switch from training_args.bin training_args.json

[not-lain]

What does this PR do?

switch from training_args.bin to training_args.json and only capture the parameters that the user passed
I'm using the same approach we are using in huggingface_hub's PyTorchModelHubMixin to store as little parameters as possible.
a minimalistic approach to test this is pr

from transformers import TrainingArguments
args = TrainingArguments(output_dir="folder",eval_strategy="no") # or any other paramters
print(args.to_json_string())
# outputs
"""
{
  "output_dir": "folder",
  "eval_strategy": "no",
  "logging_dir": "folder\\runs\\Nov29_02-44-45_iphone-laptop"
}
"""
# logging_dir is a special parameter that is always captured and added to the training_args because we want to ensure consistency

# stores the parameters into a file
args.to_json_file("training_args.json")

# loads an instance using the class directly
args2  = TrainingArguments.from_json_file("training_args.json") 

using this approach, we ensure that we only store the parameters that the user-defined manually and not the ones that have default values or the ones inferred from the system (ie cpu, cuda, tpu ... ), leaving some room for flexibility.

in a sense the parameters are mutable, meaning the user can physically alter them.

Fixes #34612

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@muellerzr @SunMarc

[not-lain]

cc @muellerzr & @SunMarc for review.
failed test is unrelated

[not-lain]

friendly tagging @muellerzr & @SunMarc
failed tests are unrelated

[muellerzr]

Very nice! I'm a fan of this change. However, we should be careful about old training resumptions from older versions and likely should do a deprecation until 5.0.0 (warning if a .bin was found but we do the shift to .json now)

How does that sound?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[not-lain]

[don't merge yet I need to fix something fundamental that was not captured by the tests]

[not-lain]

cc @muellerzr for feedback
I implemented a rollback, now default is traing_args.bin and people can switch to training_args.json by setting the TRAINER_SAFE_SERIALIZE environmental variable.

originally there was never a mechanism inside of the trainer API to load any training_args and people used to load the training arguments manually, I added something useful which is TrainingArguments.from_json_file(file_name )
but I think I will integrate this loading mechanism with the trainer api in another pr (optional new feature)

[not-lain]

friendly tagging @muellerzr here for a review.

[not-lain]

cc @muellerzr for review

[SunMarc]

Thanks for the PR ! Left a few comments !

[SunMarc]

We can remove this and the function since you took care of that in serialize_parameter. Can you add a comment in serialize_parameter to explain what is happening to torch_dtype ?

[not-lain]

I hope I did not misunderstand you, let me know if this has been resolved after the new changes

[SunMarc]

Looks fine thanks ! SInce we are not using _dict_torch_dtype_to_str method, you can also remove it in this PR

[not-lain]

cc @SunMarc for review

I've switched to dynamic serialization which something that came up with to tackle #25903 .
this will serialize the parameters needed to make the training arguments lighweight, mutable and reproducible.

here's a thorough example on what to expect :

from transformers import TrainingArguments
args = TrainingArguments(output_dir="folder",eval_strategy="no") # or any other parameters
print(args.to_json_string())
# outputs (logging_dir needs to stay consistant in case we want tensorboard to continue the plots)
"""
{
  "output_dir": "folder",
  "eval_strategy": "no",
  "logging_dir": "folder\\runs\\Dec27_21-48-40_iphone-laptop"
}
"""
# has some default parameters in the method which have been captured and serialized
# this also captures manually set parameters in either args or kwargs format.
args = args.set_dataloader() 
print(args.to_json_string())
# outputs
"""
{
  "output_dir": "folder",
  "eval_strategy": "no",
  "logging_dir": "folder\\runs\\Dec27_21-48-40_iphone-laptop",
  "auto_find_batch_size": false,
  "ignore_data_skip": false
}
"""
# save and reload
args.to_json_file("training_args.json")
args2  = TrainingArguments.from_json_file("training_args.json")
print(args2.to_json_string())
# output
"""
{
  "output_dir": "folder",
  "eval_strategy": "no",
  "logging_dir": "folder\\runs\\Dec27_21-48-40_iphone-laptop",
  "auto_find_batch_size": false,
  "ignore_data_skip": false
}
"""

let me know if you have any reviews

[SunMarc]

LGTM ! thanks for iterating ! Can you have a second look @muellerzr ?