html_url
stringlengths 51
51
| title
stringlengths 6
280
| comments
stringlengths 67
24.7k
| body
stringlengths 51
36.2k
⌀ | comment_length
int64 16
1.45k
| text
stringlengths 159
38.3k
|
|---|---|---|---|---|---|
https://github.com/huggingface/datasets/issues/2591 | Cached dataset overflowing disk space | Hi @BirgerMoell.
You have several options:
- to set caching to be stored on a different path location, other than the default one (`~/.cache/huggingface/datasets`):
- either setting the environment variable `HF_DATASETS_CACHE` with the path to the new cache location
- or by passing it with the parameter `cache_dir` when loading each of the datasets: `dataset = load_dataset(..., cache_dir=your_new_location)`
You can get all the information in the docs: https://huggingface.co/docs/datasets/loading_datasets.html#cache-directory
- I wouldn't recommend disabling caching, because current implementation generates cache files anyway, although in a temporary directory and they are deleted when the session closes. See details here: https://huggingface.co/docs/datasets/processing.html#enable-or-disable-caching
- You could alternatively load the datasets in streaming mode. This is a new feature which allows loading the datasets without downloading the entire files. More information here: https://huggingface.co/docs/datasets/dataset_streaming.html | I'm training a Swedish Wav2vec2 model on a Linux GPU and having issues that the huggingface cached dataset folder is completely filling up my disk space (I'm training on a dataset of around 500 gb).
The cache folder is 500gb (and now my disk space is full).
Is there a way to toggle caching or set the caching to be stored on a different device (I have another drive with 4 tb that could hold the caching files).
This might not technically be a bug, but I was unsure and I felt that the bug was the closest one.
Traceback (most recent call last):
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/multiprocess/pool.py", line 121, in worker
result = (True, func(*args, **kwds))
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 186, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/fingerprint.py", line 397, in wrapper
out = func(self, *args, **kwargs)
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 1983, in _map_single
writer.finalize()
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_writer.py", line 418, in finalize
self.pa_writer.close()
File "pyarrow/ipc.pxi", line 402, in pyarrow.lib._CRecordBatchWriter.close
File "pyarrow/error.pxi", line 97, in pyarrow.lib.check_status
OSError: [Errno 28] Error writing bytes to file. Detail: [errno 28] No space left on device
"""
The above exception was the direct cause of the following exception:
| 127 | Cached dataset overflowing disk space
I'm training a Swedish Wav2vec2 model on a Linux GPU and having issues that the huggingface cached dataset folder is completely filling up my disk space (I'm training on a dataset of around 500 gb).
The cache folder is 500gb (and now my disk space is full).
Is there a way to toggle caching or set the caching to be stored on a different device (I have another drive with 4 tb that could hold the caching files).
This might not technically be a bug, but I was unsure and I felt that the bug was the closest one.
Traceback (most recent call last):
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/multiprocess/pool.py", line 121, in worker
result = (True, func(*args, **kwds))
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 186, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/fingerprint.py", line 397, in wrapper
out = func(self, *args, **kwargs)
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 1983, in _map_single
writer.finalize()
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_writer.py", line 418, in finalize
self.pa_writer.close()
File "pyarrow/ipc.pxi", line 402, in pyarrow.lib._CRecordBatchWriter.close
File "pyarrow/error.pxi", line 97, in pyarrow.lib.check_status
OSError: [Errno 28] Error writing bytes to file. Detail: [errno 28] No space left on device
"""
The above exception was the direct cause of the following exception:
Hi @BirgerMoell.
You have several options:
- to set caching to be stored on a different path location, other than the default one (`~/.cache/huggingface/datasets`):
- either setting the environment variable `HF_DATASETS_CACHE` with the path to the new cache location
- or by passing it with the parameter `cache_dir` when loading each of the datasets: `dataset = load_dataset(..., cache_dir=your_new_location)`
You can get all the information in the docs: https://huggingface.co/docs/datasets/loading_datasets.html#cache-directory
- I wouldn't recommend disabling caching, because current implementation generates cache files anyway, although in a temporary directory and they are deleted when the session closes. See details here: https://huggingface.co/docs/datasets/processing.html#enable-or-disable-caching
- You could alternatively load the datasets in streaming mode. This is a new feature which allows loading the datasets without downloading the entire files. More information here: https://huggingface.co/docs/datasets/dataset_streaming.html |
https://github.com/huggingface/datasets/issues/2591 | Cached dataset overflowing disk space | Hi @BirgerMoell,
We are planning to add a new feature to datasets, which could be interesting in your case: Add the option to delete temporary files (decompressed files) from the cache directory (see: #2481, #2604).
We will ping you once this feature is implemented, so that the size of your cache directory will be considerably reduced. | I'm training a Swedish Wav2vec2 model on a Linux GPU and having issues that the huggingface cached dataset folder is completely filling up my disk space (I'm training on a dataset of around 500 gb).
The cache folder is 500gb (and now my disk space is full).
Is there a way to toggle caching or set the caching to be stored on a different device (I have another drive with 4 tb that could hold the caching files).
This might not technically be a bug, but I was unsure and I felt that the bug was the closest one.
Traceback (most recent call last):
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/multiprocess/pool.py", line 121, in worker
result = (True, func(*args, **kwds))
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 186, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/fingerprint.py", line 397, in wrapper
out = func(self, *args, **kwargs)
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 1983, in _map_single
writer.finalize()
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_writer.py", line 418, in finalize
self.pa_writer.close()
File "pyarrow/ipc.pxi", line 402, in pyarrow.lib._CRecordBatchWriter.close
File "pyarrow/error.pxi", line 97, in pyarrow.lib.check_status
OSError: [Errno 28] Error writing bytes to file. Detail: [errno 28] No space left on device
"""
The above exception was the direct cause of the following exception:
| 56 | Cached dataset overflowing disk space
I'm training a Swedish Wav2vec2 model on a Linux GPU and having issues that the huggingface cached dataset folder is completely filling up my disk space (I'm training on a dataset of around 500 gb).
The cache folder is 500gb (and now my disk space is full).
Is there a way to toggle caching or set the caching to be stored on a different device (I have another drive with 4 tb that could hold the caching files).
This might not technically be a bug, but I was unsure and I felt that the bug was the closest one.
Traceback (most recent call last):
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/multiprocess/pool.py", line 121, in worker
result = (True, func(*args, **kwds))
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 186, in wrapper
out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/fingerprint.py", line 397, in wrapper
out = func(self, *args, **kwargs)
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_dataset.py", line 1983, in _map_single
writer.finalize()
File "/home/birger/miniconda3/envs/wav2vec2/lib/python3.7/site-packages/datasets/arrow_writer.py", line 418, in finalize
self.pa_writer.close()
File "pyarrow/ipc.pxi", line 402, in pyarrow.lib._CRecordBatchWriter.close
File "pyarrow/error.pxi", line 97, in pyarrow.lib.check_status
OSError: [Errno 28] Error writing bytes to file. Detail: [errno 28] No space left on device
"""
The above exception was the direct cause of the following exception:
Hi @BirgerMoell,
We are planning to add a new feature to datasets, which could be interesting in your case: Add the option to delete temporary files (decompressed files) from the cache directory (see: #2481, #2604).
We will ping you once this feature is implemented, so that the size of your cache directory will be considerably reduced. |
https://github.com/huggingface/datasets/issues/2585 | sqaud_v2 dataset contains misalignment between the answer text and the context value at the answer index | Hi @mmajurski, thanks for reporting this issue.
Indeed this misalignment arises because the source dataset context field contains leading blank spaces (and these are counted within the answer_start), while our datasets loading script removes these leading blank spaces.
I'm going to fix our script so that all leading blank spaces in the source dataset are kept, and there is no misalignment between the answer text and the answer_start within the context. | ## Describe the bug
The built in huggingface squad_v2 dataset that you can access via datasets.load_dataset contains mis-alignment between the answers['text'] and the characters in the context at the location specified by answers['answer_start'].
For example:
id = '56d1f453e7d4791d009025bd'
answers = {'text': ['Pure Land'], 'answer_start': [146]}
However the actual text in context at location 146 is 'ure Land,'
Which is an off-by-one error from the correct answer.
## Steps to reproduce the bug
```python
import datasets
def check_context_answer_alignment(example):
for a_idx in range(len(example['answers']['text'])):
# check raw dataset for answer consistency between context and answer
answer_text = example['answers']['text'][a_idx]
a_st_idx = example['answers']['answer_start'][a_idx]
a_end_idx = a_st_idx + len(example['answers']['text'][a_idx])
answer_text_from_context = example['context'][a_st_idx:a_end_idx]
if answer_text != answer_text_from_context:
#print(example['id'])
return False
return True
dataset = datasets.load_dataset('squad_v2', split='train', keep_in_memory=True)
start_len = len(dataset)
dataset = dataset.filter(check_context_answer_alignment,
num_proc=1,
keep_in_memory=True)
end_len = len(dataset)
print('{} instances contain mis-alignment between the answer text and answer index.'.format(start_len - end_len))
```
## Expected results
This code should result in 0 rows being filtered out from the dataset.
## Actual results
This filter command results in 258 rows being flagged as containing a discrepancy between the text contained within answers['text'] and the text in example['context'] at the answers['answer_start'] location.
This code will reproduce the problem and produce the following count:
"258 instances contain mis-alignment between the answer text and answer index."
## Environment info
Steps to rebuilt the Conda environment:
```
# create a virtual environment to stuff all these packages into
conda create -n round8 python=3.8 -y
# activate the virtual environment
conda activate round8
# install pytorch (best done through conda to handle cuda dependencies)
conda install pytorch torchvision torchtext cudatoolkit=11.1 -c pytorch-lts -c nvidia
pip install jsonpickle transformers datasets matplotlib
```
OS: Ubuntu 20.04
Python 3.8
Result of `conda env export`:
```
name: round8
channels:
- pytorch-lts
- nvidia
- defaults
dependencies:
- _libgcc_mutex=0.1=main
- _openmp_mutex=4.5=1_gnu
- blas=1.0=mkl
- brotlipy=0.7.0=py38h27cfd23_1003
- bzip2=1.0.8=h7b6447c_0
- ca-certificates=2021.5.25=h06a4308_1
- certifi=2021.5.30=py38h06a4308_0
- cffi=1.14.5=py38h261ae71_0
- chardet=4.0.0=py38h06a4308_1003
- cryptography=3.4.7=py38hd23ed53_0
- cudatoolkit=11.1.74=h6bb024c_0
- ffmpeg=4.2.2=h20bf706_0
- freetype=2.10.4=h5ab3b9f_0
- gmp=6.2.1=h2531618_2
- gnutls=3.6.15=he1e5248_0
- idna=2.10=pyhd3eb1b0_0
- intel-openmp=2021.2.0=h06a4308_610
- jpeg=9b=h024ee3a_2
- lame=3.100=h7b6447c_0
- lcms2=2.12=h3be6417_0
- ld_impl_linux-64=2.35.1=h7274673_9
- libffi=3.3=he6710b0_2
- libgcc-ng=9.3.0=h5101ec6_17
- libgomp=9.3.0=h5101ec6_17
- libidn2=2.3.1=h27cfd23_0
- libopus=1.3.1=h7b6447c_0
- libpng=1.6.37=hbc83047_0
- libstdcxx-ng=9.3.0=hd4cf53a_17
- libtasn1=4.16.0=h27cfd23_0
- libtiff=4.2.0=h85742a9_0
- libunistring=0.9.10=h27cfd23_0
- libuv=1.40.0=h7b6447c_0
- libvpx=1.7.0=h439df22_0
- libwebp-base=1.2.0=h27cfd23_0
- lz4-c=1.9.3=h2531618_0
- mkl=2021.2.0=h06a4308_296
- mkl-service=2.3.0=py38h27cfd23_1
- mkl_fft=1.3.0=py38h42c9631_2
- mkl_random=1.2.1=py38ha9443f7_2
- ncurses=6.2=he6710b0_1
- nettle=3.7.3=hbbd107a_1
- ninja=1.10.2=hff7bd54_1
- numpy=1.20.2=py38h2d18471_0
- numpy-base=1.20.2=py38hfae3a4d_0
- olefile=0.46=py_0
- openh264=2.1.0=hd408876_0
- openssl=1.1.1k=h27cfd23_0
- pillow=8.2.0=py38he98fc37_0
- pip=21.1.2=py38h06a4308_0
- pycparser=2.20=py_2
- pyopenssl=20.0.1=pyhd3eb1b0_1
- pysocks=1.7.1=py38h06a4308_0
- python=3.8.10=h12debd9_8
- pytorch=1.8.1=py3.8_cuda11.1_cudnn8.0.5_0
- readline=8.1=h27cfd23_0
- requests=2.25.1=pyhd3eb1b0_0
- setuptools=52.0.0=py38h06a4308_0
- six=1.16.0=pyhd3eb1b0_0
- sqlite=3.35.4=hdfb4753_0
- tk=8.6.10=hbc83047_0
- torchtext=0.9.1=py38
- torchvision=0.9.1=py38_cu111
- typing_extensions=3.7.4.3=pyha847dfd_0
- urllib3=1.26.4=pyhd3eb1b0_0
- wheel=0.36.2=pyhd3eb1b0_0
- x264=1!157.20191217=h7b6447c_0
- xz=5.2.5=h7b6447c_0
- zlib=1.2.11=h7b6447c_3
- zstd=1.4.9=haebb681_0
- pip:
- click==8.0.1
- cycler==0.10.0
- datasets==1.8.0
- dill==0.3.4
- filelock==3.0.12
- fsspec==2021.6.0
- huggingface-hub==0.0.8
- joblib==1.0.1
- jsonpickle==2.0.0
- kiwisolver==1.3.1
- matplotlib==3.4.2
- multiprocess==0.70.12.2
- packaging==20.9
- pandas==1.2.4
- pyarrow==3.0.0
- pyparsing==2.4.7
- python-dateutil==2.8.1
- pytz==2021.1
- regex==2021.4.4
- sacremoses==0.0.45
- tokenizers==0.10.3
- tqdm==4.49.0
- transformers==4.6.1
- xxhash==2.0.2
prefix: /home/mmajurski/anaconda3/envs/round8
```
| 71 | sqaud_v2 dataset contains misalignment between the answer text and the context value at the answer index
## Describe the bug
The built in huggingface squad_v2 dataset that you can access via datasets.load_dataset contains mis-alignment between the answers['text'] and the characters in the context at the location specified by answers['answer_start'].
For example:
id = '56d1f453e7d4791d009025bd'
answers = {'text': ['Pure Land'], 'answer_start': [146]}
However the actual text in context at location 146 is 'ure Land,'
Which is an off-by-one error from the correct answer.
## Steps to reproduce the bug
```python
import datasets
def check_context_answer_alignment(example):
for a_idx in range(len(example['answers']['text'])):
# check raw dataset for answer consistency between context and answer
answer_text = example['answers']['text'][a_idx]
a_st_idx = example['answers']['answer_start'][a_idx]
a_end_idx = a_st_idx + len(example['answers']['text'][a_idx])
answer_text_from_context = example['context'][a_st_idx:a_end_idx]
if answer_text != answer_text_from_context:
#print(example['id'])
return False
return True
dataset = datasets.load_dataset('squad_v2', split='train', keep_in_memory=True)
start_len = len(dataset)
dataset = dataset.filter(check_context_answer_alignment,
num_proc=1,
keep_in_memory=True)
end_len = len(dataset)
print('{} instances contain mis-alignment between the answer text and answer index.'.format(start_len - end_len))
```
## Expected results
This code should result in 0 rows being filtered out from the dataset.
## Actual results
This filter command results in 258 rows being flagged as containing a discrepancy between the text contained within answers['text'] and the text in example['context'] at the answers['answer_start'] location.
This code will reproduce the problem and produce the following count:
"258 instances contain mis-alignment between the answer text and answer index."
## Environment info
Steps to rebuilt the Conda environment:
```
# create a virtual environment to stuff all these packages into
conda create -n round8 python=3.8 -y
# activate the virtual environment
conda activate round8
# install pytorch (best done through conda to handle cuda dependencies)
conda install pytorch torchvision torchtext cudatoolkit=11.1 -c pytorch-lts -c nvidia
pip install jsonpickle transformers datasets matplotlib
```
OS: Ubuntu 20.04
Python 3.8
Result of `conda env export`:
```
name: round8
channels:
- pytorch-lts
- nvidia
- defaults
dependencies:
- _libgcc_mutex=0.1=main
- _openmp_mutex=4.5=1_gnu
- blas=1.0=mkl
- brotlipy=0.7.0=py38h27cfd23_1003
- bzip2=1.0.8=h7b6447c_0
- ca-certificates=2021.5.25=h06a4308_1
- certifi=2021.5.30=py38h06a4308_0
- cffi=1.14.5=py38h261ae71_0
- chardet=4.0.0=py38h06a4308_1003
- cryptography=3.4.7=py38hd23ed53_0
- cudatoolkit=11.1.74=h6bb024c_0
- ffmpeg=4.2.2=h20bf706_0
- freetype=2.10.4=h5ab3b9f_0
- gmp=6.2.1=h2531618_2
- gnutls=3.6.15=he1e5248_0
- idna=2.10=pyhd3eb1b0_0
- intel-openmp=2021.2.0=h06a4308_610
- jpeg=9b=h024ee3a_2
- lame=3.100=h7b6447c_0
- lcms2=2.12=h3be6417_0
- ld_impl_linux-64=2.35.1=h7274673_9
- libffi=3.3=he6710b0_2
- libgcc-ng=9.3.0=h5101ec6_17
- libgomp=9.3.0=h5101ec6_17
- libidn2=2.3.1=h27cfd23_0
- libopus=1.3.1=h7b6447c_0
- libpng=1.6.37=hbc83047_0
- libstdcxx-ng=9.3.0=hd4cf53a_17
- libtasn1=4.16.0=h27cfd23_0
- libtiff=4.2.0=h85742a9_0
- libunistring=0.9.10=h27cfd23_0
- libuv=1.40.0=h7b6447c_0
- libvpx=1.7.0=h439df22_0
- libwebp-base=1.2.0=h27cfd23_0
- lz4-c=1.9.3=h2531618_0
- mkl=2021.2.0=h06a4308_296
- mkl-service=2.3.0=py38h27cfd23_1
- mkl_fft=1.3.0=py38h42c9631_2
- mkl_random=1.2.1=py38ha9443f7_2
- ncurses=6.2=he6710b0_1
- nettle=3.7.3=hbbd107a_1
- ninja=1.10.2=hff7bd54_1
- numpy=1.20.2=py38h2d18471_0
- numpy-base=1.20.2=py38hfae3a4d_0
- olefile=0.46=py_0
- openh264=2.1.0=hd408876_0
- openssl=1.1.1k=h27cfd23_0
- pillow=8.2.0=py38he98fc37_0
- pip=21.1.2=py38h06a4308_0
- pycparser=2.20=py_2
- pyopenssl=20.0.1=pyhd3eb1b0_1
- pysocks=1.7.1=py38h06a4308_0
- python=3.8.10=h12debd9_8
- pytorch=1.8.1=py3.8_cuda11.1_cudnn8.0.5_0
- readline=8.1=h27cfd23_0
- requests=2.25.1=pyhd3eb1b0_0
- setuptools=52.0.0=py38h06a4308_0
- six=1.16.0=pyhd3eb1b0_0
- sqlite=3.35.4=hdfb4753_0
- tk=8.6.10=hbc83047_0
- torchtext=0.9.1=py38
- torchvision=0.9.1=py38_cu111
- typing_extensions=3.7.4.3=pyha847dfd_0
- urllib3=1.26.4=pyhd3eb1b0_0
- wheel=0.36.2=pyhd3eb1b0_0
- x264=1!157.20191217=h7b6447c_0
- xz=5.2.5=h7b6447c_0
- zlib=1.2.11=h7b6447c_3
- zstd=1.4.9=haebb681_0
- pip:
- click==8.0.1
- cycler==0.10.0
- datasets==1.8.0
- dill==0.3.4
- filelock==3.0.12
- fsspec==2021.6.0
- huggingface-hub==0.0.8
- joblib==1.0.1
- jsonpickle==2.0.0
- kiwisolver==1.3.1
- matplotlib==3.4.2
- multiprocess==0.70.12.2
- packaging==20.9
- pandas==1.2.4
- pyarrow==3.0.0
- pyparsing==2.4.7
- python-dateutil==2.8.1
- pytz==2021.1
- regex==2021.4.4
- sacremoses==0.0.45
- tokenizers==0.10.3
- tqdm==4.49.0
- transformers==4.6.1
- xxhash==2.0.2
prefix: /home/mmajurski/anaconda3/envs/round8
```
Hi @mmajurski, thanks for reporting this issue.
Indeed this misalignment arises because the source dataset context field contains leading blank spaces (and these are counted within the answer_start), while our datasets loading script removes these leading blank spaces.
I'm going to fix our script so that all leading blank spaces in the source dataset are kept, and there is no misalignment between the answer text and the answer_start within the context. |
https://github.com/huggingface/datasets/issues/2585 | sqaud_v2 dataset contains misalignment between the answer text and the context value at the answer index | If you are going to be altering the data cleaning from the source Squad dataset, here is one thing to consider.
There are occasional double spaces separating words which it might be nice to get rid of.
Either way, thank you. | ## Describe the bug
The built in huggingface squad_v2 dataset that you can access via datasets.load_dataset contains mis-alignment between the answers['text'] and the characters in the context at the location specified by answers['answer_start'].
For example:
id = '56d1f453e7d4791d009025bd'
answers = {'text': ['Pure Land'], 'answer_start': [146]}
However the actual text in context at location 146 is 'ure Land,'
Which is an off-by-one error from the correct answer.
## Steps to reproduce the bug
```python
import datasets
def check_context_answer_alignment(example):
for a_idx in range(len(example['answers']['text'])):
# check raw dataset for answer consistency between context and answer
answer_text = example['answers']['text'][a_idx]
a_st_idx = example['answers']['answer_start'][a_idx]
a_end_idx = a_st_idx + len(example['answers']['text'][a_idx])
answer_text_from_context = example['context'][a_st_idx:a_end_idx]
if answer_text != answer_text_from_context:
#print(example['id'])
return False
return True
dataset = datasets.load_dataset('squad_v2', split='train', keep_in_memory=True)
start_len = len(dataset)
dataset = dataset.filter(check_context_answer_alignment,
num_proc=1,
keep_in_memory=True)
end_len = len(dataset)
print('{} instances contain mis-alignment between the answer text and answer index.'.format(start_len - end_len))
```
## Expected results
This code should result in 0 rows being filtered out from the dataset.
## Actual results
This filter command results in 258 rows being flagged as containing a discrepancy between the text contained within answers['text'] and the text in example['context'] at the answers['answer_start'] location.
This code will reproduce the problem and produce the following count:
"258 instances contain mis-alignment between the answer text and answer index."
