config_paths["vocoder"],
-                        ar_cfg_path=self.config_paths["phonetovq8192"],
-                    )
-                except:
-                    # 如果还是失败，创建最简单的管道
-                    self.pipelines["tts"] = VevoInferencePipeline(device=self.device)
+                self.pipelines["tts"] = VevoInferencePipeline(device=self.device)
             
         return self.pipelines["tts"]