生成词表

readme.txt

@@ -8,7 +8,7 @@ see https://github.com/NVIDIA-NeMo/NeMo/blob/main/tutorials/asr/ASR_Example_Comm
 
 conda create -n NeMo python==3.12 pip
 conda activate NeMo
-pip install datasets
+pip install -r requirements.txt
 
 python scripts/speech_recognition/convert_hf_dataset_to_nemo.py \
     output_dir=${OUTPUT_DIR} \
@@ -60,4 +60,19 @@ python scripts/speech_recognition/convert_to_tarred_audio_dataset.py \
   python patch_nemo.py
 
 
-  
+
+VOCAB_SIZE=128
+python scripts/tokenizers/process_asr_text_tokenizer.py \
+  --manifest=data/common_voice_11_0/ja/train_tarred_1bk/tarred_audio_manifest.json \
+  --vocab_size=${VOCAB_SIZE} \
+  --data_root=data/common_voice_11_0/ja/tokenizers \
+  --tokenizer="spe" \
+  --spe_type=bpe
+
+
+
+生成词表
+python text_tokenizer.py
+
+
+

text_tokenizer.py

@@ -0,0 +1,65 @@
+import os
+import sys
+import json
+
+# 将 scripts/tokenizers 添加到 sys.path，以便导入 process_asr_text_tokenizer
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "scripts", "tokenizers")))
+
+# 由于 process_asr_text_tokenizer 内部使用了双下划线的函数，并且在 import 的时候定义了 parser，
+# 为了避免冲突，我们直接调用其内部使用的函数，或者使用 os.system 来调用脚本
+# 考虑到动态计算 VOCAB_SIZE，我们首先读取 manifest，收集所有的字符来计算 VOCAB_SIZE
+
+def calculate_vocab_size(manifest_path):
+    unique_chars = set()
+    with open(manifest_path, 'r', encoding='utf-8') as f:
+        for line in f:
+            item = json.loads(line)
+            text = item.get('text', '')
+            for char in text:
+                unique_chars.add(char)
+    
+    # 词汇表大小 = 独立字符数 + 预留特殊 token (如 <s>, </s>, <pad>, <unk> 等) 的数量
+    # 通常预留 10-20 个位置给特殊 token 和未见过的字符，这里我们加 20
+    vocab_size = len(unique_chars) + 20
+    # 为了后续的计算，通常建议 vocab_size 为 2 的幂或特定倍数，这里直接返回计算结果
+    return vocab_size
+
+def main():
+    manifest_path = "data/common_voice_11_0/ja/train_tarred_1bk/tarred_audio_manifest.json"
+    data_root = "data/common_voice_11_0/ja/tokenizers"
+    
+    print("Calculating dynamic VOCAB_SIZE...")
+    vocab_size = calculate_vocab_size(manifest_path)
+    print(f"Calculated VOCAB_SIZE: {vocab_size}")
+
+    # 调用 process_asr_text_tokenizer
+    # 由于 process_asr_text_tokenizer 在导入时会解析命令行参数，所以我们采用 subprocess 的方式调用
+    # 这样可以避免 argparse 报错，同时为了让脚本找到 nemo，我们将当前目录加入 PYTHONPATH
+    import subprocess
+    env = os.environ.copy()
+    env["PYTHONPATH"] = os.path.abspath(os.path.dirname(__file__)) + os.pathsep + env.get("PYTHONPATH", "")
+
+    command = [
+        "python",
+        "scripts/tokenizers/process_asr_text_tokenizer.py",
+        f"--manifest={manifest_path}",
+        f"--vocab_size={vocab_size}",
+        f"--data_root={data_root}",
+        "--tokenizer=spe",
+        "--spe_type=bpe"
+    ]
+    
+    # 尝试在 conda 的 NeMo 环境中执行
+    command = ["conda", "run", "-n", "NeMo"] + command
+    
+    print(f"Running command: \n{' '.join(command)}")
+    result = subprocess.run(command, env=env)
+    exit_code = result.returncode
+    if exit_code != 0:
+        print(f"Error executing tokenizer script, exit code: {exit_code}")
+        sys.exit(1)
+    else:
+        print("Tokenizer processing completed successfully.")
+
+if __name__ == "__main__":
+    main()