## Environment info
Steps to rebuilt the Conda environment:
```
# create a virtual environment to stuff all these packages into
conda create -n round8 python=3.8 -y
# activate the virtual environment
conda activate round8
# install pytorch (best done through conda to handle cuda dependencies)
conda install pytorch torchvision torchtext cudatoolkit=11.1 -c pytorch-lts -c nvidia
pip install jsonpickle transformers datasets matplotlib
```
OS: Ubuntu 20.04
Python 3.8
Result of `conda env export`:
```
name: round8
channels:
- pytorch-lts
- nvidia
- defaults
dependencies:
- _libgcc_mutex=0.1=main
- _openmp_mutex=4.5=1_gnu
- blas=1.0=mkl
- brotlipy=0.7.0=py38h27cfd23_1003
- bzip2=1.0.8=h7b6447c_0
- ca-certificates=2021.5.25=h06a4308_1
- certifi=2021.5.30=py38h06a4308_0
- cffi=1.14.5=py38h261ae71_0
- chardet=4.0.0=py38h06a4308_1003
- cryptography=3.4.7=py38hd23ed53_0
- cudatoolkit=11.1.74=h6bb024c_0
- ffmpeg=4.2.2=h20bf706_0
- freetype=2.10.4=h5ab3b9f_0
- gmp=6.2.1=h2531618_2
- gnutls=3.6.15=he1e5248_0
- idna=2.10=pyhd3eb1b0_0
- intel-openmp=2021.2.0=h06a4308_610
- jpeg=9b=h024ee3a_2
- lame=3.100=h7b6447c_0
- lcms2=2.12=h3be6417_0
- ld_impl_linux-64=2.35.1=h7274673_9
- libffi=3.3=he6710b0_2
- libgcc-ng=9.3.0=h5101ec6_17
- libgomp=9.3.0=h5101ec6_17
- libidn2=2.3.1=h27cfd23_0
- libopus=1.3.1=h7b6447c_0
- libpng=1.6.37=hbc83047_0
- libstdcxx-ng=9.3.0=hd4cf53a_17
- libtasn1=4.16.0=h27cfd23_0
- libtiff=4.2.0=h85742a9_0
- libunistring=0.9.10=h27cfd23_0
- libuv=1.40.0=h7b6447c_0
- libvpx=1.7.0=h439df22_0
- libwebp-base=1.2.0=h27cfd23_0
- lz4-c=1.9.3=h2531618_0
- mkl=2021.2.0=h06a4308_296
- mkl-service=2.3.0=py38h27cfd23_1
- mkl_fft=1.3.0=py38h42c9631_2
- mkl_random=1.2.1=py38ha9443f7_2
- ncurses=6.2=he6710b0_1
- nettle=3.7.3=hbbd107a_1
- ninja=1.10.2=hff7bd54_1
- numpy=1.20.2=py38h2d18471_0
- numpy-base=1.20.2=py38hfae3a4d_0
- olefile=0.46=py_0
- openh264=2.1.0=hd408876_0
- openssl=1.1.1k=h27cfd23_0
- pillow=8.2.0=py38he98fc37_0
- pip=21.1.2=py38h06a4308_0
- pycparser=2.20=py_2
- pyopenssl=20.0.1=pyhd3eb1b0_1
- pysocks=1.7.1=py38h06a4308_0
- python=3.8.10=h12debd9_8
- pytorch=1.8.1=py3.8_cuda11.1_cudnn8.0.5_0
- readline=8.1=h27cfd23_0
- requests=2.25.1=pyhd3eb1b0_0
- setuptools=52.0.0=py38h06a4308_0
- six=1.16.0=pyhd3eb1b0_0
- sqlite=3.35.4=hdfb4753_0
- tk=8.6.10=hbc83047_0
- torchtext=0.9.1=py38
- torchvision=0.9.1=py38_cu111
- typing_extensions=3.7.4.3=pyha847dfd_0
- urllib3=1.26.4=pyhd3eb1b0_0
- wheel=0.36.2=pyhd3eb1b0_0
- x264=1!157.20191217=h7b6447c_0
- xz=5.2.5=h7b6447c_0
- zlib=1.2.11=h7b6447c_3
- zstd=1.4.9=haebb681_0
- pip:
- click==8.0.1
- cycler==0.10.0
- datasets==1.8.0
- dill==0.3.4
- filelock==3.0.12
- fsspec==2021.6.0
- huggingface-hub==0.0.8
- joblib==1.0.1
- jsonpickle==2.0.0
- kiwisolver==1.3.1
- matplotlib==3.4.2
- multiprocess==0.70.12.2
- packaging==20.9
- pandas==1.2.4
- pyarrow==3.0.0
- pyparsing==2.4.7
- python-dateutil==2.8.1
- pytz==2021.1
- regex==2021.4.4
- sacremoses==0.0.45
- tokenizers==0.10.3
- tqdm==4.49.0
- transformers==4.6.1
- xxhash==2.0.2
prefix: /home/mmajurski/anaconda3/envs/round8
```
| 41 | sqaud_v2 dataset contains misalignment between the answer text and the context value at the answer index
## Describe the bug
The built in huggingface squad_v2 dataset that you can access via datasets.load_dataset contains mis-alignment between the answers['text'] and the characters in the context at the location specified by answers['answer_start'].
For example:
id = '56d1f453e7d4791d009025bd'
answers = {'text': ['Pure Land'], 'answer_start': [146]}
However the actual text in context at location 146 is 'ure Land,'
Which is an off-by-one error from the correct answer.
## Steps to reproduce the bug
```python
import datasets
def check_context_answer_alignment(example):
for a_idx in range(len(example['answers']['text'])):
# check raw dataset for answer consistency between context and answer
answer_text = example['answers']['text'][a_idx]
a_st_idx = example['answers']['answer_start'][a_idx]
a_end_idx = a_st_idx + len(example['answers']['text'][a_idx])
answer_text_from_context = example['context'][a_st_idx:a_end_idx]
if answer_text != answer_text_from_context:
#print(example['id'])
return False
return True
dataset = datasets.load_dataset('squad_v2', split='train', keep_in_memory=True)
start_len = len(dataset)
dataset = dataset.filter(check_context_answer_alignment,
num_proc=1,
keep_in_memory=True)
end_len = len(dataset)
print('{} instances contain mis-alignment between the answer text and answer index.'.format(start_len - end_len))
```
## Expected results
This code should result in 0 rows being filtered out from the dataset.
## Actual results
This filter command results in 258 rows being flagged as containing a discrepancy between the text contained within answers['text'] and the text in example['context'] at the answers['answer_start'] location.
This code will reproduce the problem and produce the following count:
"258 instances contain mis-alignment between the answer text and answer index."
## Environment info
Steps to rebuilt the Conda environment:
```
# create a virtual environment to stuff all these packages into
conda create -n round8 python=3.8 -y
# activate the virtual environment
conda activate round8
# install pytorch (best done through conda to handle cuda dependencies)
conda install pytorch torchvision torchtext cudatoolkit=11.1 -c pytorch-lts -c nvidia
pip install jsonpickle transformers datasets matplotlib
```
OS: Ubuntu 20.04
Python 3.8
Result of `conda env export`:
```
name: round8
channels:
- pytorch-lts
- nvidia
- defaults
dependencies:
- _libgcc_mutex=0.1=main
- _openmp_mutex=4.5=1_gnu
- blas=1.0=mkl
- brotlipy=0.7.0=py38h27cfd23_1003
- bzip2=1.0.8=h7b6447c_0
- ca-certificates=2021.5.25=h06a4308_1
- certifi=2021.5.30=py38h06a4308_0
- cffi=1.14.5=py38h261ae71_0
- chardet=4.0.0=py38h06a4308_1003
- cryptography=3.4.7=py38hd23ed53_0
- cudatoolkit=11.1.74=h6bb024c_0
- ffmpeg=4.2.2=h20bf706_0
- freetype=2.10.4=h5ab3b9f_0
- gmp=6.2.1=h2531618_2
- gnutls=3.6.15=he1e5248_0
- idna=2.10=pyhd3eb1b0_0
- intel-openmp=2021.2.0=h06a4308_610
- jpeg=9b=h024ee3a_2
- lame=3.100=h7b6447c_0
- lcms2=2.12=h3be6417_0
- ld_impl_linux-64=2.35.1=h7274673_9
- libffi=3.3=he6710b0_2
- libgcc-ng=9.3.0=h5101ec6_17
- libgomp=9.3.0=h5101ec6_17
- libidn2=2.3.1=h27cfd23_0
- libopus=1.3.1=h7b6447c_0
- libpng=1.6.37=hbc83047_0
- libstdcxx-ng=9.3.0=hd4cf53a_17
- libtasn1=4.16.0=h27cfd23_0
- libtiff=4.2.0=h85742a9_0
- libunistring=0.9.10=h27cfd23_0
- libuv=1.40.0=h7b6447c_0
- libvpx=1.7.0=h439df22_0
- libwebp-base=1.2.0=h27cfd23_0
- lz4-c=1.9.3=h2531618_0
- mkl=2021.2.0=h06a4308_296
- mkl-service=2.3.0=py38h27cfd23_1
- mkl_fft=1.3.0=py38h42c9631_2
- mkl_random=1.2.1=py38ha9443f7_2
- ncurses=6.2=he6710b0_1
- nettle=3.7.3=hbbd107a_1
- ninja=1.10.2=hff7bd54_1
- numpy=1.20.2=py38h2d18471_0
- numpy-base=1.20.2=py38hfae3a4d_0
- olefile=0.46=py_0
- openh264=2.1.0=hd408876_0
- openssl=1.1.1k=h27cfd23_0
- pillow=8.2.0=py38he98fc37_0
- pip=21.1.2=py38h06a4308_0
- pycparser=2.20=py_2
- pyopenssl=20.0.1=pyhd3eb1b0_1
- pysocks=1.7.1=py38h06a4308_0
- python=3.8.10=h12debd9_8
- pytorch=1.8.1=py3.8_cuda11.1_cudnn8.0.5_0
- readline=8.1=h27cfd23_0
- requests=2.25.1=pyhd3eb1b0_0
- setuptools=52.0.0=py38h06a4308_0
- six=1.16.0=pyhd3eb1b0_0
- sqlite=3.35.4=hdfb4753_0
- tk=8.6.10=hbc83047_0
- torchtext=0.9.1=py38
- torchvision=0.9.1=py38_cu111
- typing_extensions=3.7.4.3=pyha847dfd_0
- urllib3=1.26.4=pyhd3eb1b0_0
- wheel=0.36.2=pyhd3eb1b0_0
- x264=1!157.20191217=h7b6447c_0
- xz=5.2.5=h7b6447c_0
- zlib=1.2.11=h7b6447c_3
- zstd=1.4.9=haebb681_0
- pip:
- click==8.0.1
- cycler==0.10.0
- datasets==1.8.0
- dill==0.3.4
- filelock==3.0.12
- fsspec==2021.6.0
- huggingface-hub==0.0.8
- joblib==1.0.1
- jsonpickle==2.0.0
- kiwisolver==1.3.1
- matplotlib==3.4.2
- multiprocess==0.70.12.2
- packaging==20.9
- pandas==1.2.4
- pyarrow==3.0.0
- pyparsing==2.4.7
- python-dateutil==2.8.1
- pytz==2021.1
- regex==2021.4.4
- sacremoses==0.0.45
- tokenizers==0.10.3
- tqdm==4.49.0
- transformers==4.6.1
- xxhash==2.0.2
prefix: /home/mmajurski/anaconda3/envs/round8
```
If you are going to be altering the data cleaning from the source Squad dataset, here is one thing to consider.
There are occasional double spaces separating words which it might be nice to get rid of.
Either way, thank you. |
https://github.com/huggingface/datasets/issues/2583 | Error iteration over IterableDataset using Torch DataLoader | Hi ! This is because you first need to format the dataset for pytorch:
```python
>>> import torch
>>> from datasets import load_dataset
>>> dataset = load_dataset('oscar', "unshuffled_deduplicated_en", split='train', streaming=True)
>>> torch_iterable_dataset = dataset.with_format("torch")
>>> assert isinstance(torch_iterable_dataset, torch.utils.data.IterableDataset)
>>> dataloader = torch.utils.data.DataLoader(torch_iterable_dataset, batch_size=4)
>>> next(iter(dataloader))
{'id': tensor([0, 1, 2, 3]), 'text': ['Mtendere Village was inspired...]}
```
This is because the pytorch dataloader expects a subclass of `torch.utils.data.IterableDataset`. Since you can't pass an arbitrary iterable to a pytorch dataloader, you first need to build an object that inherits from `torch.utils.data.IterableDataset` using `with_format("torch")` for example.
| ## Describe the bug
I have an IterableDataset (created using streaming=True) and I am trying to create batches using Torch DataLoader class by passing this IterableDataset to it. This throws error which is pasted below. I can do the same by using Torch IterableDataset. One thing I noticed is that in the former case when I look at the dataloader.sampler class I get torch.utils.data.sampler.SequentialSampler while the latter one gives torch.utils.data.dataloader._InfiniteConstantSampler.
I am not sure if this is how it is meant to be used, but that's what seemed reasonable to me.
## Steps to reproduce the bug
1. Does not work.
```python
>>> from datasets import load_dataset
>>> dataset = load_dataset('oscar', "unshuffled_deduplicated_en", split='train', streaming=True)
>>> dataloader = torch.utils.data.DataLoader(dataset, batch_size=4)
>>> dataloader.sampler
<torch.utils.data.sampler.SequentialSampler object at 0x7f245a510208>
>>> for batch in dataloader:
... print(batch)
```
2. Works.
```python
import torch
from torch.utils.data import Dataset, IterableDataset, DataLoader
class CustomIterableDataset(IterableDataset):
'Characterizes a dataset for PyTorch'
def __init__(self, data):
'Initialization'
self.data = data
def __iter__(self):
return iter(self.data)
data = list(range(12))
dataset = CustomIterableDataset(data)
dataloader = DataLoader(dataset, batch_size=4)
print("dataloader: ", dataloader.sampler)
for batch in dataloader:
print(batch)
```
## Expected results
To get batches of data with the batch size as 4. Output from the latter one (2) though Datasource is different here so actual data is different.
dataloader: <torch.utils.data.dataloader._InfiniteConstantSampler object at 0x7f1cc29e2c50>
tensor([0, 1, 2, 3])
tensor([4, 5, 6, 7])
tensor([ 8, 9, 10, 11])
## Actual results
<torch.utils.data.sampler.SequentialSampler object at 0x7f245a510208>
...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 435, in __next__
data = self._next_data()
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 474, in _next_data
index = self._next_index() # may raise StopIteration
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 427, in _next_index
return next(self._sampler_iter) # may raise StopIteration
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 227, in __iter__
for idx in self.sampler:
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 67, in __iter__
return iter(range(len(self.data_source)))
TypeError: object of type 'IterableDataset' has no len()
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: '1.8.1.dev0'
- Platform: Linux
- Python version: Python 3.6.8
- PyArrow version: '3.0.0'
| 93 | Error iteration over IterableDataset using Torch DataLoader
## Describe the bug
I have an IterableDataset (created using streaming=True) and I am trying to create batches using Torch DataLoader class by passing this IterableDataset to it. This throws error which is pasted below. I can do the same by using Torch IterableDataset. One thing I noticed is that in the former case when I look at the dataloader.sampler class I get torch.utils.data.sampler.SequentialSampler while the latter one gives torch.utils.data.dataloader._InfiniteConstantSampler.
I am not sure if this is how it is meant to be used, but that's what seemed reasonable to me.
## Steps to reproduce the bug
1. Does not work.
```python
>>> from datasets import load_dataset
>>> dataset = load_dataset('oscar', "unshuffled_deduplicated_en", split='train', streaming=True)
>>> dataloader = torch.utils.data.DataLoader(dataset, batch_size=4)
>>> dataloader.sampler
<torch.utils.data.sampler.SequentialSampler object at 0x7f245a510208>
>>> for batch in dataloader:
... print(batch)
```
2. Works.
```python
import torch
from torch.utils.data import Dataset, IterableDataset, DataLoader
class CustomIterableDataset(IterableDataset):
'Characterizes a dataset for PyTorch'
def __init__(self, data):
'Initialization'
self.data = data
def __iter__(self):
return iter(self.data)
data = list(range(12))
dataset = CustomIterableDataset(data)
dataloader = DataLoader(dataset, batch_size=4)
print("dataloader: ", dataloader.sampler)
for batch in dataloader:
print(batch)
```
## Expected results
To get batches of data with the batch size as 4. Output from the latter one (2) though Datasource is different here so actual data is different.
dataloader: <torch.utils.data.dataloader._InfiniteConstantSampler object at 0x7f1cc29e2c50>
tensor([0, 1, 2, 3])
tensor([4, 5, 6, 7])
tensor([ 8, 9, 10, 11])
## Actual results
<torch.utils.data.sampler.SequentialSampler object at 0x7f245a510208>
...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 435, in __next__
data = self._next_data()
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 474, in _next_data
index = self._next_index() # may raise StopIteration
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 427, in _next_index
return next(self._sampler_iter) # may raise StopIteration
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 227, in __iter__
for idx in self.sampler:
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 67, in __iter__
return iter(range(len(self.data_source)))
TypeError: object of type 'IterableDataset' has no len()
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: '1.8.1.dev0'
- Platform: Linux
- Python version: Python 3.6.8
- PyArrow version: '3.0.0'
Hi ! This is because you first need to format the dataset for pytorch:
```python
>>> import torch
>>> from datasets import load_dataset
>>> dataset = load_dataset('oscar', "unshuffled_deduplicated_en", split='train', streaming=True)
>>> torch_iterable_dataset = dataset.with_format("torch")
>>> assert isinstance(torch_iterable_dataset, torch.utils.data.IterableDataset)
>>> dataloader = torch.utils.data.DataLoader(torch_iterable_dataset, batch_size=4)
>>> next(iter(dataloader))
{'id': tensor([0, 1, 2, 3]), 'text': ['Mtendere Village was inspired...]}
```
This is because the pytorch dataloader expects a subclass of `torch.utils.data.IterableDataset`. Since you can't pass an arbitrary iterable to a pytorch dataloader, you first need to build an object that inherits from `torch.utils.data.IterableDataset` using `with_format("torch")` for example.
|
https://github.com/huggingface/datasets/issues/2583 | Error iteration over IterableDataset using Torch DataLoader | Thank you for that and the example!
What you said makes total sense; I just somehow missed that and assumed HF IterableDataset was a subclass of Torch IterableDataset. | ## Describe the bug
I have an IterableDataset (created using streaming=True) and I am trying to create batches using Torch DataLoader class by passing this IterableDataset to it. This throws error which is pasted below. I can do the same by using Torch IterableDataset. One thing I noticed is that in the former case when I look at the dataloader.sampler class I get torch.utils.data.sampler.SequentialSampler while the latter one gives torch.utils.data.dataloader._InfiniteConstantSampler.
I am not sure if this is how it is meant to be used, but that's what seemed reasonable to me.
## Steps to reproduce the bug
1. Does not work.
```python
>>> from datasets import load_dataset
>>> dataset = load_dataset('oscar', "unshuffled_deduplicated_en", split='train', streaming=True)
>>> dataloader = torch.utils.data.DataLoader(dataset, batch_size=4)
>>> dataloader.sampler
<torch.utils.data.sampler.SequentialSampler object at 0x7f245a510208>
>>> for batch in dataloader:
... print(batch)
```
2. Works.
```python
import torch
from torch.utils.data import Dataset, IterableDataset, DataLoader
class CustomIterableDataset(IterableDataset):
'Characterizes a dataset for PyTorch'
def __init__(self, data):
'Initialization'
self.data = data
def __iter__(self):
return iter(self.data)
data = list(range(12))
dataset = CustomIterableDataset(data)
dataloader = DataLoader(dataset, batch_size=4)
print("dataloader: ", dataloader.sampler)
for batch in dataloader:
print(batch)
```
## Expected results
To get batches of data with the batch size as 4. Output from the latter one (2) though Datasource is different here so actual data is different.
dataloader: <torch.utils.data.dataloader._InfiniteConstantSampler object at 0x7f1cc29e2c50>
tensor([0, 1, 2, 3])
tensor([4, 5, 6, 7])
tensor([ 8, 9, 10, 11])
## Actual results
<torch.utils.data.sampler.SequentialSampler object at 0x7f245a510208>
...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 435, in __next__
data = self._next_data()
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 474, in _next_data
index = self._next_index() # may raise StopIteration
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 427, in _next_index
return next(self._sampler_iter) # may raise StopIteration
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 227, in __iter__
for idx in self.sampler:
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 67, in __iter__
return iter(range(len(self.data_source)))
TypeError: object of type 'IterableDataset' has no len()
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: '1.8.1.dev0'
- Platform: Linux
- Python version: Python 3.6.8
- PyArrow version: '3.0.0'
| 28 | Error iteration over IterableDataset using Torch DataLoader
## Describe the bug
I have an IterableDataset (created using streaming=True) and I am trying to create batches using Torch DataLoader class by passing this IterableDataset to it. This throws error which is pasted below. I can do the same by using Torch IterableDataset. One thing I noticed is that in the former case when I look at the dataloader.sampler class I get torch.utils.data.sampler.SequentialSampler while the latter one gives torch.utils.data.dataloader._InfiniteConstantSampler.
I am not sure if this is how it is meant to be used, but that's what seemed reasonable to me.
## Steps to reproduce the bug
1. Does not work.
```python
>>> from datasets import load_dataset
>>> dataset = load_dataset('oscar', "unshuffled_deduplicated_en", split='train', streaming=True)
>>> dataloader = torch.utils.data.DataLoader(dataset, batch_size=4)
>>> dataloader.sampler
<torch.utils.data.sampler.SequentialSampler object at 0x7f245a510208>
>>> for batch in dataloader:
... print(batch)
```
2. Works.
```python
import torch
from torch.utils.data import Dataset, IterableDataset, DataLoader
class CustomIterableDataset(IterableDataset):
'Characterizes a dataset for PyTorch'
def __init__(self, data):
'Initialization'
self.data = data
def __iter__(self):
return iter(self.data)
data = list(range(12))
dataset = CustomIterableDataset(data)
dataloader = DataLoader(dataset, batch_size=4)
print("dataloader: ", dataloader.sampler)
for batch in dataloader:
print(batch)
```
## Expected results
To get batches of data with the batch size as 4. Output from the latter one (2) though Datasource is different here so actual data is different.
dataloader: <torch.utils.data.dataloader._InfiniteConstantSampler object at 0x7f1cc29e2c50>
tensor([0, 1, 2, 3])
tensor([4, 5, 6, 7])
tensor([ 8, 9, 10, 11])
## Actual results
<torch.utils.data.sampler.SequentialSampler object at 0x7f245a510208>
...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 435, in __next__
data = self._next_data()
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 474, in _next_data
index = self._next_index() # may raise StopIteration
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 427, in _next_index
return next(self._sampler_iter) # may raise StopIteration
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 227, in __iter__
for idx in self.sampler:
File "/data/leshekha/lib/HFDatasets/lib/python3.6/site-packages/torch/utils/data/sampler.py", line 67, in __iter__
return iter(range(len(self.data_source)))
TypeError: object of type 'IterableDataset' has no len()
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: '1.8.1.dev0'
- Platform: Linux
- Python version: Python 3.6.8
- PyArrow version: '3.0.0'
Thank you for that and the example!
What you said makes total sense; I just somehow missed that and assumed HF IterableDataset was a subclass of Torch IterableDataset. |
https://github.com/huggingface/datasets/issues/2573 | Finding right block-size with JSON loading difficult for user | This was actually a second error arising from a too small block-size in the json reader.
Finding the right block size is difficult for the layman user | As reported by @thomwolf, while loading a JSON Lines file with "json" loading script, he gets
> json.decoder.JSONDecodeError: Extra data: line 2 column 1 (char 383)
| 27 | Finding right block-size with JSON loading difficult for user
As reported by @thomwolf, while loading a JSON Lines file with "json" loading script, he gets
> json.decoder.JSONDecodeError: Extra data: line 2 column 1 (char 383)
This was actually a second error arising from a too small block-size in the json reader.
Finding the right block size is difficult for the layman user |
https://github.com/huggingface/datasets/issues/2572 | Support Zstandard compressed files | I am trying to load a dataset using Hugging Face Datasets load_dataset method. I am getting the value error as show below. Can someone help with this? I am using Windows laptop and Google Colab notebook.
```
!pip install zstandard
from datasets import load_dataset
lds = load_dataset(
"json",
data_files="https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst",
split="train",
streaming=True,
)
WARNING:datasets.builder:Using custom data configuration default-a1d9e8eaedd958cd
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
[<ipython-input-12-5b4fdcb8e6d5>](https://localhost:8080/#) in <module>
6 )
7
----> 8 next(iter(law_dataset_streamed))
17 frames
[/usr/local/lib/python3.8/dist-packages/fsspec/core.py](https://localhost:8080/#) in get_compression(urlpath, compression)
485 compression = infer_compression(urlpath)
486 if compression is not None and compression not in compr:
--> 487 raise ValueError("Compression type %s not supported" % compression)
488 return compression
489
ValueError: Compression type zstd not supported
``` | Add support for Zstandard compressed files: https://facebook.github.io/zstd/ | 115 | Support Zstandard compressed files
Add support for Zstandard compressed files: https://facebook.github.io/zstd/
I am trying to load a dataset using Hugging Face Datasets load_dataset method. I am getting the value error as show below. Can someone help with this? I am using Windows laptop and Google Colab notebook.
