html_url
stringlengths 48
51
| title
stringlengths 5
268
| comments
stringlengths 70
51.8k
| body
stringlengths 0
29.8k
| comment_length
int64 16
1.52k
| text
stringlengths 164
54.1k
|
|---|---|---|---|---|---|
https://github.com/huggingface/datasets/issues/160 | caching in map causes same result to be returned for train, validation and test | Yes, with caching on, it seems to work without the function renaming hack, I mentioned this also in the PR. Thanks! | 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
| 21 | caching in map causes same result to be returned for train, validation and test
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
Yes, with caching on, it seems to work without the function renaming hack, I mentioned this also in the PR. Thanks! |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | You can just run:
`val = nlp.load_dataset('squad')`
if you want to have just the validation script you can also do:
`val = nlp.load_dataset('squad', split="validation")` | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 24 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
You can just run:
`val = nlp.load_dataset('squad')`
if you want to have just the validation script you can also do:
`val = nlp.load_dataset('squad', split="validation")` |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | If you want to load a local dataset, make sure you include a `./` before the folder name. | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 18 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
If you want to load a local dataset, make sure you include a `./` before the folder name. |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | This happens by just doing run all cells on colab ... I assumed the colab example is broken. | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 18 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
This happens by just doing run all cells on colab ... I assumed the colab example is broken. |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | Oh I see you might have a wrong version of pyarrow install on the colab -> could you try the following. Add these lines to the beginning of your notebook, restart the runtime and run it again:
```
!pip uninstall -y -qq pyarrow
!pip uninstall -y -qq nlp
!pip install -qq git+https://github.com/huggingface/nlp.git
``` | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 53 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
Oh I see you might have a wrong version of pyarrow install on the colab -> could you try the following. Add these lines to the beginning of your notebook, restart the runtime and run it again:
```
!pip uninstall -y -qq pyarrow
!pip uninstall -y -qq nlp
!pip install -qq git+https://github.com/huggingface/nlp.git
``` |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | > Oh I see you might have a wrong version of pyarrow install on the colab -> could you try the following. Add these lines to the beginning of your notebook, restart the runtime and run it again:
>
> ```
> !pip uninstall -y -qq pyarrow
> !pip uninstall -y -qq nlp
> !pip install -qq git+https://github.com/huggingface/nlp.git
> ```
Tried, having the same error. | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 65 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
> Oh I see you might have a wrong version of pyarrow install on the colab -> could you try the following. Add these lines to the beginning of your notebook, restart the runtime and run it again:
>
> ```
> !pip uninstall -y -qq pyarrow
> !pip uninstall -y -qq nlp
> !pip install -qq git+https://github.com/huggingface/nlp.git
> ```
Tried, having the same error. |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | Can you post a link here of your colab? I'll make a copy of it and see what's wrong | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 19 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
Can you post a link here of your colab? I'll make a copy of it and see what's wrong |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | This should be fixed in the current version of the notebook. You can try it again | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 16 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
This should be fixed in the current version of the notebook. You can try it again |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | I am getting this error when running this command
```
val = nlp.load_dataset('squad', split="validation")
```
FileNotFoundError: [Errno 2] No such file or directory: '/root/.cache/huggingface/datasets/squad/plain_text/1.0.0/dataset_info.json'
Can anybody help? | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 27 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
I am getting this error when running this command
```
val = nlp.load_dataset('squad', split="validation")
```
FileNotFoundError: [Errno 2] No such file or directory: '/root/.cache/huggingface/datasets/squad/plain_text/1.0.0/dataset_info.json'
Can anybody help? |
https://github.com/huggingface/datasets/issues/157 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)" | It seems like your download was corrupted :-/ Can you run the following command:
```
rm -r /root/.cache/huggingface/datasets
```
to delete the cache completely and rerun the download? | I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a | 28 | nlp.load_dataset() gives "TypeError: list_() takes exactly one argument (2 given)"
I'm trying to load datasets from nlp but there seems to have error saying
"TypeError: list_() takes exactly one argument (2 given)"
gist can be found here
https://gist.github.com/saahiluppal/c4b878f330b10b9ab9762bc0776c0a6a
It seems like your download was corrupted :-/ Can you run the following command:
```
rm -r /root/.cache/huggingface/datasets
```
to delete the cache completely and rerun the download? |
https://github.com/huggingface/datasets/issues/156 | SyntaxError with WMT datasets | Jeez - don't know what happened there :D Should be fixed now!
Thanks a lot for reporting this @tomhosking ! | The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook | 20 | SyntaxError with WMT datasets
The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook
Jeez - don't know what happened there :D Should be fixed now!
Thanks a lot for reporting this @tomhosking ! |
https://github.com/huggingface/datasets/issues/156 | SyntaxError with WMT datasets | Hi @patrickvonplaten!
I'm now getting the below error:
```
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-28-3206959998b9> in <module>
1 import nlp
2
----> 3 dataset = nlp.load_dataset('wmt14')
4 print(dataset['train'][0])
~/.local/lib/python3.6/site-packages/nlp/load.py in load_dataset(path, name, version, data_dir, data_files, split, cache_dir, download_config, download_mode, ignore_verifications, save_infos, **config_kwargs)
507 # Instantiate the dataset builder
508 builder_instance = builder_cls(
--> 509 cache_dir=cache_dir, name=name, version=version, data_dir=data_dir, data_files=data_files, **config_kwargs,
510 )
511
TypeError: Can't instantiate abstract class Wmt with abstract methods _subsets
```
| The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook | 76 | SyntaxError with WMT datasets
The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook
Hi @patrickvonplaten!
I'm now getting the below error:
```
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-28-3206959998b9> in <module>
1 import nlp
2
----> 3 dataset = nlp.load_dataset('wmt14')
4 print(dataset['train'][0])
~/.local/lib/python3.6/site-packages/nlp/load.py in load_dataset(path, name, version, data_dir, data_files, split, cache_dir, download_config, download_mode, ignore_verifications, save_infos, **config_kwargs)
507 # Instantiate the dataset builder
508 builder_instance = builder_cls(
--> 509 cache_dir=cache_dir, name=name, version=version, data_dir=data_dir, data_files=data_files, **config_kwargs,
510 )
511
TypeError: Can't instantiate abstract class Wmt with abstract methods _subsets
```
|
https://github.com/huggingface/datasets/issues/156 | SyntaxError with WMT datasets | To correct this error I think you need the master branch of `nlp`. Can you try to install `nlp` with. `WMT` was not included at the beta release of the library.
Can you try:
`pip install git+https://github.com/huggingface/nlp.git`
and check again? | The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook | 40 | SyntaxError with WMT datasets
The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook
To correct this error I think you need the master branch of `nlp`. Can you try to install `nlp` with. `WMT` was not included at the beta release of the library.
Can you try:
`pip install git+https://github.com/huggingface/nlp.git`
and check again? |
https://github.com/huggingface/datasets/issues/156 | SyntaxError with WMT datasets | That works, thanks :)
The WMT datasets are listed in by `list_datasets()` in the beta release on pypi - it would be good to only show datasets that are actually supported by that version? | The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook | 34 | SyntaxError with WMT datasets
The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook
That works, thanks :)
The WMT datasets are listed in by `list_datasets()` in the beta release on pypi - it would be good to only show datasets that are actually supported by that version? |
https://github.com/huggingface/datasets/issues/156 | SyntaxError with WMT datasets | Usually, the idea is that a dataset can be added without releasing a new version. The problem in the case of `WMT` was that some "core" code of the library had to be changed as well.
@thomwolf @lhoestq @julien-c - How should we go about this. If we add a dataset that also requires "core" code changes, how do we handle the versioning? The moment a dataset is on AWS it will actually be listed with `list_datasets()` in all earlier versions...
Is there a way to somehow insert the `pip version` to the HfApi() and get only the datasets that were available for this version (at the date of the release of the version) @julien-c ? | The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook | 116 | SyntaxError with WMT datasets
The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook
Usually, the idea is that a dataset can be added without releasing a new version. The problem in the case of `WMT` was that some "core" code of the library had to be changed as well.
@thomwolf @lhoestq @julien-c - How should we go about this. If we add a dataset that also requires "core" code changes, how do we handle the versioning? The moment a dataset is on AWS it will actually be listed with `list_datasets()` in all earlier versions...
Is there a way to somehow insert the `pip version` to the HfApi() and get only the datasets that were available for this version (at the date of the release of the version) @julien-c ? |
https://github.com/huggingface/datasets/issues/156 | SyntaxError with WMT datasets | We plan to have something like a `requirements.txt` per dataset to prevent user from loading dataset with old version of `nlp` or any other libraries. Right now the solution is just to keep `nlp` up to date when you want to load a dataset that leverages the latests features of `nlp`.
For datasets that are on AWS but that use features that are not released yet we should be able to filter those from the `list_dataset` as soon as we have the `requirements.txt` feature on (filter datasets that need a future version of `nlp`).
Shall we rename this issue to be more explicit about the problem ?
Something like `Specify the minimum version of the nlp library required for each dataset` ? | The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook | 122 | SyntaxError with WMT datasets
The following snippet produces a syntax error:
```
import nlp
dataset = nlp.load_dataset('wmt14')
print(dataset['train'][0])
```
```
Traceback (most recent call last):
File "/home/tom/.local/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line 3326, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "<ipython-input-8-3206959998b9>", line 3, in <module>
dataset = nlp.load_dataset('wmt14')
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 505, in load_dataset
builder_cls = import_main_class(module_path, dataset=True)
File "/home/tom/.local/lib/python3.6/site-packages/nlp/load.py", line 56, in import_main_class
module = importlib.import_module(module_path)
File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 994, in _gcd_import
File "<frozen importlib._bootstrap>", line 971, in _find_and_load
File "<frozen importlib._bootstrap>", line 955, in _find_and_load_unlocked
File "<frozen importlib._bootstrap>", line 665, in _load_unlocked
File "<frozen importlib._bootstrap_external>", line 678, in exec_module
File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt14.py", line 21, in <module>
from .wmt_utils import Wmt, WmtConfig
File "/home/tom/.local/lib/python3.6/site-packages/nlp/datasets/wmt14/c258d646f4f5870b0245f783b7aa0af85c7117e06aacf1e0340bd81935094de2/wmt_utils.py", line 659
<<<<<<< HEAD
^
SyntaxError: invalid syntax
```
Python version:
`3.6.9 (default, Apr 18 2020, 01:56:04) [GCC 8.4.0]`
Running on Ubuntu 18.04, via a Jupyter notebook
We plan to have something like a `requirements.txt` per dataset to prevent user from loading dataset with old version of `nlp` or any other libraries. Right now the solution is just to keep `nlp` up to date when you want to load a dataset that leverages the latests features of `nlp`.
For datasets that are on AWS but that use features that are not released yet we should be able to filter those from the `list_dataset` as soon as we have the `requirements.txt` feature on (filter datasets that need a future version of `nlp`).
Shall we rename this issue to be more explicit about the problem ?
Something like `Specify the minimum version of the nlp library required for each dataset` ? |
https://github.com/huggingface/datasets/issues/153 | Meta-datasets (GLUE/XTREME/...) – Special care to attributions and citations | As @yoavgo suggested, there should be the possibility to call a function like nlp.bib that outputs all bibtex ref from the datasets and models actually used and eventually nlp.bib.forreadme that would output the same info + versions numbers so they can be included in a readme.md file. | Meta-datasets are interesting in terms of standardized benchmarks but they also have specific behaviors, in particular in terms of attribution and authorship. It's very important that each specific dataset inside a meta dataset is properly referenced and the citation/specific homepage/etc are very visible and accessible and not only the generic citation of the meta-dataset itself.
Let's take GLUE as an example:
The configuration has the citation for each dataset included (e.g. [here](https://github.com/huggingface/nlp/blob/master/datasets/glue/glue.py#L154-L161)) but it should be copied inside the dataset info so that, when people access `dataset.info.citation` they get both the citation for GLUE and the citation for the specific datasets inside GLUE that they have loaded. | 47 | Meta-datasets (GLUE/XTREME/...) – Special care to attributions and citations
Meta-datasets are interesting in terms of standardized benchmarks but they also have specific behaviors, in particular in terms of attribution and authorship. It's very important that each specific dataset inside a meta dataset is properly referenced and the citation/specific homepage/etc are very visible and accessible and not only the generic citation of the meta-dataset itself.
Let's take GLUE as an example:
The configuration has the citation for each dataset included (e.g. [here](https://github.com/huggingface/nlp/blob/master/datasets/glue/glue.py#L154-L161)) but it should be copied inside the dataset info so that, when people access `dataset.info.citation` they get both the citation for GLUE and the citation for the specific datasets inside GLUE that they have loaded.
As @yoavgo suggested, there should be the possibility to call a function like nlp.bib that outputs all bibtex ref from the datasets and models actually used and eventually nlp.bib.forreadme that would output the same info + versions numbers so they can be included in a readme.md file. |
https://github.com/huggingface/datasets/issues/153 | Meta-datasets (GLUE/XTREME/...) – Special care to attributions and citations | Actually, double checking with @mariamabarham, we already have this feature I think.
It's like this currently:
```python
>>> from nlp import load_dataset
>>>
>>> dataset = load_dataset('glue', 'cola', split='train')
>>> print(dataset.info.citation)
@article{warstadt2018neural,
title={Neural Network Acceptability Judgments},
author={Warstadt, Alex and Singh, Amanpreet and Bowman, Samuel R},
journal={arXiv preprint arXiv:1805.12471},
year={2018}
}
@inproceedings{wang2019glue,
title={{GLUE}: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding},
author={Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R.},
note={In the Proceedings of ICLR.},
year={2019}
}
Note that each GLUE dataset has its own citation. Please see the source to see
the correct citation for each contained dataset.
```
What do you think @dseddah? | Meta-datasets are interesting in terms of standardized benchmarks but they also have specific behaviors, in particular in terms of attribution and authorship. It's very important that each specific dataset inside a meta dataset is properly referenced and the citation/specific homepage/etc are very visible and accessible and not only the generic citation of the meta-dataset itself.
Let's take GLUE as an example:
The configuration has the citation for each dataset included (e.g. [here](https://github.com/huggingface/nlp/blob/master/datasets/glue/glue.py#L154-L161)) but it should be copied inside the dataset info so that, when people access `dataset.info.citation` they get both the citation for GLUE and the citation for the specific datasets inside GLUE that they have loaded. | 115 | Meta-datasets (GLUE/XTREME/...) – Special care to attributions and citations
Meta-datasets are interesting in terms of standardized benchmarks but they also have specific behaviors, in particular in terms of attribution and authorship. It's very important that each specific dataset inside a meta dataset is properly referenced and the citation/specific homepage/etc are very visible and accessible and not only the generic citation of the meta-dataset itself.
Let's take GLUE as an example:
The configuration has the citation for each dataset included (e.g. [here](https://github.com/huggingface/nlp/blob/master/datasets/glue/glue.py#L154-L161)) but it should be copied inside the dataset info so that, when people access `dataset.info.citation` they get both the citation for GLUE and the citation for the specific datasets inside GLUE that they have loaded.
Actually, double checking with @mariamabarham, we already have this feature I think.
It's like this currently:
```python
>>> from nlp import load_dataset
>>>
>>> dataset = load_dataset('glue', 'cola', split='train')
>>> print(dataset.info.citation)
@article{warstadt2018neural,
title={Neural Network Acceptability Judgments},
author={Warstadt, Alex and Singh, Amanpreet and Bowman, Samuel R},
journal={arXiv preprint arXiv:1805.12471},
year={2018}
}
@inproceedings{wang2019glue,
title={{GLUE}: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding},
author={Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R.},
note={In the Proceedings of ICLR.},
year={2019}
}
Note that each GLUE dataset has its own citation. Please see the source to see
the correct citation for each contained dataset.
```
What do you think @dseddah? |
https://github.com/huggingface/datasets/issues/153 | Meta-datasets (GLUE/XTREME/...) – Special care to attributions and citations | Looks good but why would there be a difference between the ref in the source and the one to be printed? | Meta-datasets are interesting in terms of standardized benchmarks but they also have specific behaviors, in particular in terms of attribution and authorship. It's very important that each specific dataset inside a meta dataset is properly referenced and the citation/specific homepage/etc are very visible and accessible and not only the generic citation of the meta-dataset itself.
