load_dataset for CSV files not working

[iliemihai]

Similar to #622, I've noticed there is a problem when trying to load a CSV file with datasets.

from datasets import load_dataset
dataset = load_dataset("csv", data_files=["./sample_data.csv"], delimiter="\t", column_names=["title", "text"], script_version="master")

Displayed error:
... ArrowInvalid: CSV parse error: Expected 2 columns, got 1

I should mention that when I've tried to read data from https://github.com/lhoestq/transformers/tree/custom-dataset-in-rag-retriever/examples/rag/test_data/my_knowledge_dataset.csv it worked without a problem. I've read that there might be some problems with /r character, so I've removed them from the custom dataset, but the problem still remains.

I've added a colab reproducing the bug, but unfortunately I cannot provide the dataset.
https://colab.research.google.com/drive/1Qzu7sC-frZVeniiWOwzoCe_UHZsrlxu8?usp=sharing

Are there any work around for it ?
Thank you

[lhoestq]

Thank you !
Could you provide a csv file that reproduces the error ?
It doesn't have to be one of your dataset. As long as it reproduces the error
That would help a lot !

[iliemihai]

I think another good example is the following:
from datasets import load_dataset
dataset = load_dataset("csv", data_files=["./sts-dev.csv"], delimiter="\t", column_names=["one", "two", "three", "four", "score", "sentence1", "sentence2"], script_version="master")
`

Displayed error CSV parse error: Expected 7 columns, got 6 even tough I put 7 columns. First four columns from the csv don't have a name, so I've named them by default. The csv file is the .dev file from STSb benchmark dataset.

[YipingNUS]

Hi, seems I also can't read csv file. I was trying with a dummy csv with only three rows.

text,label
I hate google,negative
I love Microsoft,positive
I don't like you,negative

I was using the HuggingFace image in Paperspace Gradient (datasets==1.1.3). The following code doesn't work:

from datasets import load_dataset
dataset = load_dataset('csv', script_version="master", data_files=['test_data.csv'], delimiter=",")

It outputs the following:

Using custom data configuration default
Downloading and preparing dataset csv/default-3b6254ff4dd403e5 (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /root/.cache/huggingface/datasets/csv/default-3b6254ff4dd403e5/0.0.0/2960f95a26e85d40ca41a230ac88787f715ee3003edaacb8b1f0891e9f04dda2...
Dataset csv downloaded and prepared to /root/.cache/huggingface/datasets/csv/default-3b6254ff4dd403e5/0.0.0/2960f95a26e85d40ca41a230ac88787f715ee3003edaacb8b1f0891e9f04dda2. Subsequent calls will reuse this data.

But len(dataset) gives 1 and I can't access rows with indexing dataset[0] (it gives KeyError: 0).

However, loading from pandas dataframe is working.

from datasets import Dataset
import pandas as pd
df = pd.read_csv('test_data.csv')
dataset = Dataset.from_pandas(df)

[lhoestq]

This is because load_dataset without split= returns a dictionary of split names (train/validation/test) to dataset.
You can do

from datasets import load_dataset
dataset = load_dataset('csv', script_version="master", data_files=['test_data.csv'], delimiter=",")
print(dataset["train"][0])

Or if you want to directly get the train split:

from datasets import load_dataset
dataset = load_dataset('csv', script_version="master", data_files=['test_data.csv'], delimiter=",", split="train")
print(dataset[0])

[thomwolf]

Good point

Design question for us, though: should load_dataset when no split is specified and only one split is present in the dataset (common use case with CSV/text/JSON datasets) return a Dataset instead of a DatsetDict? I feel like it's often what the user is expecting. I break a bit the paradigm of a unique return type but since this library is designed for widespread DS people more than CS people usage I would tend to think that UX should take precedence over CS reasons. What do you think?

[lhoestq]

In this case the user expects to get only one dataset object instead of the dictionary of datasets since only one csv file was specified without any split specifications.
I'm ok with returning the dataset object if no split specifications are given for text/json/csv/pandas.

For the other datasets ton the other hand the user doesn't know in advance the splits so I would keep the dictionary by default. What do you think ?

[YipingNUS]

Thanks for your quick response! I'm fine with specifying the split as @lhoestq suggested. My only concern is when I'm loading from python dict or pandas, the library returns a dataset instead of a dictionary of datasets when no split is specified. I know that they use a different function Dataset.from_dict or Dataset.from_pandas but the text/csv files use load_dataset(). However, to the user, they do the same task and we probably expect them to have the same behavior.

[z7ye]

from datasets import load_dataset
dataset = load_dataset('csv', data_files='./amazon_data/Video_Games_5.csv', delimiter=",", split=['train', 'test'])

I was running the above line, but got this error.

ValueError: Unknown split "test". Should be one of ['train'].

The data is amazon product data. I load the Video_Games_5.json.gz data into pandas and save it as csv file. and then load the csv file using the above code. I thought, split=['train', 'test'] would split the data into train and test. did I misunderstood?

Thank you!

[lhoestq]

Hi ! the split argument in load_dataset is used to select the splits you want among the available splits.
However when loading a csv with a single file as you did, only a train split is available by default.