```
!pip install zstandard
from datasets import load_dataset
lds = load_dataset(
"json",
data_files="https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst",
split="train",
streaming=True,
)
WARNING:datasets.builder:Using custom data configuration default-a1d9e8eaedd958cd
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
[<ipython-input-12-5b4fdcb8e6d5>](https://localhost:8080/#) in <module>
6 )
7
----> 8 next(iter(law_dataset_streamed))
17 frames
[/usr/local/lib/python3.8/dist-packages/fsspec/core.py](https://localhost:8080/#) in get_compression(urlpath, compression)
485 compression = infer_compression(urlpath)
486 if compression is not None and compression not in compr:
--> 487 raise ValueError("Compression type %s not supported" % compression)
488 return compression
489
ValueError: Compression type zstd not supported
``` |
https://github.com/huggingface/datasets/issues/2572 | Support Zstandard compressed files | I just tried on google colab and this works:
```python
!pip install zstandard
!pip install datasets
from datasets import load_dataset
lds = load_dataset(
"json",
data_files="https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst",
split="train",
streaming=True,
)
next(iter(lds))
```
Can you check that you have a correct installation of `zstandard` ? | Add support for Zstandard compressed files: https://facebook.github.io/zstd/ | 42 | Support Zstandard compressed files
Add support for Zstandard compressed files: https://facebook.github.io/zstd/
I just tried on google colab and this works:
```python
!pip install zstandard
!pip install datasets
from datasets import load_dataset
lds = load_dataset(
"json",
data_files="https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst",
split="train",
streaming=True,
)
next(iter(lds))
```
Can you check that you have a correct installation of `zstandard` ? |
https://github.com/huggingface/datasets/issues/2572 | Support Zstandard compressed files | > I just tried on google colab and this works:
>
> ```python
> !pip install zstandard
> !pip install datasets
> from datasets import load_dataset
>
> lds = load_dataset(
> "json",
> data_files="https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst",
> split="train",
> streaming=True,
> )
> next(iter(lds))
> ```
>
> Can you check that you have a correct installation of `zstandard` ?
I was downloading datasets first then was doing zstandard installation and that was causing the issue. This was highlighted by the Hugging Face staff and that helped. Now the issue is resolved. Thank you. | Add support for Zstandard compressed files: https://facebook.github.io/zstd/ | 93 | Support Zstandard compressed files
Add support for Zstandard compressed files: https://facebook.github.io/zstd/
> I just tried on google colab and this works:
>
> ```python
> !pip install zstandard
> !pip install datasets
> from datasets import load_dataset
>
> lds = load_dataset(
> "json",
> data_files="https://the-eye.eu/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst",
> split="train",
> streaming=True,
> )
> next(iter(lds))
> ```
>
> Can you check that you have a correct installation of `zstandard` ?
I was downloading datasets first then was doing zstandard installation and that was causing the issue. This was highlighted by the Hugging Face staff and that helped. Now the issue is resolved. Thank you. |
https://github.com/huggingface/datasets/issues/2569 | Weights of model checkpoint not initialized for RobertaModel for Bertscore | Hi @suzyahyah, thanks for reporting.
The message you get is indeed not an error message, but a warning coming from Hugging Face `transformers`. The complete warning message is:
```
Some weights of the model checkpoint at roberta-large were not used when initializing RobertaModel: ['lm_head.decoder.weight', 'lm_head.dense.weight', 'lm_head.dense.bias', 'lm_head.layer_norm.bias', 'lm_head.bias', 'lm_head.layer_norm.weight']
- This IS expected if you are initializing RobertaModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing RobertaModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
```
In this case, this behavior IS expected and you can safely ignore the warning message.
The reason is that you are just using RoBERTa to get the contextual embeddings of the input sentences/tokens, thus leaving away its head layer, whose weights are ignored.
Feel free to reopen this issue if you need further explanations. | When applying bertscore out of the box,
```Some weights of the model checkpoint at roberta-large were not used when initializing RobertaModel: ['lm_head.decoder.weight', 'lm_head.bias', 'lm_head.dense.bias', 'lm_head.layer_norm.bias', 'lm_head.dense.weight', 'lm_head.layer_norm.weight']```
Following the typical usage from https://huggingface.co/docs/datasets/loading_metrics.html
```
from datasets import load_metric
metric = load_metric('bertscore')
# Example of typical usage
for batch in dataset:
inputs, references = batch
predictions = model(inputs)
metric.add_batch(predictions=predictions, references=references)
score = metric.compute(lang="en")
#score = metric.compute(model_type="roberta-large") # gives the same error
```
I am concerned about this because my usage shouldn't require any further fine-tuning and most people would expect to use BertScore out of the box? I realised the huggingface code is a wrapper around https://github.com/Tiiiger/bert_score, but I think this repo is anyway relying on the model code and weights from huggingface repo....
## Environment info
- `datasets` version: 1.7.0
- Platform: Linux-5.4.0-1041-aws-x86_64-with-glibc2.27
- Python version: 3.9.5
- PyArrow version: 3.0.0
| 167 | Weights of model checkpoint not initialized for RobertaModel for Bertscore
When applying bertscore out of the box,
```Some weights of the model checkpoint at roberta-large were not used when initializing RobertaModel: ['lm_head.decoder.weight', 'lm_head.bias', 'lm_head.dense.bias', 'lm_head.layer_norm.bias', 'lm_head.dense.weight', 'lm_head.layer_norm.weight']```
Following the typical usage from https://huggingface.co/docs/datasets/loading_metrics.html
```
from datasets import load_metric
metric = load_metric('bertscore')
# Example of typical usage
for batch in dataset:
inputs, references = batch
predictions = model(inputs)
metric.add_batch(predictions=predictions, references=references)
score = metric.compute(lang="en")
#score = metric.compute(model_type="roberta-large") # gives the same error
```
I am concerned about this because my usage shouldn't require any further fine-tuning and most people would expect to use BertScore out of the box? I realised the huggingface code is a wrapper around https://github.com/Tiiiger/bert_score, but I think this repo is anyway relying on the model code and weights from huggingface repo....
## Environment info
- `datasets` version: 1.7.0
- Platform: Linux-5.4.0-1041-aws-x86_64-with-glibc2.27
- Python version: 3.9.5
- PyArrow version: 3.0.0
Hi @suzyahyah, thanks for reporting.
The message you get is indeed not an error message, but a warning coming from Hugging Face `transformers`. The complete warning message is:
```
Some weights of the model checkpoint at roberta-large were not used when initializing RobertaModel: ['lm_head.decoder.weight', 'lm_head.dense.weight', 'lm_head.dense.bias', 'lm_head.layer_norm.bias', 'lm_head.bias', 'lm_head.layer_norm.weight']
- This IS expected if you are initializing RobertaModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).
- This IS NOT expected if you are initializing RobertaModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).
```
In this case, this behavior IS expected and you can safely ignore the warning message.
The reason is that you are just using RoBERTa to get the contextual embeddings of the input sentences/tokens, thus leaving away its head layer, whose weights are ignored.
Feel free to reopen this issue if you need further explanations. |
https://github.com/huggingface/datasets/issues/2569 | Weights of model checkpoint not initialized for RobertaModel for Bertscore | Hi @suzyahyah, I have created a Pull Request to filter out that warning message in this specific case, since the behavior is as expected and the warning message can only cause confusion for users (as in your case). | When applying bertscore out of the box,
```Some weights of the model checkpoint at roberta-large were not used when initializing RobertaModel: ['lm_head.decoder.weight', 'lm_head.bias', 'lm_head.dense.bias', 'lm_head.layer_norm.bias', 'lm_head.dense.weight', 'lm_head.layer_norm.weight']```
Following the typical usage from https://huggingface.co/docs/datasets/loading_metrics.html
```
from datasets import load_metric
metric = load_metric('bertscore')
# Example of typical usage
for batch in dataset:
inputs, references = batch
predictions = model(inputs)
metric.add_batch(predictions=predictions, references=references)
score = metric.compute(lang="en")
#score = metric.compute(model_type="roberta-large") # gives the same error
```
I am concerned about this because my usage shouldn't require any further fine-tuning and most people would expect to use BertScore out of the box? I realised the huggingface code is a wrapper around https://github.com/Tiiiger/bert_score, but I think this repo is anyway relying on the model code and weights from huggingface repo....
## Environment info
- `datasets` version: 1.7.0
- Platform: Linux-5.4.0-1041-aws-x86_64-with-glibc2.27
- Python version: 3.9.5
- PyArrow version: 3.0.0
| 38 | Weights of model checkpoint not initialized for RobertaModel for Bertscore
When applying bertscore out of the box,
```Some weights of the model checkpoint at roberta-large were not used when initializing RobertaModel: ['lm_head.decoder.weight', 'lm_head.bias', 'lm_head.dense.bias', 'lm_head.layer_norm.bias', 'lm_head.dense.weight', 'lm_head.layer_norm.weight']```
Following the typical usage from https://huggingface.co/docs/datasets/loading_metrics.html
```
from datasets import load_metric
metric = load_metric('bertscore')
# Example of typical usage
for batch in dataset:
inputs, references = batch
predictions = model(inputs)
metric.add_batch(predictions=predictions, references=references)
score = metric.compute(lang="en")
#score = metric.compute(model_type="roberta-large") # gives the same error
```
I am concerned about this because my usage shouldn't require any further fine-tuning and most people would expect to use BertScore out of the box? I realised the huggingface code is a wrapper around https://github.com/Tiiiger/bert_score, but I think this repo is anyway relying on the model code and weights from huggingface repo....
## Environment info
- `datasets` version: 1.7.0
- Platform: Linux-5.4.0-1041-aws-x86_64-with-glibc2.27
- Python version: 3.9.5
- PyArrow version: 3.0.0
Hi @suzyahyah, I have created a Pull Request to filter out that warning message in this specific case, since the behavior is as expected and the warning message can only cause confusion for users (as in your case). |
https://github.com/huggingface/datasets/issues/2564 | concatenate_datasets for iterable datasets | It is probably worth noting here that the [documentation](https://huggingface.co/docs/datasets/process#concatenate) is misleading (indicating that it does work for IterableDatasets):
> You can also mix several datasets together by taking alternating examples from each one to create a new dataset. This is known as interleaving, and you can use it with [interleave_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.interleave_datasets). **Both [interleave_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.interleave_datasets) and [concatenate_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.concatenate_datasets) will work with regular [Dataset](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.Dataset) and [IterableDataset](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.IterableDataset) objects**. Refer to the [Stream](https://huggingface.co/docs/datasets/stream#interleave) section for an example of how it’s used. | Currently `concatenate_datasets` only works for map-style `Dataset`.
It would be nice to have it work for `IterableDataset` objects as well.
It would simply chain the iterables of the iterable datasets. | 74 | concatenate_datasets for iterable datasets
Currently `concatenate_datasets` only works for map-style `Dataset`.
It would be nice to have it work for `IterableDataset` objects as well.
It would simply chain the iterables of the iterable datasets.
It is probably worth noting here that the [documentation](https://huggingface.co/docs/datasets/process#concatenate) is misleading (indicating that it does work for IterableDatasets):
> You can also mix several datasets together by taking alternating examples from each one to create a new dataset. This is known as interleaving, and you can use it with [interleave_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.interleave_datasets). **Both [interleave_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.interleave_datasets) and [concatenate_datasets()](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.concatenate_datasets) will work with regular [Dataset](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.Dataset) and [IterableDataset](https://huggingface.co/docs/datasets/v2.2.1/en/package_reference/main_classes#datasets.IterableDataset) objects**. Refer to the [Stream](https://huggingface.co/docs/datasets/stream#interleave) section for an example of how it’s used. |
https://github.com/huggingface/datasets/issues/2561 | Existing cache for local dataset builder file updates is ignored with `ignore_verifications=True` | Hi ! I just tried to reproduce what you said:
- create a local builder class
- use `load_dataset`
- update the builder class code
- use `load_dataset` again (with or without `ignore_verifications=True`)
And it creates a new cache, as expected.
What modifications did you do to your builder's code ? | ## Describe the bug
If i have local file defining a dataset builder class and I load it using `load_dataset` functionality, the existing cache is ignored whenever the file is update even with `ignore_verifications=True`. This slows down debugging and cache generator for very large datasets.
## Steps to reproduce the bug
- Create a local dataset builder class
- load the local builder class file using `load_dataset` and let the cache build
- update the file's content
- The cache should rebuilt.
## Expected results
With `ignore_verifications=True`, `load_dataset` should pick up existing cache.
## Actual results
Creates new cache.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
| 51 | Existing cache for local dataset builder file updates is ignored with `ignore_verifications=True`
## Describe the bug
If i have local file defining a dataset builder class and I load it using `load_dataset` functionality, the existing cache is ignored whenever the file is update even with `ignore_verifications=True`. This slows down debugging and cache generator for very large datasets.
## Steps to reproduce the bug
- Create a local dataset builder class
- load the local builder class file using `load_dataset` and let the cache build
- update the file's content
- The cache should rebuilt.
## Expected results
With `ignore_verifications=True`, `load_dataset` should pick up existing cache.
## Actual results
Creates new cache.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
Hi ! I just tried to reproduce what you said:
- create a local builder class
- use `load_dataset`
- update the builder class code
- use `load_dataset` again (with or without `ignore_verifications=True`)
And it creates a new cache, as expected.
What modifications did you do to your builder's code ? |
https://github.com/huggingface/datasets/issues/2561 | Existing cache for local dataset builder file updates is ignored with `ignore_verifications=True` | Hi @lhoestq. Thanks for your reply. I just did minor modifications for which it should not regenerate cache (for e.g. Adding a print statement). Overall, regardless of cache miss, there should be an explicit option to allow reuse of existing cache if author knows cache shouldn't be affected. | ## Describe the bug
If i have local file defining a dataset builder class and I load it using `load_dataset` functionality, the existing cache is ignored whenever the file is update even with `ignore_verifications=True`. This slows down debugging and cache generator for very large datasets.
## Steps to reproduce the bug
- Create a local dataset builder class
- load the local builder class file using `load_dataset` and let the cache build
- update the file's content
- The cache should rebuilt.
## Expected results
With `ignore_verifications=True`, `load_dataset` should pick up existing cache.
## Actual results
Creates new cache.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
| 48 | Existing cache for local dataset builder file updates is ignored with `ignore_verifications=True`
## Describe the bug
If i have local file defining a dataset builder class and I load it using `load_dataset` functionality, the existing cache is ignored whenever the file is update even with `ignore_verifications=True`. This slows down debugging and cache generator for very large datasets.
## Steps to reproduce the bug
- Create a local dataset builder class
- load the local builder class file using `load_dataset` and let the cache build
- update the file's content
- The cache should rebuilt.
## Expected results
With `ignore_verifications=True`, `load_dataset` should pick up existing cache.
## Actual results
Creates new cache.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
Hi @lhoestq. Thanks for your reply. I just did minor modifications for which it should not regenerate cache (for e.g. Adding a print statement). Overall, regardless of cache miss, there should be an explicit option to allow reuse of existing cache if author knows cache shouldn't be affected. |
https://github.com/huggingface/datasets/issues/2561 | Existing cache for local dataset builder file updates is ignored with `ignore_verifications=True` | The cache is based on the hash of the dataset builder's code, so changing the code makes it recompute the cache.
You could still rename the cache directory of your previous computation to the new expected cache directory if you want to avoid having to recompute it and if you're sure that it would generate the exact same result.
The verifications are data integrity verifications: it checks the checksums of the downloaded files, as well as the size of the generated splits. | ## Describe the bug
If i have local file defining a dataset builder class and I load it using `load_dataset` functionality, the existing cache is ignored whenever the file is update even with `ignore_verifications=True`. This slows down debugging and cache generator for very large datasets.
## Steps to reproduce the bug
- Create a local dataset builder class
- load the local builder class file using `load_dataset` and let the cache build
- update the file's content
- The cache should rebuilt.
## Expected results
With `ignore_verifications=True`, `load_dataset` should pick up existing cache.
## Actual results
Creates new cache.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
| 82 | Existing cache for local dataset builder file updates is ignored with `ignore_verifications=True`
## Describe the bug
If i have local file defining a dataset builder class and I load it using `load_dataset` functionality, the existing cache is ignored whenever the file is update even with `ignore_verifications=True`. This slows down debugging and cache generator for very large datasets.
## Steps to reproduce the bug
- Create a local dataset builder class
- load the local builder class file using `load_dataset` and let the cache build
- update the file's content
- The cache should rebuilt.
## Expected results
With `ignore_verifications=True`, `load_dataset` should pick up existing cache.
## Actual results
Creates new cache.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
The cache is based on the hash of the dataset builder's code, so changing the code makes it recompute the cache.
You could still rename the cache directory of your previous computation to the new expected cache directory if you want to avoid having to recompute it and if you're sure that it would generate the exact same result.
The verifications are data integrity verifications: it checks the checksums of the downloaded files, as well as the size of the generated splits. |
https://github.com/huggingface/datasets/issues/2561 | Existing cache for local dataset builder file updates is ignored with `ignore_verifications=True` | Hi @apsdehal,
If you decide to follow @lhoestq's suggestion to rename the cache directory of your previous computation to the new expected cache directory, you can do the following to get the name of the new expected cache directory once #2500 is merged:
```python
from datasets import load_dataset_builder
dataset_builder = load_dataset_builder("path/to/your/dataset")
print(dataset_builder.cache_dir)
```
This way, you don't have to recompute the hash of the dataset script yourself each time you modify the script. | ## Describe the bug
If i have local file defining a dataset builder class and I load it using `load_dataset` functionality, the existing cache is ignored whenever the file is update even with `ignore_verifications=True`. This slows down debugging and cache generator for very large datasets.
## Steps to reproduce the bug
- Create a local dataset builder class
- load the local builder class file using `load_dataset` and let the cache build
- update the file's content
- The cache should rebuilt.
## Expected results
With `ignore_verifications=True`, `load_dataset` should pick up existing cache.
## Actual results
Creates new cache.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
| 73 | Existing cache for local dataset builder file updates is ignored with `ignore_verifications=True`
## Describe the bug
If i have local file defining a dataset builder class and I load it using `load_dataset` functionality, the existing cache is ignored whenever the file is update even with `ignore_verifications=True`. This slows down debugging and cache generator for very large datasets.
## Steps to reproduce the bug
- Create a local dataset builder class
- load the local builder class file using `load_dataset` and let the cache build
- update the file's content
- The cache should rebuilt.
## Expected results
With `ignore_verifications=True`, `load_dataset` should pick up existing cache.
## Actual results
Creates new cache.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
Hi @apsdehal,
If you decide to follow @lhoestq's suggestion to rename the cache directory of your previous computation to the new expected cache directory, you can do the following to get the name of the new expected cache directory once #2500 is merged:
```python
from datasets import load_dataset_builder
dataset_builder = load_dataset_builder("path/to/your/dataset")
print(dataset_builder.cache_dir)
```
This way, you don't have to recompute the hash of the dataset script yourself each time you modify the script. |
https://github.com/huggingface/datasets/issues/2559 | Memory usage consistently increases when processing a dataset with `.map` | Hi ! Can you share the function you pass to `map` ?
I know you mentioned it would be hard to share some code but this would really help to understand what happened | ## Describe the bug
I have a HF dataset with image paths stored in it and I am trying to load those image paths using `.map` with `num_proc=80`. I am noticing that the memory usage consistently keeps on increasing with time. I tried using `DEFAULT_WRITER_BATCH_SIZE=10` in the builder to decrease arrow writer's batch size but that doesn't seem to help.
## Steps to reproduce the bug
Providing code as it is would be hard. I can provide a MVP if that helps.
## Expected results
Memory usage should become consistent after some time following the launch of processing.
## Actual results
Memory usage keeps on increasing.
## Environment info
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0 | 33 | Memory usage consistently increases when processing a dataset with `.map`
## Describe the bug
I have a HF dataset with image paths stored in it and I am trying to load those image paths using `.map` with `num_proc=80`. I am noticing that the memory usage consistently keeps on increasing with time. I tried using `DEFAULT_WRITER_BATCH_SIZE=10` in the builder to decrease arrow writer's batch size but that doesn't seem to help.
## Steps to reproduce the bug
Providing code as it is would be hard. I can provide a MVP if that helps.
## Expected results
Memory usage should become consistent after some time following the launch of processing.
## Actual results
Memory usage keeps on increasing.
## Environment info
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
Hi ! Can you share the function you pass to `map` ?
I know you mentioned it would be hard to share some code but this would really help to understand what happened |
https://github.com/huggingface/datasets/issues/2559 | Memory usage consistently increases when processing a dataset with `.map` | This is the same behavior as in #4883, so I'm closing this issue as a duplicate. | ## Describe the bug
I have a HF dataset with image paths stored in it and I am trying to load those image paths using `.map` with `num_proc=80`. I am noticing that the memory usage consistently keeps on increasing with time. I tried using `DEFAULT_WRITER_BATCH_SIZE=10` in the builder to decrease arrow writer's batch size but that doesn't seem to help.
## Steps to reproduce the bug
Providing code as it is would be hard. I can provide a MVP if that helps.
## Expected results
Memory usage should become consistent after some time following the launch of processing.
## Actual results
Memory usage keeps on increasing.
## Environment info
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0 | 16 | Memory usage consistently increases when processing a dataset with `.map`
## Describe the bug
I have a HF dataset with image paths stored in it and I am trying to load those image paths using `.map` with `num_proc=80`. I am noticing that the memory usage consistently keeps on increasing with time. I tried using `DEFAULT_WRITER_BATCH_SIZE=10` in the builder to decrease arrow writer's batch size but that doesn't seem to help.
## Steps to reproduce the bug
Providing code as it is would be hard. I can provide a MVP if that helps.
## Expected results
Memory usage should become consistent after some time following the launch of processing.
## Actual results
Memory usage keeps on increasing.
## Environment info
- `datasets` version: 1.8.0
- Platform: Linux-5.4.0-52-generic-x86_64-with-debian-bullseye-sid
- Python version: 3.7.7
- PyArrow version: 3.0.0
This is the same behavior as in #4883, so I'm closing this issue as a duplicate. |
https://github.com/huggingface/datasets/issues/2556 | Better DuplicateKeysError error to help the user debug the issue | excuse me, my `datasets` version is `2.2.2`, but I also just see the error info like
```
DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 0
Keys should be unique and deterministic in nature
``` | As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
``` | 36 | Better DuplicateKeysError error to help the user debug the issue
As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
excuse me, my `datasets` version is `2.2.2`, but I also just see the error info like
```
DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 0
Keys should be unique and deterministic in nature
``` |
https://github.com/huggingface/datasets/issues/2556 | Better DuplicateKeysError error to help the user debug the issue | Hi ! for which dataset do you have this error ?
Also note that this issue is just about improving the error message, which is not very friendly x) | As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
``` | 29 | Better DuplicateKeysError error to help the user debug the issue
As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
Hi ! for which dataset do you have this error ?
Also note that this issue is just about improving the error message, which is not very friendly x) |
https://github.com/huggingface/datasets/issues/2556 | Better DuplicateKeysError error to help the user debug the issue | @lhoestq I would like to take a hit at improving the error message. Will open a draft PR and will reach out to you for review
| As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
``` | 26 | Better DuplicateKeysError error to help the user debug the issue
As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
@lhoestq I would like to take a hit at improving the error message. Will open a draft PR and will reach out to you for review
|
https://github.com/huggingface/datasets/issues/2556 | Better DuplicateKeysError error to help the user debug the issue | > DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
@lhoestq when you mention 42th and 1337th in the above case , are these values the examples' "id" or are they the examples' index ? | As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
``` | 37 | Better DuplicateKeysError error to help the user debug the issue
As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
> DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
@lhoestq when you mention 42th and 1337th in the above case , are these values the examples' "id" or are they the examples' index ? |
https://github.com/huggingface/datasets/issues/2556 | Better DuplicateKeysError error to help the user debug the issue | Hi ! Thanks @VijayKalmath :)
In the general case, examples don't have an "id" field, so I think it should correspond to the index | As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
``` | 24 | Better DuplicateKeysError error to help the user debug the issue
As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
Hi ! Thanks @VijayKalmath :)
In the general case, examples don't have an "id" field, so I think it should correspond to the index |
https://github.com/huggingface/datasets/issues/2556 | Better DuplicateKeysError error to help the user debug the issue | @lhoestq , I have opened a draft PR for this Issue.
I wanted to check with you if there is a way to get `<path/to/the/dataset/script>` currently or do I need to add extra code to find that.
