import load_data as ld
import numpy as np
import random
from pathlib import Path
import torch
import setproctitle
import time
import math
import os
if __name__ == "__main__":
args = ld.argparser(dataset='mnist',
metric='euclidean',
epsilon=1.58,
alpha=0.01,
delta=0.25)
# args = ld.argparser(dataset='cifar', metric='euclidean', epsilon=0.2453, alpha=0.05, delta=0)
setproctitle.setproctitle('python')
print('dataset: {dataset}\t\t'
'metric: {metric}\t\t'
'epsilon: {epsilon}\t\t'
'alpha: {alpha}\t\t'
'delta: {delta}\t\t'.format(dataset=args.dataset,
metric=args.metric,
epsilon=args.epsilon,
alpha=args.alpha,
delta=args.delta))
#### choose to use gpu if available
raise ImportError("The 'scipy' module could not be loaded. " +
"It is required for the 'brute_force' method " +
"for building a knn similarity graph.")
d = _spd.pdist(X, metric=metric)
D = _spd.squareform(d)
rank = np.argsort(D, axis=1)
neighbors = rank[:, 0:k]
k_nbr = neighbors[:, -1]
radii = D[np.arange(n), k_nbr]
return neighbors, radii
if __name__ == "__main__":
args = ld.argparser(dataset='mnist', metric='infinity', k=50)
# args = ld.argparser(dataset='cifar', metric='euclidean', alpha=0.05)
setproctitle.setproctitle('preliminary')
#### load the datasets
if args.dataset == 'mnist':
train_loader, _, valid_loader = ld.mnist_loaders(path='./data/' +
args.dataset,
seed=args.seed,
ratio=args.ratio)
for i, (X, y) in enumerate(train_loader):
train_data = X.view(-1, 28 * 28).numpy()
elif args.dataset == 'fmnist':
train_loader, _, valid_loader = ld.fashion_mnist_loaders(
path='./data/' + args.dataset, seed=args.seed, ratio=args.ratio)
####
#### strategy: cover the top-q densest images with union of rectangles as the robust region,
#### then expand them by epsilon to obtain the error region (rects-complement).
####
import load_data as ld
import numpy as np
import random
from pathlib import Path
import setproctitle
import os
if __name__ == "__main__":
args = ld.argparser(dataset='mnist',
metric='infinity',
epsilon=0.3,
q=0.703,
clusters=10)
# args = ld.argparser(dataset='cifar', metric='infinity', epsilon=0.007843, q=0.683, clusters=10)
setproctitle.setproctitle('python')
print('dataset: {dataset}\t\t'
'metric: {metric}\t\t'
'epsilon: {epsilon}\t\t'
'k: {k}\t\t'
'q: {q}\t\t'
'#clusters: {clusters}\t\t'
'#iter: {iter}\t\t'
'#repeat: {repeat}'.format(dataset=args.dataset,
metric=args.metric,
epsilon=args.epsilon,
#### l2-norm bounded perturbations
#### tune the number of clusters for the given epsilon using binary search over q
import load_data as ld
import numpy as np
import random
from pathlib import Path
import setproctitle
import os
if __name__ == "__main__":
args = ld.argparser(dataset='mnist', metric='infinity', epsilon=0.3, alpha=0.01)
# args = ld.argparser(dataset='cifar', metric='infinity', epsilon=0.007843, alpha=0.05)
setproctitle.setproctitle('python')
print('dataset: {dataset}\t\t' 'metric: {metric}\t\t' 'epsilon: {epsilon}\t\t'
'alpha: {alpha}\t\t' 'k: {k}\t\t' '#iter: {iter}\t\t'.format(
dataset=args.dataset, metric=args.metric, epsilon=args.epsilon,
alpha=args.alpha, k=args.k, iter=args.iter))
#### define results path
res_filepath = (os.path.dirname('./results/coarse_tune/'+args.proctitle)+'/alpha_'+str(args.alpha))
print("saving file to {}".format(res_filepath+'/epsilon_'+str(args.epsilon)))
if not os.path.exists(res_filepath):
os.makedirs(res_filepath)
#### load the datasets
if args.dataset == 'mnist':
train_loader, test_loader, valid_loader = ld.mnist_loaders(path='./data/'+args.dataset,
seed=args.seed,