Problem about wmt17 zh-en dataset

[winterfell2021]

It seems that in subset casia2015, some samples are like {'c[hn]':'xxx', 'en': 'aa'}.
So when using data = load_dataset('wmt17', "zh-en") to load the wmt17 zh-en dataset, which will raise the exception:

Traceback (most recent call last):
  File "train.py", line 78, in <module>
    data = load_dataset(args.dataset, "zh-en")
  File "/usr/local/lib/python3.7/dist-packages/datasets/load.py", line 1684, in load_dataset
    use_auth_token=use_auth_token,
  File "/usr/local/lib/python3.7/dist-packages/datasets/builder.py", line 705, in download_and_prepare
    dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
  File "/usr/local/lib/python3.7/dist-packages/datasets/builder.py", line 1221, in _download_and_prepare
    super()._download_and_prepare(dl_manager, verify_infos, check_duplicate_keys=verify_infos)
  File "/usr/local/lib/python3.7/dist-packages/datasets/builder.py", line 793, in _download_and_prepare
    self._prepare_split(split_generator, **prepare_split_kwargs)
  File "/usr/local/lib/python3.7/dist-packages/datasets/builder.py", line 1215, in _prepare_split
    num_examples, num_bytes = writer.finalize()
  File "/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py", line 533, in finalize
    self.write_examples_on_file()
  File "/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py", line 410, in write_examples_on_file
    self.write_batch(batch_examples=batch_examples)
  File "/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py", line 503, in write_batch
    arrays.append(pa.array(typed_sequence))
  File "pyarrow/array.pxi", line 230, in pyarrow.lib.array
  File "pyarrow/array.pxi", line 110, in pyarrow.lib._handle_arrow_array_protocol
  File "/usr/local/lib/python3.7/dist-packages/datasets/arrow_writer.py", line 198, in __arrow_array__
    out = cast_array_to_feature(out, type, allow_number_to_str=not self.trying_type)
  File "/usr/local/lib/python3.7/dist-packages/datasets/table.py", line 1675, in wrapper
    return func(array, *args, **kwargs)
  File "/usr/local/lib/python3.7/dist-packages/datasets/table.py", line 1846, in cast_array_to_feature
    return array_cast(array, feature(), allow_number_to_str=allow_number_to_str)
  File "/usr/local/lib/python3.7/dist-packages/datasets/table.py", line 1675, in wrapper
    return func(array, *args, **kwargs)
  File "/usr/local/lib/python3.7/dist-packages/datasets/table.py", line 1756, in array_cast
    raise TypeError(f"Couldn't cast array of type\n{array.type}\nto\n{pa_type}")
TypeError: Couldn't cast array of type
struct<c[hn]: string, en: string, zh: string>
to
struct<en: string, zh: string>

So the solution of this problem is to change the original array manually:

if 'c[hn]' in str(array.type):
        py_array = array.to_pylist()
        data_list = []
        for vo in py_array:
            tmp = {
                'en': vo['en'],
            }
            if 'zh' not in vo:
                tmp['zh'] = vo['c[hn]']
            else:
                tmp['zh'] = vo['zh']
            data_list.append(tmp)
        array = pa.array(data_list, type=pa.struct([
            pa.field('en', pa.string()),
            pa.field('zh', pa.string()),
        ]))

Therefore, maybe a correct version of original casia2015 file need to be updated

[VictorSanh]

Running into the same error with wmt17/zh-en, wmt18/zh-en and wmt19/zh-en.

[StevenTang1998]

@albertvillanova @lhoestq Could you take a look at this issue?

[StevenTang1998]

@winterfell2021 Hi, I wonder where the code you provided should be added. I tried to add them in the datasets/table.py in array_cast function, however, the 'zh' item is none.

[StevenTang1998]

I found some 'zh' item is none while 'c[hn]' is not.
So the code may change to:

if 'c[hn]' in str(array.type):
        py_array = array.to_pylist()
        data_list = []
        for vo in py_array:
            tmp = {
                'en': vo['en'],
            }
            if vo.get('zh'):
                tmp['zh'] = vo['zh']
            else:
                tmp['zh'] = vo['c[hn]']
            data_list.append(tmp)
        array = pa.array(data_list, type=pa.struct([
            pa.field('en', pa.string()),
            pa.field('zh', pa.string()),
        ]))

[lhoestq]

I just pushed a fix, we'll do a new release of datasets soon to include this fix. In the meantime you can use the fixed dataset by passing revision="main" to load_dataset