使用ViTFeatureExtractor执行with_transform后数据集结构异常及操作失败问题求助

我在使用ViT的特征提取器时，遇到了一些无法理解的奇怪问题，想请大家帮忙解惑。

加载数据集后，我查看训练集的特征结构如下：

ds['train'].features

{'image_file_path': Value(dtype='string', id=None),  'image':
Image(mode=None, decode=True, id=None),  'labels':
ClassLabel(names=['angular_leaf_spot', 'bean_rust', 'healthy'],
id=None)}

这时候不管是按列取标签，还是按行取样本，都能正常访问：

# 按列取前5个标签
ds['train']['labels'][0:5]
# 输出：[0, 0, 0, 0, 0]

# 取前2个样本
ds['train'][0:2]
# 输出：
{'image_file_path':
['/home/albert/.cache/huggingface/datasets/downloads/extracted/967f0d9f61a7a8de58892c6fab6f02317c06faf3e19fba6a07b0885a9a7142c7/train/angular_leaf_spot/angular_leaf_spot_train.0.jpg',
'/home/albert/.cache/huggingface/datasets/downloads/extracted/967f0d9f61a7a8de58892c6fab6f02317c06faf3e19fba6a07b0885a9a7142c7/train/angular_leaf_spot/angular_leaf_spot_train.1.jpg'],
'image': [<PIL.JpegImagePlugin.JpegImageFile image mode=RGB
size=500x500>,   <PIL.JpegImagePlugin.JpegImageFile image mode=RGB
size=500x500>],  'labels': [0, 0]}

但是当我用ViTFeatureExtractor定义转换函数，并用with_transform处理数据集后：

from transformers import ViTFeatureExtractor

model_name_or_path = 'google/vit-base-patch16-224-in21k'
feature_extractor = ViTFeatureExtractor.from_pretrained(model_name_or_path)
ds = load_dataset('beans')

def transform(example_batch):
    inputs = feature_extractor([x for x in example_batch['image']], return_tensors='pt')
    inputs['labels'] = example_batch['labels']
    return inputs

prepared_ds = ds.with_transform(transform)

查看prepared_ds['train'].features，显示的特征还是和原数据集一致，但访问样本时，返回的是处理后的pixel_values和标签：

prepared_ds['train'][0:2]
# 输出：
{'pixel_values': tensor([[[[-0.5686, -0.5686, -0.5608,  ..., -0.0275, 
0.1843, -0.2471],
...,
[-0.5843, -0.5922, -0.6078,  ...,  0.2627,  0.1608,  0.2000]],

         [[-0.7098, -0.7098, -0.7490,  ..., -0.3725, -0.1608, -0.6000],
          ...,
          [-0.8824, -0.9059, -0.9216,  ..., -0.2549, -0.2000, -0.1216]]],

        [[[-0.5137, -0.4902, -0.4196,  ..., -0.0275, -0.0039, -0.2157],
          ...,
          [-0.5216, -0.5373, -0.5451,  ..., -0.1294, -0.1529, -0.2627]],

         [[-0.1843, -0.2000, -0.1529,  ...,  0.2157,  0.2078, -0.0902],
          ...,
          [-0.7725, -0.7961, -0.8039,  ..., -0.3725, -0.4196, -0.5451]],

         [[-0.7569, -0.8510, -0.8353,  ..., -0.3255, -0.2706, -0.5608],
          ...,
          [-0.5294, -0.5529, -0.5608,  ..., -0.1686, -0.1922, -0.3333]]]]), 'labels': [0, 0]}

问题1：按列访问标签触发KeyError #

当我尝试直接按列访问标签时，执行prepared_ds['train']['labels']，却触发了KeyError，错误栈如下：

--------------------------------------------------------------------------- 
KeyError                                  Traceback (most recent call last) Cell In[32], line 1
----> 1 prepared_ds['train']['labels']

File ~/anaconda3/envs/LLM/lib/python3.12/site-packages/datasets/arrow_dataset.py:2872, in Dataset.__getitem__(self, key)    2870 def __getitem__(self, key): 
# noqa: F811    2871     """Can be used to index columns (by string names) or rows (by integer index or iterable of indices or bools)."""
-> 2872     return self._getitem(key)

