Delta-XAI

September 19, 2025 · View on GitHub

Environmental Setup

conda create -n deltaxai python=3.10.9
conda activate deltaxai
pip install -r requirements.txt

Hardware Infrastructure Specifications

  • CPU: Intel Xeon Silver 4210 @ 2.20GHz, 20 physical cores, 40 logical cores (dual-socket)
  • Memory: 256GB RAM (2 NUMA nodes)
  • GPU: 8 × NVIDIA TITAN RTX, 24GB memory per GPU (192GB total)

Reproducing Experiments

Data Preparation

Running Experiments

  • Run attribution extraction:
bash scripts/test.sh

Quick-start

  1. Load an MIMIC-III dataset:
from data.dataloader import Mimic3o
from utils.config import get_args
from run import Params

params = dict()
params['vis'] = False
params['train'] = False
params['traingen'] = False

dataset = Mimic3o(params, data_path="Mimic3o_data_path", testbs=128)
  1. Load a ExplanationRunner with initialize a target model (LSTM) applied by a Wrapper Function(PredictionDifferenceWrapper)
from run import ExplanationRunner

# output_path for result saving and ckpt_path for model checkpoint upload
runner = ExplanationRunner(args, dataset, device, output_path, ckpt_path)
runner.init_model(**model_args)
runner.load_model(True)
  1. Compute the feature attribution:
runner.clean_up(
	clean_importance=True, clean_explainer=True, clean_model=True
)
runner.get_explainers(
	args, 'SWING'
)
runner.set_model_for_explainer(set_eval=True)
runner.run_attributes()

importances = runner.importances