ctheodoris

/

Geneformer

like 300

Hi,

The

output_dataset = Dataset.from_dict(dataset_dict)

call from within the Tokenizer causes an Int32 overflow when the dataset is too big (exciting max(int32) in size):

This is due to this function in the Huggingface code:

def numpy_to_pyarrow_listarray(arr: np.ndarray, type: pa.DataType = None) -> pa.ListArray:
    """Build a PyArrow ListArray from a multidimensional NumPy array"""
    arr = np.array(arr)
    values = pa.array(arr.flatten(), type=type)
    for i in range(arr.ndim - 1):
        n_offsets = reduce(mul, arr.shape[: arr.ndim - i - 1], 1)
        step_offsets = arr.shape[arr.ndim - i - 1]
        offsets = pa.array(np.arange(n_offsets + 1) * step_offsets, type=pa.int32())
        values = pa.ListArray.from_arrays(offsets, values)
    return values

I found that in the dictionary values are saved as np.array rather than python List. This doesn't happen:

        ## DEBUG
        print('Changing lists to np.arrays..')
        dataset_dict['input_ids'] = np.array(dataset_dict['input_ids'], dtype='object')
        dataset_dict['gene'] = np.array(dataset_dict['gene'], dtype='object')
        ## DEBUG


        # create dataset
        output_dataset = Dataset.from_dict(dataset_dict)

Has someone came across this?

Kind regards,
Eyal.