If I need to find the script , I can assume that the generator function will always be in `datasets/{dataset_name}/{dataset_name}.py`. | As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
``` | 57 | Better DuplicateKeysError error to help the user debug the issue
As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
@lhoestq , I have opened a draft PR for this Issue.
I wanted to check with you if there is a way to get `<path/to/the/dataset/script>` currently or do I need to add extra code to find that.
If I need to find the script , I can assume that the generator function will always be in `datasets/{dataset_name}/{dataset_name}.py`. |
https://github.com/huggingface/datasets/issues/2556 | Better DuplicateKeysError error to help the user debug the issue | Thanks !
> I wanted to check with you if there is a way to get <path/to/the/dataset/script> currently or do I need to add extra code to find that.
You don't have access to this info inside the ArrowWriter unfortunately. This info is available in builder.py in the DatasetBuilder code that uses the ArrowWriter though, maybe a try-catch there can do the job | As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
``` | 63 | Better DuplicateKeysError error to help the user debug the issue
As mentioned in https://github.com/huggingface/datasets/issues/2552 it would be nice to improve the error message when a dataset fails to build because there are duplicate example keys.
The current one is
```python
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
and we could have something that guides the user to debugging the issue:
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
Thanks !
> I wanted to check with you if there is a way to get <path/to/the/dataset/script> currently or do I need to add extra code to find that.
You don't have access to this info inside the ArrowWriter unfortunately. This info is available in builder.py in the DatasetBuilder code that uses the ArrowWriter though, maybe a try-catch there can do the job |
https://github.com/huggingface/datasets/issues/2554 | Multilabel metrics not supported | Hi @GuillemGSubies, thanks for reporting.
I have made a PR to fix this issue and allow metrics to be computed also for multilabel classification problems. | When I try to use a metric like F1 macro I get the following error:
```
TypeError: int() argument must be a string, a bytes-like object or a number, not 'list'
```
There is an explicit casting here:
https://github.com/huggingface/datasets/blob/fc79f61cbbcfa0e8c68b28c0a8257f17e768a075/src/datasets/features.py#L274
And looks like this is because here
https://github.com/huggingface/datasets/blob/fc79f61cbbcfa0e8c68b28c0a8257f17e768a075/metrics/f1/f1.py#L88
the features can only be integers, so we cannot use that F1 for multilabel. Instead, if I create the following F1 (ints replaced with sequence of ints), it will work:
```python
class F1(datasets.Metric):
def _info(self):
return datasets.MetricInfo(
description=_DESCRIPTION,
citation=_CITATION,
inputs_description=_KWARGS_DESCRIPTION,
features=datasets.Features(
{
"predictions": datasets.Sequence(datasets.Value("int32")),
"references": datasets.Sequence(datasets.Value("int32")),
}
),
reference_urls=["https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html"],
)
def _compute(self, predictions, references, labels=None, pos_label=1, average="binary", sample_weight=None):
return {
"f1": f1_score(
references,
predictions,
labels=labels,
pos_label=pos_label,
average=average,
sample_weight=sample_weight,
),
}
```
| 25 | Multilabel metrics not supported
When I try to use a metric like F1 macro I get the following error:
```
TypeError: int() argument must be a string, a bytes-like object or a number, not 'list'
```
There is an explicit casting here:
https://github.com/huggingface/datasets/blob/fc79f61cbbcfa0e8c68b28c0a8257f17e768a075/src/datasets/features.py#L274
And looks like this is because here
https://github.com/huggingface/datasets/blob/fc79f61cbbcfa0e8c68b28c0a8257f17e768a075/metrics/f1/f1.py#L88
the features can only be integers, so we cannot use that F1 for multilabel. Instead, if I create the following F1 (ints replaced with sequence of ints), it will work:
```python
class F1(datasets.Metric):
def _info(self):
return datasets.MetricInfo(
description=_DESCRIPTION,
citation=_CITATION,
inputs_description=_KWARGS_DESCRIPTION,
features=datasets.Features(
{
"predictions": datasets.Sequence(datasets.Value("int32")),
"references": datasets.Sequence(datasets.Value("int32")),
}
),
reference_urls=["https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html"],
)
def _compute(self, predictions, references, labels=None, pos_label=1, average="binary", sample_weight=None):
return {
"f1": f1_score(
references,
predictions,
labels=labels,
pos_label=pos_label,
average=average,
sample_weight=sample_weight,
),
}
```
Hi @GuillemGSubies, thanks for reporting.
I have made a PR to fix this issue and allow metrics to be computed also for multilabel classification problems. |
https://github.com/huggingface/datasets/issues/2554 | Multilabel metrics not supported | Sorry for reopening but I just noticed that the `_compute` method for the F1 metric is still not good enough for multilabel problems:
https://github.com/huggingface/datasets/blob/92a3ee549705aa0a107c9fa5caf463b3b3da2616/metrics/f1/f1.py#L115
Somehow we should be able to change the parameter `average` at least | When I try to use a metric like F1 macro I get the following error:
```
TypeError: int() argument must be a string, a bytes-like object or a number, not 'list'
```
There is an explicit casting here:
https://github.com/huggingface/datasets/blob/fc79f61cbbcfa0e8c68b28c0a8257f17e768a075/src/datasets/features.py#L274
And looks like this is because here
https://github.com/huggingface/datasets/blob/fc79f61cbbcfa0e8c68b28c0a8257f17e768a075/metrics/f1/f1.py#L88
the features can only be integers, so we cannot use that F1 for multilabel. Instead, if I create the following F1 (ints replaced with sequence of ints), it will work:
```python
class F1(datasets.Metric):
def _info(self):
return datasets.MetricInfo(
description=_DESCRIPTION,
citation=_CITATION,
inputs_description=_KWARGS_DESCRIPTION,
features=datasets.Features(
{
"predictions": datasets.Sequence(datasets.Value("int32")),
"references": datasets.Sequence(datasets.Value("int32")),
}
),
reference_urls=["https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html"],
)
def _compute(self, predictions, references, labels=None, pos_label=1, average="binary", sample_weight=None):
return {
"f1": f1_score(
references,
predictions,
labels=labels,
pos_label=pos_label,
average=average,
sample_weight=sample_weight,
),
}
```
| 36 | Multilabel metrics not supported
When I try to use a metric like F1 macro I get the following error:
```
TypeError: int() argument must be a string, a bytes-like object or a number, not 'list'
```
There is an explicit casting here:
https://github.com/huggingface/datasets/blob/fc79f61cbbcfa0e8c68b28c0a8257f17e768a075/src/datasets/features.py#L274
And looks like this is because here
https://github.com/huggingface/datasets/blob/fc79f61cbbcfa0e8c68b28c0a8257f17e768a075/metrics/f1/f1.py#L88
the features can only be integers, so we cannot use that F1 for multilabel. Instead, if I create the following F1 (ints replaced with sequence of ints), it will work:
```python
class F1(datasets.Metric):
def _info(self):
return datasets.MetricInfo(
description=_DESCRIPTION,
citation=_CITATION,
inputs_description=_KWARGS_DESCRIPTION,
features=datasets.Features(
{
"predictions": datasets.Sequence(datasets.Value("int32")),
"references": datasets.Sequence(datasets.Value("int32")),
}
),
reference_urls=["https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html"],
)
def _compute(self, predictions, references, labels=None, pos_label=1, average="binary", sample_weight=None):
return {
"f1": f1_score(
references,
predictions,
labels=labels,
pos_label=pos_label,
average=average,
sample_weight=sample_weight,
),
}
```
Sorry for reopening but I just noticed that the `_compute` method for the F1 metric is still not good enough for multilabel problems:
https://github.com/huggingface/datasets/blob/92a3ee549705aa0a107c9fa5caf463b3b3da2616/metrics/f1/f1.py#L115
Somehow we should be able to change the parameter `average` at least |
https://github.com/huggingface/datasets/issues/2553 | load_dataset("web_nlg") NonMatchingChecksumError | Hi ! Thanks for reporting. This is due to the WebNLG repository that got updated today.
I just pushed a fix at #2558 - this shouldn't happen anymore in the future. | Hi! It seems the WebNLG dataset gives a NonMatchingChecksumError.
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset('web_nlg', name="release_v3.0_en", split="dev")
```
Gives
```
NonMatchingChecksumError: Checksums didn't match for dataset source files:
['https://gitlab.com/shimorina/webnlg-dataset/-/archive/master/webnlg-dataset-master.zip']
```
## Environment info
- `datasets` version: 1.8.0
- Platform: macOS-11.3.1-x86_64-i386-64bit
- Python version: 3.9.4
- PyArrow version: 3.0.0
Also tested on Linux, with python 3.6.8 | 31 | load_dataset("web_nlg") NonMatchingChecksumError
Hi! It seems the WebNLG dataset gives a NonMatchingChecksumError.
## Steps to reproduce the bug
```python
from datasets import load_dataset
dataset = load_dataset('web_nlg', name="release_v3.0_en", split="dev")
```
Gives
```
NonMatchingChecksumError: Checksums didn't match for dataset source files:
['https://gitlab.com/shimorina/webnlg-dataset/-/archive/master/webnlg-dataset-master.zip']
```
## Environment info
- `datasets` version: 1.8.0
- Platform: macOS-11.3.1-x86_64-i386-64bit
- Python version: 3.9.4
- PyArrow version: 3.0.0
Also tested on Linux, with python 3.6.8
Hi ! Thanks for reporting. This is due to the WebNLG repository that got updated today.
I just pushed a fix at #2558 - this shouldn't happen anymore in the future. |
https://github.com/huggingface/datasets/issues/2552 | Keys should be unique error on code_search_net | Two questions:
- with `datasets-cli env` we don't have any information on the dataset script version used. Should we give access to this somehow? Either as a note in the Error message or as an argument with the name of the dataset to `datasets-cli env`?
- I don't really understand why the id is duplicated in the code of `code_search_net`, how can I debug this actually? | ## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
| 66 | Keys should be unique error on code_search_net
## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
Two questions:
- with `datasets-cli env` we don't have any information on the dataset script version used. Should we give access to this somehow? Either as a note in the Error message or as an argument with the name of the dataset to `datasets-cli env`?
- I don't really understand why the id is duplicated in the code of `code_search_net`, how can I debug this actually? |
https://github.com/huggingface/datasets/issues/2552 | Keys should be unique error on code_search_net | Thanks for reporting. There was indeed an issue with the keys. The key was the addition of the file id and row id, which resulted in collisions. I just opened a PR to fix this at https://github.com/huggingface/datasets/pull/2555
To help users debug this kind of errors we could try to show a message like this
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
This way users who what to look for if they want to debug this issue. I opened an issue to track this: https://github.com/huggingface/datasets/issues/2556 | ## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
| 97 | Keys should be unique error on code_search_net
## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
Thanks for reporting. There was indeed an issue with the keys. The key was the addition of the file id and row id, which resulted in collisions. I just opened a PR to fix this at https://github.com/huggingface/datasets/pull/2555
To help users debug this kind of errors we could try to show a message like this
```python
DuplicateKeysError: both 42th and 1337th examples have the same keys `48`.
Please fix the dataset script at <path/to/the/dataset/script>
```
This way users who what to look for if they want to debug this issue. I opened an issue to track this: https://github.com/huggingface/datasets/issues/2556 |
https://github.com/huggingface/datasets/issues/2552 | Keys should be unique error on code_search_net | and are we sure there are not a lot of datasets which are now broken with this change? | ## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
| 18 | Keys should be unique error on code_search_net
## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
and are we sure there are not a lot of datasets which are now broken with this change? |
https://github.com/huggingface/datasets/issues/2552 | Keys should be unique error on code_search_net | Thanks to the dummy data, we know for sure that most of them work as expected.
`code_search_net` wasn't caught because the dummy data only have one dummy data file while the dataset script can actually load several of them using `os.listdir`. Let me take a look at all the other datasets that use `os.listdir` to see if the keys are alright | ## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
| 61 | Keys should be unique error on code_search_net
## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
Thanks to the dummy data, we know for sure that most of them work as expected.
`code_search_net` wasn't caught because the dummy data only have one dummy data file while the dataset script can actually load several of them using `os.listdir`. Let me take a look at all the other datasets that use `os.listdir` to see if the keys are alright |
https://github.com/huggingface/datasets/issues/2552 | Keys should be unique error on code_search_net | I found one issue on `fever` (PR here: https://github.com/huggingface/datasets/pull/2557)
All the other ones seem fine :) | ## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
| 16 | Keys should be unique error on code_search_net
## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
I found one issue on `fever` (PR here: https://github.com/huggingface/datasets/pull/2557)
All the other ones seem fine :) |
https://github.com/huggingface/datasets/issues/2552 | Keys should be unique error on code_search_net | Hi! Got same error when loading other dataset:
```python3
load_dataset('wikicorpus', 'raw_en')
```
tb:
```pytb
---------------------------------------------------------------------------
DuplicatedKeysError Traceback (most recent call last)
/opt/conda/lib/python3.8/site-packages/datasets/builder.py in _prepare_split(self, split_generator)
1109 example = self.info.features.encode_example(record)
-> 1110 writer.write(example, key)
1111 finally:
/opt/conda/lib/python3.8/site-packages/datasets/arrow_writer.py in write(self, example, key, writer_batch_size)
341 if self._check_duplicates:
--> 342 self.check_duplicate_keys()
343 # Re-intializing to empty list for next batch
/opt/conda/lib/python3.8/site-packages/datasets/arrow_writer.py in check_duplicate_keys(self)
352 if hash in tmp_record:
--> 353 raise DuplicatedKeysError(key)
354 else:
DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 519
Keys should be unique and deterministic in nature
```
Version: datasets==1.11.0 | ## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
| 91 | Keys should be unique error on code_search_net
## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
Hi! Got same error when loading other dataset:
```python3
load_dataset('wikicorpus', 'raw_en')
```
tb:
```pytb
---------------------------------------------------------------------------
DuplicatedKeysError Traceback (most recent call last)
/opt/conda/lib/python3.8/site-packages/datasets/builder.py in _prepare_split(self, split_generator)
1109 example = self.info.features.encode_example(record)
-> 1110 writer.write(example, key)
1111 finally:
/opt/conda/lib/python3.8/site-packages/datasets/arrow_writer.py in write(self, example, key, writer_batch_size)
341 if self._check_duplicates:
--> 342 self.check_duplicate_keys()
343 # Re-intializing to empty list for next batch
/opt/conda/lib/python3.8/site-packages/datasets/arrow_writer.py in check_duplicate_keys(self)
352 if hash in tmp_record:
--> 353 raise DuplicatedKeysError(key)
354 else:
DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 519
Keys should be unique and deterministic in nature
```
Version: datasets==1.11.0 |
https://github.com/huggingface/datasets/issues/2552 | Keys should be unique error on code_search_net | The wikicorpus issue has been fixed by https://github.com/huggingface/datasets/pull/2844
We'll do a new release of `datasets` soon :) | ## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
| 17 | Keys should be unique error on code_search_net
## Describe the bug
Loading `code_search_net` seems not possible at the moment.
## Steps to reproduce the bug
```python
>>> load_dataset('code_search_net')
Downloading: 8.50kB [00:00, 3.09MB/s]
Downloading: 19.1kB [00:00, 10.1MB/s]
No config specified, defaulting to: code_search_net/all
Downloading and preparing dataset code_search_net/all (download: 4.77 GiB, generated: 5.99 GiB, post-processed: Unknown size, total: 10.76 GiB) to /Users/thomwolf/.cache/huggingface/datasets/code_search_net/all/1.0.0/b3e8278faf5d67da1d06981efbeac3b76a2900693bd2239bbca7a4a3b0d6e52a...
Traceback (most recent call last):
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/builder.py", line 1067, in _prepare_split
writer.write(example, key)
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 343, in write
self.check_duplicate_keys()
File "/Users/thomwolf/Documents/GitHub/datasets/src/datasets/arrow_writer.py", line 354, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 48
Keys should be unique and deterministic in nature
```
## Environment info
- `datasets` version: 1.8.1.dev0
- Platform: macOS-10.15.7-x86_64-i386-64bit
- Python version: 3.8.5
- PyArrow version: 2.0.0
The wikicorpus issue has been fixed by https://github.com/huggingface/datasets/pull/2844
We'll do a new release of `datasets` soon :) |
https://github.com/huggingface/datasets/issues/2549 | Handling unlabeled datasets | Hi @nelson-liu,
You can pass the parameter `features` to `load_dataset`: https://huggingface.co/docs/datasets/_modules/datasets/load.html#load_dataset
If you look at the code of the MNLI script you referred in your question (https://github.com/huggingface/datasets/blob/master/datasets/multi_nli/multi_nli.py#L62-L77), you can see how the Features were originally specified.
Feel free to use it as a template, customize it and pass it to `load_dataset` using the parameter `features`. | Hi!
Is there a way for datasets to produce unlabeled instances (e.g., the `ClassLabel` can be nullable).
For example, I want to use the MNLI dataset reader ( https://github.com/huggingface/datasets/blob/master/datasets/multi_nli/multi_nli.py ) on a file that doesn't have the `gold_label` field. I tried setting `"label": data.get("gold_label")`, but got the following error:
```
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/load.py", line 748, in load_dataset
use_auth_token=use_auth_token,
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/builder.py", line 575, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/builder.py", line 652, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/builder.py", line 989, in _prepare_split
example = self.info.features.encode_example(record)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 953, in encode_example
return encode_nested_example(self, example)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 848, in encode_nested_example
k: encode_nested_example(sub_schema, sub_obj) for k, (sub_schema, sub_obj) in utils.zip_dict(schema, obj)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 848, in <dictcomp>
k: encode_nested_example(sub_schema, sub_obj) for k, (sub_schema, sub_obj) in utils.zip_dict(schema, obj)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 875, in encode_nested_example
return schema.encode_example(obj)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 653, in encode_example
if not -1 <= example_data < self.num_classes:
TypeError: '<=' not supported between instances of 'int' and 'NoneType'
```
What's the proper way to handle reading unlabeled datasets, especially for downstream usage with Transformers? | 55 | Handling unlabeled datasets
Hi!
Is there a way for datasets to produce unlabeled instances (e.g., the `ClassLabel` can be nullable).
For example, I want to use the MNLI dataset reader ( https://github.com/huggingface/datasets/blob/master/datasets/multi_nli/multi_nli.py ) on a file that doesn't have the `gold_label` field. I tried setting `"label": data.get("gold_label")`, but got the following error:
```
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/load.py", line 748, in load_dataset
use_auth_token=use_auth_token,
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/builder.py", line 575, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/builder.py", line 652, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/builder.py", line 989, in _prepare_split
example = self.info.features.encode_example(record)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 953, in encode_example
return encode_nested_example(self, example)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 848, in encode_nested_example
k: encode_nested_example(sub_schema, sub_obj) for k, (sub_schema, sub_obj) in utils.zip_dict(schema, obj)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 848, in <dictcomp>
k: encode_nested_example(sub_schema, sub_obj) for k, (sub_schema, sub_obj) in utils.zip_dict(schema, obj)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 875, in encode_nested_example
return schema.encode_example(obj)
File "/home/nfliu/miniconda3/envs/debias/lib/python3.7/site-packages/datasets/features.py", line 653, in encode_example
if not -1 <= example_data < self.num_classes:
TypeError: '<=' not supported between instances of 'int' and 'NoneType'
```
What's the proper way to handle reading unlabeled datasets, especially for downstream usage with Transformers?
Hi @nelson-liu,
You can pass the parameter `features` to `load_dataset`: https://huggingface.co/docs/datasets/_modules/datasets/load.html#load_dataset
If you look at the code of the MNLI script you referred in your question (https://github.com/huggingface/datasets/blob/master/datasets/multi_nli/multi_nli.py#L62-L77), you can see how the Features were originally specified.
Feel free to use it as a template, customize it and pass it to `load_dataset` using the parameter `features`. |
https://github.com/huggingface/datasets/issues/2548 | Field order issue in loading json | Hi @luyug, thanks for reporting.
The good news is that we fixed this issue only 9 days ago: #2507.
The patch is already in the master branch of our repository and it will be included in our next `datasets` release version 1.9.0.
Feel free to reopen the issue if the problem persists. | ## Describe the bug
The `load_dataset` function expects columns in alphabetical order when loading json files.
Similar bug was previously reported for csv in #623 and fixed in #684.
## Steps to reproduce the bug
For a json file `j.json`,
```
{"c":321, "a": 1, "b": 2}
```
Running the following,
```
f= datasets.Features({'a': Value('int32'), 'b': Value('int32'), 'c': Value('int32')})
json_data = datasets.load_dataset('json', data_files='j.json', features=f)
```
## Expected results
A successful load.
## Actual results
```
File "pyarrow/table.pxi", line 1409, in pyarrow.lib.Table.cast
ValueError: Target schema's field names are not matching the table's field names: ['c', 'a', 'b'], ['a', 'b', 'c']
```
## Environment info
- `datasets` version: 1.8.0
- Platform: Linux-3.10.0-957.1.3.el7.x86_64-x86_64-with-glibc2.10
- Python version: 3.8.8
- PyArrow version: 3.0.0
| 52 | Field order issue in loading json
## Describe the bug
The `load_dataset` function expects columns in alphabetical order when loading json files.
Similar bug was previously reported for csv in #623 and fixed in #684.
## Steps to reproduce the bug
For a json file `j.json`,
```
{"c":321, "a": 1, "b": 2}
```
Running the following,
```
f= datasets.Features({'a': Value('int32'), 'b': Value('int32'), 'c': Value('int32')})
json_data = datasets.load_dataset('json', data_files='j.json', features=f)
```
## Expected results
A successful load.
## Actual results
```
File "pyarrow/table.pxi", line 1409, in pyarrow.lib.Table.cast
ValueError: Target schema's field names are not matching the table's field names: ['c', 'a', 'b'], ['a', 'b', 'c']
```
## Environment info
- `datasets` version: 1.8.0
- Platform: Linux-3.10.0-957.1.3.el7.x86_64-x86_64-with-glibc2.10
- Python version: 3.8.8
- PyArrow version: 3.0.0
Hi @luyug, thanks for reporting.
The good news is that we fixed this issue only 9 days ago: #2507.
The patch is already in the master branch of our repository and it will be included in our next `datasets` release version 1.9.0.
Feel free to reopen the issue if the problem persists. |
https://github.com/huggingface/datasets/issues/2547 | Dataset load_from_disk is too slow | Hi ! It looks like an issue with the virtual disk you are using.
We load datasets using memory mapping. In general it makes it possible to load very big files instantaneously since it doesn't have to read the file (it just assigns virtual memory to the file on disk).
However there happens to be issues with virtual disks (for example on spot instances), for which memory mapping does a pass over the entire file, and this takes a while. We are discussing about this issue here: #2252
Memory mapping is something handled by the OS so we can't do much about it, though we're still trying to figure out what's causing this behavior exactly to see what we can do. | @lhoestq
## Describe the bug
It's not normal that I have to wait 7-8 hours for a dataset to be loaded from disk, as there are no preprocessing steps, it's only loading it with load_from_disk. I have 96 cpus, however only 1 is used for this, which is inefficient. Moreover, its usage is at 1%... This is happening in the context of a language model training, therefore I'm wasting 100$ each time I have to load the dataset from disk again (because the spot instance was stopped by aws and I need to relaunch it for example).
## Steps to reproduce the bug
Just get the oscar in spanish (around 150GGB) and try to first save in disk and then load the processed dataset. It's not dependent on the task you're doing, it just depends on the size of the text dataset.
## Expected results
I expect the dataset to be loaded in a normal time, by using the whole machine for loading it, I mean if you store the dataset in multiple files (.arrow) and then load it from multiple files, you can use multiprocessing for that and therefore don't waste so much time.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Ubuntu 18
- Python version: 3.8
I've seen you're planning to include a streaming mode for load_dataset, but that only saves the downloading and processing time, that's not being a problem for me, you cannot save the pure loading from disk time, therefore that's not a solution for my use case or for anyone who wants to use your library for training a language model. | 121 | Dataset load_from_disk is too slow
@lhoestq
## Describe the bug
It's not normal that I have to wait 7-8 hours for a dataset to be loaded from disk, as there are no preprocessing steps, it's only loading it with load_from_disk. I have 96 cpus, however only 1 is used for this, which is inefficient. Moreover, its usage is at 1%... This is happening in the context of a language model training, therefore I'm wasting 100$ each time I have to load the dataset from disk again (because the spot instance was stopped by aws and I need to relaunch it for example).