File ~/anaconda3/envs/LLM/lib/python3.12/site-packages/datasets/arrow_dataset.py:2857, in Dataset._getitem(self, key, **kwargs)    2855 formatter = get_formatter(format_type, features=self._info.features,
**format_kwargs)    2856 pa_subtable = query_table(self._data, key, indices=self._indices)
-> 2857 formatted_output = format_table(    2858     pa_subtable, key, formatter=formatter, format_columns=format_columns, output_all_columns=output_all_columns    2859 )    2860 return formatted_output

File ~/anaconda3/envs/LLM/lib/python3.12/site-packages/datasets/formatting/formatting.py:639, in format_table(table, key, formatter, format_columns, output_all_columns)
    637 python_formatter = PythonFormatter(features=formatter.features)
    638 if format_columns is None:
--> 639     return formatter(pa_table, query_type=query_type)
    640 elif query_type == "column":
    641     if key in format_columns:

File ~/anaconda3/envs/LLM/lib/python3.12/site-packages/datasets/formatting/formatting.py:405, in Formatter.__call__(self, pa_table, query_type)
    403     return self.format_row(pa_table)
    404 elif query_type == "column":
--> 405     return self.format_column(pa_table)
    406 elif query_type == "batch":
    407     return self.format_batch(pa_table)

File ~/anaconda3/envs/LLM/lib/python3.12/site-packages/datasets/formatting/formatting.py:501, in CustomFormatter.format_column(self, pa_table)
    500 def format_column(self, pa_table: pa.Table) -> ColumnFormat:
--> 501     formatted_batch = self.format_batch(pa_table)
    502     if hasattr(formatted_batch, "keys"):
    503         if len(formatted_batch.keys()) > 1:

File ~/anaconda3/envs/LLM/lib/python3.12/site-packages/datasets/formatting/formatting.py:522, in CustomFormatter.format_batch(self, pa_table)
    520 batch = self.python_arrow_extractor().extract_batch(pa_table)
    521 batch = self.python_features_decoder.decode_batch(batch)
--> 522 return self.transform(batch)

Cell In[12], line 5, in transform(example_batch)
      3 def transform(example_batch):
      4     # Take a list of PIL images and turn them to pixel values
--> 5     inputs = feature_extractor([x for x in example_batch['image']], return_tensors='pt')
      7     # Don't forget to include the labels!
      8     inputs['labels'] = example_batch['labels']

KeyError: 'image'

看起来错误是因为当我按列访问时，转换函数被触发了，但此时的example_batch里没有image字段，这让我很困惑——为什么只是访问标签，会重新执行转换函数？

问题2：保存处理后的数据集触发TypeError #

另外，当我尝试把处理后的数据集保存到磁盘时，执行prepared_ds.save_to_disk(img_path)，又遇到了TypeError：

--------------------------------------------------------------------------- 
TypeError                                 Traceback (most recent call last) Cell In[21], line 1
----> 1 dataset.save_to_disk(img_path)

File ~/anaconda3/envs/LLM/lib/python3.13/site-packages/datasets/arrow_dataset.py:1503, in Dataset.save_to_disk(self, dataset_path, max_shard_size, num_shards, num_proc, storage_options)    1501         json.dumps(state["_format_kwargs"][k])    1502     except TypeError as e:
-> 1503         raise TypeError(    1504             str(e) + f"\nThe format kwargs must be JSON serializable, but key '{k}' isn't."    1505 ) from None    1506 # Get json serializable dataset info    1507 dataset_info = asdict(self._info)

TypeError: Object of type function is not JSON serializable The format kwargs must be JSON serializable, but key 'transform' isn't.

我的疑问 #

需要说明的是，原示例里的训练、评估等流程都是正常工作的，我只是在尝试探索数据集结构、保存处理后的数据集时才遇到这些问题。

我想请教大家：

• 为什么执行with_transform()或set_transform()后，数据集的访问方式不能和原来保持一致？
• 为什么只是尝试访问某个特征，会触发转换函数的执行？
• 有没有办法在应用转换后，还能正常访问原有的特征，或者顺利保存数据集？

希望有人能帮我理清这个行为背后的原因，谢谢！

备注：内容来源于stack exchange，提问作者hamagust