Let's take GLUE as an example:
The configuration has the citation for each dataset included (e.g. [here](https://github.com/huggingface/nlp/blob/master/datasets/glue/glue.py#L154-L161)) but it should be copied inside the dataset info so that, when people access `dataset.info.citation` they get both the citation for GLUE and the citation for the specific datasets inside GLUE that they have loaded. | 21 | Meta-datasets (GLUE/XTREME/...) – Special care to attributions and citations
Meta-datasets are interesting in terms of standardized benchmarks but they also have specific behaviors, in particular in terms of attribution and authorship. It's very important that each specific dataset inside a meta dataset is properly referenced and the citation/specific homepage/etc are very visible and accessible and not only the generic citation of the meta-dataset itself.
Let's take GLUE as an example:
The configuration has the citation for each dataset included (e.g. [here](https://github.com/huggingface/nlp/blob/master/datasets/glue/glue.py#L154-L161)) but it should be copied inside the dataset info so that, when people access `dataset.info.citation` they get both the citation for GLUE and the citation for the specific datasets inside GLUE that they have loaded.
Looks good but why would there be a difference between the ref in the source and the one to be printed? |
https://github.com/huggingface/datasets/issues/153 | Meta-datasets (GLUE/XTREME/...) – Special care to attributions and citations | Yes, I think we should remove this warning @mariamabarham.
It's probably a relic of tfds which didn't have the same way to access citations. | Meta-datasets are interesting in terms of standardized benchmarks but they also have specific behaviors, in particular in terms of attribution and authorship. It's very important that each specific dataset inside a meta dataset is properly referenced and the citation/specific homepage/etc are very visible and accessible and not only the generic citation of the meta-dataset itself.
Let's take GLUE as an example:
The configuration has the citation for each dataset included (e.g. [here](https://github.com/huggingface/nlp/blob/master/datasets/glue/glue.py#L154-L161)) but it should be copied inside the dataset info so that, when people access `dataset.info.citation` they get both the citation for GLUE and the citation for the specific datasets inside GLUE that they have loaded. | 24 | Meta-datasets (GLUE/XTREME/...) – Special care to attributions and citations
Meta-datasets are interesting in terms of standardized benchmarks but they also have specific behaviors, in particular in terms of attribution and authorship. It's very important that each specific dataset inside a meta dataset is properly referenced and the citation/specific homepage/etc are very visible and accessible and not only the generic citation of the meta-dataset itself.
Let's take GLUE as an example:
The configuration has the citation for each dataset included (e.g. [here](https://github.com/huggingface/nlp/blob/master/datasets/glue/glue.py#L154-L161)) but it should be copied inside the dataset info so that, when people access `dataset.info.citation` they get both the citation for GLUE and the citation for the specific datasets inside GLUE that they have loaded.
Yes, I think we should remove this warning @mariamabarham.
It's probably a relic of tfds which didn't have the same way to access citations. |
https://github.com/huggingface/datasets/issues/149 | [Feature request] Add Ubuntu Dialogue Corpus dataset | @AlphaMycelium the Ubuntu Dialogue Corpus [version 2]( https://github.com/rkadlec/ubuntu-ranking-dataset-creator) is added. Note that it requires a manual download by following the download instructions in the [repos]( https://github.com/rkadlec/ubuntu-ranking-dataset-creator).
Maybe we can close this issue for now? | https://github.com/rkadlec/ubuntu-ranking-dataset-creator or http://dataset.cs.mcgill.ca/ubuntu-corpus-1.0/ | 34 | [Feature request] Add Ubuntu Dialogue Corpus dataset
https://github.com/rkadlec/ubuntu-ranking-dataset-creator or http://dataset.cs.mcgill.ca/ubuntu-corpus-1.0/
@AlphaMycelium the Ubuntu Dialogue Corpus [version 2]( https://github.com/rkadlec/ubuntu-ranking-dataset-creator) is added. Note that it requires a manual download by following the download instructions in the [repos]( https://github.com/rkadlec/ubuntu-ranking-dataset-creator).
Maybe we can close this issue for now? |
https://github.com/huggingface/datasets/issues/143 | ArrowTypeError in squad metrics | There was an issue in the format, thanks.
Now you can do
```python3
squad_dset = nlp.load_dataset("squad")
squad_metric = nlp.load_metric("/Users/quentinlhoest/Desktop/hf/nlp-bis/metrics/squad")
predictions = [
{"id": v["id"], "prediction_text": v["answers"]["text"][0]} # take first possible answer
for v in squad_dset["validation"]
]
squad_metric.compute(predictions, squad_dset["validation"])
```
and the expected format is
```
Args:
predictions: List of question-answers dictionaries with the following key-values:
- 'id': id of the question-answer pair as given in the references (see below)
- 'prediction_text': the text of the answer
references: List of question-answers dictionaries with the following key-values:
- 'id': id of the question-answer pair (see above),
- 'answers': a Dict {'text': list of possible texts for the answer, as a list of strings}
``` | `squad_metric.compute` is giving following error
```
ArrowTypeError: Could not convert [{'text': 'Denver Broncos'}, {'text': 'Denver Broncos'}, {'text': 'Denver Broncos'}] with type list: was not a dict, tuple, or recognized null value for conversion to struct type
```
This is how my predictions and references look like
```
predictions[0]
# {'id': '56be4db0acb8001400a502ec', 'prediction_text': 'Denver Broncos'}
```
```
references[0]
# {'answers': [{'text': 'Denver Broncos'},
{'text': 'Denver Broncos'},
{'text': 'Denver Broncos'}],
'id': '56be4db0acb8001400a502ec'}
```
These are structured as per the `squad_metric.compute` help string. | 112 | ArrowTypeError in squad metrics
`squad_metric.compute` is giving following error
```
ArrowTypeError: Could not convert [{'text': 'Denver Broncos'}, {'text': 'Denver Broncos'}, {'text': 'Denver Broncos'}] with type list: was not a dict, tuple, or recognized null value for conversion to struct type
```
This is how my predictions and references look like
```
predictions[0]
# {'id': '56be4db0acb8001400a502ec', 'prediction_text': 'Denver Broncos'}
```
```
references[0]
# {'answers': [{'text': 'Denver Broncos'},
{'text': 'Denver Broncos'},
{'text': 'Denver Broncos'}],
'id': '56be4db0acb8001400a502ec'}
```
These are structured as per the `squad_metric.compute` help string.
There was an issue in the format, thanks.
Now you can do
```python3
squad_dset = nlp.load_dataset("squad")
squad_metric = nlp.load_metric("/Users/quentinlhoest/Desktop/hf/nlp-bis/metrics/squad")
predictions = [
{"id": v["id"], "prediction_text": v["answers"]["text"][0]} # take first possible answer
for v in squad_dset["validation"]
]
squad_metric.compute(predictions, squad_dset["validation"])
```
and the expected format is
```
Args:
predictions: List of question-answers dictionaries with the following key-values:
- 'id': id of the question-answer pair as given in the references (see below)
- 'prediction_text': the text of the answer
references: List of question-answers dictionaries with the following key-values:
- 'id': id of the question-answer pair (see above),
- 'answers': a Dict {'text': list of possible texts for the answer, as a list of strings}
``` |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | I would suggest `nlds`. NLP is a very general, broad and ambiguous term, the library is not about NLP (as in processing) per se, it is about accessing Natural Language related datasets. So the name should reflect its purpose.
| Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 39 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
I would suggest `nlds`. NLP is a very general, broad and ambiguous term, the library is not about NLP (as in processing) per se, it is about accessing Natural Language related datasets. So the name should reflect its purpose.
|
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | Chiming in to second everything @honnibal said, and to add that I think the current name is going to impact the discoverability of this library. People who are looking for "NLP Datasets" through a search engine are going to see a library called `nlp` and think it's too broad. People who are looking to do NLP in python are going to search "Python NLP" and end up here, confused that this is a collection of datasets.
The names of the other huggingface libraries work because they're the only game in town: there are not very many robust, distinct libraries for `tokenizers` or `transformers` in python, for example. But there are several options for NLP in python, and adding this as a possible search result for "python nlp" when datasets are likely not what someone is searching for adds noise and frustrates potential users. | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 143 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
Chiming in to second everything @honnibal said, and to add that I think the current name is going to impact the discoverability of this library. People who are looking for "NLP Datasets" through a search engine are going to see a library called `nlp` and think it's too broad. People who are looking to do NLP in python are going to search "Python NLP" and end up here, confused that this is a collection of datasets.
The names of the other huggingface libraries work because they're the only game in town: there are not very many robust, distinct libraries for `tokenizers` or `transformers` in python, for example. But there are several options for NLP in python, and adding this as a possible search result for "python nlp" when datasets are likely not what someone is searching for adds noise and frustrates potential users. |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | I'm also not sure whether the naming of `nlp` is the problem itself, as long as it comes with the appropriate identifier, so maybe something like `huggingface_nlp`? This is analogous to what @honnibal and spacy are doing for `spacy-transformers`. Of course, this is a "step back" from the recent changes/renaming of transformers, but may be some middle ground between a complete rebranding, and keeping it identifiable. | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 66 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
I'm also not sure whether the naming of `nlp` is the problem itself, as long as it comes with the appropriate identifier, so maybe something like `huggingface_nlp`? This is analogous to what @honnibal and spacy are doing for `spacy-transformers`. Of course, this is a "step back" from the recent changes/renaming of transformers, but may be some middle ground between a complete rebranding, and keeping it identifiable. |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | Interesting, thanks for sharing your thoughts.
As we’ll move toward a first non-beta release, we will pool the community of contributors/users of the library for their opinions on a good final name (like when we renamed the beautifully (?) named `pytorch-pretrained-bert`)
In the meantime, using `from nlp import load_dataset, load_metric` should work 😉 | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 53 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
Interesting, thanks for sharing your thoughts.
As we’ll move toward a first non-beta release, we will pool the community of contributors/users of the library for their opinions on a good final name (like when we renamed the beautifully (?) named `pytorch-pretrained-bert`)
In the meantime, using `from nlp import load_dataset, load_metric` should work 😉 |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | I feel like we are conflating two distinct subjects here:
1. @honnibal's point is that using `nlp` as a package name might break existing code and bring developer usability issues in the future
2. @pmbaumgartner's point is that the `nlp` package name is too broad and shouldn't be used by a package that exposes only datasets and metrics
(let me know if I mischaracterize your point)
I'll chime in to say that the first point is a bit silly IMO. As Python developers due to the limitations of the import system we already have to share:
- a single flat namespace for packages
- which also conflicts with local modules i.e. local files
If we add the constraint that this flat namespace also be shared with variable names this gets untractable pretty fast :)
I also think all Python software developers/ML engineers/scientists are capable of at least a subset of:
- importing only the methods that they need like @thomwolf suggested
- aliasing their import
- renaming a local variable | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 170 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
I feel like we are conflating two distinct subjects here:
1. @honnibal's point is that using `nlp` as a package name might break existing code and bring developer usability issues in the future
2. @pmbaumgartner's point is that the `nlp` package name is too broad and shouldn't be used by a package that exposes only datasets and metrics
(let me know if I mischaracterize your point)
I'll chime in to say that the first point is a bit silly IMO. As Python developers due to the limitations of the import system we already have to share:
- a single flat namespace for packages
- which also conflicts with local modules i.e. local files
If we add the constraint that this flat namespace also be shared with variable names this gets untractable pretty fast :)
I also think all Python software developers/ML engineers/scientists are capable of at least a subset of:
- importing only the methods that they need like @thomwolf suggested
- aliasing their import
- renaming a local variable |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | By the way, `nlp` will very likely not be only about datasets, and not even just about datasets and metrics.
I see it as a laboratory for testing several long-term ideas about how we could do NLP in terms of research as well as open-source and community sharing, most of these ideas being too experimental/big to fit in `transformers`.
Some of the directions we would like to explore are about sharing, traceability and more experimental models, as well as seeing a model as the community-based process of creating a composite entity from data, optimization, and code.
We'll see how these ideas end up being implemented and we'll better know how we should define the library when we start to dive into these topics. I'll try to get the `nlp` team to draft a roadmap on these topics at some point. | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 140 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
By the way, `nlp` will very likely not be only about datasets, and not even just about datasets and metrics.
I see it as a laboratory for testing several long-term ideas about how we could do NLP in terms of research as well as open-source and community sharing, most of these ideas being too experimental/big to fit in `transformers`.
Some of the directions we would like to explore are about sharing, traceability and more experimental models, as well as seeing a model as the community-based process of creating a composite entity from data, optimization, and code.
We'll see how these ideas end up being implemented and we'll better know how we should define the library when we start to dive into these topics. I'll try to get the `nlp` team to draft a roadmap on these topics at some point. |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | > If we add the constraint that this flat namespace also be shared with variable names this gets untractable pretty fast :)
I'm sort of confused by your point here. The namespace *is* shared by variable names. You should not use local variables that are named the same as modules, because then you cannot use the module within the scope of your function.
For instance,
```python
import nlp
import transformers
nlp = transformers.pipeline("sentiment-analysis")
```
This is a bug: you've just overwritten the module, so now you can't use it. Or instead:
```python
import transformers
nlp = transformers.pipeline("sentiment-analysis")
# (Later, e.g. in a notebook)
import nlp
```
This is also a bug: you've overwritten your variable with an import.
If you have a module named `nlp`, you should avoid using `nlp` as a variable, or you'll have bugs in some contexts and inconsistencies in other contexts. You'll have situations where you need to import differently in one module vs another, or name variables differently in one context vs another, which is bad.
> importing only the methods that they need like @thomwolf suggested
Okay but the same logic applies to naming the module *literally anything else*. There's absolutely no point in having a module name that's 3 letters if you always plan to do `import from`! It would be entirely better to name it `nlp_datasets` if you don't want people to do `import nlp`.
And finally:
> By the way, nlp will very likely not be only about datasets, and not even just about datasets and metrics.
So...it isn't a datasets library? https://twitter.com/Thom_Wolf/status/1261282491622731781
I'm confused 😕 | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 265 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
> If we add the constraint that this flat namespace also be shared with variable names this gets untractable pretty fast :)
I'm sort of confused by your point here. The namespace *is* shared by variable names. You should not use local variables that are named the same as modules, because then you cannot use the module within the scope of your function.
For instance,
```python
import nlp
import transformers
nlp = transformers.pipeline("sentiment-analysis")
```
This is a bug: you've just overwritten the module, so now you can't use it. Or instead:
```python
import transformers
nlp = transformers.pipeline("sentiment-analysis")
# (Later, e.g. in a notebook)
import nlp
```
This is also a bug: you've overwritten your variable with an import.
If you have a module named `nlp`, you should avoid using `nlp` as a variable, or you'll have bugs in some contexts and inconsistencies in other contexts. You'll have situations where you need to import differently in one module vs another, or name variables differently in one context vs another, which is bad.
> importing only the methods that they need like @thomwolf suggested
Okay but the same logic applies to naming the module *literally anything else*. There's absolutely no point in having a module name that's 3 letters if you always plan to do `import from`! It would be entirely better to name it `nlp_datasets` if you don't want people to do `import nlp`.
And finally:
> By the way, nlp will very likely not be only about datasets, and not even just about datasets and metrics.
So...it isn't a datasets library? https://twitter.com/Thom_Wolf/status/1261282491622731781
I'm confused 😕 |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | Dropping by as I noticed that the library has been renamed `datasets` so I wonder if the conversation above is settled (`nlp` not used anymore) :) | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 26 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
Dropping by as I noticed that the library has been renamed `datasets` so I wonder if the conversation above is settled (`nlp` not used anymore) :) |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | I'd argue that `datasets` is worse than `nlp`. Datasets should be a user specific decision and not encapsulate all of python (`pip install datasets`). If this package contained every dataset in the world (NLP / vision / etc) then it would make sense =/ | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 44 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
I'd argue that `datasets` is worse than `nlp`. Datasets should be a user specific decision and not encapsulate all of python (`pip install datasets`). If this package contained every dataset in the world (NLP / vision / etc) then it would make sense =/ |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | I can't speak for the HF team @jramapuram, but as member of the community it looks to me that HF wanted to avoid the past path of changing names as scope broadened over time:
Remember
https://github.com/huggingface/pytorch-openai-transformer-lm
https://github.com/huggingface/pytorch-pretrained-BERT
https://github.com/huggingface/pytorch-transformers
and now
https://github.com/huggingface/transformers
;)
Jokes aside, seems that the library is growing in a multi-modal direction (https://github.com/huggingface/datasets/pull/363) so the current name is not that implausible. Possibly HF ambition is really to grow its community and bring here a large chunk of datasets of the world (including tabular / vision / audio?). | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 89 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
I can't speak for the HF team @jramapuram, but as member of the community it looks to me that HF wanted to avoid the past path of changing names as scope broadened over time:
Remember
https://github.com/huggingface/pytorch-openai-transformer-lm
https://github.com/huggingface/pytorch-pretrained-BERT
https://github.com/huggingface/pytorch-transformers
and now
https://github.com/huggingface/transformers
;)
Jokes aside, seems that the library is growing in a multi-modal direction (https://github.com/huggingface/datasets/pull/363) so the current name is not that implausible. Possibly HF ambition is really to grow its community and bring here a large chunk of datasets of the world (including tabular / vision / audio?). |
https://github.com/huggingface/datasets/issues/138 | Consider renaming to nld | Yea I see your point. However, wouldn't scoping solve the entire problem?
