Add jvnv jp models, ver 2.0.1

style_gen.py

@@ -1,5 +1,5 @@
 import argparse
-import concurrent.futures
+from concurrent.futures import ThreadPoolExecutor
 import warnings
 
 import numpy as np
@@ -20,6 +20,12 @@ device = torch.device(config.style_gen_config.device)
 inference.to(device)
 
 
+class NaNValueError(ValueError):
+    """カスタム例外クラス。NaN値が見つかった場合に使用されます。"""
+
+    pass
+
+
 # 推論時にインポートするために短いが関数を書く
 def get_style_vector(wav_path):
     return inference(wav_path)
@@ -29,10 +35,22 @@ def save_style_vector(wav_path):
     try:
         style_vec = get_style_vector(wav_path)
     except Exception as e:
-        logger.error(f"\nError occurred with file: {wav_path}, Details:\n{e}\n")
+        logger.error(f"Error occurred with file: {wav_path}, Details:\n{e}\n")
         raise
+    # 値にNaNが含まれていると悪影響なのでチェックする
+    if np.isnan(style_vec).any():
+        logger.warning(f"NaN value found in style vector: {wav_path}")
+        raise NaNValueError(f"NaN value found in style vector: {wav_path}")
     np.save(f"{wav_path}.npy", style_vec)  # `test.wav` -> `test.wav.npy`
-    return style_vec
+
+
+def process_line(line):
+    wavname = line.split("|")[0]
+    try:
+        save_style_vector(wavname)
+        return line, None
+    except NaNValueError:
+        return line, "nan_error"
 
 
 def save_average_style_vector(style_vectors, filename="style_vectors.npy"):
@@ -56,22 +74,53 @@ if __name__ == "__main__":
 
     device = config.style_gen_config.device
 
-    lines = []
+    training_lines = []
     with open(hps.data.training_files, encoding="utf-8") as f:
-        lines.extend(f.readlines())
-
-    with open(hps.data.validation_files, encoding="utf-8") as f:
-        lines.extend(f.readlines())
-
-    wavnames = [line.split("|")[0] for line in lines]
-
-    with concurrent.futures.ThreadPoolExecutor(max_workers=num_processes) as executor:
-        list(
+        training_lines.extend(f.readlines())
+    with ThreadPoolExecutor(max_workers=num_processes) as executor:
+        training_results = list(
             tqdm(
-                executor.map(save_style_vector, wavnames),
-                total=len(wavnames),
+                executor.map(process_line, training_lines),
+                total=len(training_lines),
                 file=SAFE_STDOUT,
             )
         )
+    ok_training_lines = [line for line, error in training_results if error is None]
+    nan_training_lines = [
+        line for line, error in training_results if error == "nan_error"
+    ]
+    if nan_training_lines:
+        nan_files = [line.split("|")[0] for line in nan_training_lines]
+        logger.warning(
+            f"Found NaN value in {len(nan_training_lines)} files: {nan_files}, so they will be deleted from training data."
+        )
 
-    logger.info(f"Finished generating style vectors! total: {len(wavnames)} npy files.")
+    val_lines = []
+    with open(hps.data.validation_files, encoding="utf-8") as f:
+        val_lines.extend(f.readlines())
+
+    with ThreadPoolExecutor(max_workers=num_processes) as executor:
+        val_results = list(
+            tqdm(
+                executor.map(process_line, val_lines),
+                total=len(val_lines),
+                file=SAFE_STDOUT,
+            )
+        )
+    ok_val_lines = [line for line, error in val_results if error is None]
+    nan_val_lines = [line for line, error in val_results if error == "nan_error"]
+    if nan_val_lines:
+        nan_files = [line.split("|")[0] for line in nan_val_lines]
+        logger.warning(
+            f"Found NaN value in {len(nan_val_lines)} files: {nan_files}, so they will be deleted from validation data."
+        )
+
+    with open(hps.data.training_files, "w", encoding="utf-8") as f:
+        f.writelines(ok_training_lines)
+
+    with open(hps.data.validation_files, "w", encoding="utf-8") as f:
+        f.writelines(ok_val_lines)
+
+    ok_num = len(ok_training_lines) + len(ok_val_lines)
+
+    logger.info(f"Finished generating style vectors! total: {ok_num} npy files.")