def weight_diff(self, x_tensor, model, clss, clss_num):
set_random_seed()
def predict_fun(x):
rst = model.forward(x)
return rst
weight_diff, samples_tensor = self.explainer.explain_instance(
x_tensor, predict_fun, clss, labels=list(range(clss_num)))
return weight_diff, self.dist, samples_tensor
def weight_diff(self, x_tensor, model, clss, clss_num):
set_random_seed()
dist = 1
w_diff_est_tensor = None
samples_tensor = None
counter = 0
error = None
while dist > self.threshold:
counter += 1
try:
samples_tensor = unif_sample(x_tensor, self.sample_num, dist)
w_diff_est_tensor = reverse(x_tensor, model, samples_tensor,
clss)
break
except Exception as e:
error = e
dist /= 2.0
logger.debug("Counter: {}".format(counter))
if w_diff_est_tensor is None:
logger.exception(error)
raise Exception(
"Due to the limitations of python's computing error, we cannot estimate the gradient of x given the distance threshold."
)
return w_diff_est_tensor, dist, samples_tensor
import torch
import matplotlib
matplotlib.use('agg')
from matplotlib.backends.backend_pdf import PdfPages
from openapi.exp.abs_exp import AbsExp
from openapi.utils import set_random_seed, PathManager, Plot
from openapi.log import getLogger
from openapi import config
######################################################################
# LMT related module must be imported to make the pickle load successfully
from openapi.mdl.lmt import LinearModelTree, Node, LocalLasso
######################################################################
set_random_seed()
logger = getLogger(__name__)
class Exp2(AbsExp):
def __init__(self, model_name, dataset, expln_name, datasize):
AbsExp.__init__(self, model_name, dataset, "2", expln_name, datasize)
@staticmethod
def plot(model_name, dataset, datasize):
exp = "2"
path_manager = PathManager(model_name, dataset)
dist = defaultdict(list)
for expl_name in AbsExp.get_method_names():
result_file = path_manager.result_json_path(
