fix 数据加载问题

examples/asr/asr_ctc/speech_to_text_ctc_bpe.py

@@ -107,6 +107,7 @@ if __name__ == '__main__':
         'model.train_ds.tarred_audio_filepaths=data/common_voice_11_0/ja/train_tarred_1bk/audio__OP_0..1023_CL_.tar',
         'model.train_ds.is_tarred=true',
         '++model.train_ds.tarred_dataset_resolve_paths=false',
+        '++model.train_ds.is_tarred_audio=true',
         'model.train_ds.manifest_filepath=data/common_voice_11_0/ja/train_tarred_1bk/tarred_audio_manifest.json',
         'model.validation_ds.manifest_filepath=data/common_voice_11_0/ja/validation/validation_common_voice_11_0_manifest.json',
         'model.test_ds.manifest_filepath=data/common_voice_11_0/ja/test/test_common_voice_11_0_manifest.json',

patch_nemo.py

@@ -1,28 +1,47 @@
-"""
-修复 nemo-toolkit 中的 SIGKILL 问题
-File "C:/Users/o/.conda/envs/NeMo/Lib/site-packages/nemo/utils/exp_manager.py", line 176, in FaultToleranceParams
-    rank_termination_signal: signal.Signals = signal.SIGKILL
-                                              ^^^^^^^^^^^^^^
-AttributeError: module 'signal' has no attribute 'SIGKILL'. Did you mean: 'SIGILL'?  是不是 signal 版本不对，conda activate NeMo 再安装正确版本？
-"""
+import sys
 
-import os
-import re
-
-file_path = r"C:\Users\o\.conda\envs\NeMo\Lib\site-packages\nemo\utils\exp_manager.py"
+file_path = r"C:\Users\o\.conda\envs\NeMo\Lib\site-packages\nemo\collections\asr\data\audio_to_text.py"
 
 with open(file_path, "r", encoding="utf-8") as f:
     content = f.read()
 
-# Replace signal.SIGKILL with getattr(signal, 'SIGKILL', signal.SIGTERM)
-new_content = content.replace(
-    "rank_termination_signal: signal.Signals = signal.SIGKILL",
-    "rank_termination_signal: signal.Signals = getattr(signal, 'SIGKILL', signal.SIGTERM)"
-)
-
-with open(file_path, "w", encoding="utf-8") as f:
-    f.write(new_content)
-
-print("Patch applied successfully.")
-
-
+old_str = """            def __next__(self):
+                if self.current_fn is None:
+                    self.current_bytes, self.current_fn = next(self.iterator)
+                    self.offset_id = 0
+                else:
+                    offset_list = self.collection.mapping[self.current_fn]
+                    if len(offset_list) == self.offset_id + 1:
+                        self.current_bytes, self.current_fn = next(self.iterator)
+                        self.offset_id = 0
+                    else:
+                        self.offset_id += 1
+
+                return self.current_bytes, self.current_fn, self.offset_id"""
+
+new_str = """            def __next__(self):
+                if self.current_fn is None:
+                    self.current_bytes, self.current_fn = next(self.iterator)
+                    self.offset_id = 0
+                else:
+                    import os
+                    file_id, _ = os.path.splitext(os.path.basename(self.current_fn))
+                    offset_list = self.collection.mapping[file_id]
+                    if len(offset_list) == self.offset_id + 1:
+                        self.current_bytes, self.current_fn = next(self.iterator)
+                        self.offset_id = 0
+                    else:
+                        self.offset_id += 1
+
+                return self.current_bytes, self.current_fn, self.offset_id"""
+
+if old_str in content:
+    content = content.replace(old_str, new_str)
+    with open(file_path, "w", encoding="utf-8") as f:
+        f.write(content)
+    print("Successfully patched audio_to_text.py!")
+else:
+    if new_str in content:
+        print("Already patched!")
+    else:
+        print("Could not find the target string to patch.")