## Steps to reproduce the bug
Just get the oscar in spanish (around 150GGB) and try to first save in disk and then load the processed dataset. It's not dependent on the task you're doing, it just depends on the size of the text dataset.
## Expected results
I expect the dataset to be loaded in a normal time, by using the whole machine for loading it, I mean if you store the dataset in multiple files (.arrow) and then load it from multiple files, you can use multiprocessing for that and therefore don't waste so much time.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Ubuntu 18
- Python version: 3.8
I've seen you're planning to include a streaming mode for load_dataset, but that only saves the downloading and processing time, that's not being a problem for me, you cannot save the pure loading from disk time, therefore that's not a solution for my use case or for anyone who wants to use your library for training a language model.
Hi ! It looks like an issue with the virtual disk you are using.
We load datasets using memory mapping. In general it makes it possible to load very big files instantaneously since it doesn't have to read the file (it just assigns virtual memory to the file on disk).
However there happens to be issues with virtual disks (for example on spot instances), for which memory mapping does a pass over the entire file, and this takes a while. We are discussing about this issue here: #2252
Memory mapping is something handled by the OS so we can't do much about it, though we're still trying to figure out what's causing this behavior exactly to see what we can do. |
https://github.com/huggingface/datasets/issues/2547 | Dataset load_from_disk is too slow | Okay, that's exactly my case, with spot instances... Therefore this isn't something we can change in any way to be able to load the dataset faster? I mean, what do you do internally at huggingface for being able to use spot instances with datasets efficiently? | @lhoestq
## Describe the bug
It's not normal that I have to wait 7-8 hours for a dataset to be loaded from disk, as there are no preprocessing steps, it's only loading it with load_from_disk. I have 96 cpus, however only 1 is used for this, which is inefficient. Moreover, its usage is at 1%... This is happening in the context of a language model training, therefore I'm wasting 100$ each time I have to load the dataset from disk again (because the spot instance was stopped by aws and I need to relaunch it for example).
## Steps to reproduce the bug
Just get the oscar in spanish (around 150GGB) and try to first save in disk and then load the processed dataset. It's not dependent on the task you're doing, it just depends on the size of the text dataset.
## Expected results
I expect the dataset to be loaded in a normal time, by using the whole machine for loading it, I mean if you store the dataset in multiple files (.arrow) and then load it from multiple files, you can use multiprocessing for that and therefore don't waste so much time.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Ubuntu 18
- Python version: 3.8
I've seen you're planning to include a streaming mode for load_dataset, but that only saves the downloading and processing time, that's not being a problem for me, you cannot save the pure loading from disk time, therefore that's not a solution for my use case or for anyone who wants to use your library for training a language model. | 45 | Dataset load_from_disk is too slow
@lhoestq
## Describe the bug
It's not normal that I have to wait 7-8 hours for a dataset to be loaded from disk, as there are no preprocessing steps, it's only loading it with load_from_disk. I have 96 cpus, however only 1 is used for this, which is inefficient. Moreover, its usage is at 1%... This is happening in the context of a language model training, therefore I'm wasting 100$ each time I have to load the dataset from disk again (because the spot instance was stopped by aws and I need to relaunch it for example).
## Steps to reproduce the bug
Just get the oscar in spanish (around 150GGB) and try to first save in disk and then load the processed dataset. It's not dependent on the task you're doing, it just depends on the size of the text dataset.
## Expected results
I expect the dataset to be loaded in a normal time, by using the whole machine for loading it, I mean if you store the dataset in multiple files (.arrow) and then load it from multiple files, you can use multiprocessing for that and therefore don't waste so much time.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Ubuntu 18
- Python version: 3.8
I've seen you're planning to include a streaming mode for load_dataset, but that only saves the downloading and processing time, that's not being a problem for me, you cannot save the pure loading from disk time, therefore that's not a solution for my use case or for anyone who wants to use your library for training a language model.
Okay, that's exactly my case, with spot instances... Therefore this isn't something we can change in any way to be able to load the dataset faster? I mean, what do you do internally at huggingface for being able to use spot instances with datasets efficiently? |
https://github.com/huggingface/datasets/issues/2547 | Dataset load_from_disk is too slow | There are no solutions yet unfortunately.
We're still trying to figure out a way to make the loading instantaneous on such disks, I'll keep you posted | @lhoestq
## Describe the bug
It's not normal that I have to wait 7-8 hours for a dataset to be loaded from disk, as there are no preprocessing steps, it's only loading it with load_from_disk. I have 96 cpus, however only 1 is used for this, which is inefficient. Moreover, its usage is at 1%... This is happening in the context of a language model training, therefore I'm wasting 100$ each time I have to load the dataset from disk again (because the spot instance was stopped by aws and I need to relaunch it for example).
## Steps to reproduce the bug
Just get the oscar in spanish (around 150GGB) and try to first save in disk and then load the processed dataset. It's not dependent on the task you're doing, it just depends on the size of the text dataset.
## Expected results
I expect the dataset to be loaded in a normal time, by using the whole machine for loading it, I mean if you store the dataset in multiple files (.arrow) and then load it from multiple files, you can use multiprocessing for that and therefore don't waste so much time.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Ubuntu 18
- Python version: 3.8
I've seen you're planning to include a streaming mode for load_dataset, but that only saves the downloading and processing time, that's not being a problem for me, you cannot save the pure loading from disk time, therefore that's not a solution for my use case or for anyone who wants to use your library for training a language model. | 26 | Dataset load_from_disk is too slow
@lhoestq
## Describe the bug
It's not normal that I have to wait 7-8 hours for a dataset to be loaded from disk, as there are no preprocessing steps, it's only loading it with load_from_disk. I have 96 cpus, however only 1 is used for this, which is inefficient. Moreover, its usage is at 1%... This is happening in the context of a language model training, therefore I'm wasting 100$ each time I have to load the dataset from disk again (because the spot instance was stopped by aws and I need to relaunch it for example).
## Steps to reproduce the bug
Just get the oscar in spanish (around 150GGB) and try to first save in disk and then load the processed dataset. It's not dependent on the task you're doing, it just depends on the size of the text dataset.
## Expected results
I expect the dataset to be loaded in a normal time, by using the whole machine for loading it, I mean if you store the dataset in multiple files (.arrow) and then load it from multiple files, you can use multiprocessing for that and therefore don't waste so much time.
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.8.0
- Platform: Ubuntu 18
- Python version: 3.8
I've seen you're planning to include a streaming mode for load_dataset, but that only saves the downloading and processing time, that's not being a problem for me, you cannot save the pure loading from disk time, therefore that's not a solution for my use case or for anyone who wants to use your library for training a language model.
There are no solutions yet unfortunately.
We're still trying to figure out a way to make the loading instantaneous on such disks, I'll keep you posted |
https://github.com/huggingface/datasets/issues/2543 | switching some low-level log.info's to log.debug? | Hi @stas00, thanks for pointing out this issue with logging.
I agree that `datasets` can sometimes be too verbose... I can create a PR and we could discuss there the choice of the log levels for different parts of the code. | In https://github.com/huggingface/transformers/pull/12276 we are now changing the examples to have `datasets` on the same log level as `transformers`, so that one setting can do a consistent logging across all involved components.
The trouble is that now we get a ton of these:
```
06/23/2021 12:15:31 - INFO - datasets.utils.filelock - Lock 139627640431136 acquired on /home/stas/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow.lock
06/23/2021 12:15:31 - INFO - datasets.arrow_writer - Done writing 50 examples in 12280 bytes /home/stas/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow.
06/23/2021 12:15:31 - INFO - datasets.arrow_dataset - Set __getitem__(key) output type to python objects for no columns (when key is int or slice) and don't output other (un-formatted) columns.
06/23/2021 12:15:31 - INFO - datasets.utils.filelock - Lock 139627640431136 released on /home/stas/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow.lock
```
May I suggest that these can be `log.debug` as it's no informative to the user.
More examples: these are not informative - too much information:
```
06/23/2021 12:14:26 - INFO - datasets.load - Checking /home/stas/.cache/huggingface/datasets/downloads/459933f1fe47711fad2f6ff8110014ff189120b45ad159ef5b8e90ea43a174fa.e23e7d1259a8c6274a82a42a8936dd1b87225302c6dc9b7261beb3bc2daac640.py for additional imports.
06/23/2021 12:14:27 - INFO - datasets.builder - Constructing Dataset for split train, validation, test, from /home/stas/.cache/huggingface/datasets/wmt16/ro-en/1.0.0/0d9fb3e814712c785176ad8cdb9f465fbe6479000ee6546725db30ad8a8b5f8a
```
While these are:
```
06/23/2021 12:14:27 - INFO - datasets.info - Loading Dataset Infos from /home/stas/.cache/huggingface/modules/datasets_modules/datasets/wmt16/0d9fb3e814712c785176ad8cdb9f465fbe6479000ee6546725db30ad8a8b5f8a
06/23/2021 12:14:27 - WARNING - datasets.builder - Reusing dataset wmt16 (/home/stas/.cache/huggingface/datasets/wmt16/ro-en/1.0.0/0d9fb3e814712c785176ad8cdb9f465fbe6479000ee6546725db30ad8a8b5f8a)
```
I also realize that `transformers` examples don't have do use `info` for `datasets` to let the default `warning` keep logging to less noisy.
But I think currently the log levels are slightly misused and skewed by 1 level. Many `warnings` will better be `info`s and most `info`s be `debug`.
e.g.:
```
06/23/2021 12:14:27 - WARNING - datasets.builder - Reusing dataset wmt16 (/home/stas/.cache/huggingface/datasets/wmt16/ro-en/1.0.0/0d9fb3e814712c785176ad8cdb9f465fbe6479000ee6546725db30ad8a8b5f8a)
```
why is this a warning? it is informing me that the cache is used, there is nothing to be worried about. I'd have it as `info`.
Warnings are typically something that's bordering error or the first thing to check when things don't work as expected.
infrequent info is there to inform of the different stages or important events.
Everything else is debug.
At least the way I understand things.
| 41 | switching some low-level log.info's to log.debug?
In https://github.com/huggingface/transformers/pull/12276 we are now changing the examples to have `datasets` on the same log level as `transformers`, so that one setting can do a consistent logging across all involved components.
The trouble is that now we get a ton of these:
```
06/23/2021 12:15:31 - INFO - datasets.utils.filelock - Lock 139627640431136 acquired on /home/stas/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow.lock
06/23/2021 12:15:31 - INFO - datasets.arrow_writer - Done writing 50 examples in 12280 bytes /home/stas/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow.
06/23/2021 12:15:31 - INFO - datasets.arrow_dataset - Set __getitem__(key) output type to python objects for no columns (when key is int or slice) and don't output other (un-formatted) columns.
06/23/2021 12:15:31 - INFO - datasets.utils.filelock - Lock 139627640431136 released on /home/stas/.cache/huggingface/metrics/sacrebleu/default/default_experiment-1-0.arrow.lock
```
May I suggest that these can be `log.debug` as it's no informative to the user.
More examples: these are not informative - too much information:
```
06/23/2021 12:14:26 - INFO - datasets.load - Checking /home/stas/.cache/huggingface/datasets/downloads/459933f1fe47711fad2f6ff8110014ff189120b45ad159ef5b8e90ea43a174fa.e23e7d1259a8c6274a82a42a8936dd1b87225302c6dc9b7261beb3bc2daac640.py for additional imports.
06/23/2021 12:14:27 - INFO - datasets.builder - Constructing Dataset for split train, validation, test, from /home/stas/.cache/huggingface/datasets/wmt16/ro-en/1.0.0/0d9fb3e814712c785176ad8cdb9f465fbe6479000ee6546725db30ad8a8b5f8a
```
While these are:
```
06/23/2021 12:14:27 - INFO - datasets.info - Loading Dataset Infos from /home/stas/.cache/huggingface/modules/datasets_modules/datasets/wmt16/0d9fb3e814712c785176ad8cdb9f465fbe6479000ee6546725db30ad8a8b5f8a
06/23/2021 12:14:27 - WARNING - datasets.builder - Reusing dataset wmt16 (/home/stas/.cache/huggingface/datasets/wmt16/ro-en/1.0.0/0d9fb3e814712c785176ad8cdb9f465fbe6479000ee6546725db30ad8a8b5f8a)
```
I also realize that `transformers` examples don't have do use `info` for `datasets` to let the default `warning` keep logging to less noisy.
But I think currently the log levels are slightly misused and skewed by 1 level. Many `warnings` will better be `info`s and most `info`s be `debug`.
e.g.:
```
06/23/2021 12:14:27 - WARNING - datasets.builder - Reusing dataset wmt16 (/home/stas/.cache/huggingface/datasets/wmt16/ro-en/1.0.0/0d9fb3e814712c785176ad8cdb9f465fbe6479000ee6546725db30ad8a8b5f8a)
```
why is this a warning? it is informing me that the cache is used, there is nothing to be worried about. I'd have it as `info`.
Warnings are typically something that's bordering error or the first thing to check when things don't work as expected.
infrequent info is there to inform of the different stages or important events.
Everything else is debug.
At least the way I understand things.
Hi @stas00, thanks for pointing out this issue with logging.
I agree that `datasets` can sometimes be too verbose... I can create a PR and we could discuss there the choice of the log levels for different parts of the code. |
https://github.com/huggingface/datasets/issues/2542 | `datasets.keyhash.DuplicatedKeysError` for `drop` and `adversarial_qa/adversarialQA` | Hi @VictorSanh, thank you for reporting this issue with duplicated keys.
- The issue with "adversarial_qa" was fixed 23 days ago: #2433. Current version of `datasets` (1.8.0) includes the patch.
- I am investigating the issue with `drop`. I'll ping you to keep you informed. | ## Describe the bug
Failure to generate the datasets (`drop` and subset `adversarialQA` from `adversarial_qa`) because of duplicate keys.
## Steps to reproduce the bug
```python
from datasets import load_dataset
load_dataset("drop")
load_dataset("adversarial_qa", "adversarialQA")
```
## Expected results
The examples keys should be unique.
## Actual results
```bash
>>> load_dataset("drop")
Using custom data configuration default
Downloading and preparing dataset drop/default (download: 7.92 MiB, generated: 111.88 MiB, post-processed: Unknown size, total: 119.80 MiB) to /home/hf/.cache/huggingface/datasets/drop/default/0.1.0/7a94f1e2bb26c4b5c75f89857c06982967d7416e5af935a9374b9bccf5068026...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/load.py", line 751, in load_dataset
use_auth_token=use_auth_token,
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 575, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 652, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 992, in _prepare_split
num_examples, num_bytes = writer.finalize()
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/arrow_writer.py", line 409, in finalize
self.check_duplicate_keys()
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/arrow_writer.py", line 349, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 28553293-d719-441b-8f00-ce3dc6df5398
Keys should be unique and deterministic in nature
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.7.0
- Platform: Linux-5.4.0-1044-gcp-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.10
- PyArrow version: 3.0.0
| 45 | `datasets.keyhash.DuplicatedKeysError` for `drop` and `adversarial_qa/adversarialQA`
## Describe the bug
Failure to generate the datasets (`drop` and subset `adversarialQA` from `adversarial_qa`) because of duplicate keys.
## Steps to reproduce the bug
```python
from datasets import load_dataset
load_dataset("drop")
load_dataset("adversarial_qa", "adversarialQA")
```
## Expected results
The examples keys should be unique.
## Actual results
```bash
>>> load_dataset("drop")
Using custom data configuration default
Downloading and preparing dataset drop/default (download: 7.92 MiB, generated: 111.88 MiB, post-processed: Unknown size, total: 119.80 MiB) to /home/hf/.cache/huggingface/datasets/drop/default/0.1.0/7a94f1e2bb26c4b5c75f89857c06982967d7416e5af935a9374b9bccf5068026...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/load.py", line 751, in load_dataset
use_auth_token=use_auth_token,
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 575, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 652, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 992, in _prepare_split
num_examples, num_bytes = writer.finalize()
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/arrow_writer.py", line 409, in finalize
self.check_duplicate_keys()
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/arrow_writer.py", line 349, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 28553293-d719-441b-8f00-ce3dc6df5398
Keys should be unique and deterministic in nature
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.7.0
- Platform: Linux-5.4.0-1044-gcp-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.10
- PyArrow version: 3.0.0
Hi @VictorSanh, thank you for reporting this issue with duplicated keys.
- The issue with "adversarial_qa" was fixed 23 days ago: #2433. Current version of `datasets` (1.8.0) includes the patch.
- I am investigating the issue with `drop`. I'll ping you to keep you informed. |
https://github.com/huggingface/datasets/issues/2542 | `datasets.keyhash.DuplicatedKeysError` for `drop` and `adversarial_qa/adversarialQA` | Hi @VictorSanh, the issue is already fixed and merged into master branch and will be included in our next release version 1.9.0. | ## Describe the bug
Failure to generate the datasets (`drop` and subset `adversarialQA` from `adversarial_qa`) because of duplicate keys.
## Steps to reproduce the bug
```python
from datasets import load_dataset
load_dataset("drop")
load_dataset("adversarial_qa", "adversarialQA")
```
## Expected results
The examples keys should be unique.
## Actual results
```bash
>>> load_dataset("drop")
Using custom data configuration default
Downloading and preparing dataset drop/default (download: 7.92 MiB, generated: 111.88 MiB, post-processed: Unknown size, total: 119.80 MiB) to /home/hf/.cache/huggingface/datasets/drop/default/0.1.0/7a94f1e2bb26c4b5c75f89857c06982967d7416e5af935a9374b9bccf5068026...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/load.py", line 751, in load_dataset
use_auth_token=use_auth_token,
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 575, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 652, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 992, in _prepare_split
num_examples, num_bytes = writer.finalize()
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/arrow_writer.py", line 409, in finalize
self.check_duplicate_keys()
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/arrow_writer.py", line 349, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 28553293-d719-441b-8f00-ce3dc6df5398
Keys should be unique and deterministic in nature
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.7.0
- Platform: Linux-5.4.0-1044-gcp-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.10
- PyArrow version: 3.0.0
| 22 | `datasets.keyhash.DuplicatedKeysError` for `drop` and `adversarial_qa/adversarialQA`
## Describe the bug
Failure to generate the datasets (`drop` and subset `adversarialQA` from `adversarial_qa`) because of duplicate keys.
## Steps to reproduce the bug
```python
from datasets import load_dataset
load_dataset("drop")
load_dataset("adversarial_qa", "adversarialQA")
```
## Expected results
The examples keys should be unique.
## Actual results
```bash
>>> load_dataset("drop")
Using custom data configuration default
Downloading and preparing dataset drop/default (download: 7.92 MiB, generated: 111.88 MiB, post-processed: Unknown size, total: 119.80 MiB) to /home/hf/.cache/huggingface/datasets/drop/default/0.1.0/7a94f1e2bb26c4b5c75f89857c06982967d7416e5af935a9374b9bccf5068026...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/load.py", line 751, in load_dataset
use_auth_token=use_auth_token,
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 575, in download_and_prepare
dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 652, in _download_and_prepare
self._prepare_split(split_generator, **prepare_split_kwargs)
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/builder.py", line 992, in _prepare_split
num_examples, num_bytes = writer.finalize()
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/arrow_writer.py", line 409, in finalize
self.check_duplicate_keys()
File "/home/hf/dev/promptsource/.venv/lib/python3.7/site-packages/datasets/arrow_writer.py", line 349, in check_duplicate_keys
raise DuplicatedKeysError(key)
datasets.keyhash.DuplicatedKeysError: FAILURE TO GENERATE DATASET !
Found duplicate Key: 28553293-d719-441b-8f00-ce3dc6df5398
Keys should be unique and deterministic in nature
```
## Environment info
<!-- You can run the command `datasets-cli env` and copy-and-paste its output below. -->
- `datasets` version: 1.7.0
- Platform: Linux-5.4.0-1044-gcp-x86_64-with-Ubuntu-18.04-bionic
- Python version: 3.7.10
- PyArrow version: 3.0.0
Hi @VictorSanh, the issue is already fixed and merged into master branch and will be included in our next release version 1.9.0. |
https://github.com/huggingface/datasets/issues/2538 | Loading partial dataset when debugging | Hi ! `load_dataset` downloads the full dataset once and caches it, so that subsequent calls to `load_dataset` just reloads the dataset from your disk.
Then when you specify a `split` in `load_dataset`, it will just load the requested split from the disk. If your specified split is a sliced split (e.g. `"train[:10]"`), then it will load the 10 first rows of the train split that you have on disk.
Therefore, as long as you don't delete your cache, all your calls to `load_dataset` will be very fast. Except the first call that downloads the dataset of course ^^ | I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks! | 98 | Loading partial dataset when debugging
I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks!
Hi ! `load_dataset` downloads the full dataset once and caches it, so that subsequent calls to `load_dataset` just reloads the dataset from your disk.
Then when you specify a `split` in `load_dataset`, it will just load the requested split from the disk. If your specified split is a sliced split (e.g. `"train[:10]"`), then it will load the 10 first rows of the train split that you have on disk.
Therefore, as long as you don't delete your cache, all your calls to `load_dataset` will be very fast. Except the first call that downloads the dataset of course ^^ |
https://github.com/huggingface/datasets/issues/2538 | Loading partial dataset when debugging | Hi @reachtarunhere.
Besides the above insights provided by @lhoestq and @thomwolf, there is also a Dataset feature in progress (I plan to finish it this week): #2249, which will allow you, when calling `load_dataset`, to pass the option to download/preprocess/cache only some specific split(s), which will definitely speed up your workflow.
If this feature is interesting for you, I can ping you once it will be merged into the master branch. | I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks! | 71 | Loading partial dataset when debugging
I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks!
Hi @reachtarunhere.
Besides the above insights provided by @lhoestq and @thomwolf, there is also a Dataset feature in progress (I plan to finish it this week): #2249, which will allow you, when calling `load_dataset`, to pass the option to download/preprocess/cache only some specific split(s), which will definitely speed up your workflow.
If this feature is interesting for you, I can ping you once it will be merged into the master branch. |
https://github.com/huggingface/datasets/issues/2538 | Loading partial dataset when debugging | Thanks all for responding.
Hey @albertvillanova
Thanks. Yes, I would be interested.
@lhoestq I think even if a small split is specified it loads up the full dataset from the disk (please correct me if this is not the case). Because it does seem to be slow to me even on subsequent calls. There is no repeated downloading so it seems that the cache is working.
I am not aware of the streaming feature @thomwolf mentioned. So I might need to read up on it. | I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks! | 85 | Loading partial dataset when debugging
I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks!
Thanks all for responding.
Hey @albertvillanova
Thanks. Yes, I would be interested.
@lhoestq I think even if a small split is specified it loads up the full dataset from the disk (please correct me if this is not the case). Because it does seem to be slow to me even on subsequent calls. There is no repeated downloading so it seems that the cache is working.
I am not aware of the streaming feature @thomwolf mentioned. So I might need to read up on it. |
https://github.com/huggingface/datasets/issues/2538 | Loading partial dataset when debugging | If I want to create a dataset, containing only the 10 elements of a given dataset (slice it), how do I do that? | I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks! | 23 | Loading partial dataset when debugging
I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks!
If I want to create a dataset, containing only the 10 elements of a given dataset (slice it), how do I do that? |
https://github.com/huggingface/datasets/issues/2538 | Loading partial dataset when debugging |
> ```python
> small_ds = ds.select(range(10))
> ```
Thanks, but this doesn't help me to save time during initial loading, right? | I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks! | 21 | Loading partial dataset when debugging
I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks!
> ```python
> small_ds = ds.select(range(10))
> ```
Thanks, but this doesn't help me to save time during initial loading, right? |
https://github.com/huggingface/datasets/issues/2538 | Loading partial dataset when debugging | Indeed by default load_dataset would download and prepare everything as Arrow files. And passing `split=train[:10]` memory maps only the beginning of the full dataset that has been prepared on disk.
If you don't want to download everything, you can use streaming :
```python
ids = load_dataset(..., streaming=True)
first_samples = list(ids["train"].take(10))
```
To get a Dataset you can use
```python
ds = Dataset.from_generator(ids.take(10).__iter__)
```
edit: fixed small bug | I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks! | 67 | Loading partial dataset when debugging
I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks!