```python
import huggingface.datasets as D
import huggingface.transformers as T
```
Calling something `datasets` is akin to saying I'm going to name my package `python` --> `import python` | Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p. | 39 | Consider renaming to nld
Hey :)
Just making a thread here recording what I said on Twitter, as it's impossible to follow discussion there. It's also just really not a good way to talk about this sort of thing.
The issue is that modules go into the global namespace, so you shouldn't use variable names that conflict with module names. This means the package makes `nlp` a bad variable name everywhere in the codebase. I've always used `nlp` as the canonical variable name of spaCy's `Language` objects, and this is a convention that a lot of other code has followed (Stanza, flair, etc). And actually, your `transformers` library uses `nlp` as the name for its `Pipeline` instance in your readme.
If you stick with the `nlp` name for this package, if anyone uses it then they should rewrite all of that code. If `nlp` is a bad choice of variable anywhere, it's a bad choice of variable everywhere --- because you shouldn't have to notice whether some other function uses a module when you're naming variables within a function. You want to have one convention that you can stick to everywhere.
If people use your `nlp` package and continue to use the `nlp` variable name, they'll find themselves with confusing bugs. There will be many many bits of code cut-and-paste from tutorials that give confusing results when combined with the data loading from the `nlp` library. The problem will be especially bad for shadowed modules (people might reasonably have a module named `nlp.py` within their codebase) and notebooks, as people might run notebook cells for data loading out-of-order.
I don't think it's an exaggeration to say that if your library becomes popular, we'll all be answering issues around this about once a week for the next few years. That seems pretty unideal, so I do hope you'll reconsider.
I suggest `nld` as a better name. It more accurately represents what the package actually does. It's pretty unideal to have a package named `nlp` that doesn't do any processing, and contains data about natural language generation or other non-NLP tasks. The name is equally short, and is sort of a visual pun on `nlp`, since a d is a rotated p.
Yea I see your point. However, wouldn't scoping solve the entire problem?
```python
import huggingface.datasets as D
import huggingface.transformers as T
```
Calling something `datasets` is akin to saying I'm going to name my package `python` --> `import python` |
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | I second this request. The bottom line is that **scores produced with different reference tokenizations are not comparable**. To discourage (even inadvertent) cheating, the user should never touch the reference. The `v13a` tokenization standard is not ideal, but at least it has been consistently used at matrix.statmt.org, facilitating comparisons.
Sacrebleu exposes [all its data sources](https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/dataset.py) and additionally provides [an API](https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/__init__.py) to accessing the references, which seem to fit within the spirit of your codebase. | https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 74 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
I second this request. The bottom line is that **scores produced with different reference tokenizations are not comparable**. To discourage (even inadvertent) cheating, the user should never touch the reference. The `v13a` tokenization standard is not ideal, but at least it has been consistently used at matrix.statmt.org, facilitating comparisons.
Sacrebleu exposes [all its data sources](https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/dataset.py) and additionally provides [an API](https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/__init__.py) to accessing the references, which seem to fit within the spirit of your codebase. |
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | Didn't we have a slide and discussion at WMT admitting that, for production-quality models, BLEU doesn't correlate with human eval anyway?
| https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 21 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
Didn't we have a slide and discussion at WMT admitting that, for production-quality models, BLEU doesn't correlate with human eval anyway?
|
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | Yes, there are slides like that at WMT every year :) BLEU correlates with human judgment only at coarse levels, and it seems to be getting worse when people try to use it to do model selection among high-performing neural systems.
However, the point isn't whether BLEU is a good metric, but whether your BLEU score can be compared to other BLEU scores. They only can be compared if you use the same reference tokenization (similar to how you [can't compare LM perplexities across different segmentations](https://sjmielke.com/comparing-perplexities.htm)). sacrebleu was an attempt to get everyone to use WMT's reference tokenization (meaning, your system has to first remove its own tokenization) so that you could just compare across papers. This also prevents scores from being gamed. | https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 123 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
Yes, there are slides like that at WMT every year :) BLEU correlates with human judgment only at coarse levels, and it seems to be getting worse when people try to use it to do model selection among high-performing neural systems.
However, the point isn't whether BLEU is a good metric, but whether your BLEU score can be compared to other BLEU scores. They only can be compared if you use the same reference tokenization (similar to how you [can't compare LM perplexities across different segmentations](https://sjmielke.com/comparing-perplexities.htm)). sacrebleu was an attempt to get everyone to use WMT's reference tokenization (meaning, your system has to first remove its own tokenization) so that you could just compare across papers. This also prevents scores from being gamed. |
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | I do not consider as a sufficient solution switching this library's default metric from BLEU to the wrapper around SacreBLEU.
As currently implemented, the wrapper allows end users to toggle SacreBLEU options, but doesn't pass along the SacreBLEU signature. As @mjpost showed in [Post18](https://www.aclweb.org/anthology/W18-6319.pdf), it's simply not credible to assume that people will stick to the defaults, therefore, the signature is necessary to be explicit about what options were used.
In addition to the `v13a` or `intl` options for the SacreBLEU `tokenize` argument, which was pointed out earlier, papers frequently differ on whether they lowercase text before scoring (`lowercase`) and the smoothing method used (`smooth_method`). BLEU scores can differ substantially (over 1 BLEU) just by changing these options.
Losing the SacreBLEU signature is a regression in reproducibility and clarity.
(Perhaps this should belong in a separate issue?) | https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 137 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
I do not consider as a sufficient solution switching this library's default metric from BLEU to the wrapper around SacreBLEU.
As currently implemented, the wrapper allows end users to toggle SacreBLEU options, but doesn't pass along the SacreBLEU signature. As @mjpost showed in [Post18](https://www.aclweb.org/anthology/W18-6319.pdf), it's simply not credible to assume that people will stick to the defaults, therefore, the signature is necessary to be explicit about what options were used.
In addition to the `v13a` or `intl` options for the SacreBLEU `tokenize` argument, which was pointed out earlier, papers frequently differ on whether they lowercase text before scoring (`lowercase`) and the smoothing method used (`smooth_method`). BLEU scores can differ substantially (over 1 BLEU) just by changing these options.
Losing the SacreBLEU signature is a regression in reproducibility and clarity.
(Perhaps this should belong in a separate issue?) |
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | Thanks for sharing your thoughts. This is a very important discussion.
Also one of the first items on our mid-term roadmap (we will try to clean it and share it soon) is to introduce mechanisms to get high-quality traceability and reproducibility for all the processes related to the library.
So having the signature for `sacrebleu` is really important!
Regarding BLEU, I guess we can just remove it from the canonical metrics included in the repo itself (it won't prevent people to add it as "user-metrics" but at least we won't be promoting it).
On a more general note (definitely too large for the scope of this issue) we are wondering, with @srush in particular, how we could handle the selection of metrics/datasets with the most community-based and bottom-up approach possible. If you have opinions on this, please share! | https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 138 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
Thanks for sharing your thoughts. This is a very important discussion.
Also one of the first items on our mid-term roadmap (we will try to clean it and share it soon) is to introduce mechanisms to get high-quality traceability and reproducibility for all the processes related to the library.
So having the signature for `sacrebleu` is really important!
Regarding BLEU, I guess we can just remove it from the canonical metrics included in the repo itself (it won't prevent people to add it as "user-metrics" but at least we won't be promoting it).
On a more general note (definitely too large for the scope of this issue) we are wondering, with @srush in particular, how we could handle the selection of metrics/datasets with the most community-based and bottom-up approach possible. If you have opinions on this, please share! |
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | Yeah, I would love to have discussions about ways this project can have an community-based, transparent process to arrive at strong default metrics. @kpu / @mjpost do you have any suggestions of how that might work or pointers to places where this is done right? Perhaps this question can be template for what is likely to be repeated for other datasets. | https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 61 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
Yeah, I would love to have discussions about ways this project can have an community-based, transparent process to arrive at strong default metrics. @kpu / @mjpost do you have any suggestions of how that might work or pointers to places where this is done right? Perhaps this question can be template for what is likely to be repeated for other datasets. |
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | I think @bittlingmayer is referring to Figure 6 in http://statmt.org/wmt19/pdf/53/WMT02.pdf . When you look at Appendix A there are some cases where metrics fall apart at the high end and some where they correlate well. en-zh is arguably production-quality.
This could evolve into a metrics Bazaar where the value add is really the packaging and consistency: it installs/compiles the metrics for me, gives a reproducible name to use in publication (involve the authors; you don't want a different sacrebleu hash system), a version number, and evaluation of the metrics like http://ufallab.ms.mff.cuni.cz/~bojar/wmt19-metrics-task-package.tgz but run when code changes rather than once a year. | https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 101 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
I think @bittlingmayer is referring to Figure 6 in http://statmt.org/wmt19/pdf/53/WMT02.pdf . When you look at Appendix A there are some cases where metrics fall apart at the high end and some where they correlate well. en-zh is arguably production-quality.
This could evolve into a metrics Bazaar where the value add is really the packaging and consistency: it installs/compiles the metrics for me, gives a reproducible name to use in publication (involve the authors; you don't want a different sacrebleu hash system), a version number, and evaluation of the metrics like http://ufallab.ms.mff.cuni.cz/~bojar/wmt19-metrics-task-package.tgz but run when code changes rather than once a year. |
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | While a Bazaar setup works for models / datasets, I am not sure it is ideal for metrics ? Ideal from my perspective would be to have tasks with metrics moderated by experts who document, cite, and codify known pitchfalls (as above^) and make it non-trivial for beginners to mess it up. | https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 52 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
While a Bazaar setup works for models / datasets, I am not sure it is ideal for metrics ? Ideal from my perspective would be to have tasks with metrics moderated by experts who document, cite, and codify known pitchfalls (as above^) and make it non-trivial for beginners to mess it up. |
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | @srush @thomwolf
ModelFront could provide (automated, "QE-based") evaluation for all the pretrained translation models you host. Not bottom-up and not valid for claiming SoTA, but independent, practical for builders and not top-down.
For that I would also suggest some diverse benchmarks (so split it out into datasets with only user-generated data, or only constants, or only UI strings, or only READMEs) which tease out known trade-offs. Even hypothetical magic eval is limited if we always reduce it to a single number.
Realistically people want to know how a model compares to an API like Google Translate, Microsoft Translator, DeepL or Yandex (especially for a language pair like EN:RU, or for the many languages that only Yandex supports), and that could be done too.
| https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 123 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
@srush @thomwolf
ModelFront could provide (automated, "QE-based") evaluation for all the pretrained translation models you host. Not bottom-up and not valid for claiming SoTA, but independent, practical for builders and not top-down.
For that I would also suggest some diverse benchmarks (so split it out into datasets with only user-generated data, or only constants, or only UI strings, or only READMEs) which tease out known trade-offs. Even hypothetical magic eval is limited if we always reduce it to a single number.
Realistically people want to know how a model compares to an API like Google Translate, Microsoft Translator, DeepL or Yandex (especially for a language pair like EN:RU, or for the many languages that only Yandex supports), and that could be done too.
|
https://github.com/huggingface/datasets/issues/137 | Tokenized BLEU considered harmful - Discussion on community-based process | Very important discussion.
I am trying to understand the effects of tokenization.
I wanted to ask which is a good practice.
Sacrebleu should be used on top of the tokenized output, or detokenized(raw) text? | https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too. | 34 | Tokenized BLEU considered harmful - Discussion on community-based process
https://github.com/huggingface/nlp/blob/7d1526dfeeb29248d832f1073192dbf03ad642da/metrics/bleu/bleu.py#L76 assumes the inputs are tokenized by the user. This is bad practice because the user's tokenizer is usually not the same as the one used by `mteval-v13a.pl`, the closest thing we have to a standard. Moreover, tokenizers are like window managers: they can be endlessly customized and nobody has quite the same options.
As @mjpost reported in https://www.aclweb.org/anthology/W18-6319.pdf BLEU configurations can vary by 1.8. Yet people are incorrectly putting non-comparable BLEU scores in the same table, such as Table 1 in https://arxiv.org/abs/2004.04902 .
There are a few use cases for tokenized BLEU like Thai. For Chinese, people seem to use character BLEU for better or worse.
The default easy option should be the one that's correct more often. And that is sacrebleu. Please don't make it easy for people to run what is usually the wrong option; it definitely shouldn't be `bleu`.
Also, I know this is inherited from TensorFlow and, paging @lmthang, they should discourage it too.
Very important discussion.
I am trying to understand the effects of tokenization.
I wanted to ask which is a good practice.
Sacrebleu should be used on top of the tokenized output, or detokenized(raw) text? |
https://github.com/huggingface/datasets/issues/133 | [Question] Using/adding a local dataset | Hi @zphang,
So you can just give the local path to a dataset script file and it should work.
Here is an example:
- you can download one of the scripts in the `datasets` folder of the present repo (or clone the repo)
- then you can load it with `load_dataset('PATH/TO/YOUR/LOCAL/SCRIPT.py')`
Does it make sense? | Users may want to either create/modify a local copy of a dataset, or use a custom-built dataset with the same `Dataset` API as externally downloaded datasets.
It appears to be possible to point to a local dataset path rather than downloading the external ones, but I'm not exactly sure how to go about doing this.
A notebook/example script demonstrating this would be very helpful. | 55 | [Question] Using/adding a local dataset
Users may want to either create/modify a local copy of a dataset, or use a custom-built dataset with the same `Dataset` API as externally downloaded datasets.
It appears to be possible to point to a local dataset path rather than downloading the external ones, but I'm not exactly sure how to go about doing this.
A notebook/example script demonstrating this would be very helpful.
Hi @zphang,
So you can just give the local path to a dataset script file and it should work.
Here is an example:
- you can download one of the scripts in the `datasets` folder of the present repo (or clone the repo)
- then you can load it with `load_dataset('PATH/TO/YOUR/LOCAL/SCRIPT.py')`
Does it make sense? |
https://github.com/huggingface/datasets/issues/133 | [Question] Using/adding a local dataset | Could you give a more concrete example, please?
I looked up wikitext dataset script from the repo. Should I just overwrite the `data_file` on line 98 to point to the local dataset directory? Would it work for different configurations of wikitext (wikitext2, wikitext103 etc.)?
Or maybe we can use DownloadManager to specify local dataset location? In that case, where do we use DownloadManager instance?
Thanks | Users may want to either create/modify a local copy of a dataset, or use a custom-built dataset with the same `Dataset` API as externally downloaded datasets.
It appears to be possible to point to a local dataset path rather than downloading the external ones, but I'm not exactly sure how to go about doing this.
A notebook/example script demonstrating this would be very helpful. | 65 | [Question] Using/adding a local dataset
Users may want to either create/modify a local copy of a dataset, or use a custom-built dataset with the same `Dataset` API as externally downloaded datasets.
It appears to be possible to point to a local dataset path rather than downloading the external ones, but I'm not exactly sure how to go about doing this.
A notebook/example script demonstrating this would be very helpful.
Could you give a more concrete example, please?
I looked up wikitext dataset script from the repo. Should I just overwrite the `data_file` on line 98 to point to the local dataset directory? Would it work for different configurations of wikitext (wikitext2, wikitext103 etc.)?
Or maybe we can use DownloadManager to specify local dataset location? In that case, where do we use DownloadManager instance?
Thanks |
https://github.com/huggingface/datasets/issues/133 | [Question] Using/adding a local dataset | Hi @MaveriQ , although what I am doing is to commit a new dataset, but I think looking at imdb script might help.
You may want to use `dl_manager.download_custom`, give it a url(arbitrary string), a custom_download(arbitrary function) and return a path, and finally use _get sample to fetch a sample. | Users may want to either create/modify a local copy of a dataset, or use a custom-built dataset with the same `Dataset` API as externally downloaded datasets.
It appears to be possible to point to a local dataset path rather than downloading the external ones, but I'm not exactly sure how to go about doing this.
A notebook/example script demonstrating this would be very helpful. | 50 | [Question] Using/adding a local dataset
Users may want to either create/modify a local copy of a dataset, or use a custom-built dataset with the same `Dataset` API as externally downloaded datasets.
It appears to be possible to point to a local dataset path rather than downloading the external ones, but I'm not exactly sure how to go about doing this.
