Improve: Convert ONNX version of BERT models to FP16

I have found that half-precision has little effect on speech synthesis quality and, depending on the environment, can reduce file size and memory usage by half, so I have decided to use FP16.

bert/bert_models.json

@@ -5,7 +5,7 @@
     },
     "deberta-v2-large-japanese-char-wwm-onnx": {
         "repo_id": "tsukumijima/deberta-v2-large-japanese-char-wwm-onnx",
-        "files": ["model.onnx"]
+        "files": ["model_fp16.onnx"]
     },
     "chinese-roberta-wwm-ext-large": {
         "repo_id": "hfl/chinese-roberta-wwm-ext-large",
@@ -13,7 +13,7 @@
     },
     "chinese-roberta-wwm-ext-large-onnx": {
         "repo_id": "tsukumijima/chinese-roberta-wwm-ext-large-onnx",
-        "files": ["model.onnx"]
+        "files": ["model_fp16.onnx"]
     },
     "deberta-v3-large": {
         "repo_id": "microsoft/deberta-v3-large",
@@ -21,6 +21,6 @@
     },
     "deberta-v3-large-onnx": {
         "repo_id": "tsukumijima/deberta-v3-large-onnx",
-        "files": ["spm.model", "model.onnx"]
+        "files": ["spm.model", "model_fp16.onnx"]
     }
 }

style_bert_vits2/nlp/onnx_bert_models.py

@@ -82,7 +82,7 @@ def load_model(
         model_path = Path(
             hf_hub_download(
                 repo_id=pretrained_model_name_or_path,
-                filename="model.onnx",
+                filename="model_fp16.onnx",
                 cache_dir=cache_dir,
                 revision=revision,
             )
@@ -99,7 +99,7 @@ def load_model(
     # pretrained_model_name_or_path にファイルパスが指定された場合:
     # 既にダウンロード済みという前提のもと、モデルへのローカルパスを model_path に格納する
     else:
-        model_path = Path(pretrained_model_name_or_path).resolve() / "model.onnx"
+        model_path = Path(pretrained_model_name_or_path).resolve() / "model_fp16.onnx"
 
     start_time = time.time()
     sess_options = onnxruntime.SessionOptions()