Indeed by default load_dataset would download and prepare everything as Arrow files. And passing `split=train[:10]` memory maps only the beginning of the full dataset that has been prepared on disk.
If you don't want to download everything, you can use streaming :
```python
ids = load_dataset(..., streaming=True)
first_samples = list(ids["train"].take(10))
```
To get a Dataset you can use
```python
ds = Dataset.from_generator(ids.take(10).__iter__)
```
edit: fixed small bug |
https://github.com/huggingface/datasets/issues/2538 | Loading partial dataset when debugging | Thanks @lhoestq, but I don't think it is 100% accurate, as it doesn't keep the dataset structure exactly the same.
To load the full dataset, I do:
```
data = load_dataset("json", data_files="a.json")
train_data = data["train"].shuffle()
```
But when I am changing it as per your instructions:
```
ids = load_dataset("json", data_files="a.json", streaming=True)
data = Dataset.from_generator(ids["train"].take(1).__iter__)
train_data = data["train"].shuffle()
```
It throws KeyError.
I need a simple way, like you suggested, to have a subset of a Dataset, which exactly the same attributes.
| I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks! | 82 | Loading partial dataset when debugging
I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks!
Thanks @lhoestq, but I don't think it is 100% accurate, as it doesn't keep the dataset structure exactly the same.
To load the full dataset, I do:
```
data = load_dataset("json", data_files="a.json")
train_data = data["train"].shuffle()
```
But when I am changing it as per your instructions:
```
ids = load_dataset("json", data_files="a.json", streaming=True)
data = Dataset.from_generator(ids["train"].take(1).__iter__)
train_data = data["train"].shuffle()
```
It throws KeyError.
I need a simple way, like you suggested, to have a subset of a Dataset, which exactly the same attributes.
|
https://github.com/huggingface/datasets/issues/2538 | Loading partial dataset when debugging | Whoops I fixed my code sorry
```diff
- ds = Dataset.from_generator(ids["train"].take(10).__iter__)
+ ds = Dataset.from_generator(ids.take(10).__iter__)
```
in your case that means running
```python
train_data = data.shuffle()
```
without `["train"]` | I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks! | 29 | Loading partial dataset when debugging
I am using PyTorch Lightning along with datasets (thanks for so many datasets already prepared and the great splits).
Every time I execute load_dataset for the imdb dataset it takes some time even if I specify a split involving very few samples. I guess this due to hashing as per the other issues.
Is there a way to only load part of the dataset on load_dataset? This would really speed up my workflow.
Something like a debug mode would really help. Thanks!
Whoops I fixed my code sorry
```diff
- ds = Dataset.from_generator(ids["train"].take(10).__iter__)
+ ds = Dataset.from_generator(ids.take(10).__iter__)
```
in your case that means running
```python
train_data = data.shuffle()
```
without `["train"]` |
https://github.com/huggingface/datasets/issues/2532 | Tokenizer's normalization preprocessor cause misalignment in return_offsets_mapping for tokenizer classification task | Hi @jerryIsHere, thanks for reporting the issue. But are you sure this is a bug in HuggingFace **Datasets**? | [This colab notebook](https://colab.research.google.com/drive/151gKyo0YIwnlznrOHst23oYH_a3mAe3Z?usp=sharing) implements a token classification input pipeline extending the logic from [this hugging example](https://huggingface.co/transformers/custom_datasets.html#tok-ner).
The pipeline works fine with most instance in different languages, but unfortunately, [the Japanese Kana ligature (a form of abbreviation? I don't know Japanese well)](https://en.wikipedia.org/wiki/Kana_ligature) break the alignment of `return_offsets_mapping`:
Without the try catch block, it riase `ValueError: NumPy boolean array indexing assignment cannot assign 88 input values to the 87 output values where the mask is true`, example shown here [(another colab notebook)](https://colab.research.google.com/drive/1MmOqf3ppzzdKKyMWkn0bJy6DqzOO0SSm?usp=sharing)
It is clear that the normalizer is the process that break the alignment, as it is observed that `tokenizer._tokenizer.normalizer.normalize_str('ヿ')` return 'コト'.
One workaround is to include `tokenizer._tokenizer.normalizer.normalize_str` before the tokenizer preprocessing pipeline, which is also provided in the [first colab notebook](https://colab.research.google.com/drive/151gKyo0YIwnlznrOHst23oYH_a3mAe3Z?usp=sharing) with the name `udposTestDatasetWorkaround`.
I guess similar logics should be included inside the tokenizer and the offsets_mapping generation process such that user don't need to include them in their code. But I don't understand the code of tokenizer well that I think I am not able to do this.
p.s.
**I am using my own dataset building script in the provided example, but the script should be equivalent to the changes made by this [update](https://github.com/huggingface/datasets/pull/2466)**
`get_dataset `is just a simple wrapping for `load_dataset`
and the `tokenizer` is just `XLMRobertaTokenizerFast.from_pretrained("xlm-roberta-large")` | 18 | Tokenizer's normalization preprocessor cause misalignment in return_offsets_mapping for tokenizer classification task
[This colab notebook](https://colab.research.google.com/drive/151gKyo0YIwnlznrOHst23oYH_a3mAe3Z?usp=sharing) implements a token classification input pipeline extending the logic from [this hugging example](https://huggingface.co/transformers/custom_datasets.html#tok-ner).
The pipeline works fine with most instance in different languages, but unfortunately, [the Japanese Kana ligature (a form of abbreviation? I don't know Japanese well)](https://en.wikipedia.org/wiki/Kana_ligature) break the alignment of `return_offsets_mapping`:
Without the try catch block, it riase `ValueError: NumPy boolean array indexing assignment cannot assign 88 input values to the 87 output values where the mask is true`, example shown here [(another colab notebook)](https://colab.research.google.com/drive/1MmOqf3ppzzdKKyMWkn0bJy6DqzOO0SSm?usp=sharing)
It is clear that the normalizer is the process that break the alignment, as it is observed that `tokenizer._tokenizer.normalizer.normalize_str('ヿ')` return 'コト'.
One workaround is to include `tokenizer._tokenizer.normalizer.normalize_str` before the tokenizer preprocessing pipeline, which is also provided in the [first colab notebook](https://colab.research.google.com/drive/151gKyo0YIwnlznrOHst23oYH_a3mAe3Z?usp=sharing) with the name `udposTestDatasetWorkaround`.
I guess similar logics should be included inside the tokenizer and the offsets_mapping generation process such that user don't need to include them in their code. But I don't understand the code of tokenizer well that I think I am not able to do this.
p.s.
**I am using my own dataset building script in the provided example, but the script should be equivalent to the changes made by this [update](https://github.com/huggingface/datasets/pull/2466)**
`get_dataset `is just a simple wrapping for `load_dataset`
and the `tokenizer` is just `XLMRobertaTokenizerFast.from_pretrained("xlm-roberta-large")`
Hi @jerryIsHere, thanks for reporting the issue. But are you sure this is a bug in HuggingFace **Datasets**? |
https://github.com/huggingface/datasets/issues/2532 | Tokenizer's normalization preprocessor cause misalignment in return_offsets_mapping for tokenizer classification task | > Hi @jerryIsHere, thanks for reporting the issue. But are you sure this is a bug in HuggingFace **Datasets**?
Oh, I am sorry
I would reopen the post on huggingface/transformers | [This colab notebook](https://colab.research.google.com/drive/151gKyo0YIwnlznrOHst23oYH_a3mAe3Z?usp=sharing) implements a token classification input pipeline extending the logic from [this hugging example](https://huggingface.co/transformers/custom_datasets.html#tok-ner).
The pipeline works fine with most instance in different languages, but unfortunately, [the Japanese Kana ligature (a form of abbreviation? I don't know Japanese well)](https://en.wikipedia.org/wiki/Kana_ligature) break the alignment of `return_offsets_mapping`:
Without the try catch block, it riase `ValueError: NumPy boolean array indexing assignment cannot assign 88 input values to the 87 output values where the mask is true`, example shown here [(another colab notebook)](https://colab.research.google.com/drive/1MmOqf3ppzzdKKyMWkn0bJy6DqzOO0SSm?usp=sharing)
It is clear that the normalizer is the process that break the alignment, as it is observed that `tokenizer._tokenizer.normalizer.normalize_str('ヿ')` return 'コト'.
One workaround is to include `tokenizer._tokenizer.normalizer.normalize_str` before the tokenizer preprocessing pipeline, which is also provided in the [first colab notebook](https://colab.research.google.com/drive/151gKyo0YIwnlznrOHst23oYH_a3mAe3Z?usp=sharing) with the name `udposTestDatasetWorkaround`.
I guess similar logics should be included inside the tokenizer and the offsets_mapping generation process such that user don't need to include them in their code. But I don't understand the code of tokenizer well that I think I am not able to do this.
p.s.
**I am using my own dataset building script in the provided example, but the script should be equivalent to the changes made by this [update](https://github.com/huggingface/datasets/pull/2466)**
`get_dataset `is just a simple wrapping for `load_dataset`
and the `tokenizer` is just `XLMRobertaTokenizerFast.from_pretrained("xlm-roberta-large")` | 30 | Tokenizer's normalization preprocessor cause misalignment in return_offsets_mapping for tokenizer classification task
[This colab notebook](https://colab.research.google.com/drive/151gKyo0YIwnlznrOHst23oYH_a3mAe3Z?usp=sharing) implements a token classification input pipeline extending the logic from [this hugging example](https://huggingface.co/transformers/custom_datasets.html#tok-ner).
The pipeline works fine with most instance in different languages, but unfortunately, [the Japanese Kana ligature (a form of abbreviation? I don't know Japanese well)](https://en.wikipedia.org/wiki/Kana_ligature) break the alignment of `return_offsets_mapping`:
Without the try catch block, it riase `ValueError: NumPy boolean array indexing assignment cannot assign 88 input values to the 87 output values where the mask is true`, example shown here [(another colab notebook)](https://colab.research.google.com/drive/1MmOqf3ppzzdKKyMWkn0bJy6DqzOO0SSm?usp=sharing)
It is clear that the normalizer is the process that break the alignment, as it is observed that `tokenizer._tokenizer.normalizer.normalize_str('ヿ')` return 'コト'.
One workaround is to include `tokenizer._tokenizer.normalizer.normalize_str` before the tokenizer preprocessing pipeline, which is also provided in the [first colab notebook](https://colab.research.google.com/drive/151gKyo0YIwnlznrOHst23oYH_a3mAe3Z?usp=sharing) with the name `udposTestDatasetWorkaround`.
I guess similar logics should be included inside the tokenizer and the offsets_mapping generation process such that user don't need to include them in their code. But I don't understand the code of tokenizer well that I think I am not able to do this.
p.s.
**I am using my own dataset building script in the provided example, but the script should be equivalent to the changes made by this [update](https://github.com/huggingface/datasets/pull/2466)**
`get_dataset `is just a simple wrapping for `load_dataset`
and the `tokenizer` is just `XLMRobertaTokenizerFast.from_pretrained("xlm-roberta-large")`
> Hi @jerryIsHere, thanks for reporting the issue. But are you sure this is a bug in HuggingFace **Datasets**?
Oh, I am sorry
I would reopen the post on huggingface/transformers |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | I'm currently adding it, the entire dataset is quite big around 30 GB so I add splits separately. You can take a look here https://huggingface.co/datasets/merve/coco | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 25 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
I'm currently adding it, the entire dataset is quite big around 30 GB so I add splits separately. You can take a look here https://huggingface.co/datasets/merve/coco |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | I talked to @lhoestq and it's best if I download this dataset through TensorFlow datasets instead, so I'll be implementing that one really soon.
@NielsRogge | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 25 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
I talked to @lhoestq and it's best if I download this dataset through TensorFlow datasets instead, so I'll be implementing that one really soon.
@NielsRogge |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | I started adding COCO, will be done tomorrow EOD
my work so far https://github.com/merveenoyan/datasets (my fork) | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 16 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
I started adding COCO, will be done tomorrow EOD
my work so far https://github.com/merveenoyan/datasets (my fork) |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | Hi Merve @merveenoyan , thank you so much for your great contribution! May I ask about the current progress of your implementation? Cuz I see the pull request is still in progess here. Or can I just run the COCO scripts in your fork repo? | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 45 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Hi Merve @merveenoyan , thank you so much for your great contribution! May I ask about the current progress of your implementation? Cuz I see the pull request is still in progess here. Or can I just run the COCO scripts in your fork repo? |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | Hello @yixuanren I had another prioritized project about to be merged, but I'll start continuing today will finish up soon. | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 20 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Hello @yixuanren I had another prioritized project about to be merged, but I'll start continuing today will finish up soon. |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | > Hello @yixuanren I had another prioritized project about to be merged, but I'll start continuing today will finish up soon.
It's really nice of you!! I see you've commited another version just now | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 34 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
> Hello @yixuanren I had another prioritized project about to be merged, but I'll start continuing today will finish up soon.
It's really nice of you!! I see you've commited another version just now |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | Hi @NielsRogge and @merveenoyan, did you find a way to load a dataset with COCO annotations to HF's hub?
I have a panoptic segmentation dataset in COCO format and would like to share it with the community.
Thanks in advance :) | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 41 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Hi @NielsRogge and @merveenoyan, did you find a way to load a dataset with COCO annotations to HF's hub?
I have a panoptic segmentation dataset in COCO format and would like to share it with the community.
Thanks in advance :) |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | The COCO format is not supported out of the box in the HF's hub - you'd need to reformat it to an [ImageFolder](https://huggingface.co/docs/datasets/image_dataset#imagefolder) with metadata format, or write a [loading script](https://huggingface.co/docs/datasets/image_dataset#loading-script) | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 31 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
The COCO format is not supported out of the box in the HF's hub - you'd need to reformat it to an [ImageFolder](https://huggingface.co/docs/datasets/image_dataset#imagefolder) with metadata format, or write a [loading script](https://huggingface.co/docs/datasets/image_dataset#loading-script) |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | > The COCO format is not supported out of the box in the HF's hub - you'd need to reformat it to an [ImageFolder](https://huggingface.co/docs/datasets/image_dataset#imagefolder) with metadata format, or write a [loading script](https://huggingface.co/docs/datasets/image_dataset#loading-script)
Hi @lhoestq , thank you for your quick reply.
I've correctly created a metadata.jsonl file for a dataset with instance segmentation annotations [here](https://huggingface.co/datasets/lombardata/data_2017)
but do not understand how I can integrate panoptic annotations with the metadata format of ImageFolder datasets. The "problem" with panoptic annotations is that we have a folder with images, a json file with annotations and another folder with png annotations.
I checked between all the datasets already published on HuggingFace and, the only one who has uploaded a correct panoptic dataset is @NielsRogge [here](https://huggingface.co/datasets/nielsr/coco-panoptic-val2017) and [here](https://huggingface.co/datasets/nielsr/ade20k-panoptic-demo). Indeed he accomplished to have three fields :
1.image (image)
2.label (image)
3.segments_info (list)
but I not find the corresponding code that allows to upload a panoptic dataset from this 3 sources.
Can you please share an example code?
Thanks ! | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 163 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
> The COCO format is not supported out of the box in the HF's hub - you'd need to reformat it to an [ImageFolder](https://huggingface.co/docs/datasets/image_dataset#imagefolder) with metadata format, or write a [loading script](https://huggingface.co/docs/datasets/image_dataset#loading-script)
Hi @lhoestq , thank you for your quick reply.
I've correctly created a metadata.jsonl file for a dataset with instance segmentation annotations [here](https://huggingface.co/datasets/lombardata/data_2017)
but do not understand how I can integrate panoptic annotations with the metadata format of ImageFolder datasets. The "problem" with panoptic annotations is that we have a folder with images, a json file with annotations and another folder with png annotations.
I checked between all the datasets already published on HuggingFace and, the only one who has uploaded a correct panoptic dataset is @NielsRogge [here](https://huggingface.co/datasets/nielsr/coco-panoptic-val2017) and [here](https://huggingface.co/datasets/nielsr/ade20k-panoptic-demo). Indeed he accomplished to have three fields :
1.image (image)
2.label (image)
3.segments_info (list)
but I not find the corresponding code that allows to upload a panoptic dataset from this 3 sources.
Can you please share an example code?
Thanks ! |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | Both were uploaded using `ds.push_to_hub()` :)
You can get a Dataset from a python dictionary using `ds = Dataset.from_dict(...)` and casts the paths to images to the `Image()` type using `ds = ds.cast_column("image", Image())`.
```python
from datasets import Dataset, Image
ds = Dataset.from_dict(...)
ds = ds.cast_column("image", Image())
ds = ds.cast_column("label", Image())
ds.push_to_hub(...)
``` | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 53 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Both were uploaded using `ds.push_to_hub()` :)
You can get a Dataset from a python dictionary using `ds = Dataset.from_dict(...)` and casts the paths to images to the `Image()` type using `ds = ds.cast_column("image", Image())`.
```python
from datasets import Dataset, Image
ds = Dataset.from_dict(...)
ds = ds.cast_column("image", Image())
ds = ds.cast_column("label", Image())
ds.push_to_hub(...)
``` |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | > Both were uploaded using `ds.push_to_hub()` :)
>
> You can get a Dataset from a python dictionary using `ds = Dataset.from_dict(...)` and casts the paths to images to the `Image()` type using `ds = ds.cast_column("image", Image())`.
>
> ```python
> from datasets import Dataset, Image
>
> ds = Dataset.from_dict(...)
> ds = ds.cast_column("image", Image())
> ds = ds.cast_column("label", Image())
> ds.push_to_hub(...)
> ```
Thank you very much @lhoestq , I succesfully created a hf dataset [here](https://huggingface.co/datasets/lombardata/panoptic_2023_06_21) with the two fields :
1.image (image)
2.label (image)
following your suggestions. Now still remain the problem of uploading **segments_info** information to the dataset.
There is a function that easily imports the _panoptic_coco_annotation.json_ file to a segment_info field?
I think we must define a **list_of_segment**, i.e. a list of lists of this type :
```python
[ { "area": 214858, "bbox": [ 0, 0, 511, 760 ], "category_id": 0, "id": 7895160, "iscrowd": 0 }, { "area": 73067, "bbox": [ 98, 719, 413, 253 ], "category_id": 3, "id": 3289680, "iscrowd": 0 }, { "area": 832, "bbox": [ 53, 0, 101, 16 ], "category_id": 5, "id": 5273720, "iscrowd": 0 }, { "area": 70668, "bbox": [ 318, 60, 191, 392 ], "category_id": 8, "id": 15132390, "iscrowd": 0 }, { "area": 32696, "bbox": [ 0, 100, 78, 872 ], "category_id": 18, "id": 472063, "iscrowd": 0 }, { "area": 76045, "bbox": [ 42, 48, 264, 924 ], "category_id": 37, "id": 16713830, "iscrowd": 0 }, { "area": 27103, "bbox": [ 288, 482, 216, 306 ], "category_id": 47, "id": 16753408, "iscrowd": 0 } ]
```
and then apply again the **cast_column** function [here](https://github.com/huggingface/datasets/blob/2.13.0/src/datasets/arrow_dataset.py#L2060) but with a list as a second argument, like :
```python
from datasets import Dataset, Image
ds = ds.cast_column("image", Image())
ds = ds.cast_column("label", Image())
ds = ds.cast_column("segments_info", list)
```
but I do not see how to transfer the information of the _panoptic_coco_annotation.json_ to a list of lists of this type :
```python
[ { "area": 214858, "bbox": [ 0, 0, 511, 760 ], "category_id": 0, "id": 7895160, "iscrowd": 0 }, { "area": 73067, "bbox": [ 98, 719, 413, 253 ], "category_id": 3, "id": 3289680, "iscrowd": 0 }, { "area": 832, "bbox": [ 53, 0, 101, 16 ], "category_id": 5, "id": 5273720, "iscrowd": 0 }, { "area": 70668, "bbox": [ 318, 60, 191, 392 ], "category_id": 8, "id": 15132390, "iscrowd": 0 }, { "area": 32696, "bbox": [ 0, 100, 78, 872 ], "category_id": 18, "id": 472063, "iscrowd": 0 }, { "area": 76045, "bbox": [ 42, 48, 264, 924 ], "category_id": 37, "id": 16713830, "iscrowd": 0 }, { "area": 27103, "bbox": [ 288, 482, 216, 306 ], "category_id": 47, "id": 16753408, "iscrowd": 0 } ]
```
like @NielsRogge has done [here](https://huggingface.co/datasets/nielsr/coco-panoptic-val2017) and [here](https://huggingface.co/datasets/nielsr/ade20k-panoptic-demo).
Thank you again for your help and have a good day ! | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 456 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
> Both were uploaded using `ds.push_to_hub()` :)
>
> You can get a Dataset from a python dictionary using `ds = Dataset.from_dict(...)` and casts the paths to images to the `Image()` type using `ds = ds.cast_column("image", Image())`.
>
> ```python
> from datasets import Dataset, Image
>
> ds = Dataset.from_dict(...)
> ds = ds.cast_column("image", Image())
> ds = ds.cast_column("label", Image())
> ds.push_to_hub(...)
> ```
Thank you very much @lhoestq , I succesfully created a hf dataset [here](https://huggingface.co/datasets/lombardata/panoptic_2023_06_21) with the two fields :
1.image (image)
2.label (image)
following your suggestions. Now still remain the problem of uploading **segments_info** information to the dataset.
There is a function that easily imports the _panoptic_coco_annotation.json_ file to a segment_info field?
I think we must define a **list_of_segment**, i.e. a list of lists of this type :
```python
[ { "area": 214858, "bbox": [ 0, 0, 511, 760 ], "category_id": 0, "id": 7895160, "iscrowd": 0 }, { "area": 73067, "bbox": [ 98, 719, 413, 253 ], "category_id": 3, "id": 3289680, "iscrowd": 0 }, { "area": 832, "bbox": [ 53, 0, 101, 16 ], "category_id": 5, "id": 5273720, "iscrowd": 0 }, { "area": 70668, "bbox": [ 318, 60, 191, 392 ], "category_id": 8, "id": 15132390, "iscrowd": 0 }, { "area": 32696, "bbox": [ 0, 100, 78, 872 ], "category_id": 18, "id": 472063, "iscrowd": 0 }, { "area": 76045, "bbox": [ 42, 48, 264, 924 ], "category_id": 37, "id": 16713830, "iscrowd": 0 }, { "area": 27103, "bbox": [ 288, 482, 216, 306 ], "category_id": 47, "id": 16753408, "iscrowd": 0 } ]
```
and then apply again the **cast_column** function [here](https://github.com/huggingface/datasets/blob/2.13.0/src/datasets/arrow_dataset.py#L2060) but with a list as a second argument, like :
```python
from datasets import Dataset, Image
ds = ds.cast_column("image", Image())
ds = ds.cast_column("label", Image())
ds = ds.cast_column("segments_info", list)
```
but I do not see how to transfer the information of the _panoptic_coco_annotation.json_ to a list of lists of this type :
```python
[ { "area": 214858, "bbox": [ 0, 0, 511, 760 ], "category_id": 0, "id": 7895160, "iscrowd": 0 }, { "area": 73067, "bbox": [ 98, 719, 413, 253 ], "category_id": 3, "id": 3289680, "iscrowd": 0 }, { "area": 832, "bbox": [ 53, 0, 101, 16 ], "category_id": 5, "id": 5273720, "iscrowd": 0 }, { "area": 70668, "bbox": [ 318, 60, 191, 392 ], "category_id": 8, "id": 15132390, "iscrowd": 0 }, { "area": 32696, "bbox": [ 0, 100, 78, 872 ], "category_id": 18, "id": 472063, "iscrowd": 0 }, { "area": 76045, "bbox": [ 42, 48, 264, 924 ], "category_id": 37, "id": 16713830, "iscrowd": 0 }, { "area": 27103, "bbox": [ 288, 482, 216, 306 ], "category_id": 47, "id": 16753408, "iscrowd": 0 } ]
```
like @NielsRogge has done [here](https://huggingface.co/datasets/nielsr/coco-panoptic-val2017) and [here](https://huggingface.co/datasets/nielsr/ade20k-panoptic-demo).
