Error running test code `AttributeError: type object 'TranslationDataset' has no attribute 'name'`

[Rj7]

When I try to load data a translation dataset using TranslationDataset class I get this error. There is no attribute named name in Dataset class or TranslationDataset.

I tried do as specified in one of the test cases. https://github.com/pytorch/text/blob/master/test/translation.py#L75

data_set = datasets.TranslationDataset.splits(root='../data/',train='training.10k', validation='newstest2012', test='newstest2015',exts=('.de', '.en'), fields=(DE, EN))

~/anaconda3/envs/opennmt/lib/python3.6/site-packages/torchtext/datasets/translation.py in splits(cls, exts, fields, root, train, validation, test, **kwargs)
     57                 Dataset.
     58         """
---> 59         path = cls.download(root)
     60 
     61         train_data = None if train is None else cls(

~/anaconda3/envs/opennmt/lib/python3.6/site-packages/torchtext/data/dataset.py in download(cls, root, check)
    109             dataset_path (str): Path to extracted dataset.
    110         """
--> 111         path = os.path.join(root, cls.name)
    112         check = path if check is None else check
    113         if not os.path.isdir(check):

AttributeError: type object 'TranslationDataset' has no attribute 'name'

[mttk]

@Rj7 the TranslationDataset class is not an actual dataset, and not passing the path argument (which loads the dataset) you are forcing the dataset to download itself (which it can't, because it's just a superclass for the other datasets in torchtext.datasets.translation.py).

This works in the test code because in line https://github.com/pytorch/text/blob/master/test/translation.py#L52 the Multi30k dataset is downloaded first, and then the path it was downloaded to (default: '.data/multi30k') is passed as an argument to the TranslationDataset in line https://github.com/pytorch/text/blob/master/test/translation.py#L80

To summarize:

• if you want to use the superclass TranslationDataset you need to pass the path argument to the location of an already downloaded dataset
• otherwise, use a concrete implementation Multi30k without path and the dataset will be downloaded.

Hope this helps.

[Rj7]

@mttk Awesome. It makes sense now. Thank you!