Fix

config.py

@@ -260,20 +260,9 @@ with open(os.path.join("configs", "paths.yml"), "r", encoding="utf-8") as f:
     # - dataset_root: the root directory of the dataset, default to "Data"
     # - assets_root: the root directory of the assets, default to "model_assets"
 
-parser = argparse.ArgumentParser()
-# 为避免与以前的config.json起冲突，将其更名如下
-parser.add_argument(
-    "-y",
-    "--yml_config",
-    type=str,
-    default="config.yml",
-    help="Path to setting yaml file.",
-)
-
-args = parser.parse_args()
 
 try:
-    config = Config(args.yml_config, path_config)
+    config = Config("config.yml", path_config)
 except TypeError:
     logger.warning("Old config.yml found. Replace it with default_config.yml.")
     shutil.copy(src="default_config.yml", dst="config.yml")

speech_mos.py

@@ -3,6 +3,8 @@ import csv
 import warnings
 from pathlib import Path
 
+import matplotlib.pyplot as plt
+import pandas as pd
 import torch
 from tqdm import tqdm
 
@@ -71,9 +73,6 @@ for model_file in tqdm(safetensors_files):
         score = predictor(torch.from_numpy(audio).unsqueeze(0), sr).item()
         scores.append(score)
         logger.info(f"score: {score}")
-    mean = sum(scores) / len(scores)
-    logger.success(f"mean: {mean}")
-    scores.append(mean)
     # `test_e10_s1000.safetensors`` -> 1000を取り出す
     step = int(model_file.stem.split("_")[2][1:])
     results.append((model_file.name, step, scores))
@@ -90,3 +89,36 @@ with open(args.output, "w", encoding="utf-8", newline="") as f:
     writer.writerow(["model_path"] + ["step"] + test_texts + ["mean"])
     for model_file, step, scores in results:
         writer.writerow([model_file] + [step] + scores)
+
+# step countと各MOSの値を格納するリストを初期化
+steps = []
+mos_values = []
+
+# resultsからデータを抽出
+for _, step, scores in results:
+    steps.append(step)
+    mos_values.append(scores)  # scores は MOS1, MOS2, MOS3,... のリスト
+
+# DataFrame形式に変換
+df = pd.DataFrame(mos_values, index=steps)
+# ステップ数でソート
+df = df.sort_index()
+
+# 各MOSと平均値についての折れ線グラフを描画
+for col in df.columns:
+    plt.plot(df.index, df[col], label=f"MOS{col + 1}")
+
+# 平均値の計算と描画
+df["mean"] = df.mean(axis=1)
+plt.plot(df.index, df["mean"], label="Mean", color="black", linewidth=2)
+
+# グラフのタイトルと軸ラベルを設定
+plt.title("TTS Model Naturalness MOS")
+plt.xlabel("Step Count")
+plt.ylabel("MOS")
+
+# 凡例を表示
+plt.legend()
+
+# グラフを表示
+plt.show()

transcribe.py

@@ -70,6 +70,7 @@ if __name__ == "__main__":
         language = Languages.ZH
     else:
         raise ValueError(f"{language} is not supported.")
+    logger.debug(f"Initial prompt: {initial_prompt}")
     with open(output_file, "w", encoding="utf-8") as f:
         for wav_file in tqdm(wav_files, file=SAFE_STDOUT):
             file_name = os.path.basename(wav_file)

webui_dataset.py

@@ -56,7 +56,7 @@ def do_transcribe(model_name, whisper_model, compute_type, language, initial_pro
             "--language",
             language,
             "--initial_prompt",
-            '"initial_prompt"',
+            f'"{initial_prompt}"',
         ]
     )
     return "音声の文字起こしが完了しました。"

webui_train.py

@@ -302,9 +302,9 @@ if __name__ == "__main__":
                     label="エポック数",
                     info="100もあれば十分そうだけどもっと回すと質が上がるかもしれない",
                     value=100,
-                    minimum=1,
+                    minimum=10,
                     maximum=1000,
-                    step=1,
+                    step=10,
                 )
                 save_every_steps = gr.Slider(
                     label="何ステップごとに結果を保存するか",