Refactor: don't keep models in bert_feature module for each language

BERT models and tokenizers are already stored and managed in the bert_models module and should not be stored here.
In addition, since there may be situations where the user would like to use cpu instead of mps for inference when using it as a library, the automatic switching process to mps was removed.

style_bert_vits2/nlp/chinese/bert_feature.py

@@ -1,16 +1,11 @@
-import sys
 from typing import Optional
 
 import torch
-from transformers import PreTrainedModel
 
 from style_bert_vits2.constants import Languages
 from style_bert_vits2.nlp import bert_models
 
 
-__models: dict[str, PreTrainedModel] = {}
-
-
 def extract_bert_feature(
     text: str,
     word2ph: list[int],
@@ -32,18 +27,9 @@ def extract_bert_feature(
         torch.Tensor: BERT の特徴量
     """
 
-    if (
-        sys.platform == "darwin"
-        and torch.backends.mps.is_available()
-        and device == "cpu"
-    ):
-        device = "mps"
-    if not device:
-        device = "cuda"
     if device == "cuda" and not torch.cuda.is_available():
         device = "cpu"
-    if device not in __models.keys():
+    model = bert_models.load_model(Languages.ZH).to(device)  # type: ignore
-        __models[device] = bert_models.load_model(Languages.ZH).to(device)  # type: ignore
 
     style_res_mean = None
     with torch.no_grad():
@@ -51,13 +37,13 @@ def extract_bert_feature(
         inputs = tokenizer(text, return_tensors="pt")
         for i in inputs:
             inputs[i] = inputs[i].to(device)  # type: ignore
-        res = __models[device](**inputs, output_hidden_states=True)
+        res = model(**inputs, output_hidden_states=True)
         res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
         if assist_text:
             style_inputs = tokenizer(assist_text, return_tensors="pt")
             for i in style_inputs:
                 style_inputs[i] = style_inputs[i].to(device)  # type: ignore
-            style_res = __models[device](**style_inputs, output_hidden_states=True)
+            style_res = model(**style_inputs, output_hidden_states=True)
             style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
             style_res_mean = style_res.mean(0)
 

style_bert_vits2/nlp/english/bert_feature.py

@@ -1,16 +1,11 @@
-import sys
 from typing import Optional
 
 import torch
-from transformers import PreTrainedModel
 
 from style_bert_vits2.constants import Languages
 from style_bert_vits2.nlp import bert_models
 
 
-__models: dict[str, PreTrainedModel] = {}
-
-
 def extract_bert_feature(
     text: str,
     word2ph: list[int],
@@ -32,18 +27,9 @@ def extract_bert_feature(
         torch.Tensor: BERT の特徴量
     """
 
-    if (
-        sys.platform == "darwin"
-        and torch.backends.mps.is_available()
-        and device == "cpu"
-    ):
-        device = "mps"
-    if not device:
-        device = "cuda"
     if device == "cuda" and not torch.cuda.is_available():
         device = "cpu"
-    if device not in __models.keys():
+    model = bert_models.load_model(Languages.EN).to(device)  # type: ignore
-        __models[device] = bert_models.load_model(Languages.EN).to(device)  # type: ignore
 
     style_res_mean = None
     with torch.no_grad():
@@ -51,13 +37,13 @@ def extract_bert_feature(
         inputs = tokenizer(text, return_tensors="pt")
         for i in inputs:
             inputs[i] = inputs[i].to(device)  # type: ignore
-        res = __models[device](**inputs, output_hidden_states=True)
+        res = model(**inputs, output_hidden_states=True)
         res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
         if assist_text:
             style_inputs = tokenizer(assist_text, return_tensors="pt")
             for i in style_inputs:
                 style_inputs[i] = style_inputs[i].to(device)  # type: ignore
-            style_res = __models[device](**style_inputs, output_hidden_states=True)
+            style_res = model(**style_inputs, output_hidden_states=True)
             style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
             style_res_mean = style_res.mean(0)
 

style_bert_vits2/nlp/japanese/bert_feature.py

@@ -1,17 +1,12 @@
-import sys
 from typing import Optional
 
 import torch
-from transformers import PreTrainedModel
 
 from style_bert_vits2.constants import Languages
 from style_bert_vits2.nlp import bert_models
 from style_bert_vits2.nlp.japanese.g2p import text_to_sep_kata
 
 
-__models: dict[str, PreTrainedModel] = {}
-
-
 def extract_bert_feature(
     text: str,
     word2ph: list[int],
@@ -36,21 +31,12 @@ def extract_bert_feature(
     # 各単語が何文字かを作る `word2ph` を使う必要があるので、読めない文字は必ず無視する
     # でないと `word2ph` の結果とテキストの文字数結果が整合性が取れない
     text = "".join(text_to_sep_kata(text, raise_yomi_error=False)[0])
-
     if assist_text:
         assist_text = "".join(text_to_sep_kata(assist_text, raise_yomi_error=False)[0])
-    if (
+
-        sys.platform == "darwin"
-        and torch.backends.mps.is_available()
-        and device == "cpu"
-    ):
-        device = "mps"
-    if not device:
-        device = "cuda"
     if device == "cuda" and not torch.cuda.is_available():
         device = "cpu"
-    if device not in __models.keys():
+    model = bert_models.load_model(Languages.JP).to(device)  # type: ignore
-        __models[device] = bert_models.load_model(Languages.JP).to(device)  # type: ignore
 
     style_res_mean = None
     with torch.no_grad():
@@ -58,13 +44,13 @@ def extract_bert_feature(
         inputs = tokenizer(text, return_tensors="pt")
         for i in inputs:
             inputs[i] = inputs[i].to(device)  # type: ignore
-        res = __models[device](**inputs, output_hidden_states=True)
+        res = model(**inputs, output_hidden_states=True)
         res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()
         if assist_text:
             style_inputs = tokenizer(assist_text, return_tensors="pt")
             for i in style_inputs:
                 style_inputs[i] = style_inputs[i].to(device)  # type: ignore
-            style_res = __models[device](**style_inputs, output_hidden_states=True)
+            style_res = model(**style_inputs, output_hidden_states=True)
             style_res = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
             style_res_mean = style_res.mean(0)