A notebook/example script demonstrating this would be very helpful.
Hi @MaveriQ , although what I am doing is to commit a new dataset, but I think looking at imdb script might help.
You may want to use `dl_manager.download_custom`, give it a url(arbitrary string), a custom_download(arbitrary function) and return a path, and finally use _get sample to fetch a sample. |
https://github.com/huggingface/datasets/issues/133 | [Question] Using/adding a local dataset | The download manager supports local directories. You can specify a local directory instead of a url and it should work. | Users may want to either create/modify a local copy of a dataset, or use a custom-built dataset with the same `Dataset` API as externally downloaded datasets.
It appears to be possible to point to a local dataset path rather than downloading the external ones, but I'm not exactly sure how to go about doing this.
A notebook/example script demonstrating this would be very helpful. | 20 | [Question] Using/adding a local dataset
Users may want to either create/modify a local copy of a dataset, or use a custom-built dataset with the same `Dataset` API as externally downloaded datasets.
It appears to be possible to point to a local dataset path rather than downloading the external ones, but I'm not exactly sure how to go about doing this.
A notebook/example script demonstrating this would be very helpful.
The download manager supports local directories. You can specify a local directory instead of a url and it should work. |
https://github.com/huggingface/datasets/issues/131 | [Feature request] Add Toronto BookCorpus dataset | As far as I understand, `wikitext` is refer to `WikiText-103` and `WikiText-2` that created by researchers in Salesforce, and mostly used in traditional language modeling.
You might want to say `wikipedia`, a dump from wikimedia foundation.
Also I would like to have Toronto BookCorpus too ! Though it involves copyright problem... | I know the copyright/distribution of this one is complex, but it would be great to have! That, combined with the existing `wikitext`, would provide a complete dataset for pretraining models like BERT. | 51 | [Feature request] Add Toronto BookCorpus dataset
I know the copyright/distribution of this one is complex, but it would be great to have! That, combined with the existing `wikitext`, would provide a complete dataset for pretraining models like BERT.
As far as I understand, `wikitext` is refer to `WikiText-103` and `WikiText-2` that created by researchers in Salesforce, and mostly used in traditional language modeling.
You might want to say `wikipedia`, a dump from wikimedia foundation.
Also I would like to have Toronto BookCorpus too ! Though it involves copyright problem... |
https://github.com/huggingface/datasets/issues/130 | Loading GLUE dataset loads CoLA by default | As a follow-up to this: It looks like the actual GLUE task name is supplied as the `name` argument. Is there a way to check what `name`s/sub-datasets are available under a grouping like GLUE? That information doesn't seem to be readily available in info from `nlp.list_datasets()`.
Edit: I found the info under `Glue.BUILDER_CONFIGS` | If I run:
```python
dataset = nlp.load_dataset('glue')
```
The resultant dataset seems to be CoLA be default, without throwing any error. This is in contrast to calling:
```python
metric = nlp.load_metric("glue")
```
which throws an error telling the user that they need to specify a task in GLUE. Should the same apply for loading datasets? | 53 | Loading GLUE dataset loads CoLA by default
If I run:
```python
dataset = nlp.load_dataset('glue')
```
The resultant dataset seems to be CoLA be default, without throwing any error. This is in contrast to calling:
```python
metric = nlp.load_metric("glue")
```
which throws an error telling the user that they need to specify a task in GLUE. Should the same apply for loading datasets?
As a follow-up to this: It looks like the actual GLUE task name is supplied as the `name` argument. Is there a way to check what `name`s/sub-datasets are available under a grouping like GLUE? That information doesn't seem to be readily available in info from `nlp.list_datasets()`.
Edit: I found the info under `Glue.BUILDER_CONFIGS` |
https://github.com/huggingface/datasets/issues/130 | Loading GLUE dataset loads CoLA by default | Yes so the first config is loaded by default when no `name` is supplied but for GLUE this should probably throw an error indeed.
We can probably just add an `__init__` at the top of the `class Glue(nlp.GeneratorBasedBuilder)` in the `glue.py` script which does this check:
```
class Glue(nlp.GeneratorBasedBuilder):
def __init__(self, *args, **kwargs):
assert 'name' in kwargs and kwargs[name] is not None, "Glue has to be called with a configuration name"
super(Glue, self).__init__(*args, **kwargs)
``` | If I run:
```python
dataset = nlp.load_dataset('glue')
```
The resultant dataset seems to be CoLA be default, without throwing any error. This is in contrast to calling:
```python
metric = nlp.load_metric("glue")
```
which throws an error telling the user that they need to specify a task in GLUE. Should the same apply for loading datasets? | 75 | Loading GLUE dataset loads CoLA by default
If I run:
```python
dataset = nlp.load_dataset('glue')
```
The resultant dataset seems to be CoLA be default, without throwing any error. This is in contrast to calling:
```python
metric = nlp.load_metric("glue")
```
which throws an error telling the user that they need to specify a task in GLUE. Should the same apply for loading datasets?
Yes so the first config is loaded by default when no `name` is supplied but for GLUE this should probably throw an error indeed.
We can probably just add an `__init__` at the top of the `class Glue(nlp.GeneratorBasedBuilder)` in the `glue.py` script which does this check:
```
class Glue(nlp.GeneratorBasedBuilder):
def __init__(self, *args, **kwargs):
assert 'name' in kwargs and kwargs[name] is not None, "Glue has to be called with a configuration name"
super(Glue, self).__init__(*args, **kwargs)
``` |
https://github.com/huggingface/datasets/issues/130 | Loading GLUE dataset loads CoLA by default | An error is raised if the sub-dataset is not specified :)
```
ValueError: Config name is missing.
Please pick one among the available configs: ['cola', 'sst2', 'mrpc', 'qqp', 'stsb', 'mnli', 'mnli_mismatched', 'mnli_matched', 'qnli', 'rte', 'wnli', 'ax']
Example of usage:
`load_dataset('glue', 'cola')`
``` | If I run:
```python
dataset = nlp.load_dataset('glue')
```
The resultant dataset seems to be CoLA be default, without throwing any error. This is in contrast to calling:
```python
metric = nlp.load_metric("glue")
```
which throws an error telling the user that they need to specify a task in GLUE. Should the same apply for loading datasets? | 42 | Loading GLUE dataset loads CoLA by default
If I run:
```python
dataset = nlp.load_dataset('glue')
```
The resultant dataset seems to be CoLA be default, without throwing any error. This is in contrast to calling:
```python
metric = nlp.load_metric("glue")
```
which throws an error telling the user that they need to specify a task in GLUE. Should the same apply for loading datasets?
An error is raised if the sub-dataset is not specified :)
```
ValueError: Config name is missing.
Please pick one among the available configs: ['cola', 'sst2', 'mrpc', 'qqp', 'stsb', 'mnli', 'mnli_mismatched', 'mnli_matched', 'qnli', 'rte', 'wnli', 'ax']
Example of usage:
`load_dataset('glue', 'cola')`
``` |
https://github.com/huggingface/datasets/issues/129 | [Feature request] Add Google Natural Question dataset | Still work in progress :)
The idea is to have the dataset already processed somewhere so that the user only have to download the processed files. I'm also doing it for wikipedia. | Would be great to have https://github.com/google-research-datasets/natural-questions as an alternative to SQuAD. | 32 | [Feature request] Add Google Natural Question dataset
Would be great to have https://github.com/google-research-datasets/natural-questions as an alternative to SQuAD.
Still work in progress :)
The idea is to have the dataset already processed somewhere so that the user only have to download the processed files. I'm also doing it for wikipedia. |
https://github.com/huggingface/datasets/issues/129 | [Feature request] Add Google Natural Question dataset | Super appreciate your hard work !!
I'll cross my fingers and hope easily loadable wikipedia dataset will come soon. | Would be great to have https://github.com/google-research-datasets/natural-questions as an alternative to SQuAD. | 19 | [Feature request] Add Google Natural Question dataset
Would be great to have https://github.com/google-research-datasets/natural-questions as an alternative to SQuAD.
Super appreciate your hard work !!
I'll cross my fingers and hope easily loadable wikipedia dataset will come soon. |
https://github.com/huggingface/datasets/issues/129 | [Feature request] Add Google Natural Question dataset | Quick update on NQ: due to some limitations I met using apache beam + parquet I was not able to use the dataset in a nested parquet structure in python to convert it to our Apache Arrow format yet.
However we had planned to change this conversion step anyways so we'll make just sure that it enables to process and convert the NQ dataset to arrow. | Would be great to have https://github.com/google-research-datasets/natural-questions as an alternative to SQuAD. | 66 | [Feature request] Add Google Natural Question dataset
Would be great to have https://github.com/google-research-datasets/natural-questions as an alternative to SQuAD.
Quick update on NQ: due to some limitations I met using apache beam + parquet I was not able to use the dataset in a nested parquet structure in python to convert it to our Apache Arrow format yet.
However we had planned to change this conversion step anyways so we'll make just sure that it enables to process and convert the NQ dataset to arrow. |
https://github.com/huggingface/datasets/issues/128 | Some error inside nlp.load_dataset() | Google colab has an old version of Apache Arrow built-in.
Be sure you execute the "pip install" cell and restart the notebook environment if the colab asks for it. | First of all, nice work!
I am going through [this overview notebook](https://colab.research.google.com/github/huggingface/nlp/blob/master/notebooks/Overview.ipynb)
In simple step `dataset = nlp.load_dataset('squad', split='validation[:10%]')`
I get an error, which is connected with some inner code, I think:
```
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-8-d848d3a99b8c> in <module>()
1 # Downloading and loading a dataset
2
----> 3 dataset = nlp.load_dataset('squad', split='validation[:10%]')
8 frames
/usr/local/lib/python3.6/dist-packages/nlp/load.py in load_dataset(path, name, version, data_dir, data_files, split, cache_dir, download_config, download_mode, ignore_verifications, save_infos, **config_kwargs)
515 download_mode=download_mode,
516 ignore_verifications=ignore_verifications,
--> 517 save_infos=save_infos,
518 )
519
/usr/local/lib/python3.6/dist-packages/nlp/builder.py in download_and_prepare(self, download_config, download_mode, ignore_verifications, save_infos, dl_manager, **download_and_prepare_kwargs)
361 verify_infos = not save_infos and not ignore_verifications
362 self._download_and_prepare(
--> 363 dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
364 )
365 # Sync info
/usr/local/lib/python3.6/dist-packages/nlp/builder.py in _download_and_prepare(self, dl_manager, verify_infos, **prepare_split_kwargs)
414 try:
415 # Prepare split will record examples associated to the split
--> 416 self._prepare_split(split_generator, **prepare_split_kwargs)
417 except OSError:
418 raise OSError("Cannot find data file. " + (self.MANUAL_DOWNLOAD_INSTRUCTIONS or ""))
/usr/local/lib/python3.6/dist-packages/nlp/builder.py in _prepare_split(self, split_generator)
585 fname = "{}-{}.arrow".format(self.name, split_generator.name)
586 fpath = os.path.join(self._cache_dir, fname)
--> 587 examples_type = self.info.features.type
588 writer = ArrowWriter(data_type=examples_type, path=fpath, writer_batch_size=self._writer_batch_size)
589
/usr/local/lib/python3.6/dist-packages/nlp/features.py in type(self)
460 @property
461 def type(self):
--> 462 return get_nested_type(self)
463
464 @classmethod
/usr/local/lib/python3.6/dist-packages/nlp/features.py in get_nested_type(schema)
370 # Nested structures: we allow dict, list/tuples, sequences
371 if isinstance(schema, dict):
--> 372 return pa.struct({key: get_nested_type(value) for key, value in schema.items()})
373 elif isinstance(schema, (list, tuple)):
374 assert len(schema) == 1, "We defining list feature, you should just provide one example of the inner type"
/usr/local/lib/python3.6/dist-packages/nlp/features.py in <dictcomp>(.0)
370 # Nested structures: we allow dict, list/tuples, sequences
371 if isinstance(schema, dict):
--> 372 return pa.struct({key: get_nested_type(value) for key, value in schema.items()})
373 elif isinstance(schema, (list, tuple)):
374 assert len(schema) == 1, "We defining list feature, you should just provide one example of the inner type"
/usr/local/lib/python3.6/dist-packages/nlp/features.py in get_nested_type(schema)
379 # We allow to reverse list of dict => dict of list for compatiblity with tfds
380 if isinstance(inner_type, pa.StructType):
--> 381 return pa.struct(dict((f.name, pa.list_(f.type, schema.length)) for f in inner_type))
382 return pa.list_(inner_type, schema.length)
383
/usr/local/lib/python3.6/dist-packages/nlp/features.py in <genexpr>(.0)
379 # We allow to reverse list of dict => dict of list for compatiblity with tfds
380 if isinstance(inner_type, pa.StructType):
--> 381 return pa.struct(dict((f.name, pa.list_(f.type, schema.length)) for f in inner_type))
382 return pa.list_(inner_type, schema.length)
383
TypeError: list_() takes exactly one argument (2 given)
``` | 29 | Some error inside nlp.load_dataset()
First of all, nice work!
I am going through [this overview notebook](https://colab.research.google.com/github/huggingface/nlp/blob/master/notebooks/Overview.ipynb)
In simple step `dataset = nlp.load_dataset('squad', split='validation[:10%]')`
I get an error, which is connected with some inner code, I think:
```
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-8-d848d3a99b8c> in <module>()
1 # Downloading and loading a dataset
2
----> 3 dataset = nlp.load_dataset('squad', split='validation[:10%]')
8 frames
/usr/local/lib/python3.6/dist-packages/nlp/load.py in load_dataset(path, name, version, data_dir, data_files, split, cache_dir, download_config, download_mode, ignore_verifications, save_infos, **config_kwargs)
515 download_mode=download_mode,
516 ignore_verifications=ignore_verifications,
--> 517 save_infos=save_infos,
518 )
519
/usr/local/lib/python3.6/dist-packages/nlp/builder.py in download_and_prepare(self, download_config, download_mode, ignore_verifications, save_infos, dl_manager, **download_and_prepare_kwargs)
361 verify_infos = not save_infos and not ignore_verifications
362 self._download_and_prepare(
--> 363 dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
364 )
365 # Sync info
/usr/local/lib/python3.6/dist-packages/nlp/builder.py in _download_and_prepare(self, dl_manager, verify_infos, **prepare_split_kwargs)
414 try:
415 # Prepare split will record examples associated to the split
--> 416 self._prepare_split(split_generator, **prepare_split_kwargs)
417 except OSError:
418 raise OSError("Cannot find data file. " + (self.MANUAL_DOWNLOAD_INSTRUCTIONS or ""))
/usr/local/lib/python3.6/dist-packages/nlp/builder.py in _prepare_split(self, split_generator)
585 fname = "{}-{}.arrow".format(self.name, split_generator.name)
586 fpath = os.path.join(self._cache_dir, fname)
--> 587 examples_type = self.info.features.type
588 writer = ArrowWriter(data_type=examples_type, path=fpath, writer_batch_size=self._writer_batch_size)
589
/usr/local/lib/python3.6/dist-packages/nlp/features.py in type(self)
460 @property
461 def type(self):
--> 462 return get_nested_type(self)
463
464 @classmethod
/usr/local/lib/python3.6/dist-packages/nlp/features.py in get_nested_type(schema)
370 # Nested structures: we allow dict, list/tuples, sequences
371 if isinstance(schema, dict):
--> 372 return pa.struct({key: get_nested_type(value) for key, value in schema.items()})
373 elif isinstance(schema, (list, tuple)):
374 assert len(schema) == 1, "We defining list feature, you should just provide one example of the inner type"
/usr/local/lib/python3.6/dist-packages/nlp/features.py in <dictcomp>(.0)
370 # Nested structures: we allow dict, list/tuples, sequences
371 if isinstance(schema, dict):
--> 372 return pa.struct({key: get_nested_type(value) for key, value in schema.items()})
373 elif isinstance(schema, (list, tuple)):
374 assert len(schema) == 1, "We defining list feature, you should just provide one example of the inner type"
/usr/local/lib/python3.6/dist-packages/nlp/features.py in get_nested_type(schema)
379 # We allow to reverse list of dict => dict of list for compatiblity with tfds
380 if isinstance(inner_type, pa.StructType):
--> 381 return pa.struct(dict((f.name, pa.list_(f.type, schema.length)) for f in inner_type))
382 return pa.list_(inner_type, schema.length)
383
/usr/local/lib/python3.6/dist-packages/nlp/features.py in <genexpr>(.0)
379 # We allow to reverse list of dict => dict of list for compatiblity with tfds
380 if isinstance(inner_type, pa.StructType):
--> 381 return pa.struct(dict((f.name, pa.list_(f.type, schema.length)) for f in inner_type))
382 return pa.list_(inner_type, schema.length)
383
TypeError: list_() takes exactly one argument (2 given)
```
Google colab has an old version of Apache Arrow built-in.
