Added fix for mldata.org bug

Added a workaround to download MNIST data since mldata.org keeps going down (scikit-learn/scikit-learn#8588)

linear_svm/mnist_workaround.py

@@ -0,0 +1,21 @@
+"""
+A workaround to download MNIST data since mldata.org appears to be unstable
+Taken from https://github.com/scikit-learn/scikit-learn/issues/8588#issuecomment-292634781
+"""
+
+from shutil import copyfileobj
+from six.moves import urllib
+from sklearn.datasets.base import get_data_home
+import os
+
+def fetch_mnist():
+	mnist_alternative_url = "https://github.com/amplab/datascience-sp14/raw/master/lab7/mldata/mnist-original.mat"
+	data_home = get_data_home()
+	data_home = os.path.join(data_home, 'mldata')
+	if not os.path.exists(data_home):
+	    os.makedirs(data_home)
+	mnist_save_path = os.path.join(data_home, "mnist-original.mat")
+	if not os.path.exists(mnist_save_path):
+	    mnist_url = urllib.request.urlopen(mnist_alternative_url)
+	    with open(mnist_save_path, "wb") as matlab_file:
+	        copyfileobj(mnist_url, matlab_file)

linear_svm/sp_func_svm.py

@@ -15,9 +15,7 @@ def sp_frontend(images, rho=0.02, wavelet='bior4.4', mode='periodization', max_l
     :param wavelet: Wavelet to use in the transform. See https://pywavelets.readthedocs.io/ for more details.
     :param mode: Signal extension mode. See https://pywavelets.readthedocs.io/ for more details.
     :param max_lev: Maximum allowed level of decomposition.
-    """
-	# Input is assumed to be in the range [-1, 1] and of shape [num_samples, 784] 
-	# Projects input onto 
+    """ 
 	num_samples = images.shape[0]
 	num_features = images.shape[1]
 	num_features_per_dim = np.int(np.sqrt(num_features))

linear_svm/test_defense_svm.py

@@ -10,12 +10,14 @@
 import pywt
 from sklearn import datasets, utils, model_selection, metrics, svm
 from sp_func_svm import sp_project, sp_frontend
+from mnist_workaround import fetch_mnist
 
 epsilon = 0.25 # L-infinity attack budget. Images are assumed to be in the range [-1, 1].
 rho = 0.02 # Sparsity level used in the defense, in the range [0, 1].
 digit_1 = 3
 digit_2 = 7
 
+fetch_mnist()
 mnist = datasets.fetch_mldata("MNIST original")
 digit_1_data = 2.0*mnist.data[mnist.target==digit_1]/255.0 - 1.0
 digit_2_data = 2.0*mnist.data[mnist.target==digit_2]/255.0 - 1.0