Training code for LDS benchmark

[enkeejunior1]

Hello!

I'm working on implementing LDS for a new dataset, but I'm having trouble finding the code for training models on 50% subsets. Could someone point me to the relevant script or file that handles subset model training for LDS? I found main/experiments/gpt2_wikitext but just wanted to ask is there any code for more general case.

Thanks in advance!

[TheaperDeng]

Hi @enkeejunior1. Thank you for the issue!

• We do have a function for the model retraining for LDS here:

  dattri/dattri/model_util/retrain.py

  Lines 134 to 145 in e909f76

  	def retrain_lds(
  	train_func: Callable,
  	dataloader: torch.utils.data.DataLoader,
  	path: str,
  	num_subsets: int = 100,
  	subset_ratio: float = 0.5,
  	num_runs_per_subset: int = 1,
  	start_id: int = 0,
  	total_num_subsets: int = 0,
  	seed: Optional[int] = None,
  	**kwargs,
  	) -> None:

  . You may use it for your new model or new dataset. The function is also well-documented here: https://trais-lab.github.io/dattri/api/benchmark.html#dattri.model_util.retrain.retrain_lds

• If you want to retrain the model in dattri's benchmark settings, you may use https://github.com/TRAIS-Lab/dattri/blob/main/dattri/script/dattri_retrain.py#L33 or dattri_retrain --help.

Please feel free to ask here if you still have any other questions.