Be sure you execute the "pip install" cell and restart the notebook environment if the colab asks for it. |
https://github.com/huggingface/datasets/issues/120 | 🐛 `map` not working | I didn't assign the output 🤦♂️
```python
dataset.map(test)
```
should be :
```python
dataset = dataset.map(test)
``` | I'm trying to run a basic example (mapping function to add a prefix).
[Here is the colab notebook I'm using.](https://colab.research.google.com/drive/1YH4JCAy0R1MMSc-k_Vlik_s1LEzP_t1h?usp=sharing)
```python
import nlp
dataset = nlp.load_dataset('squad', split='validation[:10%]')
def test(sample):
sample['title'] = "test prefix @@@ " + sample["title"]
return sample
print(dataset[0]['title'])
dataset.map(test)
print(dataset[0]['title'])
```
Output :
> Super_Bowl_50
Super_Bowl_50
Expected output :
> Super_Bowl_50
test prefix @@@ Super_Bowl_50 | 17 | 🐛 `map` not working
I'm trying to run a basic example (mapping function to add a prefix).
[Here is the colab notebook I'm using.](https://colab.research.google.com/drive/1YH4JCAy0R1MMSc-k_Vlik_s1LEzP_t1h?usp=sharing)
```python
import nlp
dataset = nlp.load_dataset('squad', split='validation[:10%]')
def test(sample):
sample['title'] = "test prefix @@@ " + sample["title"]
return sample
print(dataset[0]['title'])
dataset.map(test)
print(dataset[0]['title'])
```
Output :
> Super_Bowl_50
Super_Bowl_50
Expected output :
> Super_Bowl_50
test prefix @@@ Super_Bowl_50
I didn't assign the output 🤦♂️
```python
dataset.map(test)
```
should be :
```python
dataset = dataset.map(test)
``` |
https://github.com/huggingface/datasets/issues/119 | 🐛 Colab : type object 'pyarrow.lib.RecordBatch' has no attribute 'from_struct_array' | It's strange, after installing `nlp` on Colab, the `pyarrow` version seems fine from `pip` but not from python :
```python
import pyarrow
!pip show pyarrow
print("version = {}".format(pyarrow.__version__))
```
> Name: pyarrow
Version: 0.17.0
Summary: Python library for Apache Arrow
Home-page: https://arrow.apache.org/
Author: None
Author-email: None
License: Apache License, Version 2.0
Location: /usr/local/lib/python3.6/dist-packages
Requires: numpy
Required-by: nlp, feather-format
>
> version = 0.14.1 | I'm trying to load CNN/DM dataset on Colab.
[Colab notebook](https://colab.research.google.com/drive/11Mf7iNhIyt6GpgA1dBEtg3cyMHmMhtZS?usp=sharing)
But I meet this error :
> AttributeError: type object 'pyarrow.lib.RecordBatch' has no attribute 'from_struct_array'
| 63 | 🐛 Colab : type object 'pyarrow.lib.RecordBatch' has no attribute 'from_struct_array'
I'm trying to load CNN/DM dataset on Colab.
[Colab notebook](https://colab.research.google.com/drive/11Mf7iNhIyt6GpgA1dBEtg3cyMHmMhtZS?usp=sharing)
But I meet this error :
> AttributeError: type object 'pyarrow.lib.RecordBatch' has no attribute 'from_struct_array'
It's strange, after installing `nlp` on Colab, the `pyarrow` version seems fine from `pip` but not from python :
```python
import pyarrow
!pip show pyarrow
print("version = {}".format(pyarrow.__version__))
```
> Name: pyarrow
Version: 0.17.0
Summary: Python library for Apache Arrow
Home-page: https://arrow.apache.org/
Author: None
Author-email: None
License: Apache License, Version 2.0
Location: /usr/local/lib/python3.6/dist-packages
Requires: numpy
Required-by: nlp, feather-format
>
> version = 0.14.1 |
https://github.com/huggingface/datasets/issues/119 | 🐛 Colab : type object 'pyarrow.lib.RecordBatch' has no attribute 'from_struct_array' | Ok I just had to restart the runtime after installing `nlp`. After restarting, the version of `pyarrow` is fine. | I'm trying to load CNN/DM dataset on Colab.
[Colab notebook](https://colab.research.google.com/drive/11Mf7iNhIyt6GpgA1dBEtg3cyMHmMhtZS?usp=sharing)
But I meet this error :
> AttributeError: type object 'pyarrow.lib.RecordBatch' has no attribute 'from_struct_array'
| 19 | 🐛 Colab : type object 'pyarrow.lib.RecordBatch' has no attribute 'from_struct_array'
I'm trying to load CNN/DM dataset on Colab.
[Colab notebook](https://colab.research.google.com/drive/11Mf7iNhIyt6GpgA1dBEtg3cyMHmMhtZS?usp=sharing)
But I meet this error :
> AttributeError: type object 'pyarrow.lib.RecordBatch' has no attribute 'from_struct_array'
Ok I just had to restart the runtime after installing `nlp`. After restarting, the version of `pyarrow` is fine. |
https://github.com/huggingface/datasets/issues/116 | 🐛 Trying to use ROUGE metric : pyarrow.lib.ArrowInvalid: Column 1 named references expected length 534 but got length 323 | Sure, [here is a Colab notebook](https://colab.research.google.com/drive/1uiS89fnHMG7HV_cYxp3r-_LqJQvNNKs9?usp=sharing) reproducing the error.
> ArrowInvalid: Column 1 named references expected length 36 but got length 56 | I'm trying to use rouge metric.
I have to files : `test.pred.tokenized` and `test.gold.tokenized` with each line containing a sentence.
I tried :
```python
import nlp
rouge = nlp.load_metric('rouge')
with open("test.pred.tokenized") as p, open("test.gold.tokenized") as g:
for lp, lg in zip(p, g):
rouge.add(lp, lg)
```
But I meet following error :
> pyarrow.lib.ArrowInvalid: Column 1 named references expected length 534 but got length 323
---
Full stack-trace :
```
Traceback (most recent call last):
File "<stdin>", line 3, in <module>
File "/home/me/.venv/transformers/lib/python3.6/site-packages/nlp/metric.py", line 224, in add
self.writer.write_batch(batch)
File "/home/me/.venv/transformers/lib/python3.6/site-packages/nlp/arrow_writer.py", line 148, in write_batch
pa_table: pa.Table = pa.Table.from_pydict(batch_examples, schema=self._schema)
File "pyarrow/table.pxi", line 1550, in pyarrow.lib.Table.from_pydict
File "pyarrow/table.pxi", line 1503, in pyarrow.lib.Table.from_arrays
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 1 named references expected length 534 but got length 323
```
(`nlp` installed from source) | 22 | 🐛 Trying to use ROUGE metric : pyarrow.lib.ArrowInvalid: Column 1 named references expected length 534 but got length 323
I'm trying to use rouge metric.
I have to files : `test.pred.tokenized` and `test.gold.tokenized` with each line containing a sentence.
I tried :
```python
import nlp
rouge = nlp.load_metric('rouge')
with open("test.pred.tokenized") as p, open("test.gold.tokenized") as g:
for lp, lg in zip(p, g):
rouge.add(lp, lg)
```
But I meet following error :
> pyarrow.lib.ArrowInvalid: Column 1 named references expected length 534 but got length 323
---
Full stack-trace :
```
Traceback (most recent call last):
File "<stdin>", line 3, in <module>
File "/home/me/.venv/transformers/lib/python3.6/site-packages/nlp/metric.py", line 224, in add
self.writer.write_batch(batch)
File "/home/me/.venv/transformers/lib/python3.6/site-packages/nlp/arrow_writer.py", line 148, in write_batch
pa_table: pa.Table = pa.Table.from_pydict(batch_examples, schema=self._schema)
File "pyarrow/table.pxi", line 1550, in pyarrow.lib.Table.from_pydict
File "pyarrow/table.pxi", line 1503, in pyarrow.lib.Table.from_arrays
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 1 named references expected length 534 but got length 323
```
(`nlp` installed from source)
Sure, [here is a Colab notebook](https://colab.research.google.com/drive/1uiS89fnHMG7HV_cYxp3r-_LqJQvNNKs9?usp=sharing) reproducing the error.
> ArrowInvalid: Column 1 named references expected length 36 but got length 56 |
https://github.com/huggingface/datasets/issues/116 | 🐛 Trying to use ROUGE metric : pyarrow.lib.ArrowInvalid: Column 1 named references expected length 534 but got length 323 | This is because `add` takes as input a batch of elements and you provided only one. I think we should have `add` for one prediction/reference and `add_batch` for a batch of predictions/references. This would make it more coherent with the way we use Arrow.
Let me do this change | I'm trying to use rouge metric.
I have to files : `test.pred.tokenized` and `test.gold.tokenized` with each line containing a sentence.
I tried :
```python
import nlp
rouge = nlp.load_metric('rouge')
with open("test.pred.tokenized") as p, open("test.gold.tokenized") as g:
for lp, lg in zip(p, g):
rouge.add(lp, lg)
```
But I meet following error :
> pyarrow.lib.ArrowInvalid: Column 1 named references expected length 534 but got length 323
---
Full stack-trace :
```
Traceback (most recent call last):
File "<stdin>", line 3, in <module>
File "/home/me/.venv/transformers/lib/python3.6/site-packages/nlp/metric.py", line 224, in add
self.writer.write_batch(batch)
File "/home/me/.venv/transformers/lib/python3.6/site-packages/nlp/arrow_writer.py", line 148, in write_batch
pa_table: pa.Table = pa.Table.from_pydict(batch_examples, schema=self._schema)
File "pyarrow/table.pxi", line 1550, in pyarrow.lib.Table.from_pydict
File "pyarrow/table.pxi", line 1503, in pyarrow.lib.Table.from_arrays
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 1 named references expected length 534 but got length 323
```
(`nlp` installed from source) | 49 | 🐛 Trying to use ROUGE metric : pyarrow.lib.ArrowInvalid: Column 1 named references expected length 534 but got length 323
I'm trying to use rouge metric.
I have to files : `test.pred.tokenized` and `test.gold.tokenized` with each line containing a sentence.
I tried :
```python
import nlp
rouge = nlp.load_metric('rouge')
with open("test.pred.tokenized") as p, open("test.gold.tokenized") as g:
for lp, lg in zip(p, g):
rouge.add(lp, lg)
```
But I meet following error :
> pyarrow.lib.ArrowInvalid: Column 1 named references expected length 534 but got length 323
---
Full stack-trace :
```
Traceback (most recent call last):
File "<stdin>", line 3, in <module>
File "/home/me/.venv/transformers/lib/python3.6/site-packages/nlp/metric.py", line 224, in add
self.writer.write_batch(batch)
File "/home/me/.venv/transformers/lib/python3.6/site-packages/nlp/arrow_writer.py", line 148, in write_batch
pa_table: pa.Table = pa.Table.from_pydict(batch_examples, schema=self._schema)
File "pyarrow/table.pxi", line 1550, in pyarrow.lib.Table.from_pydict
File "pyarrow/table.pxi", line 1503, in pyarrow.lib.Table.from_arrays
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 1 named references expected length 534 but got length 323
```
(`nlp` installed from source)
This is because `add` takes as input a batch of elements and you provided only one. I think we should have `add` for one prediction/reference and `add_batch` for a batch of predictions/references. This would make it more coherent with the way we use Arrow.
Let me do this change |
https://github.com/huggingface/datasets/issues/115 | AttributeError: 'dict' object has no attribute 'info' | I could access the info by first accessing the different splits :
```python
import nlp
cnn_dm = nlp.load_dataset('cnn_dailymail')
print(cnn_dm['train'].info)
```
Information seems to be duplicated between the subsets :
```python
print(cnn_dm["train"].info == cnn_dm["test"].info == cnn_dm["validation"].info)
# True
```
Is it expected ? | I'm trying to access the information of CNN/DM dataset :
```python
cnn_dm = nlp.load_dataset('cnn_dailymail')
print(cnn_dm.info)
```
returns :
> AttributeError: 'dict' object has no attribute 'info' | 42 | AttributeError: 'dict' object has no attribute 'info'
I'm trying to access the information of CNN/DM dataset :
```python
cnn_dm = nlp.load_dataset('cnn_dailymail')
print(cnn_dm.info)
```
returns :
> AttributeError: 'dict' object has no attribute 'info'
I could access the info by first accessing the different splits :
```python
import nlp
cnn_dm = nlp.load_dataset('cnn_dailymail')
print(cnn_dm['train'].info)
```
Information seems to be duplicated between the subsets :
```python
print(cnn_dm["train"].info == cnn_dm["test"].info == cnn_dm["validation"].info)
# True
```
Is it expected ? |
https://github.com/huggingface/datasets/issues/115 | AttributeError: 'dict' object has no attribute 'info' | Good point @Colanim ! What happens under the hood when running:
```python
import nlp
cnn_dm = nlp.load_dataset('cnn_dailymail')
```
is that for every split in `cnn_dailymail`, a different dataset object (which all holds the same info) is created. This has the advantages that the datasets are easily separable in a training setup.
Also note that you can load e.g. only the `train` split of the dataset via:
```python
cnn_dm_train = nlp.load_dataset('cnn_dailymail', split="train")
print(cnn_dm_train.info)
```
I think we should make the `info` object slightly different when creating the dataset for each split - at the moment it contains for example the variable `splits` which should maybe be renamed to `split` and contain only one `SplitInfo` object ...
| I'm trying to access the information of CNN/DM dataset :
```python
cnn_dm = nlp.load_dataset('cnn_dailymail')
print(cnn_dm.info)
```
returns :
> AttributeError: 'dict' object has no attribute 'info' | 115 | AttributeError: 'dict' object has no attribute 'info'
I'm trying to access the information of CNN/DM dataset :
```python
cnn_dm = nlp.load_dataset('cnn_dailymail')
print(cnn_dm.info)
```
returns :
> AttributeError: 'dict' object has no attribute 'info'
Good point @Colanim ! What happens under the hood when running:
```python
import nlp
cnn_dm = nlp.load_dataset('cnn_dailymail')
```
is that for every split in `cnn_dailymail`, a different dataset object (which all holds the same info) is created. This has the advantages that the datasets are easily separable in a training setup.
