'cp950' codec error from load_dataset('xtreme', 'tydiqa')

[jerryIsHere]

I guess the error is related to python source encoding issue that my PC is trying to decode the source code with wrong encoding-decoding tools, perhaps :
https://www.python.org/dev/peps/pep-0263/

I guess the error was triggered by the code " module = importlib.import_module(module_path)" at line 57 in the source code: nlp/src/nlp/load.py / (https://github.com/huggingface/nlp/blob/911d5596f9b500e39af8642fe3d1b891758999c7/src/nlp/load.py#L51)

Any ideas?

p.s. tried the same code on colab, that runs perfectly

[patpizio]

This is probably a Windows issue, we need to specify the encoding when load_dataset() reads the original CSV file.
Try to find the open() statement called by load_dataset() and add an encoding='utf-8' parameter.
See issues #242 and #307

[lhoestq]

It should be in xtreme.py:L755:

        if self.config.name == "tydiqa" or self.config.name.startswith("MLQA") or self.config.name == "SQuAD":
            with open(filepath) as f:
                data = json.load(f)

Could you try to add the encoding parameter:

open(filepath, encoding='utf-8')

[lhoestq]

Hello @jerryIsHere :) Did it work ?
If so we may change the dataset script to force the utf-8 encoding

[jerryIsHere]

@lhoestq sorry for being that late, I found 4 copy of xtreme.py. I did the changes as what has been told to all of them.
The problem is not solved

[lhoestq]

Could you provide a better error message so that we can make sure it comes from the opening of the tydiqa's json files ?

[jerryIsHere]

@lhoestq
The error message is same as before:
Exception has occurred: UnicodeDecodeError
'cp950' codec can't decode byte 0xe2 in position 111: illegal multibyte sequence
File "D:\python\test\test.py", line 3, in
dataset = load_dataset('xtreme', 'tydiqa')

I said that I found 4 copy of xtreme.py and add the 「, encoding='utf-8'」 parameter to the open() function
these python script was found under this directory
C:\Users\USER\AppData\Local\Programs\Python\Python37\Lib\site-packages\nlp\datasets\xtreme

[ghazi-f]

Hi there !
I encountered the same issue with the IMDB dataset on windows. It threw an error about charmap not being able to decode a symbol during the first time I tried to download it. I checked on a remote linux machine I have, and it can't be reproduced.
I added encoding='UTF-8' to both lines that have open in imdb.py (108 and 114) and it worked for me.
Thank you !

[lhoestq]

Hi there !
I encountered the same issue with the IMDB dataset on windows. It threw an error about charmap not being able to decode a symbol during the first time I tried to download it. I checked on a remote linux machine I have, and it can't be reproduced.
I added encoding='UTF-8' to both lines that have open in imdb.py (108 and 114) and it worked for me.
Thank you !

Hello !
Glad you managed to fix this issue on your side.
Do you mind opening a PR for IMDB ?

[jerryIsHere]

This is probably a Windows issue, we need to specify the encoding when load_dataset() reads the original CSV file.
Try to find the open() statement called by load_dataset() and add an encoding='utf-8' parameter.
See issues #242 and #307

Sorry for not responding for about a month.
I have just found that it is necessary to change / add the environment variable as what was told in #242.
Everything works after I add the new environment variable and restart my PC.

I think the encoding issue for windows isn't limited to the open() function call specific to few dataset, but actually in the entire library, depends on the machine / os you use.

[lhoestq]

Since #481 we shouldn't have other issues with encodings as they need to be set to "utf-8" be default.

Closing this one, but feel free to re-open if you gave other questions