Thank you again for your help and have a good day ! |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | You can pass this data in .from_dict() - no need to cast anything for this column
```python
ds = Dataset.from_dict({
"image": [...],
"label": [...],
"segments_info": [...],
)}
```
where `segments_info` is the list of the segment_infos of all the examples in the dataset, and therefore is a list of lists of dicts. | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 52 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
You can pass this data in .from_dict() - no need to cast anything for this column
```python
ds = Dataset.from_dict({
"image": [...],
"label": [...],
"segments_info": [...],
)}
```
where `segments_info` is the list of the segment_infos of all the examples in the dataset, and therefore is a list of lists of dicts. |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | > You can pass this data in .from_dict() - no need to cast anything for this column
>
> ```python
> ds = Dataset.from_dict({
> "image": [...],
> "label": [...],
> "segments_info": [...],
> )}
> ```
>
> where `segments_info` is the list of the segment_infos of all the examples in the dataset, and therefore is a list of lists of dicts.
Thank you for the quick reply @lhoestq , but then how to generate the `segments_info` list of lists of dicts starting from a _panoptic_coco_annotation.json_ file ?
| ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 89 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
> You can pass this data in .from_dict() - no need to cast anything for this column
>
> ```python
> ds = Dataset.from_dict({
> "image": [...],
> "label": [...],
> "segments_info": [...],
> )}
> ```
>
> where `segments_info` is the list of the segment_infos of all the examples in the dataset, and therefore is a list of lists of dicts.
Thank you for the quick reply @lhoestq , but then how to generate the `segments_info` list of lists of dicts starting from a _panoptic_coco_annotation.json_ file ?
|
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | You read the JSON file and transform the data yourself. I don't think there's an automatic converter somewhere | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 18 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
You read the JSON file and transform the data yourself. I don't think there's an automatic converter somewhere |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | > You read the JSON file and transform the data yourself. I don't think there's an automatic converter somewhere
Perfect, I've done it and succesfully uploaded a new dataset [here](https://huggingface.co/datasets/lombardata/panoptic_2023_06_22), but I've (I hope) a last problem.
The dataset has currently 302 images and, when I upload it to the hub, only the first page of images is correctly uploaded.
When I try to see the second/third/fourth page of items of my dataset, I can see that the fields **segments_info** and **image_name** are correctly uploaded, while the images are not (the "null" string is shown everywhere).
I've checked the path of images that are not uploaded and they exists, is there a problem with the size of the dataset ?
How can I upload the whole dataset to the hub ?
Thank you again @lhoestq and have a good day ! | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 141 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
> You read the JSON file and transform the data yourself. I don't think there's an automatic converter somewhere
Perfect, I've done it and succesfully uploaded a new dataset [here](https://huggingface.co/datasets/lombardata/panoptic_2023_06_22), but I've (I hope) a last problem.
The dataset has currently 302 images and, when I upload it to the hub, only the first page of images is correctly uploaded.
When I try to see the second/third/fourth page of items of my dataset, I can see that the fields **segments_info** and **image_name** are correctly uploaded, while the images are not (the "null" string is shown everywhere).
I've checked the path of images that are not uploaded and they exists, is there a problem with the size of the dataset ?
How can I upload the whole dataset to the hub ?
Thank you again @lhoestq and have a good day ! |
https://github.com/huggingface/datasets/issues/2526 | Add COCO datasets | Awesome ! Your dataset looks all good 🤗
The `null` in the viewer is a bug on our side, let me investigate | ## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 22 | Add COCO datasets
## Adding a Dataset
- **Name:** COCO
- **Description:** COCO is a large-scale object detection, segmentation, and captioning dataset.
- **Paper + website:** https://cocodataset.org/#home
- **Data:** https://cocodataset.org/#download
- **Motivation:** It would be great to have COCO available in HuggingFace datasets, as we are moving beyond just text. COCO includes multi-modalities (images + text), as well as a huge amount of images annotated with objects, segmentation masks, keypoints etc., on which models like DETR (which I recently added to HuggingFace Transformers) are trained. Currently, one needs to download everything from the website and place it in a local folder, but it would be much easier if we can directly access it through the datasets API.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Awesome ! Your dataset looks all good 🤗
The `null` in the viewer is a bug on our side, let me investigate |
https://github.com/huggingface/datasets/issues/2522 | Documentation Mistakes in Dataset: emotion | Hi,
this issue has been already reported in the dataset repo (https://github.com/dair-ai/emotion_dataset/issues/2), so this is a bug on their side. | As per documentation,
Dataset: emotion
Homepage: https://github.com/dair-ai/emotion_dataset
Dataset: https://github.com/huggingface/datasets/blob/master/datasets/emotion/emotion.py
Permalink: https://huggingface.co/datasets/viewer/?dataset=emotion
Emotion is a dataset of English Twitter messages with eight basic emotions: anger, anticipation, disgust, fear, joy, sadness, surprise, and trust. For more detailed information please refer to the paper.
But when we view the data, there are only 6 emotions, anger, fear, joy, sadness, surprise, and trust. | 20 | Documentation Mistakes in Dataset: emotion
As per documentation,
Dataset: emotion
Homepage: https://github.com/dair-ai/emotion_dataset
Dataset: https://github.com/huggingface/datasets/blob/master/datasets/emotion/emotion.py
Permalink: https://huggingface.co/datasets/viewer/?dataset=emotion
Emotion is a dataset of English Twitter messages with eight basic emotions: anger, anticipation, disgust, fear, joy, sadness, surprise, and trust. For more detailed information please refer to the paper.
But when we view the data, there are only 6 emotions, anger, fear, joy, sadness, surprise, and trust.
Hi,
this issue has been already reported in the dataset repo (https://github.com/dair-ai/emotion_dataset/issues/2), so this is a bug on their side. |
https://github.com/huggingface/datasets/issues/2522 | Documentation Mistakes in Dataset: emotion | The documentation has another bug in the dataset card [here](https://huggingface.co/datasets/emotion).
In the dataset summary **six** emotions are mentioned: *"six basic emotions: anger, fear, joy, love, sadness, and surprise"*, however, in the datafields section we have only **five**:
```
label: a classification label, with possible values including sadness (0), joy (1), love (2), anger (3), fear (4).
``` | As per documentation,
Dataset: emotion
Homepage: https://github.com/dair-ai/emotion_dataset
Dataset: https://github.com/huggingface/datasets/blob/master/datasets/emotion/emotion.py
Permalink: https://huggingface.co/datasets/viewer/?dataset=emotion
Emotion is a dataset of English Twitter messages with eight basic emotions: anger, anticipation, disgust, fear, joy, sadness, surprise, and trust. For more detailed information please refer to the paper.
But when we view the data, there are only 6 emotions, anger, fear, joy, sadness, surprise, and trust. | 57 | Documentation Mistakes in Dataset: emotion
As per documentation,
Dataset: emotion
Homepage: https://github.com/dair-ai/emotion_dataset
Dataset: https://github.com/huggingface/datasets/blob/master/datasets/emotion/emotion.py
Permalink: https://huggingface.co/datasets/viewer/?dataset=emotion
Emotion is a dataset of English Twitter messages with eight basic emotions: anger, anticipation, disgust, fear, joy, sadness, surprise, and trust. For more detailed information please refer to the paper.
But when we view the data, there are only 6 emotions, anger, fear, joy, sadness, surprise, and trust.
The documentation has another bug in the dataset card [here](https://huggingface.co/datasets/emotion).
In the dataset summary **six** emotions are mentioned: *"six basic emotions: anger, fear, joy, love, sadness, and surprise"*, however, in the datafields section we have only **five**:
```
label: a classification label, with possible values including sadness (0), joy (1), love (2), anger (3), fear (4).
``` |
https://github.com/huggingface/datasets/issues/2522 | Documentation Mistakes in Dataset: emotion | @GDGauravDutta the dataset author replied in their issue: https://github.com/dair-ai/emotion_dataset/issues/2
> The dataset released is a preprocessed six emotion variant released for educational and research purposes.
@albertovilla the dataset card was fixed with 6 emotions. | As per documentation,
Dataset: emotion
Homepage: https://github.com/dair-ai/emotion_dataset
Dataset: https://github.com/huggingface/datasets/blob/master/datasets/emotion/emotion.py
Permalink: https://huggingface.co/datasets/viewer/?dataset=emotion
Emotion is a dataset of English Twitter messages with eight basic emotions: anger, anticipation, disgust, fear, joy, sadness, surprise, and trust. For more detailed information please refer to the paper.
But when we view the data, there are only 6 emotions, anger, fear, joy, sadness, surprise, and trust. | 34 | Documentation Mistakes in Dataset: emotion
As per documentation,
Dataset: emotion
Homepage: https://github.com/dair-ai/emotion_dataset
Dataset: https://github.com/huggingface/datasets/blob/master/datasets/emotion/emotion.py
Permalink: https://huggingface.co/datasets/viewer/?dataset=emotion
Emotion is a dataset of English Twitter messages with eight basic emotions: anger, anticipation, disgust, fear, joy, sadness, surprise, and trust. For more detailed information please refer to the paper.
But when we view the data, there are only 6 emotions, anger, fear, joy, sadness, surprise, and trust.
@GDGauravDutta the dataset author replied in their issue: https://github.com/dair-ai/emotion_dataset/issues/2
> The dataset released is a preprocessed six emotion variant released for educational and research purposes.
@albertovilla the dataset card was fixed with 6 emotions. |
https://github.com/huggingface/datasets/issues/2520 | Datasets with tricky task templates | The `task_templates` API is deprecated in favor of the `train-eval-index` YAML field, so I'm closing this issue. | I'm collecting a list of datasets here that don't follow the "standard" taxonomy and require further investigation to implement task templates for.
## Text classification
* [hatexplain](https://huggingface.co/datasets/hatexplain): ostensibly a form of text classification, but not in the standard `(text, target)` format and each sample appears to be tokenized.
* [muchocine](https://huggingface.co/datasets/muchocine): contains two candidate text columns (long-form and summary) which in principle requires two `TextClassification` templates which is not currently supported | 17 | Datasets with tricky task templates
I'm collecting a list of datasets here that don't follow the "standard" taxonomy and require further investigation to implement task templates for.
## Text classification
* [hatexplain](https://huggingface.co/datasets/hatexplain): ostensibly a form of text classification, but not in the standard `(text, target)` format and each sample appears to be tokenized.
* [muchocine](https://huggingface.co/datasets/muchocine): contains two candidate text columns (long-form and summary) which in principle requires two `TextClassification` templates which is not currently supported
The `task_templates` API is deprecated in favor of the `train-eval-index` YAML field, so I'm closing this issue. |
https://github.com/huggingface/datasets/issues/2516 | datasets.map pickle issue resulting in invalid mapping function | Hi ! `map` calls `__getstate__` using `dill` to hash your map function. This is used by the caching mechanism to recover previously computed results. That's why you don't see any `__setstate__` call.
Why do you change an attribute of your tokenizer when `__getstate__` is called ? | I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
| 46 | datasets.map pickle issue resulting in invalid mapping function
I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
Hi ! `map` calls `__getstate__` using `dill` to hash your map function. This is used by the caching mechanism to recover previously computed results. That's why you don't see any `__setstate__` call.
Why do you change an attribute of your tokenizer when `__getstate__` is called ? |
https://github.com/huggingface/datasets/issues/2516 | datasets.map pickle issue resulting in invalid mapping function | @lhoestq because if I try to pickle my custom tokenizer (it contains a pure python pretokenization step in an otherwise rust backed tokenizer) I get
> Exception: Error while attempting to pickle Tokenizer: Custom PreTokenizer cannot be serialized
So I remove the Custom PreTokenizer in `__getstate__` and then restore it in `__setstate__` (since it doesn't contain any state). This is what my `__getstate__` / `__setstate__` looks like:
def __getstate__(self):
"""
Removes pre_tokenizer since it cannot be pickled
"""
logger.debug("Copy state dict")
out = self.__dict__.copy()
logger.debug("Detaching pre_tokenizer")
out['_tokenizer'].pre_tokenizer = tokenizers.pre_tokenizers.Sequence([])
return out
def __setstate__(self, d):
"""
Reinstates pre_tokenizer
"""
logger.debug("Reattaching pre_tokenizer")
self.__dict__ = d
self.backend_tokenizer.pre_tokenizer = self._pre_tokenizer()
If this is the case can you think of another way of avoiding my issue? | I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
| 121 | datasets.map pickle issue resulting in invalid mapping function
I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
@lhoestq because if I try to pickle my custom tokenizer (it contains a pure python pretokenization step in an otherwise rust backed tokenizer) I get
> Exception: Error while attempting to pickle Tokenizer: Custom PreTokenizer cannot be serialized
So I remove the Custom PreTokenizer in `__getstate__` and then restore it in `__setstate__` (since it doesn't contain any state). This is what my `__getstate__` / `__setstate__` looks like:
def __getstate__(self):
"""
Removes pre_tokenizer since it cannot be pickled
"""
logger.debug("Copy state dict")
out = self.__dict__.copy()
logger.debug("Detaching pre_tokenizer")
out['_tokenizer'].pre_tokenizer = tokenizers.pre_tokenizers.Sequence([])
return out
def __setstate__(self, d):
"""
Reinstates pre_tokenizer
"""
logger.debug("Reattaching pre_tokenizer")
self.__dict__ = d
self.backend_tokenizer.pre_tokenizer = self._pre_tokenizer()
If this is the case can you think of another way of avoiding my issue? |
https://github.com/huggingface/datasets/issues/2516 | datasets.map pickle issue resulting in invalid mapping function | Actually, maybe I need to deep copy `self.__dict__`? That way `self` isn't modified. That was my intention and I thought it was working - I'll double-check after the weekend. | I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
| 29 | datasets.map pickle issue resulting in invalid mapping function
I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
Actually, maybe I need to deep copy `self.__dict__`? That way `self` isn't modified. That was my intention and I thought it was working - I'll double-check after the weekend. |
https://github.com/huggingface/datasets/issues/2516 | datasets.map pickle issue resulting in invalid mapping function | Doing a deep copy results in the warning:
> 06/20/2021 16:02:15 - WARNING - datasets.fingerprint - Parameter 'function'=<function tokenize_function at 0x7f1e95f05d40> of the transform datasets.arrow_dataset.Dataset._map_single couldn't be hashed properly, a random hash was used instead. Make sure your transforms and parameters are serializable with pickle or dill for the dataset fingerprinting and caching to work. If you reuse this transform, the caching mechanism will consider it to be different from the previous calls and recompute everything. This warning is only showed once. Subsequent hashing failures won't be showed.
```
def __getstate__(self):
"""
Removes pre_tokenizer since it cannot be pickled
"""
logger.debug("Copy state dict")
out = copy.deepcopy(self.__dict__)
logger.debug("Detaching pre_tokenizer")
out['_tokenizer'].pre_tokenizer = tokenizers.pre_tokenizers.Sequence([])
return out
``` | I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
| 114 | datasets.map pickle issue resulting in invalid mapping function
I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
Doing a deep copy results in the warning:
> 06/20/2021 16:02:15 - WARNING - datasets.fingerprint - Parameter 'function'=<function tokenize_function at 0x7f1e95f05d40> of the transform datasets.arrow_dataset.Dataset._map_single couldn't be hashed properly, a random hash was used instead. Make sure your transforms and parameters are serializable with pickle or dill for the dataset fingerprinting and caching to work. If you reuse this transform, the caching mechanism will consider it to be different from the previous calls and recompute everything. This warning is only showed once. Subsequent hashing failures won't be showed.
```
def __getstate__(self):
"""
Removes pre_tokenizer since it cannot be pickled
"""
logger.debug("Copy state dict")
out = copy.deepcopy(self.__dict__)
logger.debug("Detaching pre_tokenizer")
out['_tokenizer'].pre_tokenizer = tokenizers.pre_tokenizers.Sequence([])
return out
``` |
https://github.com/huggingface/datasets/issues/2516 | datasets.map pickle issue resulting in invalid mapping function | Looks like there is still an object that is not pickable in your `tokenize_function` function.
You can test if an object can be pickled and hashed by using
```python
from datasets.fingerprint import Hasher
Hasher.hash(my_object)
```
Under the hood it pickles the object to compute its hash, so it calls `__getstate__` when applicable. | I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
| 52 | datasets.map pickle issue resulting in invalid mapping function
I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
Looks like there is still an object that is not pickable in your `tokenize_function` function.
You can test if an object can be pickled and hashed by using
```python
from datasets.fingerprint import Hasher
Hasher.hash(my_object)
```
Under the hood it pickles the object to compute its hash, so it calls `__getstate__` when applicable. |
https://github.com/huggingface/datasets/issues/2516 | datasets.map pickle issue resulting in invalid mapping function | I figured it out, the problem is deep copy itself uses pickle (unless you implement `__deepcopy__`). So when I changed `__getstate__` it started throwing an error.
I'm sure there's a better way of doing this, but in order to return the `__dict__` without the non-pikelable pre-tokeniser and without modifying self I removed the pre-tokenizers, did a deep copy and then re-generated it.
It does work - although I noticed Hasher doesn't call `__hash__` if the object being hashed implements it which I feel it should? If it did I could return a hash of the tokenizers.json file instead.
```
def __getstate__(self):
"""
Removes pre_tokenizer since it cannot be pickled
"""
logger.debug("Copy state dict")
self.backend_tokenizer.pre_tokenizer = tokenizers.pre_tokenizers.Sequence([])
out = copy.deepcopy(self.__dict__) #self.__dict__.copy()
self.backend_tokenizer.pre_tokenizer = self._pre_tokenizer()
return out
```
| I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
| 126 | datasets.map pickle issue resulting in invalid mapping function
I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
I figured it out, the problem is deep copy itself uses pickle (unless you implement `__deepcopy__`). So when I changed `__getstate__` it started throwing an error.
I'm sure there's a better way of doing this, but in order to return the `__dict__` without the non-pikelable pre-tokeniser and without modifying self I removed the pre-tokenizers, did a deep copy and then re-generated it.
It does work - although I noticed Hasher doesn't call `__hash__` if the object being hashed implements it which I feel it should? If it did I could return a hash of the tokenizers.json file instead.
```
def __getstate__(self):
"""
Removes pre_tokenizer since it cannot be pickled
"""
logger.debug("Copy state dict")
self.backend_tokenizer.pre_tokenizer = tokenizers.pre_tokenizers.Sequence([])
out = copy.deepcopy(self.__dict__) #self.__dict__.copy()
self.backend_tokenizer.pre_tokenizer = self._pre_tokenizer()
return out
```
|
https://github.com/huggingface/datasets/issues/2516 | datasets.map pickle issue resulting in invalid mapping function | I'm glad you figured something out :)
Regarding hashing: we're not using hashing for the same purpose as the python `__hash__` purpose (which is in general for dictionary lookups). For example it is allowed for python hashing to not return the same hash across sessions, while our hashing must return the same hashes across sessions for the caching to work properly. | I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
| 61 | datasets.map pickle issue resulting in invalid mapping function
I trained my own tokenizer, and I needed to use a python custom class. Because of this I have to detach the custom step before saving and reattach after restore. I did this using the standard pickle `__get_state__` / `__set_state__` mechanism. I think it's correct but it fails when I use it inside a function which is mapped to a dataset, i.e. in the manner of run_mlm.py and other huggingface scripts.
The following reproduces the issue - most likely I'm missing something
A simulated tokeniser which can be pickled
```
class CustomTokenizer:
def __init__(self):
self.state = "init"
def __getstate__(self):
print("__getstate__ called")
out = self.__dict__.copy()
self.state = "pickled"
return out
def __setstate__(self, d):
print("__setstate__ called")
self.__dict__ = d
self.state = "restored"
tokenizer = CustomTokenizer()
```
Test that it actually works - prints "__getstate__ called" and "__setstate__ called"
```
import pickle
serialized = pickle.dumps(tokenizer)
restored = pickle.loads(serialized)
assert restored.state == "restored"
```
Simulate a function that tokenises examples, when dataset.map is called, this function
```
def tokenize_function(examples):
assert tokenizer.state == "restored" # this shouldn't fail but it does
output = tokenizer(examples) # this will fail as tokenizer isn't really a tokenizer
return output
```
Use map to simulate tokenization
```
import glob
from datasets import load_dataset
assert tokenizer.state == "restored"
train_files = glob.glob('train*.csv')
validation_files = glob.glob('validation*.csv')
datasets = load_dataset("csv", data_files=dict(train=train_files, validation=validation_files))
tokenized_datasets = datasets.map(
tokenize_function,
batched=True,
)
```
What's happening is I can see that __getstate__ is called but not __setstate__, so the state of `tokenize_function` is invalid at the point that it's actually executed. This doesn't matter as far as I can see for the standard tokenizers as they don't use __getstate__ / __setstate__. I'm not sure if there's another hook I'm supposed to implement as well?
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-22-a2aef4f74aaa> in <module>
8 tokenized_datasets = datasets.map(
9 tokenize_function,
---> 10 batched=True,
11 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in map(self, function, with_indices, input_columns, batched, batch_size, remove_columns, keep_in_memory, load_from_cache_file, cache_file_names, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, desc)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/dataset_dict.py in <dictcomp>(.0)
487 desc=desc,
488 )
--> 489 for k, dataset in self.items()
490 }
491 )
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in map(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, num_proc, suffix_template, new_fingerprint, desc)
1633 fn_kwargs=fn_kwargs,
1634 new_fingerprint=new_fingerprint,
-> 1635 desc=desc,
1636 )
1637 else:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in wrapper(*args, **kwargs)
184 }
185 # apply actual function
--> 186 out: Union["Dataset", "DatasetDict"] = func(self, *args, **kwargs)
187 datasets: List["Dataset"] = list(out.values()) if isinstance(out, dict) else [out]
188 # re-apply format to the output
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/fingerprint.py in wrapper(*args, **kwargs)
395 # Call actual function
396
--> 397 out = func(self, *args, **kwargs)
398
399 # Update fingerprint of in-place transforms + update in-place history of transforms
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in _map_single(self, function, with_indices, input_columns, batched, batch_size, drop_last_batch, remove_columns, keep_in_memory, load_from_cache_file, cache_file_name, writer_batch_size, features, disable_nullable, fn_kwargs, new_fingerprint, rank, offset, desc)
1961 indices,
1962 check_same_num_examples=len(input_dataset.list_indexes()) > 0,
-> 1963 offset=offset,
1964 )
1965 except NumExamplesMismatch:
~/.pyenv/versions/3.7.6/envs/xxx/lib/python3.7/site-packages/datasets/arrow_dataset.py in apply_function_on_filtered_inputs(inputs, indices, check_same_num_examples, offset)
1853 effective_indices = [i + offset for i in indices] if isinstance(indices, list) else indices + offset
1854 processed_inputs = (
-> 1855 function(*fn_args, effective_indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)
1856 )
1857 if update_data is None:
<ipython-input-21-8ee4a8ba5b1b> in tokenize_function(examples)
1 def tokenize_function(examples):
----> 2 assert tokenizer.state == "restored"
3 tokenizer(examples)
4 return examples
I'm glad you figured something out :)
Regarding hashing: we're not using hashing for the same purpose as the python `__hash__` purpose (which is in general for dictionary lookups). For example it is allowed for python hashing to not return the same hash across sessions, while our hashing must return the same hashes across sessions for the caching to work properly. |
https://github.com/huggingface/datasets/issues/2514 | Can datasets remove duplicated rows? | Hi ! For now this is probably the best option.
We might add a feature like this in the feature as well.
Do you know any deduplication method that works on arbitrary big datasets without filling up RAM ?
Otherwise we can have do the deduplication in memory like pandas but I feel like this is going to be limiting for some cases | **Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no | 63 | Can datasets remove duplicated rows?
**Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no
Hi ! For now this is probably the best option.
We might add a feature like this in the feature as well.
Do you know any deduplication method that works on arbitrary big datasets without filling up RAM ?
Otherwise we can have do the deduplication in memory like pandas but I feel like this is going to be limiting for some cases |
https://github.com/huggingface/datasets/issues/2514 | Can datasets remove duplicated rows? | Yes, I'd like to work on this feature once I'm done with #2500, but first I have to do some research, and see if the implementation wouldn't be too complex.
In the meantime, maybe [this lib](https://github.com/TomScheffers/pyarrow_ops) can help. However, note that this lib operates directly on pyarrow tables and relies only on `hash` to find duplicates (e.g. `-1` and `-2` have the same hash in Python 3, so this lib will treat them as duplicates), which doesn't make much sense. | **Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no | 80 | Can datasets remove duplicated rows?
**Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no
Yes, I'd like to work on this feature once I'm done with #2500, but first I have to do some research, and see if the implementation wouldn't be too complex.
In the meantime, maybe [this lib](https://github.com/TomScheffers/pyarrow_ops) can help. However, note that this lib operates directly on pyarrow tables and relies only on `hash` to find duplicates (e.g. `-1` and `-2` have the same hash in Python 3, so this lib will treat them as duplicates), which doesn't make much sense. |
https://github.com/huggingface/datasets/issues/2514 | Can datasets remove duplicated rows? | > Hi ! For now this is probably the best option.
> We might add a feature like this in the feature as well.
>
> Do you know any deduplication method that works on arbitrary big datasets without filling up RAM ?
> Otherwise we can have do the deduplication in memory like pandas but I feel like this is going to be limiting for some cases
Great if this is can be done. Thanks!!
Not sure if you are asking me. In any case I don't know of any unfortunately :( in practice if data is really large we normally do it with spark (only for info. I understand this is not useful in developing this library..) | **Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no | 119 | Can datasets remove duplicated rows?
**Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no
> Hi ! For now this is probably the best option.
> We might add a feature like this in the feature as well.
>
> Do you know any deduplication method that works on arbitrary big datasets without filling up RAM ?
> Otherwise we can have do the deduplication in memory like pandas but I feel like this is going to be limiting for some cases
Great if this is can be done. Thanks!!
Not sure if you are asking me. In any case I don't know of any unfortunately :( in practice if data is really large we normally do it with spark (only for info. I understand this is not useful in developing this library..) |
https://github.com/huggingface/datasets/issues/2514 | Can datasets remove duplicated rows? | Hello,
I'm also interested in this feature.
Has there been progress on this issue?
Could we use a similar trick as above, but with a better hashing algorithm like SHA?
We could also use a [bloom filter](https://en.wikipedia.org/wiki/Bloom_filter), should we care a lot about collision in this case? | **Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no | 47 | Can datasets remove duplicated rows?
**Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no
Hello,
I'm also interested in this feature.
Has there been progress on this issue?
Could we use a similar trick as above, but with a better hashing algorithm like SHA?
We could also use a [bloom filter](https://en.wikipedia.org/wiki/Bloom_filter), should we care a lot about collision in this case? |
https://github.com/huggingface/datasets/issues/2514 | Can datasets remove duplicated rows? | For reference, we can get a solution fairly easily if we assume that we can hold in memory all unique values.
```python
from datasets import Dataset
from itertools import cycle
from functools import partial
memory = set()
def is_unique(elem:Any , column: str, memory: set) -> bool:
if elem[column] in memory:
return False
else:
memory.add(elem[column])
return True
# Example dataset
ds = Dataset.from_dict({"col1" : [sent for i, sent in zip(range(10), cycle(["apple", "orange", "pear"]))],
"col2": [i % 5 for i in range(10)]})
# Drop duplicates in `ds` on "col1"
ds2 = ds.filter(partial(is_unique, column="col1", memory=memory))
```
Of course, we can improve the API so that we can introduce `Dataset.drop_duplicates`.
For the parallel version, we can use a shared memory set. | **Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no | 117 | Can datasets remove duplicated rows?
**Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no
For reference, we can get a solution fairly easily if we assume that we can hold in memory all unique values.
```python
from datasets import Dataset
from itertools import cycle
from functools import partial
memory = set()
def is_unique(elem:Any , column: str, memory: set) -> bool:
if elem[column] in memory:
return False
else:
memory.add(elem[column])
return True
# Example dataset
ds = Dataset.from_dict({"col1" : [sent for i, sent in zip(range(10), cycle(["apple", "orange", "pear"]))],
"col2": [i % 5 for i in range(10)]})
# Drop duplicates in `ds` on "col1"
ds2 = ds.filter(partial(is_unique, column="col1", memory=memory))
```
Of course, we can improve the API so that we can introduce `Dataset.drop_duplicates`.
For the parallel version, we can use a shared memory set. |
https://github.com/huggingface/datasets/issues/2514 | Can datasets remove duplicated rows? | An approach that works assuming you can hold the all the unique document hashes in memory:
```python
from datasets import load_dataset
def get_hash(example):
"""Get hash of content field."""
return {"hash": hash(example["content"])} # can use any hashing function here
def check_uniques(example, uniques):
"""Check if current hash is still in set of unique hashes and remove if true."""
if example["hash"] in uniques:
uniques.remove(example["hash"])
return True
else:
return False
ds = load_dataset("some_dataset")
ds = ds.map(get_hash)
uniques = set(ds.unique("hash"))
ds_filter = ds.filter(check_uniques, fn_kwargs={"uniques": uniques})
```
If the `uniques` could be stored in arrow then no additional memory would used at all but I don't know if this is possible.
| **Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no | 105 | Can datasets remove duplicated rows?
**Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no
An approach that works assuming you can hold the all the unique document hashes in memory:
```python
from datasets import load_dataset
def get_hash(example):
"""Get hash of content field."""
return {"hash": hash(example["content"])} # can use any hashing function here
def check_uniques(example, uniques):
"""Check if current hash is still in set of unique hashes and remove if true."""
if example["hash"] in uniques:
uniques.remove(example["hash"])
return True
else:
return False
ds = load_dataset("some_dataset")
ds = ds.map(get_hash)
uniques = set(ds.unique("hash"))
ds_filter = ds.filter(check_uniques, fn_kwargs={"uniques": uniques})
```
If the `uniques` could be stored in arrow then no additional memory would used at all but I don't know if this is possible.
|
https://github.com/huggingface/datasets/issues/2514 | Can datasets remove duplicated rows? | @lvwerra hey, could you tell me how reliable is this deduplication method. i am currently using the same deduplication strategy to deduplicate a large text corpus to pretrain LLMs ~ 11B to 20B. just needed to ensure if this strategy would be fine on large datasets for LLMs pretraining. | **Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no | 49 | Can datasets remove duplicated rows?
**Is your feature request related to a problem? Please describe.**
i find myself more and more relying on datasets just to do all the preprocessing. One thing however, for removing duplicated rows, I couldn't find out how and am always converting datasets to pandas to do that..
**Describe the solution you'd like**
have a functionality of " remove duplicated rows"
**Describe alternatives you've considered**
convert dataset to pandas, remove duplicate, and convert back...
**Additional context**
no
@lvwerra hey, could you tell me how reliable is this deduplication method. i am currently using the same deduplication strategy to deduplicate a large text corpus to pretrain LLMs ~ 11B to 20B. just needed to ensure if this strategy would be fine on large datasets for LLMs pretraining. |
https://github.com/huggingface/datasets/issues/2511 | Add C4 | Update on this: I'm computing the checksums of the data files. It will be available soon | ## Adding a Dataset
- **Name:** *C4*
- **Description:** *https://github.com/allenai/allennlp/discussions/5056*
- **Paper:** *https://arxiv.org/abs/1910.10683*
- **Data:** *https://huggingface.co/datasets/allenai/c4*
- **Motivation:** *Used a lot for pretraining*
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Should fix https://github.com/huggingface/datasets/issues/1710 | 16 | Add C4
## Adding a Dataset
- **Name:** *C4*
- **Description:** *https://github.com/allenai/allennlp/discussions/5056*
- **Paper:** *https://arxiv.org/abs/1910.10683*
- **Data:** *https://huggingface.co/datasets/allenai/c4*
- **Motivation:** *Used a lot for pretraining*
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Should fix https://github.com/huggingface/datasets/issues/1710
Update on this: I'm computing the checksums of the data files. It will be available soon |
https://github.com/huggingface/datasets/issues/2508 | Load Image Classification Dataset from Local | Hi ! Is this folder structure a standard, a bit like imagenet ?
In this case maybe we can consider having a dataset loader for cifar-like, imagenet-like, squad-like, conll-like etc. datasets ?
```python
from datasets import load_dataset
my_custom_cifar = load_dataset("cifar_like", data_dir="path/to/data/dir")
```
Let me know what you think | **Is your feature request related to a problem? Please describe.**
Yes - we would like to load an image classification dataset with datasets without having to write a custom data loader.
**Describe the solution you'd like**
Given a folder structure with images of each class in each folder, the ability to load these folders into a HuggingFace dataset like "cifar10".
**Describe alternatives you've considered**
Implement ViT training outside of the HuggingFace Trainer and without datasets (we did this but prefer to stay on the main path)
Write custom data loader logic
**Additional context**
We're training ViT on custom dataset
| 48 | Load Image Classification Dataset from Local
**Is your feature request related to a problem? Please describe.**
Yes - we would like to load an image classification dataset with datasets without having to write a custom data loader.
**Describe the solution you'd like**
Given a folder structure with images of each class in each folder, the ability to load these folders into a HuggingFace dataset like "cifar10".
**Describe alternatives you've considered**
Implement ViT training outside of the HuggingFace Trainer and without datasets (we did this but prefer to stay on the main path)
Write custom data loader logic
**Additional context**
We're training ViT on custom dataset
Hi ! Is this folder structure a standard, a bit like imagenet ?
In this case maybe we can consider having a dataset loader for cifar-like, imagenet-like, squad-like, conll-like etc. datasets ?
```python
from datasets import load_dataset
my_custom_cifar = load_dataset("cifar_like", data_dir="path/to/data/dir")
```
Let me know what you think |
https://github.com/huggingface/datasets/issues/2508 | Load Image Classification Dataset from Local | @lhoestq I think we'll want a generic `image-folder` dataset (same as 'imagenet-like'). This is like `torchvision.datasets.ImageFolder`, and is something vision folks are used to seeing. | **Is your feature request related to a problem? Please describe.**
Yes - we would like to load an image classification dataset with datasets without having to write a custom data loader.
**Describe the solution you'd like**
Given a folder structure with images of each class in each folder, the ability to load these folders into a HuggingFace dataset like "cifar10".
**Describe alternatives you've considered**
Implement ViT training outside of the HuggingFace Trainer and without datasets (we did this but prefer to stay on the main path)
Write custom data loader logic
**Additional context**
We're training ViT on custom dataset
| 25 | Load Image Classification Dataset from Local
**Is your feature request related to a problem? Please describe.**
Yes - we would like to load an image classification dataset with datasets without having to write a custom data loader.
**Describe the solution you'd like**
Given a folder structure with images of each class in each folder, the ability to load these folders into a HuggingFace dataset like "cifar10".
**Describe alternatives you've considered**
Implement ViT training outside of the HuggingFace Trainer and without datasets (we did this but prefer to stay on the main path)
Write custom data loader logic
**Additional context**
We're training ViT on custom dataset
@lhoestq I think we'll want a generic `image-folder` dataset (same as 'imagenet-like'). This is like `torchvision.datasets.ImageFolder`, and is something vision folks are used to seeing. |
https://github.com/huggingface/datasets/issues/2508 | Load Image Classification Dataset from Local | Opening this back up, since I'm planning on tackling this. Already posted a quick version of it on my account on the hub.
```python
from datasets import load_dataset
ds = load_dataset('nateraw/image-folder', data_files='PetImages/')
``` | **Is your feature request related to a problem? Please describe.**
Yes - we would like to load an image classification dataset with datasets without having to write a custom data loader.
**Describe the solution you'd like**
Given a folder structure with images of each class in each folder, the ability to load these folders into a HuggingFace dataset like "cifar10".
**Describe alternatives you've considered**
Implement ViT training outside of the HuggingFace Trainer and without datasets (we did this but prefer to stay on the main path)
Write custom data loader logic
**Additional context**
We're training ViT on custom dataset
| 33 | Load Image Classification Dataset from Local
**Is your feature request related to a problem? Please describe.**
Yes - we would like to load an image classification dataset with datasets without having to write a custom data loader.
**Describe the solution you'd like**
Given a folder structure with images of each class in each folder, the ability to load these folders into a HuggingFace dataset like "cifar10".
**Describe alternatives you've considered**
Implement ViT training outside of the HuggingFace Trainer and without datasets (we did this but prefer to stay on the main path)
Write custom data loader logic
**Additional context**
We're training ViT on custom dataset
Opening this back up, since I'm planning on tackling this. Already posted a quick version of it on my account on the hub.
```python
from datasets import load_dataset
ds = load_dataset('nateraw/image-folder', data_files='PetImages/')
``` |
https://github.com/huggingface/datasets/issues/2503 | SubjQA wrong boolean values in entries | @arnaudstiegler I have just checked that these mismatches are already present in the original dataset: https://github.com/megagonlabs/SubjQA
We are going to contact the dataset owners to report this. | ## Describe the bug
SubjQA seems to have a boolean that's consistently wrong.
It defines:
- question_subj_level: The subjectiviy level of the question (on a 1 to 5 scale with 1 being the most subjective).
- is_ques_subjective: A boolean subjectivity label derived from question_subj_level (i.e., scores below 4 are considered as subjective)
However, `is_ques_subjective` seems to have wrong values in the entire dataset.
For instance, in the example in the dataset card, we have:
- "question_subj_level": 2
- "is_ques_subjective": false
However, according to the description, the question should be subjective since the `question_subj_level` is below 4
| 27 | SubjQA wrong boolean values in entries
## Describe the bug
SubjQA seems to have a boolean that's consistently wrong.
It defines:
- question_subj_level: The subjectiviy level of the question (on a 1 to 5 scale with 1 being the most subjective).
- is_ques_subjective: A boolean subjectivity label derived from question_subj_level (i.e., scores below 4 are considered as subjective)
However, `is_ques_subjective` seems to have wrong values in the entire dataset.
For instance, in the example in the dataset card, we have:
- "question_subj_level": 2
- "is_ques_subjective": false
However, according to the description, the question should be subjective since the `question_subj_level` is below 4
@arnaudstiegler I have just checked that these mismatches are already present in the original dataset: https://github.com/megagonlabs/SubjQA
We are going to contact the dataset owners to report this. |
https://github.com/huggingface/datasets/issues/2503 | SubjQA wrong boolean values in entries | I have:
- opened an issue in their repo: https://github.com/megagonlabs/SubjQA/issues/3
- written an email to all the paper authors | ## Describe the bug
SubjQA seems to have a boolean that's consistently wrong.
It defines:
- question_subj_level: The subjectiviy level of the question (on a 1 to 5 scale with 1 being the most subjective).
- is_ques_subjective: A boolean subjectivity label derived from question_subj_level (i.e., scores below 4 are considered as subjective)
However, `is_ques_subjective` seems to have wrong values in the entire dataset.
For instance, in the example in the dataset card, we have:
- "question_subj_level": 2
- "is_ques_subjective": false
However, according to the description, the question should be subjective since the `question_subj_level` is below 4
| 19 | SubjQA wrong boolean values in entries
## Describe the bug
SubjQA seems to have a boolean that's consistently wrong.
It defines:
- question_subj_level: The subjectiviy level of the question (on a 1 to 5 scale with 1 being the most subjective).
- is_ques_subjective: A boolean subjectivity label derived from question_subj_level (i.e., scores below 4 are considered as subjective)
However, `is_ques_subjective` seems to have wrong values in the entire dataset.
For instance, in the example in the dataset card, we have:
- "question_subj_level": 2
- "is_ques_subjective": false
However, according to the description, the question should be subjective since the `question_subj_level` is below 4
I have:
- opened an issue in their repo: https://github.com/megagonlabs/SubjQA/issues/3
- written an email to all the paper authors |
https://github.com/huggingface/datasets/issues/2499 | Python Programming Puzzles | Thanks @VictorSanh!
There's also a [notebook](https://aka.ms/python_puzzles) and [demo](https://aka.ms/python_puzzles_study) available now to try out some of the puzzles | ## Adding a Dataset
- **Name:** Python Programming Puzzles
- **Description:** Programming challenge called programming puzzles, as an objective and comprehensive evaluation of program synthesis
- **Paper:** https://arxiv.org/pdf/2106.05784.pdf
- **Data:** https://github.com/microsoft/PythonProgrammingPuzzles ([Scrolling through the data](https://github.com/microsoft/PythonProgrammingPuzzles/blob/main/problems/README.md))
- **Motivation:** Spans a large range of difficulty, problems, and domains. A useful resource for evaluation as we don't have a clear understanding of the abilities and skills of extremely large LMs.
Note: it's a growing dataset (contributions are welcome), so we'll need careful versioning for this dataset.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
| 17 | Python Programming Puzzles
## Adding a Dataset
- **Name:** Python Programming Puzzles
- **Description:** Programming challenge called programming puzzles, as an objective and comprehensive evaluation of program synthesis
- **Paper:** https://arxiv.org/pdf/2106.05784.pdf
- **Data:** https://github.com/microsoft/PythonProgrammingPuzzles ([Scrolling through the data](https://github.com/microsoft/PythonProgrammingPuzzles/blob/main/problems/README.md))
- **Motivation:** Spans a large range of difficulty, problems, and domains. A useful resource for evaluation as we don't have a clear understanding of the abilities and skills of extremely large LMs.
Note: it's a growing dataset (contributions are welcome), so we'll need careful versioning for this dataset.
Instructions to add a new dataset can be found [here](https://github.com/huggingface/datasets/blob/master/ADD_NEW_DATASET.md).
Thanks @VictorSanh!
There's also a [notebook](https://aka.ms/python_puzzles) and [demo](https://aka.ms/python_puzzles_study) available now to try out some of the puzzles |
https://github.com/huggingface/datasets/issues/2498 | Improve torch formatting performance | That’s interesting thanks, let’s see what we can do. Can you detail your last sentence? I’m not sure I understand it well. | **Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
| 22 | Improve torch formatting performance
**Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
That’s interesting thanks, let’s see what we can do. Can you detail your last sentence? I’m not sure I understand it well. |
https://github.com/huggingface/datasets/issues/2498 | Improve torch formatting performance | Hi ! I just re-ran a quick benchmark and using `to_numpy()` seems to be faster now:
```python
import pyarrow as pa # I used pyarrow 3.0.0
import numpy as np
n, max_length = 1_000, 512
low, high, size = 0, 2 << 16, (n, max_length)
table = pa.Table.from_pydict({
"input_ids": np.random.default_rng(42).integers(low=low, high=high, size=size).tolist()
})
%%timeit
_ = table.to_pandas()["input_ids"].to_numpy()
# 1.44 ms ± 80.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
%%timeit
_ = table["input_ids"].to_pandas().to_numpy()
# 461 µs ± 14.2 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
%%timeit
_ = table["input_ids"].to_numpy()
# 317 µs ± 5.06 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```
Currently the conversion from arrow to numpy is done in the NumpyArrowExtractor here:
https://github.com/huggingface/datasets/blob/d6d0ede9486ffad7944642ca9a326e058b676788/src/datasets/formatting/formatting.py#L143-L166
Let's update the NumpyArrowExtractor to call `to_numpy` directly and see how our github benchmarks evolve ?__ | **Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
| 150 | Improve torch formatting performance
**Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
Hi ! I just re-ran a quick benchmark and using `to_numpy()` seems to be faster now:
```python
import pyarrow as pa # I used pyarrow 3.0.0
import numpy as np
n, max_length = 1_000, 512
low, high, size = 0, 2 << 16, (n, max_length)
table = pa.Table.from_pydict({
"input_ids": np.random.default_rng(42).integers(low=low, high=high, size=size).tolist()
})
%%timeit
_ = table.to_pandas()["input_ids"].to_numpy()
# 1.44 ms ± 80.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
%%timeit
_ = table["input_ids"].to_pandas().to_numpy()
# 461 µs ± 14.2 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
%%timeit
_ = table["input_ids"].to_numpy()
# 317 µs ± 5.06 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```
Currently the conversion from arrow to numpy is done in the NumpyArrowExtractor here:
https://github.com/huggingface/datasets/blob/d6d0ede9486ffad7944642ca9a326e058b676788/src/datasets/formatting/formatting.py#L143-L166
Let's update the NumpyArrowExtractor to call `to_numpy` directly and see how our github benchmarks evolve ?__ |
https://github.com/huggingface/datasets/issues/2498 | Improve torch formatting performance | Sounds like a plan @lhoestq If you create a PR I'll pick it up and try it out right away! | **Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
| 20 | Improve torch formatting performance
**Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
Sounds like a plan @lhoestq If you create a PR I'll pick it up and try it out right away! |
https://github.com/huggingface/datasets/issues/2498 | Improve torch formatting performance | I’m not exactly sure how to read the graph but it seems that to_categorical take a lot of time here. Could you share more informations on the features/stats of your datasets so we could maybe design a synthetic datasets that looks more similar for debugging testing? | **Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
| 46 | Improve torch formatting performance
**Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
I’m not exactly sure how to read the graph but it seems that to_categorical take a lot of time here. Could you share more informations on the features/stats of your datasets so we could maybe design a synthetic datasets that looks more similar for debugging testing? |
https://github.com/huggingface/datasets/issues/2498 | Improve torch formatting performance | > I’m not exactly sure how to read the graph but it seems that to_categorical take a lot of time here. Could you share more informations on the features/stats of your datasets so we could maybe design a synthetic datasets that looks more similar for debugging testing?
@thomwolf starting from the top, each rectangle represents the cumulative amount of it takes to execute the method call. Therefore, format_batch in torch_formatter.py takes ~20 sec, and the largest portion of that call is taken by to_pandas call and the smaller portion (grey rectangle) by the other method invocation(s) in format_batch (series_to_numpy etc).
Features of the dataset are BERT pre-training model input columns i.e:
```
f = Features({
"input_ids": Sequence(feature=Value(dtype="int32")),
"attention_mask": Sequence(feature=Value(dtype="int8")),
"token_type_ids": Sequence(feature=Value(dtype="int8")),
"labels": Sequence(feature=Value(dtype="int32")),
"next_sentence_label": Value(dtype="int8")
})
```
I'll work with @lhoestq till we get to the bottom of this one.
| **Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
| 140 | Improve torch formatting performance
**Is your feature request related to a problem? Please describe.**
It would be great, if possible, to further improve read performance of raw encoded datasets and their subsequent conversion to torch tensors.
A bit more background. I am working on LM pre-training using HF ecosystem. We use encoded HF Wikipedia and BookCorpus datasets. The training machines are similar to DGX-1 workstations. We use HF trainer torch.distributed training approach on a single machine with 8 GPUs.
The current performance is about 30% slower than NVidia optimized BERT [examples](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling) baseline. Quite a bit of customized code and training loop tricks were used to achieve the baseline performance. It would be great to achieve the same performance while using nothing more than off the shelf HF ecosystem. Perhaps, in the future, with @stas00 work on deepspeed integration, it could even be exceeded.
**Describe the solution you'd like**
Using profiling tools we've observed that appx. 25% of cumulative run time is spent on data loader next call.
As you can observe most of the data loader next call is spent in HF datasets torch_formatter.py format_batch call.
Digging a bit deeper into format_batch we can see the following profiler data:
Once again, a lot of time is spent in pyarrow table conversion to pandas which seems like an intermediary step. Offline @lhoestq told me that this approach was, for some unknown reason, faster than direct to numpy conversion.
**Describe alternatives you've considered**
I am not familiar with pyarrow and have not yet considered the alternatives to the current approach.
Most of the online advice around data loader performance improvements revolve around increasing number of workers, using pin memory for copying tensors from host device to gpus but we've already tried these avenues without much performance improvement. Weights & Biases dashboard for the pre-training task reports CPU utilization of ~ 10%, GPUs are completely saturated (GPU utilization is above 95% on all GPUs), while disk utilization is above 90%.
> I’m not exactly sure how to read the graph but it seems that to_categorical take a lot of time here. Could you share more informations on the features/stats of your datasets so we could maybe design a synthetic datasets that looks more similar for debugging testing?
@thomwolf starting from the top, each rectangle represents the cumulative amount of it takes to execute the method call. Therefore, format_batch in torch_formatter.py takes ~20 sec, and the largest portion of that call is taken by to_pandas call and the smaller portion (grey rectangle) by the other method invocation(s) in format_batch (series_to_numpy etc).
Features of the dataset are BERT pre-training model input columns i.e:
```
f = Features({
"input_ids": Sequence(feature=Value(dtype="int32")),
"attention_mask": Sequence(feature=Value(dtype="int8")),
"token_type_ids": Sequence(feature=Value(dtype="int8")),
"labels": Sequence(feature=Value(dtype="int32")),
"next_sentence_label": Value(dtype="int8")
})
```
I'll work with @lhoestq till we get to the bottom of this one.
|