readme.txt

@@ -6,6 +6,37 @@ see https://github.com/NVIDIA-NeMo/NeMo/blob/main/tutorials/asr/ASR_Example_Comm
     从零开始训练或从已有模型微调
 
 
+
+这个 KeyError 报错是由于 WebDataset 解析 tar 包中的路径与 NeMo 的 Manifest 映射字典键值不匹配 导致的。
+
+### 问题分析
+在使用 Tarred Dataset 模式训练时：
+
+1. tar 文件中音频文件的路径带了多级目录结构，例如 huggingface_echodict/NeMo/data/.../common_voice_ja_26015575.wav 。
+2. 当 WebDataset 迭代解压时，它会将不带扩展名的完整路径（如 .../common_voice_ja_26015575 ）作为字典的 __key__ 返回。
+3. 然而，NeMo 的内部实现（ ASRManifestProcessor ）是通过获取 audio_filepath 的 os.path.basename （文件名本身） 作为映射的字典键。这导致字典里的键是 common_voice_ja_26015575 。
+4. 随后，在 nemo.collections.asr.data.audio_to_text.py 的 _loop_offsets 方法中，代码直接使用了 WebDataset 返回的长路径作为键去 self.collection.mapping 里查找，因为没做 basename 处理，从而引发了 KeyError 报错。
+### 修复过程
+为了解决这个问题，我编写了一个脚本 patch_nemo.py 来直接修补你的 Conda 环境中 NeMo 安装包下的源码文件 nemo/collections/asr/data/audio_to_text.py 。
+
+在代码中，将如下原本的错误逻辑：
+
+```
+offset_list = self.collection.mapping[self.current_fn]
+```
+替换为了带有 basename 提取的逻辑：
+
+```
+import os
+file_id, _ = os.path.splitext(os.path.basename(self.current_fn))
+offset_list = self.collection.mapping[file_id]
+```
+我已经自动执行了这个脚本，成功对你环境里的源码打上了补丁。这样 NeMo 就能正确地去掉冗长的目录路径，提取出对应的 file_id 去字典里查询索引。
+
+你现在可以再次运行训练脚本， KeyError 应该已经完全解决了。
+
+
+
 conda create -n NeMo python==3.12 pip
 conda activate NeMo
 pip install -r requirements.txt
@@ -82,3 +113,25 @@ python text_tokenizer.py
 
 
 
+TOKENIZER=data/common_voice_11_0/ja/tokenizers/tokenizer_spe_bpe_v2491
+TRAIN_MANIFEST=data/common_voice_11_0/ja/train_tarred_1bk/tarred_audio_manifest.json
+TARRED_AUDIO_FILEPATHS=data/common_voice_11_0/ja/train_tarred_1bk/audio_0..1023.tar # "_0..1023" is the range for the batch of files audio_0.tar, audio_1.tar, ..., audio_1023.tar
+DEV_MANIFEST=data/common_voice_11_0/ja/validation/validation_common_voice_11_0_manifest.json
+TEST_MANIFEST=data/common_voice_11_0/ja/test/test_common_voice_11_0_manifest.json
+
+python examples/asr/asr_ctc/speech_to_text_ctc_bpe.py \
+  --config-path=../conf/conformer/ \
+  --config-name=conformer_ctc_bpe \
+  exp_manager.name="Name of our experiment" \
+  exp_manager.resume_if_exists=true \
+  exp_manager.resume_ignore_no_checkpoint=true \
+  exp_manager.exp_dir=results/ \
+  ++model.encoder.conv_norm_type=layer_norm \
+  model.tokenizer.dir=$TOKENIZER \
+  model.train_ds.is_tarred=true \
+  model.train_ds.tarred_audio_filepaths=$TARRED_AUDIO_FILEPATHS \
+  model.train_ds.manifest_filepath=$TRAIN_MANIFEST \
+  model.validation_ds.manifest_filepath=$DEV_MANIFEST \
+  model.test_ds.manifest_filepath=$TEST_MANIFEST
+
+