AttributeError: 'BartConfig' object has no attribute '_to_dict_new'

[LRY1994]

Environment info

• adapter-transformers version:3.0.0
• Platform: Unbuntu
• Python version: 3.8
• PyTorch version (GPU?): 1.11.0
• Using GPU in script?: yes
• Using distributed or parallel set-up in script?: no

Information

Model I am using : BartForConditionalGeneration

Language I am using the model on: English

Adapter setup I am using (if any): no

The problem arises when using:

• torch.save(self.model, self.args.best_model_dir + "model.bin")
• model = torch.load(self.args.best_model_dir + "model.bin")
• Traceback (most recent call last):
  File "/home/simon/桌面/closed-book-prompt-qa/src/test.py", line 3, in
  model = torch.load( "src/temp/model_1653153641/model.bin")
  File "/home/simon/anaconda3/envs/qa/lib/python3.8/site-packages/torch/serialization.py", line 712, in load
  return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)
  File "/home/simon/anaconda3/envs/qa/lib/python3.8/site-packages/torch/serialization.py", line 1046, in _load
  result = unpickler.load()
  File "/home/simon/anaconda3/envs/qa/lib/python3.8/site-packages/transformers/configuration_utils.py", line 253, in getattribute
  return super().getattribute(key)
  AttributeError: 'BartConfig' object has no attribute '_to_dict_new'

Expected behavior

successfully loaded for my downstream task

[LRY1994]

when I downgrade to version 2.3.0 , it works .

[calpt]

Hey @LRY1994, this behavior is indeed introduced by a change in v3.x of the library. The recommended way of saving & loading model checkpoints using (adapter-)transformers is via the save_pretrained()/ from_pretrained() methods:

model.save_pretrained(self.args.best_model_dir)
...
model = BartForConditionalGeneration.from_pretrained(self.args.best_model_dir)

We'll look into solving the issues using torch.save()/load() you described.

[calpt]

This issue should be fixed with #406 and the next adapter-transformers release (v3.1.0).