Refactor: moved the module for extracting BERT features from text in each language to style_bert_vits2/text_processing/(language)/bert_feature.py

style_bert_vits2/text_processing/japanese/bert_feature.py

@@ -0,0 +1,86 @@
+import sys
+
+import torch
+from transformers import PreTrainedModel
+
+from style_bert_vits2.constants import Languages
+from style_bert_vits2.text_processing import bert_models
+from style_bert_vits2.text_processing.japanese.g2p import text_to_sep_kata
+
+
+models: dict[str, PreTrainedModel] = {}
+
+
+def extract_bert_feature(
+    text: str,
+    word2ph: list[int],
+    device: torch.device | str,
+    assist_text: str | None = None,
+    assist_text_weight: float = 0.7,
+) -> torch.Tensor:
+    """
+    日本語のテキストから BERT の特徴量を抽出する
+
+    Args:
+        text (str): 日本語のテキスト
+        word2ph (list[int]): 元のテキストの各文字に音素が何個割り当てられるかを表すリスト
+        device (torch.device | str): 推論に利用するデバイス
+        assist_text (str | None, optional): 補助テキスト (デフォルト: None)
+        assist_text_weight (float, optional): 補助テキストの重み (デフォルト: 0.7)
+
+    Returns:
+        torch.Tensor: BERT の特徴量
+    """
+
+    # 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():
+        models[device] = bert_models.load_model(Languages.JP).to(device)  # type: ignore
+
+    style_res_mean = None
+    with torch.no_grad():
+        tokenizer = bert_models.load_tokenizer(Languages.JP)
+        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 = 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 = torch.cat(style_res["hidden_states"][-3:-2], -1)[0].cpu()
+            style_res_mean = style_res.mean(0)
+
+    assert len(word2ph) == len(text) + 2, text
+    word2phone = word2ph
+    phone_level_feature = []
+    for i in range(len(word2phone)):
+        if assist_text:
+            assert style_res_mean is not None
+            repeat_feature = (
+                res[i].repeat(word2phone[i], 1) * (1 - assist_text_weight)
+                + style_res_mean.repeat(word2phone[i], 1) * assist_text_weight
+            )
+        else:
+            repeat_feature = res[i].repeat(word2phone[i], 1)
+        phone_level_feature.append(repeat_feature)
+
+    phone_level_feature = torch.cat(phone_level_feature, dim=0)
+
+    return phone_level_feature.T