Also note that you can load e.g. only the `train` split of the dataset via:
```python
cnn_dm_train = nlp.load_dataset('cnn_dailymail', split="train")
print(cnn_dm_train.info)
```
I think we should make the `info` object slightly different when creating the dataset for each split - at the moment it contains for example the variable `splits` which should maybe be renamed to `split` and contain only one `SplitInfo` object ...
|
https://github.com/huggingface/datasets/issues/38 | [Checksums] Error for some datasets | Fixed with 06882b4
Now your command works :)
Note that you can also do
```
nlp-cli test datasets/nlp/xnli --save_checksums
```
So that it will save the checksums directly in the right directory. | The checksums command works very nicely for `squad`. But for `crime_and_punish` and `xnli`,
the same bug happens:
When running:
```
python nlp-cli nlp-cli test xnli --save_checksums
```
leads to:
```
File "nlp-cli", line 33, in <module>
service.run()
File "/home/patrick/python_bin/nlp/commands/test.py", line 61, in run
ignore_checksums=self._ignore_checksums,
File "/home/patrick/python_bin/nlp/builder.py", line 383, in download_and_prepare
self._download_and_prepare(dl_manager=dl_manager, download_config=download_config)
File "/home/patrick/python_bin/nlp/builder.py", line 627, in _download_and_prepare
dl_manager=dl_manager, max_examples_per_split=download_config.max_examples_per_split,
File "/home/patrick/python_bin/nlp/builder.py", line 431, in _download_and_prepare
split_generators = self._split_generators(dl_manager, **split_generators_kwargs)
File "/home/patrick/python_bin/nlp/datasets/xnli/8bf4185a2da1ef2a523186dd660d9adcf0946189e7fa5942ea31c63c07b68a7f/xnli.py", line 95, in _split_generators
dl_dir = dl_manager.download_and_extract(_DATA_URL)
File "/home/patrick/python_bin/nlp/utils/download_manager.py", line 246, in download_and_extract
return self.extract(self.download(url_or_urls))
File "/home/patrick/python_bin/nlp/utils/download_manager.py", line 186, in download
self._record_sizes_checksums(url_or_urls, downloaded_path_or_paths)
File "/home/patrick/python_bin/nlp/utils/download_manager.py", line 166, in _record_sizes_checksums
self._recorded_sizes_checksums[url] = get_size_checksum(path)
File "/home/patrick/python_bin/nlp/utils/checksums_utils.py", line 81, in get_size_checksum
with open(path, "rb") as f:
TypeError: expected str, bytes or os.PathLike object, not tuple
```
| 32 | [Checksums] Error for some datasets
The checksums command works very nicely for `squad`. But for `crime_and_punish` and `xnli`,
the same bug happens:
When running:
```
python nlp-cli nlp-cli test xnli --save_checksums
```
leads to:
```
File "nlp-cli", line 33, in <module>
service.run()
File "/home/patrick/python_bin/nlp/commands/test.py", line 61, in run
ignore_checksums=self._ignore_checksums,
File "/home/patrick/python_bin/nlp/builder.py", line 383, in download_and_prepare
self._download_and_prepare(dl_manager=dl_manager, download_config=download_config)
File "/home/patrick/python_bin/nlp/builder.py", line 627, in _download_and_prepare
dl_manager=dl_manager, max_examples_per_split=download_config.max_examples_per_split,
File "/home/patrick/python_bin/nlp/builder.py", line 431, in _download_and_prepare
split_generators = self._split_generators(dl_manager, **split_generators_kwargs)
File "/home/patrick/python_bin/nlp/datasets/xnli/8bf4185a2da1ef2a523186dd660d9adcf0946189e7fa5942ea31c63c07b68a7f/xnli.py", line 95, in _split_generators
dl_dir = dl_manager.download_and_extract(_DATA_URL)
File "/home/patrick/python_bin/nlp/utils/download_manager.py", line 246, in download_and_extract
return self.extract(self.download(url_or_urls))
File "/home/patrick/python_bin/nlp/utils/download_manager.py", line 186, in download
self._record_sizes_checksums(url_or_urls, downloaded_path_or_paths)
File "/home/patrick/python_bin/nlp/utils/download_manager.py", line 166, in _record_sizes_checksums
self._recorded_sizes_checksums[url] = get_size_checksum(path)
File "/home/patrick/python_bin/nlp/utils/checksums_utils.py", line 81, in get_size_checksum
with open(path, "rb") as f:
TypeError: expected str, bytes or os.PathLike object, not tuple
```
Fixed with 06882b4
Now your command works :)
Note that you can also do
```
nlp-cli test datasets/nlp/xnli --save_checksums
```
So that it will save the checksums directly in the right directory. |
https://github.com/huggingface/datasets/issues/6 | Error when citation is not given in the DatasetInfo | Yes looks good to me.
Note that we may refactor quite strongly the `info.py` to make it a lot simpler (it's very complicated for basically a dictionary of info I think) | The following error is raised when the `citation` parameter is missing when we instantiate a `DatasetInfo`:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/jplu/dev/jplu/datasets/src/nlp/info.py", line 338, in __repr__
citation_pprint = _indent('"""{}"""'.format(self.citation.strip()))
AttributeError: 'NoneType' object has no attribute 'strip'
```
I propose to do the following change in the `info.py` file. The method:
```python
def __repr__(self):
splits_pprint = _indent("\n".join(["{"] + [
" '{}': {},".format(k, split.num_examples)
for k, split in sorted(self.splits.items())
] + ["}"]))
features_pprint = _indent(repr(self.features))
citation_pprint = _indent('"""{}"""'.format(self.citation.strip()))
return INFO_STR.format(
name=self.name,
version=self.version,
description=self.description,
total_num_examples=self.splits.total_num_examples,
features=features_pprint,
splits=splits_pprint,
citation=citation_pprint,
homepage=self.homepage,
supervised_keys=self.supervised_keys,
# Proto add a \n that we strip.
license=str(self.license).strip())
```
Becomes:
```python
def __repr__(self):
splits_pprint = _indent("\n".join(["{"] + [
" '{}': {},".format(k, split.num_examples)
for k, split in sorted(self.splits.items())
] + ["}"]))
features_pprint = _indent(repr(self.features))
## the strip is done only is the citation is given
citation_pprint = self.citation
if self.citation:
citation_pprint = _indent('"""{}"""'.format(self.citation.strip()))
return INFO_STR.format(
name=self.name,
version=self.version,
description=self.description,
total_num_examples=self.splits.total_num_examples,
features=features_pprint,
splits=splits_pprint,
citation=citation_pprint,
homepage=self.homepage,
supervised_keys=self.supervised_keys,
# Proto add a \n that we strip.
license=str(self.license).strip())
```
And now it is ok. @thomwolf are you ok with this fix? | 31 | Error when citation is not given in the DatasetInfo
The following error is raised when the `citation` parameter is missing when we instantiate a `DatasetInfo`:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/jplu/dev/jplu/datasets/src/nlp/info.py", line 338, in __repr__
citation_pprint = _indent('"""{}"""'.format(self.citation.strip()))
AttributeError: 'NoneType' object has no attribute 'strip'
```
I propose to do the following change in the `info.py` file. The method:
```python
def __repr__(self):
splits_pprint = _indent("\n".join(["{"] + [
" '{}': {},".format(k, split.num_examples)
for k, split in sorted(self.splits.items())
] + ["}"]))
features_pprint = _indent(repr(self.features))
citation_pprint = _indent('"""{}"""'.format(self.citation.strip()))
return INFO_STR.format(
name=self.name,
version=self.version,
description=self.description,
total_num_examples=self.splits.total_num_examples,
features=features_pprint,
splits=splits_pprint,
citation=citation_pprint,
homepage=self.homepage,
supervised_keys=self.supervised_keys,
# Proto add a \n that we strip.
license=str(self.license).strip())
```
Becomes:
```python
def __repr__(self):
splits_pprint = _indent("\n".join(["{"] + [
" '{}': {},".format(k, split.num_examples)
for k, split in sorted(self.splits.items())
] + ["}"]))
features_pprint = _indent(repr(self.features))
## the strip is done only is the citation is given
citation_pprint = self.citation
if self.citation:
citation_pprint = _indent('"""{}"""'.format(self.citation.strip()))
return INFO_STR.format(
name=self.name,
version=self.version,
description=self.description,
total_num_examples=self.splits.total_num_examples,
features=features_pprint,
splits=splits_pprint,
citation=citation_pprint,
homepage=self.homepage,
supervised_keys=self.supervised_keys,
# Proto add a \n that we strip.
license=str(self.license).strip())
```
And now it is ok. @thomwolf are you ok with this fix?
Yes looks good to me.
Note that we may refactor quite strongly the `info.py` to make it a lot simpler (it's very complicated for basically a dictionary of info I think) |
https://github.com/huggingface/datasets/issues/5 | ValueError when a split is empty | To fix this I propose to modify only the file `arrow_reader.py` with few updates. First update, the following method:
```python
def _make_file_instructions_from_absolutes(
name,
name2len,
absolute_instructions,
):
"""Returns the files instructions from the absolute instructions list."""
# For each split, return the files instruction (skip/take)
file_instructions = []
num_examples = 0
for abs_instr in absolute_instructions:
length = name2len[abs_instr.splitname]
if not length:
raise ValueError(
'Split empty. This might means that dataset hasn\'t been generated '
'yet and info not restored from GCS, or that legacy dataset is used.')
filename = filename_for_dataset_split(
dataset_name=name,
split=abs_instr.splitname,
filetype_suffix='arrow')
from_ = 0 if abs_instr.from_ is None else abs_instr.from_
to = length if abs_instr.to is None else abs_instr.to
num_examples += to - from_
single_file_instructions = [{"filename": filename, "skip": from_, "take": to - from_}]
file_instructions.extend(single_file_instructions)
return FileInstructions(
num_examples=num_examples,
file_instructions=file_instructions,
)
```
Becomes:
```python
def _make_file_instructions_from_absolutes(
name,
name2len,
absolute_instructions,
):
"""Returns the files instructions from the absolute instructions list."""
# For each split, return the files instruction (skip/take)
file_instructions = []
num_examples = 0
for abs_instr in absolute_instructions:
length = name2len[abs_instr.splitname]
## Delete the if not length and the raise
filename = filename_for_dataset_split(
dataset_name=name,
split=abs_instr.splitname,
filetype_suffix='arrow')
from_ = 0 if abs_instr.from_ is None else abs_instr.from_
to = length if abs_instr.to is None else abs_instr.to
num_examples += to - from_
single_file_instructions = [{"filename": filename, "skip": from_, "take": to - from_}]
file_instructions.extend(single_file_instructions)
return FileInstructions(
num_examples=num_examples,
file_instructions=file_instructions,
)
```
Second update the following method:
```python
def _read_files(files, info):
"""Returns Dataset for given file instructions.
Args:
files: List[dict(filename, skip, take)], the files information.
The filenames contain the absolute path, not relative.
skip/take indicates which example read in the file: `ds.slice(skip, take)`
"""
pa_batches = []
for f_dict in files:
pa_table: pa.Table = _get_dataset_from_filename(f_dict)
pa_batches.extend(pa_table.to_batches())
pa_table = pa.Table.from_batches(pa_batches)
ds = Dataset(arrow_table=pa_table, data_files=files, info=info)
return ds
```
Becomes:
```python
def _read_files(files, info):
"""Returns Dataset for given file instructions.
Args:
files: List[dict(filename, skip, take)], the files information.
The filenames contain the absolute path, not relative.
skip/take indicates which example read in the file: `ds.slice(skip, take)`
"""
pa_batches = []
for f_dict in files:
pa_table: pa.Table = _get_dataset_from_filename(f_dict)
pa_batches.extend(pa_table.to_batches())
## we modify the table only if there are some batches
if pa_batches:
pa_table = pa.Table.from_batches(pa_batches)
ds = Dataset(arrow_table=pa_table, data_files=files, info=info)
return ds
```
With these two updates it works now. @thomwolf are you ok with this changes? | When a split is empty either TEST, VALIDATION or TRAIN I get the following error:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/jplu/dev/jplu/datasets/src/nlp/load.py", line 295, in load
ds = dbuilder.as_dataset(**as_dataset_kwargs)
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 587, in as_dataset
datasets = utils.map_nested(build_single_dataset, split, map_tuple=True)
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 158, in map_nested
for k, v in data_struct.items()
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 158, in <dictcomp>
for k, v in data_struct.items()
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 172, in map_nested
return function(data_struct)
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 601, in _build_single_dataset
split=split,
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 625, in _as_dataset
split_infos=self.info.splits.values(),
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 200, in read
return py_utils.map_nested(_read_instruction_to_ds, instructions)
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 172, in map_nested
return function(data_struct)
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 191, in _read_instruction_to_ds
file_instructions = make_file_instructions(name, split_infos, instruction)
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 104, in make_file_instructions
absolute_instructions=absolute_instructions,
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 122, in _make_file_instructions_from_absolutes
'Split empty. This might means that dataset hasn\'t been generated '
ValueError: Split empty. This might means that dataset hasn't been generated yet and info not restored from GCS, or that legacy dataset is used.
```
How to reproduce:
```python
import csv
import nlp
class Bbc(nlp.GeneratorBasedBuilder):
VERSION = nlp.Version("1.0.0")
def __init__(self, **config):
self.train = config.pop("train", None)
self.validation = config.pop("validation", None)
super(Bbc, self).__init__(**config)
def _info(self):
return nlp.DatasetInfo(builder=self, description="bla", features=nlp.features.FeaturesDict({"id": nlp.int32, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": self.train}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": self.validation}),
nlp.SplitGenerator(name=nlp.Split.TEST, gen_kwargs={"filepath": None})]
def _generate_examples(self, filepath):
if not filepath:
return None, {}
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"id": idx, "text": line[1], "label": line[0]}
```
```python
import nlp
dataset = nlp.load("bbc", builder_kwargs={"train": "bbc/data/train.csv", "validation": "bbc/data/test.csv"})
``` | 379 | ValueError when a split is empty
When a split is empty either TEST, VALIDATION or TRAIN I get the following error:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/jplu/dev/jplu/datasets/src/nlp/load.py", line 295, in load
ds = dbuilder.as_dataset(**as_dataset_kwargs)
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 587, in as_dataset
datasets = utils.map_nested(build_single_dataset, split, map_tuple=True)
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 158, in map_nested
for k, v in data_struct.items()
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 158, in <dictcomp>
for k, v in data_struct.items()
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 172, in map_nested
return function(data_struct)
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 601, in _build_single_dataset
split=split,
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 625, in _as_dataset
split_infos=self.info.splits.values(),
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 200, in read
return py_utils.map_nested(_read_instruction_to_ds, instructions)
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 172, in map_nested
return function(data_struct)
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 191, in _read_instruction_to_ds
file_instructions = make_file_instructions(name, split_infos, instruction)
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 104, in make_file_instructions
absolute_instructions=absolute_instructions,
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 122, in _make_file_instructions_from_absolutes
'Split empty. This might means that dataset hasn\'t been generated '
ValueError: Split empty. This might means that dataset hasn't been generated yet and info not restored from GCS, or that legacy dataset is used.
```
How to reproduce:
```python
import csv
import nlp
class Bbc(nlp.GeneratorBasedBuilder):
VERSION = nlp.Version("1.0.0")
def __init__(self, **config):
self.train = config.pop("train", None)
self.validation = config.pop("validation", None)
super(Bbc, self).__init__(**config)
def _info(self):
return nlp.DatasetInfo(builder=self, description="bla", features=nlp.features.FeaturesDict({"id": nlp.int32, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": self.train}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": self.validation}),
nlp.SplitGenerator(name=nlp.Split.TEST, gen_kwargs={"filepath": None})]
def _generate_examples(self, filepath):
if not filepath:
return None, {}
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"id": idx, "text": line[1], "label": line[0]}
```
```python
import nlp
dataset = nlp.load("bbc", builder_kwargs={"train": "bbc/data/train.csv", "validation": "bbc/data/test.csv"})
```
To fix this I propose to modify only the file `arrow_reader.py` with few updates. First update, the following method:
```python
def _make_file_instructions_from_absolutes(
name,
name2len,
absolute_instructions,
):
"""Returns the files instructions from the absolute instructions list."""
# For each split, return the files instruction (skip/take)
file_instructions = []
num_examples = 0
for abs_instr in absolute_instructions:
length = name2len[abs_instr.splitname]
if not length:
raise ValueError(
'Split empty. This might means that dataset hasn\'t been generated '
'yet and info not restored from GCS, or that legacy dataset is used.')
filename = filename_for_dataset_split(
dataset_name=name,
split=abs_instr.splitname,
filetype_suffix='arrow')
from_ = 0 if abs_instr.from_ is None else abs_instr.from_
to = length if abs_instr.to is None else abs_instr.to
num_examples += to - from_
single_file_instructions = [{"filename": filename, "skip": from_, "take": to - from_}]
file_instructions.extend(single_file_instructions)
return FileInstructions(
num_examples=num_examples,
file_instructions=file_instructions,
)
```
Becomes:
```python
def _make_file_instructions_from_absolutes(
name,
name2len,
absolute_instructions,
):
"""Returns the files instructions from the absolute instructions list."""
# For each split, return the files instruction (skip/take)
file_instructions = []
num_examples = 0
for abs_instr in absolute_instructions:
length = name2len[abs_instr.splitname]
## Delete the if not length and the raise
filename = filename_for_dataset_split(
dataset_name=name,
split=abs_instr.splitname,
filetype_suffix='arrow')
from_ = 0 if abs_instr.from_ is None else abs_instr.from_
to = length if abs_instr.to is None else abs_instr.to
num_examples += to - from_
single_file_instructions = [{"filename": filename, "skip": from_, "take": to - from_}]
file_instructions.extend(single_file_instructions)
return FileInstructions(
num_examples=num_examples,
file_instructions=file_instructions,
)
```
Second update the following method:
```python
def _read_files(files, info):
"""Returns Dataset for given file instructions.
Args:
files: List[dict(filename, skip, take)], the files information.
The filenames contain the absolute path, not relative.
skip/take indicates which example read in the file: `ds.slice(skip, take)`
"""
pa_batches = []
for f_dict in files:
pa_table: pa.Table = _get_dataset_from_filename(f_dict)
pa_batches.extend(pa_table.to_batches())
pa_table = pa.Table.from_batches(pa_batches)
ds = Dataset(arrow_table=pa_table, data_files=files, info=info)
return ds
```
Becomes:
```python
def _read_files(files, info):
"""Returns Dataset for given file instructions.
Args:
files: List[dict(filename, skip, take)], the files information.
