update

modules/SynthesizeSegments.py

@@ -1,15 +1,15 @@
-import numpy as np
 from pydub import AudioSegment
 from typing import Any, List, Dict, Union
 from scipy.io.wavfile import write
 import io
+from modules.utils import rng
 from modules.utils.audio import time_stretch, pitch_shift
 from modules import generate_audio
 from modules.normalization import text_normalize
 import logging
 import json
-import random
 import copy
+import numpy as np
 
 from modules.speaker import Speaker
 
@@ -55,8 +55,8 @@ def to_number(value, t, default=0):
 
 
 class SynthesizeSegments:
-    batch_default_spk_seed = int(np.random.randint(0, 2**32 - 1))
-    batch_default_infer_seed = int(np.random.randint(0, 2**32 - 1))
+    batch_default_spk_seed = rng.np_rng()
+    batch_default_infer_seed = rng.np_rng()
 
     def __init__(self, batch_size: int = 8):
         self.batch_size = batch_size

modules/utils/SeedContext.py

@@ -1,14 +1,16 @@
 import torch
 import random
 import numpy as np
+from modules.utils import rng
 
 
 def deterministic(seed=0):
     random.seed(seed)
     np.random.seed(seed)
-    torch.manual_seed(seed)
+    torch_rn = rng.convert_np_to_torch(seed)
+    torch.manual_seed(torch_rn)
     if torch.cuda.is_available():
-        torch.cuda.manual_seed_all(seed)
+        torch.cuda.manual_seed_all(torch_rn)
     torch.backends.cudnn.deterministic = True
     torch.backends.cudnn.benchmark = False
 

modules/utils/rng.py

@@ -0,0 +1,35 @@
+import numpy as np
+import torch
+import random
+
+TORCH_RNG_MAX = -0x8000000000000000
+TORCH_RNG_MIN = 0xFFFFFFFFFFFFFFFF
+
+NP_RNG_MAX = np.iinfo(np.uint32).max
+NP_RNG_MIN = 0
+
+
+def troch_rng(seed: int):
+    torch.manual_seed(seed)
+    random_float = torch.empty(1).uniform_().item()
+    torch_rn = int(random_float * (TORCH_RNG_MAX - TORCH_RNG_MIN) + TORCH_RNG_MIN)
+    np_rn = int(random_float * (NP_RNG_MAX - NP_RNG_MIN) + NP_RNG_MIN)
+    return torch_rn, np_rn
+
+
+def convert_np_to_torch(np_rn: int):
+    random_float = (np_rn - NP_RNG_MIN) / (NP_RNG_MAX - NP_RNG_MIN)
+    torch_rn = int(random_float * (TORCH_RNG_MAX - TORCH_RNG_MIN) + TORCH_RNG_MIN)
+    return torch_rn
+
+
+def np_rng():
+    return int(np.random.randint(NP_RNG_MIN, NP_RNG_MAX, dtype=np.uint32))
+
+
+if __name__ == "__main__":
+    import random
+
+    s1 = np_rng()
+    s2 = troch_rng(s1)
+    print(f"s1 {s1}  => s2: {s2}")

webui.py

@@ -69,6 +69,8 @@ def segments_length_limit(segments, total_max: int):
     ret_segments = []
     total_len = 0
     for seg in segments:
+        if "text" not in seg:
+            continue
         total_len += len(seg["text"])
         if total_len > total_max:
             break