Indeed since data_files='./amazon_data/Video_Games_5.csv' is equivalent to data_files={"train": './amazon_data/Video_Games_5.csv'}, you can get a dataset with

from datasets import load_dataset
dataset = load_dataset('csv', data_files='./amazon_data/Video_Games_5.csv', delimiter=",", split="train")

And then to get both a train and test split you can do

dataset = dataset.train_test_split()
print(dataset.keys())
# ['train', 'test']

Also note that a csv dataset may have several available splits if it is defined this way:

from datasets import load_dataset
dataset = load_dataset('csv', data_files={
    "train": './amazon_data/Video_Games_5_train.csv',
    "test": './amazon_data/Video_Games_5_test.csv'
})
print(dataset.keys())
# ['train', 'test']

[thomwolf]

In this case the user expects to get only one dataset object instead of the dictionary of datasets since only one csv file was specified without any split specifications.
I'm ok with returning the dataset object if no split specifications are given for text/json/csv/pandas.

For the other datasets ton the other hand the user doesn't know in advance the splits so I would keep the dictionary by default. What do you think ?

Yes maybe this would be good. I think having to select 'train' from the resulting object why the user gave no split information is a confusing and not intuitive behavior.

[kauvinlucas]

Similar to #622, I've noticed there is a problem when trying to load a CSV file with datasets.

from datasets import load_dataset
dataset = load_dataset("csv", data_files=["./sample_data.csv"], delimiter="\t", column_names=["title", "text"], script_version="master")

Displayed error:
... ArrowInvalid: CSV parse error: Expected 2 columns, got 1

I'm also facing the same issue when trying to load from a csv file locally:

from nlp import load_dataset
dataset = load_dataset('csv', data_files='sample_data.csv')

Error when executed from Google Colab:

ArrowInvalid                              Traceback (most recent call last)
<ipython-input-34-79a8d4f65ed6> in <module>()
      1 from nlp import load_dataset
----> 2 dataset = load_dataset('csv', data_files='sample_data.csv')

9 frames
/usr/local/lib/python3.7/dist-packages/nlp/load.py in load_dataset(path, name, version, data_dir, data_files, split, cache_dir, features, download_config, download_mode, ignore_verifications, save_infos, **config_kwargs)
    547     # Download and prepare data
    548     builder_instance.download_and_prepare(
--> 549         download_config=download_config, download_mode=download_mode, ignore_verifications=ignore_verifications,
    550     )
    551 

/usr/local/lib/python3.7/dist-packages/nlp/builder.py in download_and_prepare(self, download_config, download_mode, ignore_verifications, try_from_hf_gcs, dl_manager, **download_and_prepare_kwargs)
    461                 if not downloaded_from_gcs:
    462                     self._download_and_prepare(
--> 463                         dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
    464                     )
    465                 # Sync info

/usr/local/lib/python3.7/dist-packages/nlp/builder.py in _download_and_prepare(self, dl_manager, verify_infos, **prepare_split_kwargs)
    535             try:
    536                 # Prepare split will record examples associated to the split
--> 537                 self._prepare_split(split_generator, **prepare_split_kwargs)
    538             except OSError:
    539                 raise OSError("Cannot find data file. " + (self.manual_download_instructions or ""))

/usr/local/lib/python3.7/dist-packages/nlp/builder.py in _prepare_split(self, split_generator)
    863 
    864         generator = self._generate_tables(**split_generator.gen_kwargs)
--> 865         for key, table in utils.tqdm(generator, unit=" tables", leave=False):
    866             writer.write_table(table)
    867         num_examples, num_bytes = writer.finalize()

/usr/local/lib/python3.7/dist-packages/tqdm/notebook.py in __iter__(self, *args, **kwargs)
    213     def __iter__(self, *args, **kwargs):
    214         try:
--> 215             for obj in super(tqdm_notebook, self).__iter__(*args, **kwargs):
    216                 # return super(tqdm...) will not catch exception
    217                 yield obj

/usr/local/lib/python3.7/dist-packages/tqdm/std.py in __iter__(self)
   1102                 fp_write=getattr(self.fp, 'write', sys.stderr.write))
   1103 
-> 1104         for obj in iterable:
   1105             yield obj
   1106             # Update and possibly print the progressbar.

/usr/local/lib/python3.7/dist-packages/nlp/datasets/csv/ede98314803c971fef04bcee45d660c62f3332e8a74491e0b876106f3d99bd9b/csv.py in _generate_tables(self, files)
     78                 read_options=self.config.pa_read_options,
     79                 parse_options=self.config.pa_parse_options,
---> 80                 convert_options=self.config.convert_options,
     81             )
     82             yield i, pa_table

/usr/local/lib/python3.7/dist-packages/pyarrow/_csv.pyx in pyarrow._csv.read_csv()

/usr/local/lib/python3.7/dist-packages/pyarrow/error.pxi in pyarrow.lib.pyarrow_internal_check_status()

/usr/local/lib/python3.7/dist-packages/pyarrow/error.pxi in pyarrow.lib.check_status()

ArrowInvalid: CSV parse error: Expected 1 columns, got 8

Version:

nlp==0.4.0

[lhoestq]

Hi @kauvinlucas

You can use the latest versions of datasets to do this.
To do so, just pip install datasets instead of nlp (the library was renamed) and then

from datasets import load_dataset
dataset = load_dataset('csv', data_files='sample_data.csv')

[Valerieps]

Hi
I'm having a different problem with loading local csv.

from datasets import load_dataset  
dataset = load_dataset('csv', data_files='sample.csv')  

gives ValueError: Specified named and prefix; you can only specify one. error

versions:

• datasets: 1.1.3
• python: 3.9.6
• pyarrow: 2.0.0

[Valerieps]

Oh.. I figured it out. According to issue #42387 from pandas, this new version does not accept None for both parameters (which was being done by the repo I'm testing). Dowgrading Pandas==1.0.4 and Python==3.8 worked