The filenames contain the absolute path, not relative.
skip/take indicates which example read in the file: `ds.slice(skip, take)`
"""
pa_batches = []
for f_dict in files:
pa_table: pa.Table = _get_dataset_from_filename(f_dict)
pa_batches.extend(pa_table.to_batches())
## we modify the table only if there are some batches
if pa_batches:
pa_table = pa.Table.from_batches(pa_batches)
ds = Dataset(arrow_table=pa_table, data_files=files, info=info)
return ds
```
With these two updates it works now. @thomwolf are you ok with this changes? |
https://github.com/huggingface/datasets/issues/5 | ValueError when a split is empty | Yes sounds good to me!
Do you want to make a PR? or I can do it as well | When a split is empty either TEST, VALIDATION or TRAIN I get the following error:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/jplu/dev/jplu/datasets/src/nlp/load.py", line 295, in load
ds = dbuilder.as_dataset(**as_dataset_kwargs)
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 587, in as_dataset
datasets = utils.map_nested(build_single_dataset, split, map_tuple=True)
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 158, in map_nested
for k, v in data_struct.items()
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 158, in <dictcomp>
for k, v in data_struct.items()
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 172, in map_nested
return function(data_struct)
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 601, in _build_single_dataset
split=split,
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 625, in _as_dataset
split_infos=self.info.splits.values(),
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 200, in read
return py_utils.map_nested(_read_instruction_to_ds, instructions)
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 172, in map_nested
return function(data_struct)
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 191, in _read_instruction_to_ds
file_instructions = make_file_instructions(name, split_infos, instruction)
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 104, in make_file_instructions
absolute_instructions=absolute_instructions,
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 122, in _make_file_instructions_from_absolutes
'Split empty. This might means that dataset hasn\'t been generated '
ValueError: Split empty. This might means that dataset hasn't been generated yet and info not restored from GCS, or that legacy dataset is used.
```
How to reproduce:
```python
import csv
import nlp
class Bbc(nlp.GeneratorBasedBuilder):
VERSION = nlp.Version("1.0.0")
def __init__(self, **config):
self.train = config.pop("train", None)
self.validation = config.pop("validation", None)
super(Bbc, self).__init__(**config)
def _info(self):
return nlp.DatasetInfo(builder=self, description="bla", features=nlp.features.FeaturesDict({"id": nlp.int32, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": self.train}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": self.validation}),
nlp.SplitGenerator(name=nlp.Split.TEST, gen_kwargs={"filepath": None})]
def _generate_examples(self, filepath):
if not filepath:
return None, {}
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"id": idx, "text": line[1], "label": line[0]}
```
```python
import nlp
dataset = nlp.load("bbc", builder_kwargs={"train": "bbc/data/train.csv", "validation": "bbc/data/test.csv"})
``` | 19 | ValueError when a split is empty
When a split is empty either TEST, VALIDATION or TRAIN I get the following error:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/jplu/dev/jplu/datasets/src/nlp/load.py", line 295, in load
ds = dbuilder.as_dataset(**as_dataset_kwargs)
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 587, in as_dataset
datasets = utils.map_nested(build_single_dataset, split, map_tuple=True)
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 158, in map_nested
for k, v in data_struct.items()
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 158, in <dictcomp>
for k, v in data_struct.items()
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 172, in map_nested
return function(data_struct)
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 601, in _build_single_dataset
split=split,
File "/home/jplu/dev/jplu/datasets/src/nlp/builder.py", line 625, in _as_dataset
split_infos=self.info.splits.values(),
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 200, in read
return py_utils.map_nested(_read_instruction_to_ds, instructions)
File "/home/jplu/dev/jplu/datasets/src/nlp/utils/py_utils.py", line 172, in map_nested
return function(data_struct)
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 191, in _read_instruction_to_ds
file_instructions = make_file_instructions(name, split_infos, instruction)
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 104, in make_file_instructions
absolute_instructions=absolute_instructions,
File "/home/jplu/dev/jplu/datasets/src/nlp/arrow_reader.py", line 122, in _make_file_instructions_from_absolutes
'Split empty. This might means that dataset hasn\'t been generated '
ValueError: Split empty. This might means that dataset hasn't been generated yet and info not restored from GCS, or that legacy dataset is used.
```
How to reproduce:
```python
import csv
import nlp
class Bbc(nlp.GeneratorBasedBuilder):
VERSION = nlp.Version("1.0.0")
def __init__(self, **config):
self.train = config.pop("train", None)
self.validation = config.pop("validation", None)
super(Bbc, self).__init__(**config)
def _info(self):
return nlp.DatasetInfo(builder=self, description="bla", features=nlp.features.FeaturesDict({"id": nlp.int32, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": self.train}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": self.validation}),
nlp.SplitGenerator(name=nlp.Split.TEST, gen_kwargs={"filepath": None})]
def _generate_examples(self, filepath):
if not filepath:
return None, {}
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"id": idx, "text": line[1], "label": line[0]}
```
```python
import nlp
dataset = nlp.load("bbc", builder_kwargs={"train": "bbc/data/train.csv", "validation": "bbc/data/test.csv"})
```
Yes sounds good to me!
Do you want to make a PR? or I can do it as well |
https://github.com/huggingface/datasets/issues/4 | [Feature] Keep the list of labels of a dataset as metadata | Yes! I see mostly two options for this:
- a `Feature` approach like currently (but we might deprecate features)
- wrapping in a smart way the Dictionary arrays of Arrow: https://arrow.apache.org/docs/python/data.html?highlight=dictionary%20encode#dictionary-arrays | It would be useful to keep the list of the labels of a dataset as metadata. Either directly in the `DatasetInfo` or in the Arrow metadata. | 31 | [Feature] Keep the list of labels of a dataset as metadata
It would be useful to keep the list of the labels of a dataset as metadata. Either directly in the `DatasetInfo` or in the Arrow metadata.
Yes! I see mostly two options for this:
- a `Feature` approach like currently (but we might deprecate features)
- wrapping in a smart way the Dictionary arrays of Arrow: https://arrow.apache.org/docs/python/data.html?highlight=dictionary%20encode#dictionary-arrays |
https://github.com/huggingface/datasets/issues/4 | [Feature] Keep the list of labels of a dataset as metadata | This should be accessible now as a feature in dataset.info.features (and even have the mapping methods). | It would be useful to keep the list of the labels of a dataset as metadata. Either directly in the `DatasetInfo` or in the Arrow metadata. | 16 | [Feature] Keep the list of labels of a dataset as metadata
It would be useful to keep the list of the labels of a dataset as metadata. Either directly in the `DatasetInfo` or in the Arrow metadata.
This should be accessible now as a feature in dataset.info.features (and even have the mapping methods). |
https://github.com/huggingface/datasets/issues/4 | [Feature] Keep the list of labels of a dataset as metadata | Hi,
I hope we could get a better documentation.
It took me more than 1 hour to found this way to get the label information. | It would be useful to keep the list of the labels of a dataset as metadata. Either directly in the `DatasetInfo` or in the Arrow metadata. | 25 | [Feature] Keep the list of labels of a dataset as metadata
It would be useful to keep the list of the labels of a dataset as metadata. Either directly in the `DatasetInfo` or in the Arrow metadata.
Hi,
I hope we could get a better documentation.
It took me more than 1 hour to found this way to get the label information. |
https://github.com/huggingface/datasets/issues/4 | [Feature] Keep the list of labels of a dataset as metadata | Yes we are working on the doc right now, should be in the next release quite soon. | It would be useful to keep the list of the labels of a dataset as metadata. Either directly in the `DatasetInfo` or in the Arrow metadata. | 17 | [Feature] Keep the list of labels of a dataset as metadata
It would be useful to keep the list of the labels of a dataset as metadata. Either directly in the `DatasetInfo` or in the Arrow metadata.
Yes we are working on the doc right now, should be in the next release quite soon. |
https://github.com/huggingface/datasets/issues/3 | [Feature] More dataset outputs | Yes!
- pandas will be a one-liner in `arrow_dataset`: https://arrow.apache.org/docs/python/generated/pyarrow.Table.html#pyarrow.Table.to_pandas
- for Spark I have no idea. let's investigate that at some point | Add the following dataset outputs:
- Spark
- Pandas | 23 | [Feature] More dataset outputs
Add the following dataset outputs:
- Spark
- Pandas
Yes!
- pandas will be a one-liner in `arrow_dataset`: https://arrow.apache.org/docs/python/generated/pyarrow.Table.html#pyarrow.Table.to_pandas
- for Spark I have no idea. let's investigate that at some point |
https://github.com/huggingface/datasets/issues/3 | [Feature] More dataset outputs | For Spark it looks to be pretty straightforward as well https://spark.apache.org/docs/latest/sql-pyspark-pandas-with-arrow.html but looks to be having a dependency to Spark is necessary, then nevermind we can skip it | Add the following dataset outputs:
- Spark
- Pandas | 28 | [Feature] More dataset outputs
Add the following dataset outputs:
- Spark
- Pandas
For Spark it looks to be pretty straightforward as well https://spark.apache.org/docs/latest/sql-pyspark-pandas-with-arrow.html but looks to be having a dependency to Spark is necessary, then nevermind we can skip it |
https://github.com/huggingface/datasets/issues/2 | Issue to read a local dataset | Ok, there are some news, most good than bad :laughing:
The dataset script now became:
```python
import csv
import nlp
class Bbc(nlp.GeneratorBasedBuilder):
VERSION = nlp.Version("1.0.0")
def __init__(self, **config):
self.train = config.pop("train", None)
self.validation = config.pop("validation", None)
super(Bbc, self).__init__(**config)
def _info(self):
return nlp.DatasetInfo(builder=self, description="bla", features=nlp.features.FeaturesDict({"id": nlp.int32, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": self.train}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": self.validation})]
def _generate_examples(self, filepath):
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"id": idx, "text": line[1], "label": line[0]}
```
And the dataset folder becomes:
```
.
├── bbc
│ ├── bbc.py
│ └── data
│ ├── test.csv
│ └── train.csv
```
I can load the dataset by using the keywords arguments like this:
```python
import nlp
dataset = nlp.load("bbc", builder_kwargs={"train": "bbc/data/train.csv", "validation": "bbc/data/test.csv"})
```
That was the good part ^^ Because it took me some time to understand that the script itself is put in cache in `datasets/src/nlp/datasets/some-hash/bbc.py` which is very difficult to discover without checking the source code. It means that doesn't matter the changes you do to your original script it is taken into account. I think instead of doing a hash on the name (I suppose it is the name), a hash on the content of the script itself should be a better solution.
Then by diving a bit in the code I found the `force_reload` parameter [here](https://github.com/huggingface/datasets/blob/master/src/nlp/load.py#L50) but the call of this `load_dataset` method is done with the `builder_kwargs` as seen [here](https://github.com/huggingface/datasets/blob/master/src/nlp/load.py#L166) which is ok until the call to the builder is done as the builder do not have this `force_reload` parameter. To show as example, the previous load becomes:
```python
import nlp
dataset = nlp.load("bbc", builder_kwargs={"train": "bbc/data/train.csv", "validation": "bbc/data/test.csv", "force_reload": True})
```
Raises
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/jplu/dev/jplu/datasets/src/nlp/load.py", line 283, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/home/jplu/dev/jplu/datasets/src/nlp/load.py", line 170, in builder
builder_instance = builder_cls(**builder_kwargs)
File "/home/jplu/dev/jplu/datasets/src/nlp/datasets/84d638d2a8ca919d1021a554e741766f50679dc6553d5a0612b6094311babd39/bbc.py", line 12, in __init__
super(Bbc, self).__init__(**config)
TypeError: __init__() got an unexpected keyword argument 'force_reload'
```
So yes the cache is refreshed with the new script but then raises this error. | Hello,
As proposed by @thomwolf, I open an issue to explain what I'm trying to do without success. What I want to do is to create and load a local dataset, the script I have done is the following:
```python
import os
import csv
import nlp
class BbcConfig(nlp.BuilderConfig):
def __init__(self, **kwargs):
super(BbcConfig, self).__init__(**kwargs)
class Bbc(nlp.GeneratorBasedBuilder):
_DIR = "./data"
_DEV_FILE = "test.csv"
_TRAINING_FILE = "train.csv"
BUILDER_CONFIGS = [BbcConfig(name="bbc", version=nlp.Version("1.0.0"))]
def _info(self):
return nlp.DatasetInfo(builder=self, features=nlp.features.FeaturesDict({"id": nlp.string, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
files = {"train": os.path.join(self._DIR, self._TRAINING_FILE), "dev": os.path.join(self._DIR, self._DEV_FILE)}
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": files["train"]}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": files["dev"]})]
def _generate_examples(self, filepath):
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"idx": idx, "text": line[1], "label": line[0]}
```
The dataset is attached to this issue as well:
[data.zip](https://github.com/huggingface/datasets/files/4476928/data.zip)
Now the steps to reproduce what I would like to do:
1. unzip data locally (I know the nlp lib can detect and extract archives but I want to reduce and facilitate the reproduction as much as possible)
2. create the `bbc.py` script as above at the same location than the unziped `data` folder.
Now I try to load the dataset in three different ways and none works, the first one with the name of the dataset like I would do with TFDS:
```python
import nlp
from bbc import Bbc
dataset = nlp.load("bbc")
```
I get:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 88, in load_dataset
local_files_only=local_files_only,
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/utils/file_utils.py", line 214, in cached_path
if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
File "/opt/anaconda3/envs/transformers/lib/python3.7/zipfile.py", line 203, in is_zipfile
with open(filename, "rb") as fp:
TypeError: expected str, bytes or os.PathLike object, not NoneType
```
But @thomwolf told me that no need to import the script, just put the path of it, then I tried three different way to do:
```python
import nlp
dataset = nlp.load("bbc.py")
```
And
```python
import nlp
dataset = nlp.load("./bbc.py")
```
And
```python
import nlp
dataset = nlp.load("/absolute/path/to/bbc.py")
```
These three ways gives me:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 124, in load_dataset
dataset_module = importlib.import_module(module_path)
File "/opt/anaconda3/envs/transformers/lib/python3.7/importlib/__init__.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
File "<frozen importlib._bootstrap>", line 983, in _find_and_load
File "<frozen importlib._bootstrap>", line 965, in _find_and_load_unlocked
ModuleNotFoundError: No module named 'nlp.datasets.2fd72627d92c328b3e9c4a3bf7ec932c48083caca09230cebe4c618da6e93688.bbc'
```
Any idea of what I'm missing? or I might have spot a bug :) | 353 | Issue to read a local dataset
Hello,
As proposed by @thomwolf, I open an issue to explain what I'm trying to do without success. What I want to do is to create and load a local dataset, the script I have done is the following:
```python
import os
import csv
import nlp
class BbcConfig(nlp.BuilderConfig):
def __init__(self, **kwargs):
super(BbcConfig, self).__init__(**kwargs)
class Bbc(nlp.GeneratorBasedBuilder):
_DIR = "./data"
_DEV_FILE = "test.csv"
_TRAINING_FILE = "train.csv"
BUILDER_CONFIGS = [BbcConfig(name="bbc", version=nlp.Version("1.0.0"))]
def _info(self):
return nlp.DatasetInfo(builder=self, features=nlp.features.FeaturesDict({"id": nlp.string, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
files = {"train": os.path.join(self._DIR, self._TRAINING_FILE), "dev": os.path.join(self._DIR, self._DEV_FILE)}
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": files["train"]}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": files["dev"]})]
def _generate_examples(self, filepath):
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"idx": idx, "text": line[1], "label": line[0]}
```
The dataset is attached to this issue as well:
[data.zip](https://github.com/huggingface/datasets/files/4476928/data.zip)
Now the steps to reproduce what I would like to do:
1. unzip data locally (I know the nlp lib can detect and extract archives but I want to reduce and facilitate the reproduction as much as possible)
2. create the `bbc.py` script as above at the same location than the unziped `data` folder.
