caching in map causes same result to be returned for train, validation and test

[dpressel]

hello,

I am working on a program that uses the nlp library with the SST2 dataset.

The rough outline of the program is:

import nlp as nlp_datasets
...
parser.add_argument('--dataset', help='HuggingFace Datasets id', default=['glue', 'sst2'], nargs='+')
...
dataset = nlp_datasets.load_dataset(*args.dataset)
...
# Create feature vocabs
vocabs = create_vocabs(dataset.values(), vectorizers)
...
# Create a function to vectorize based on vectorizers and vocabs:

print('TS', train_set.num_rows)
print('VS', valid_set.num_rows)
print('ES', test_set.num_rows)

# factory method to create a `convert_to_features` function based on vocabs
convert_to_features = create_featurizer(vectorizers, vocabs)
train_set = train_set.map(convert_to_features, batched=True)
train_set.set_format(type='torch', columns=list(vectorizers.keys()) + ['y', 'lengths'])
train_loader = torch.utils.data.DataLoader(train_set, batch_size=args.batchsz)

valid_set = valid_set.map(convert_to_features, batched=True)
valid_set.set_format(type='torch', columns=list(vectorizers.keys()) + ['y', 'lengths'])
valid_loader = torch.utils.data.DataLoader(valid_set, batch_size=args.batchsz)

test_set = test_set.map(convert_to_features, batched=True)
test_set.set_format(type='torch', columns=list(vectorizers.keys()) + ['y', 'lengths'])
test_loader = torch.utils.data.DataLoader(test_set, batch_size=args.batchsz)

print('TS', train_set.num_rows)
print('VS', valid_set.num_rows)
print('ES', test_set.num_rows)

Im not sure if Im using it incorrectly, but the results are not what I expect. Namely, the .map() seems to grab the datset from the cache and then loses track of what the specific dataset is, instead using my training data for all datasets:

TS 67349
VS 872
ES 1821
TS 67349
VS 67349
ES 67349

The behavior changes if I turn off the caching but then the results fail:

train_set = train_set.map(convert_to_features, batched=True, load_from_cache_file=False)
...
valid_set = valid_set.map(convert_to_features, batched=True, load_from_cache_file=False)
...
test_set = test_set.map(convert_to_features, batched=True, load_from_cache_file=False)

Now I get the right set of features back...

TS 67349
VS 872
ES 1821
100%|██████████| 68/68 [00:00<00:00, 92.78it/s]
100%|██████████| 1/1 [00:00<00:00, 75.47it/s]
  0%|          | 0/2 [00:00<?, ?it/s]TS 67349
VS 872
ES 1821
100%|██████████| 2/2 [00:00<00:00, 77.19it/s]

but I think its losing track of the original training set:

Traceback (most recent call last):
  File "/home/dpressel/dev/work/baseline/api-examples/layers-classify-hf-datasets.py", line 148, in <module>
    for x in train_loader:
  File "/home/dpressel/anaconda3/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 345, in __next__
    data = self._next_data()
  File "/home/dpressel/anaconda3/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 385, in _next_data
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
  File "/home/dpressel/anaconda3/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/dpressel/anaconda3/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 44, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/home/dpressel/anaconda3/lib/python3.7/site-packages/nlp/arrow_dataset.py", line 338, in __getitem__
    output_all_columns=self._output_all_columns,
  File "/home/dpressel/anaconda3/lib/python3.7/site-packages/nlp/arrow_dataset.py", line 294, in _getitem
    outputs = self._unnest(self._data.slice(key, 1).to_pydict())
  File "pyarrow/table.pxi", line 1211, in pyarrow.lib.Table.slice
  File "pyarrow/public-api.pxi", line 390, in pyarrow.lib.pyarrow_wrap_table
  File "pyarrow/error.pxi", line 85, in pyarrow.lib.check_status
pyarrow.lib.ArrowInvalid: Column 3: In chunk 0: Invalid: Length spanned by list offsets (15859698) larger than values array (length 100000)

Process finished with exit code 1

The full-example program (minus the print stmts) is here:
https://github.com/dpressel/mead-baseline/pull/620/files

[patrickvonplaten]

Hi @dpressel,

thanks for posting your issue! Can you maybe add a complete code snippet that we can copy paste to reproduce the error? For example, I'm not sure where the variable train_set comes from in your code and it seems like you are loading multiple datasets at once?

[dpressel]

Hi, the full example was listed in the PR above, but here is the exact link:

https://github.com/dpressel/mead-baseline/blob/3c1aa3ca062cb23f303ca98ac40b6652b37ee971/api-examples/layers-classify-hf-datasets.py

The problem is coming from

            if cache_file_name is None:
                # we create a unique hash from the function, current dataset file and the mapping args
                cache_kwargs = {
                    "with_indices": with_indices,
                    "batched": batched,
                    "batch_size": batch_size,
                    "remove_columns": remove_columns,
                    "keep_in_memory": keep_in_memory,
                    "load_from_cache_file": load_from_cache_file,
                    "cache_file_name": cache_file_name,
                    "writer_batch_size": writer_batch_size,
                    "arrow_schema": arrow_schema,
                    "disable_nullable": disable_nullable,
                }
                cache_file_name = self._get_cache_file_path(function, cache_kwargs)

The cached value is always the same, but I was able to change that by just renaming the function each time which seems to fix the issue.

[patrickvonplaten]

Ok, I think @lhoestq has already found a solution :-) Maybe you can chime in @lhoestq

[dpressel]

This fixed my issue (I think)

mead-ml/mead-baseline@48aa8ec

[dpressel]

Ok, I think @lhoestq has already found a solution :-) Maybe you can chime in @lhoestq

Oh, awesome! I see the PR, Ill check it out

[lhoestq]

The PR should prevent the cache from losing track of the of the dataset type (based on the location of its data). Not sure about your second problem though (cache off).

[dpressel]

Yes, with caching on, it seems to work without the function renaming hack, I mentioned this also in the PR. Thanks!