Dataset Viewer not available on features of type datasets.Array2D(shape=(None, 768), dtype='float64')

🤗Datasets
1

The data we upload to huggingface contains some features that have a flexible time axis. For example, for each sample it can be of shape (T, 768), T varies. According to the official doc, huggingface does support this type of 2D array by leaving the first dim as None, as I put here:

features = datasets.Features(
    {
        "video_id": datasets.Value("string"),
        "label": datasets.Value("string"),
        "visual_feature": datasets.Array2D(shape=(None, 768), dtype='float64'),                         # (T, 768), T can vary
    },
)

But when I submit it to huggingface, I get this error on the dataset viewer panel:

Error code:   StreamingRowsError
Exception:    ValueError
Message:      shape=(768,) and ndims=2 don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 99, in get_rows_or_raise
                  return get_rows(
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 77, in get_rows
                  rows_plus_one = list(itertools.islice(ds, rows_max_number + 1))
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 2270, in __iter__
                  for key, example in ex_iterable:
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 1856, in __iter__
                  for key, pa_table in self._iter_arrow():
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 1879, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 476, in _iter_arrow
                  for key, pa_table in iterator:
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 323, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/parquet/parquet.py", line 106, in _generate_tables
                  yield f"{file_idx}_{batch_idx}", self._cast_table(pa_table)
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/parquet/parquet.py", line 73, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/features/features.py", line 1826, in arrow_schema
                  return pa.schema(self.type).with_metadata({"huggingface": json.dumps(hf_metadata)})
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/features/features.py", line 1815, in type
                  return get_nested_type(self)
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/features/features.py", line 1253, in get_nested_type
                  {key: get_nested_type(schema[key]) for key in schema}
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/features/features.py", line 1253, in <dictcomp>
                  {key: get_nested_type(schema[key]) for key in schema}
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/features/features.py", line 1277, in get_nested_type
                  return schema()
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/features/features.py", line 551, in __call__
                  pa_type = globals()[self.__class__.__name__ + "ExtensionType"](self.shape, self.dtype)
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/features/features.py", line 665, in __init__
                  raise ValueError(f"shape={shape} and ndims={self.ndims} don't match")
              ValueError: shape=(768,) and ndims=2 don't match

Is it an expected behavior that Array2D with first dim being None cannot display and the whole dataset viewer panel stops functioning? Or I am doing something wrong with my dataset uploading script?

1 Like
2

The Dataset library anticipates cases where the dimensions of data change dynamically, but the DatasetViewer library does not yet anticipate this. This is probably a ToDo rather than a bug. @lhoestq

The array type also allows the first dimension of the array to be dynamic. This is useful for handling sequences with variable lengths such as sentences, without having to pad or truncate the input to a uniform shape.
features = Features({‘a’: Array3D(shape=(None, 5, 2), dtype=‘int32’)})

3

Hi thanks for confirming that! Also want to know will this array2d of flexible first dim prevent huggingface from automatically generating croissant meta? I’ve made several attempts on this, and notice that if the dataset contains one such array2d then after a short delay even the viewer still shows the error I can get croissant meta, but when the dataset contains multiple, say 3, each of dim (None, 768) (None, 1024) (None, 1024), then I get no croissant and no auto generated parquet branch

1 Like
4

If the data can be converted to parquet format, I think metadata will probably be created eventually, but maybe there is an error in the converter…:thinking: