fix: Add weight whitelist support for torch 2.6 (#110)

configs/__init__.py

@@ -1 +1 @@
-from .config import singleton_variable, Config, CPUConfig
+from .config import Singleton, Config, CPUConfig

configs/config.py

@@ -22,18 +22,16 @@ version_config_list = [
 ]
 
 
-def singleton_variable(func):
-    def wrapper(*args, **kwargs):
-        if wrapper.instance is None:
-            wrapper.instance = func(*args, **kwargs)
-        return wrapper.instance
+class Singleton(type):
+    _instances = {}
 
-    wrapper.instance = None
-    return wrapper
+    def __call__(cls, *args, **kwargs):
+        if cls not in cls._instances:
+            cls._instances[cls] = super(Singleton, cls).__call__(*args, **kwargs)
+        return cls._instances[cls]
 
 
-@singleton_variable
-class Config:
+class Config(metaclass=Singleton):
     def __init__(self):
         self.device = "cuda:0"
         self.is_half = True
@@ -129,6 +127,16 @@ class Config:
         else:
             return False
 
+    @staticmethod
+    def use_insecure_load():
+        try:
+            from fairseq.data.dictionary import Dictionary
+
+            torch.serialization.add_safe_globals([Dictionary])
+            logging.warning("Using insecure weight loading for fairseq dictionary")
+        except AttributeError:
+            pass
+
     def use_fp32_config(self):
         for config_file in version_config_list:
             self.json_config[config_file]["train"]["fp16_run"] = False
@@ -210,15 +218,20 @@ class Config:
         else:
             if self.instead:
                 logger.info(f"Use {self.instead} instead")
+
         logger.info(
             "Half-precision floating-point: %s, device: %s"
             % (self.is_half, self.device)
         )
+
+        # Check if the pytorch is 2.6 or higher
+        if tuple(map(int, torch.__version__.split("+")[0].split("."))) >= (2, 6, 0):
+            self.use_insecure_load()
+
         return x_pad, x_query, x_center, x_max
 
 
-@singleton_variable
-class CPUConfig:
+class CPUConfig(metaclass=Singleton):
     def __init__(self):
         self.device = "cpu"
         self.is_half = False

i18n/i18n.py

@@ -1,7 +1,7 @@
 import json
 import locale
 import os
-from configs import singleton_variable
+from configs import Singleton
 
 
 def load_language_list(language):
@@ -10,8 +10,7 @@ def load_language_list(language):
     return language_list
 
 
-@singleton_variable
-class I18nAuto:
+class I18nAuto(metaclass=Singleton):
     def __init__(self, language=None):
         if language in ["Auto", None]:
             language = locale.getdefaultlocale(

infer/modules/train/extract_f0_print.py

@@ -96,6 +96,10 @@ if __name__ == "__main__":
     # exp_dir=r"E:\codes\py39\dataset\mi-test"
     # n_p=16
     # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+
+    from configs import Config
+    Config.use_insecure_load()
+    
     printt(" ".join(sys.argv))
     featureInput = FeatureInput(is_half, device)
     paths = []

infer/modules/train/extract_feature_print.py

@@ -23,11 +23,15 @@ else:
     os.environ["CUDA_VISIBLE_DEVICES"] = str(i_gpu)
     version = sys.argv[6]
     is_half = sys.argv[7].lower() == "true"
+
 import fairseq
 import numpy as np
 import torch
 import torch.nn.functional as F
 
+from configs import Config
+Config.use_insecure_load()
+
 if "privateuseone" not in device:
     device = "cpu"
     if torch.cuda.is_available():

infer/modules/train/preprocess.py

@@ -142,4 +142,6 @@ def preprocess_trainset(inp_root, sr, n_p, exp_dir, per):
 
 
 if __name__ == "__main__":
+    from configs import Config
+    Config.use_insecure_load()
     preprocess_trainset(inp_root, sr, n_p, exp_dir, per)

infer/modules/vc/hash.py

@@ -2,10 +2,12 @@ import numpy as np
 import torch
 import hashlib
 import pathlib
+
+from functools import lru_cache
 from scipy.fft import fft
 from pybase16384 import encode_to_string, decode_from_string
 
-from configs import CPUConfig, singleton_variable
+from configs import CPUConfig
 from rvc.synthesizer import get_synthesizer
 
 from .pipeline import Pipeline
@@ -29,27 +31,27 @@ half_hash_len = 512
 expand_factor = 65536 * 8
 
 
-@singleton_variable
+@lru_cache(None)  # None 表示无限缓存
 def original_audio_storage():
     return np.load(pathlib.Path(__file__).parent / "lgdsng.npz")
 
 
-@singleton_variable
+@lru_cache(None)
 def original_audio():
     return original_audio_storage()["a"]
 
 
-@singleton_variable
+@lru_cache(None)
 def original_audio_time_minus():
     return original_audio_storage()["t"]
 
 
-@singleton_variable
+@lru_cache(None)
 def original_audio_freq_minus():
     return original_audio_storage()["f"]
 
 
-@singleton_variable
+@lru_cache(None)
 def original_rmvpe_f0():
     x = original_audio_storage()
     return x["pitch"], x["pitchf"]