Fix TorchModuleWrapper serialization issue

[Surya2k1]

Currently model with TorchModuleWrapper fails to save with the following error.

UnicodeDecodeError: 'utf-8' codec can't decode byte 0x80 in position 64: invalid start byte

I have tried with different encoding as experimental and found 'latin-1' works fine for saving and reloading along with some code changes to TorchModuleWrapper class. Atleast this change worked for the minimal code snippet as mentioned in #20860

May fixes #20860

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 27.27273% with 8 lines in your changes missing coverage. Please review.

Project coverage is 76.74%. Comparing base (c04cf9d) to head (ba24025).
Report is 2 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/src/utils/torch_utils.py	0.00%	6 Missing ⚠️
keras/src/saving/serialization_lib.py	60.00%	2 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (c04cf9d) and HEAD (ba24025). Click for more details.

HEAD has 2 uploads less than BASE

Flag	BASE (c04cf9d)	HEAD (ba24025)
keras	5	4
keras-torch	1	0

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #20869      +/-   ##
==========================================
- Coverage   82.24%   76.74%   -5.50%     
==========================================
  Files         561      561              
  Lines       52633    52640       +7     
  Branches     8137     8139       +2     
==========================================
- Hits        43288    40400    -2888     
- Misses       7340    10282    +2942     
+ Partials     2005     1958      -47     

Flag	Coverage Δ
keras	76.65% <27.27%> (-5.40%)	⬇️
keras-jax	64.22% <27.27%> (-0.01%)	⬇️
keras-numpy	59.03% <27.27%> (-0.01%)	⬇️
keras-openvino	32.52% <0.00%> (-0.01%)	⬇️
keras-tensorflow	64.84% <27.27%> (-0.01%)	⬇️
keras-torch	?

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[MicheleCattaneo]

It seems like you are trying to replace utf-8 with latin-1. I think that it would not fix the issue as both encodings are meant for text rather than arbitrary data. I wonder, what could have been the reason for using utf-8 to begin with?

[fchollet]

Thanks for the PR. I think this issue calls for a different fix -- what's the best way to serialize bytes (to go in a JSON file) in the general case? Most online solutions seem to recommend utf-8, but clearly this isn't fully general.

[MicheleCattaneo]

Thanks for the PR. I think this issue calls for a different fix -- what's the best way to serialize bytes (to go in a JSON file) in the general case? Most online solutions seem to recommend utf-8, but clearly this isn't fully general.

Is it possible that we need to first use something like base64 in get_config? The value of buffer.getvalue() are raw bytes that are placed in a dictionary, they later will be forced into a JSON, which I think should only accept strings (sequences of unicode characters) rather than bytes sequences. Once that is a valid string, then writing the JSON string to a file should hopefully work fine with utf-8.
What do you think?

[fchollet]

That sounds like something we can try. If we do that, we should introduce a new kind of object (not __bytes__) since we need to be backwards compatible with already saved files.