Now I try to load the dataset in three different ways and none works, the first one with the name of the dataset like I would do with TFDS:
```python
import nlp
from bbc import Bbc
dataset = nlp.load("bbc")
```
I get:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 88, in load_dataset
local_files_only=local_files_only,
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/utils/file_utils.py", line 214, in cached_path
if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
File "/opt/anaconda3/envs/transformers/lib/python3.7/zipfile.py", line 203, in is_zipfile
with open(filename, "rb") as fp:
TypeError: expected str, bytes or os.PathLike object, not NoneType
```
But @thomwolf told me that no need to import the script, just put the path of it, then I tried three different way to do:
```python
import nlp
dataset = nlp.load("bbc.py")
```
And
```python
import nlp
dataset = nlp.load("./bbc.py")
```
And
```python
import nlp
dataset = nlp.load("/absolute/path/to/bbc.py")
```
These three ways gives me:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 124, in load_dataset
dataset_module = importlib.import_module(module_path)
File "/opt/anaconda3/envs/transformers/lib/python3.7/importlib/__init__.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
File "<frozen importlib._bootstrap>", line 983, in _find_and_load
File "<frozen importlib._bootstrap>", line 965, in _find_and_load_unlocked
ModuleNotFoundError: No module named 'nlp.datasets.2fd72627d92c328b3e9c4a3bf7ec932c48083caca09230cebe4c618da6e93688.bbc'
```
Any idea of what I'm missing? or I might have spot a bug :)
Ok, there are some news, most good than bad :laughing:
The dataset script now became:
```python
import csv
import nlp
class Bbc(nlp.GeneratorBasedBuilder):
VERSION = nlp.Version("1.0.0")
def __init__(self, **config):
self.train = config.pop("train", None)
self.validation = config.pop("validation", None)
super(Bbc, self).__init__(**config)
def _info(self):
return nlp.DatasetInfo(builder=self, description="bla", features=nlp.features.FeaturesDict({"id": nlp.int32, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": self.train}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": self.validation})]
def _generate_examples(self, filepath):
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"id": idx, "text": line[1], "label": line[0]}
```
And the dataset folder becomes:
```
.
├── bbc
│ ├── bbc.py
│ └── data
│ ├── test.csv
│ └── train.csv
```
I can load the dataset by using the keywords arguments like this:
```python
import nlp
dataset = nlp.load("bbc", builder_kwargs={"train": "bbc/data/train.csv", "validation": "bbc/data/test.csv"})
```
That was the good part ^^ Because it took me some time to understand that the script itself is put in cache in `datasets/src/nlp/datasets/some-hash/bbc.py` which is very difficult to discover without checking the source code. It means that doesn't matter the changes you do to your original script it is taken into account. I think instead of doing a hash on the name (I suppose it is the name), a hash on the content of the script itself should be a better solution.
Then by diving a bit in the code I found the `force_reload` parameter [here](https://github.com/huggingface/datasets/blob/master/src/nlp/load.py#L50) but the call of this `load_dataset` method is done with the `builder_kwargs` as seen [here](https://github.com/huggingface/datasets/blob/master/src/nlp/load.py#L166) which is ok until the call to the builder is done as the builder do not have this `force_reload` parameter. To show as example, the previous load becomes:
```python
import nlp
dataset = nlp.load("bbc", builder_kwargs={"train": "bbc/data/train.csv", "validation": "bbc/data/test.csv", "force_reload": True})
```
Raises
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/jplu/dev/jplu/datasets/src/nlp/load.py", line 283, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/home/jplu/dev/jplu/datasets/src/nlp/load.py", line 170, in builder
builder_instance = builder_cls(**builder_kwargs)
File "/home/jplu/dev/jplu/datasets/src/nlp/datasets/84d638d2a8ca919d1021a554e741766f50679dc6553d5a0612b6094311babd39/bbc.py", line 12, in __init__
super(Bbc, self).__init__(**config)
TypeError: __init__() got an unexpected keyword argument 'force_reload'
```
So yes the cache is refreshed with the new script but then raises this error. |
https://github.com/huggingface/datasets/issues/2 | Issue to read a local dataset | Ok great, so as discussed today, let's:
- have a main dataset directory inside the lib with sub-directories hashed by the content of the file
- keep a cache for downloading the scripts from S3 for now
- later: add methods to list and clean the local versions of the datasets (and the distant versions on S3 as well)
Side question: do you often use `builder_kwargs` for other things than supplying file paths? I was thinking about having a more easy to read and remember `data_files` argument maybe. | Hello,
As proposed by @thomwolf, I open an issue to explain what I'm trying to do without success. What I want to do is to create and load a local dataset, the script I have done is the following:
```python
import os
import csv
import nlp
class BbcConfig(nlp.BuilderConfig):
def __init__(self, **kwargs):
super(BbcConfig, self).__init__(**kwargs)
class Bbc(nlp.GeneratorBasedBuilder):
_DIR = "./data"
_DEV_FILE = "test.csv"
_TRAINING_FILE = "train.csv"
BUILDER_CONFIGS = [BbcConfig(name="bbc", version=nlp.Version("1.0.0"))]
def _info(self):
return nlp.DatasetInfo(builder=self, features=nlp.features.FeaturesDict({"id": nlp.string, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
files = {"train": os.path.join(self._DIR, self._TRAINING_FILE), "dev": os.path.join(self._DIR, self._DEV_FILE)}
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": files["train"]}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": files["dev"]})]
def _generate_examples(self, filepath):
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"idx": idx, "text": line[1], "label": line[0]}
```
The dataset is attached to this issue as well:
[data.zip](https://github.com/huggingface/datasets/files/4476928/data.zip)
Now the steps to reproduce what I would like to do:
1. unzip data locally (I know the nlp lib can detect and extract archives but I want to reduce and facilitate the reproduction as much as possible)
2. create the `bbc.py` script as above at the same location than the unziped `data` folder.
Now I try to load the dataset in three different ways and none works, the first one with the name of the dataset like I would do with TFDS:
```python
import nlp
from bbc import Bbc
dataset = nlp.load("bbc")
```
I get:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 88, in load_dataset
local_files_only=local_files_only,
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/utils/file_utils.py", line 214, in cached_path
if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
File "/opt/anaconda3/envs/transformers/lib/python3.7/zipfile.py", line 203, in is_zipfile
with open(filename, "rb") as fp:
TypeError: expected str, bytes or os.PathLike object, not NoneType
```
But @thomwolf told me that no need to import the script, just put the path of it, then I tried three different way to do:
```python
import nlp
dataset = nlp.load("bbc.py")
```
And
```python
import nlp
dataset = nlp.load("./bbc.py")
```
And
```python
import nlp
dataset = nlp.load("/absolute/path/to/bbc.py")
```
These three ways gives me:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 124, in load_dataset
dataset_module = importlib.import_module(module_path)
File "/opt/anaconda3/envs/transformers/lib/python3.7/importlib/__init__.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
File "<frozen importlib._bootstrap>", line 983, in _find_and_load
File "<frozen importlib._bootstrap>", line 965, in _find_and_load_unlocked
ModuleNotFoundError: No module named 'nlp.datasets.2fd72627d92c328b3e9c4a3bf7ec932c48083caca09230cebe4c618da6e93688.bbc'
```
Any idea of what I'm missing? or I might have spot a bug :) | 88 | Issue to read a local dataset
Hello,
As proposed by @thomwolf, I open an issue to explain what I'm trying to do without success. What I want to do is to create and load a local dataset, the script I have done is the following:
```python
import os
import csv
import nlp
class BbcConfig(nlp.BuilderConfig):
def __init__(self, **kwargs):
super(BbcConfig, self).__init__(**kwargs)
class Bbc(nlp.GeneratorBasedBuilder):
_DIR = "./data"
_DEV_FILE = "test.csv"
_TRAINING_FILE = "train.csv"
BUILDER_CONFIGS = [BbcConfig(name="bbc", version=nlp.Version("1.0.0"))]
def _info(self):
return nlp.DatasetInfo(builder=self, features=nlp.features.FeaturesDict({"id": nlp.string, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
files = {"train": os.path.join(self._DIR, self._TRAINING_FILE), "dev": os.path.join(self._DIR, self._DEV_FILE)}
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": files["train"]}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": files["dev"]})]
def _generate_examples(self, filepath):
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"idx": idx, "text": line[1], "label": line[0]}
```
The dataset is attached to this issue as well:
[data.zip](https://github.com/huggingface/datasets/files/4476928/data.zip)
Now the steps to reproduce what I would like to do:
1. unzip data locally (I know the nlp lib can detect and extract archives but I want to reduce and facilitate the reproduction as much as possible)
2. create the `bbc.py` script as above at the same location than the unziped `data` folder.
Now I try to load the dataset in three different ways and none works, the first one with the name of the dataset like I would do with TFDS:
```python
import nlp
from bbc import Bbc
dataset = nlp.load("bbc")
```
I get:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 88, in load_dataset
local_files_only=local_files_only,
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/utils/file_utils.py", line 214, in cached_path
if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
File "/opt/anaconda3/envs/transformers/lib/python3.7/zipfile.py", line 203, in is_zipfile
with open(filename, "rb") as fp:
TypeError: expected str, bytes or os.PathLike object, not NoneType
```
But @thomwolf told me that no need to import the script, just put the path of it, then I tried three different way to do:
```python
import nlp
dataset = nlp.load("bbc.py")
```
And
```python
import nlp
dataset = nlp.load("./bbc.py")
```
And
```python
import nlp
dataset = nlp.load("/absolute/path/to/bbc.py")
```
These three ways gives me:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 124, in load_dataset
dataset_module = importlib.import_module(module_path)
File "/opt/anaconda3/envs/transformers/lib/python3.7/importlib/__init__.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
File "<frozen importlib._bootstrap>", line 983, in _find_and_load
File "<frozen importlib._bootstrap>", line 965, in _find_and_load_unlocked
ModuleNotFoundError: No module named 'nlp.datasets.2fd72627d92c328b3e9c4a3bf7ec932c48083caca09230cebe4c618da6e93688.bbc'
```
Any idea of what I'm missing? or I might have spot a bug :)
Ok great, so as discussed today, let's:
- have a main dataset directory inside the lib with sub-directories hashed by the content of the file
- keep a cache for downloading the scripts from S3 for now
- later: add methods to list and clean the local versions of the datasets (and the distant versions on S3 as well)
Side question: do you often use `builder_kwargs` for other things than supplying file paths? I was thinking about having a more easy to read and remember `data_files` argument maybe. |
https://github.com/huggingface/datasets/issues/2 | Issue to read a local dataset | Good plan!
Yes I do use `builder_kwargs` for other things such as:
- dataset name
- properties to know how to properly read a CSV file: do I have to skip the first line in a CSV, which delimiter is used, and the columns ids to use.
- properties to know how to properly read a JSON file: which properties in a JSON object to read | Hello,
As proposed by @thomwolf, I open an issue to explain what I'm trying to do without success. What I want to do is to create and load a local dataset, the script I have done is the following:
```python
import os
import csv
import nlp
class BbcConfig(nlp.BuilderConfig):
def __init__(self, **kwargs):
super(BbcConfig, self).__init__(**kwargs)
class Bbc(nlp.GeneratorBasedBuilder):
_DIR = "./data"
_DEV_FILE = "test.csv"
_TRAINING_FILE = "train.csv"
BUILDER_CONFIGS = [BbcConfig(name="bbc", version=nlp.Version("1.0.0"))]
def _info(self):
return nlp.DatasetInfo(builder=self, features=nlp.features.FeaturesDict({"id": nlp.string, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
files = {"train": os.path.join(self._DIR, self._TRAINING_FILE), "dev": os.path.join(self._DIR, self._DEV_FILE)}
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": files["train"]}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": files["dev"]})]
def _generate_examples(self, filepath):
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"idx": idx, "text": line[1], "label": line[0]}
```
The dataset is attached to this issue as well:
[data.zip](https://github.com/huggingface/datasets/files/4476928/data.zip)
Now the steps to reproduce what I would like to do:
1. unzip data locally (I know the nlp lib can detect and extract archives but I want to reduce and facilitate the reproduction as much as possible)
2. create the `bbc.py` script as above at the same location than the unziped `data` folder.
Now I try to load the dataset in three different ways and none works, the first one with the name of the dataset like I would do with TFDS:
```python
import nlp
from bbc import Bbc
dataset = nlp.load("bbc")
```
I get:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 88, in load_dataset
local_files_only=local_files_only,
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/utils/file_utils.py", line 214, in cached_path
if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
File "/opt/anaconda3/envs/transformers/lib/python3.7/zipfile.py", line 203, in is_zipfile
with open(filename, "rb") as fp:
TypeError: expected str, bytes or os.PathLike object, not NoneType
```
But @thomwolf told me that no need to import the script, just put the path of it, then I tried three different way to do:
```python
import nlp
dataset = nlp.load("bbc.py")
```
And
```python
import nlp
dataset = nlp.load("./bbc.py")
```
And
```python
import nlp
dataset = nlp.load("/absolute/path/to/bbc.py")
```
These three ways gives me:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 124, in load_dataset
dataset_module = importlib.import_module(module_path)
File "/opt/anaconda3/envs/transformers/lib/python3.7/importlib/__init__.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
File "<frozen importlib._bootstrap>", line 983, in _find_and_load
File "<frozen importlib._bootstrap>", line 965, in _find_and_load_unlocked
ModuleNotFoundError: No module named 'nlp.datasets.2fd72627d92c328b3e9c4a3bf7ec932c48083caca09230cebe4c618da6e93688.bbc'
```
Any idea of what I'm missing? or I might have spot a bug :) | 66 | Issue to read a local dataset
Hello,
As proposed by @thomwolf, I open an issue to explain what I'm trying to do without success. What I want to do is to create and load a local dataset, the script I have done is the following:
```python
import os
import csv
import nlp
class BbcConfig(nlp.BuilderConfig):
def __init__(self, **kwargs):
super(BbcConfig, self).__init__(**kwargs)
class Bbc(nlp.GeneratorBasedBuilder):
_DIR = "./data"
_DEV_FILE = "test.csv"
_TRAINING_FILE = "train.csv"
BUILDER_CONFIGS = [BbcConfig(name="bbc", version=nlp.Version("1.0.0"))]
def _info(self):
return nlp.DatasetInfo(builder=self, features=nlp.features.FeaturesDict({"id": nlp.string, "text": nlp.string, "label": nlp.string}))
def _split_generators(self, dl_manager):
files = {"train": os.path.join(self._DIR, self._TRAINING_FILE), "dev": os.path.join(self._DIR, self._DEV_FILE)}
return [nlp.SplitGenerator(name=nlp.Split.TRAIN, gen_kwargs={"filepath": files["train"]}),
nlp.SplitGenerator(name=nlp.Split.VALIDATION, gen_kwargs={"filepath": files["dev"]})]
def _generate_examples(self, filepath):
with open(filepath) as f:
reader = csv.reader(f, delimiter=',', quotechar="\"")
lines = list(reader)[1:]
for idx, line in enumerate(lines):
yield idx, {"idx": idx, "text": line[1], "label": line[0]}
```
The dataset is attached to this issue as well:
[data.zip](https://github.com/huggingface/datasets/files/4476928/data.zip)
Now the steps to reproduce what I would like to do:
1. unzip data locally (I know the nlp lib can detect and extract archives but I want to reduce and facilitate the reproduction as much as possible)
2. create the `bbc.py` script as above at the same location than the unziped `data` folder.
Now I try to load the dataset in three different ways and none works, the first one with the name of the dataset like I would do with TFDS:
```python
import nlp
from bbc import Bbc
dataset = nlp.load("bbc")
```
I get:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 88, in load_dataset
local_files_only=local_files_only,
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/utils/file_utils.py", line 214, in cached_path
if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
File "/opt/anaconda3/envs/transformers/lib/python3.7/zipfile.py", line 203, in is_zipfile
with open(filename, "rb") as fp:
TypeError: expected str, bytes or os.PathLike object, not NoneType
```
But @thomwolf told me that no need to import the script, just put the path of it, then I tried three different way to do:
```python
import nlp
dataset = nlp.load("bbc.py")
```
And
```python
import nlp
dataset = nlp.load("./bbc.py")
```
And
```python
import nlp
dataset = nlp.load("/absolute/path/to/bbc.py")
```
These three ways gives me:
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 280, in load
dbuilder: DatasetBuilder = builder(path, name, data_dir=data_dir, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 166, in builder
builder_cls = load_dataset(path, name=name, **builder_kwargs)
File "/opt/anaconda3/envs/transformers/lib/python3.7/site-packages/nlp/load.py", line 124, in load_dataset
dataset_module = importlib.import_module(module_path)
File "/opt/anaconda3/envs/transformers/lib/python3.7/importlib/__init__.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
File "<frozen importlib._bootstrap>", line 983, in _find_and_load
File "<frozen importlib._bootstrap>", line 965, in _find_and_load_unlocked
ModuleNotFoundError: No module named 'nlp.datasets.2fd72627d92c328b3e9c4a3bf7ec932c48083caca09230cebe4c618da6e93688.bbc'
```
Any idea of what I'm missing? or I might have spot a bug :)
Good plan!
Yes I do use `builder_kwargs` for other things such as:
- dataset name
- properties to know how to properly read a CSV file: do I have to skip the first line in a CSV, which delimiter is used, and the columns ids to use.
- properties to know how to properly read a JSON file: which properties in a JSON object to read |