def _parse_defences(self, defences):
"""Apply defences to the classifier
:param defences: (string) names of the defences to add, supports "featsqueeze[1-8]" and "labsmooth"
"""
self.defences = defences
if defences:
pattern = re.compile("featsqueeze[1-8]?")
for d in defences:
# Add feature squeezing
if pattern.match(d):
try:
bit_depth = int(d[-1])
self.feature_squeeze = FeatureSqueezing(
bit_depth=bit_depth)
except:
raise ValueError(
"You must specify the bit depth for feature squeezing: featsqueeze[1-8]"
)
# Add label smoothing
if d == "labsmooth":
self.label_smooth = LabelSmoothing()
def test_default(self):
m, n = 1000, 20
y = np.zeros((m, n))
y[(range(m), np.random.choice(range(n), m))] = 1.
ls = LabelSmoothing()
_, smooth_y = ls(None, y)
self.assertTrue(np.isclose(np.sum(smooth_y, axis=1), np.ones(m)).all())
self.assertTrue((np.max(smooth_y, axis=1) == np.ones(m) * 0.9).all())
def test_customizing(self):
m, n = 1000, 20
y = np.zeros((m, n))
y[(range(m), np.random.choice(range(n), m))] = 1.
ls = LabelSmoothing(max_value=1.0 / n)
_, smooth_y = ls(None, y)
self.assertTrue(np.isclose(np.sum(smooth_y, axis=1), np.ones(m)).all())
self.assertTrue((np.max(smooth_y, axis=1) == np.ones(m) / n).all())
self.assertTrue(np.isclose(smooth_y, np.ones((m, n)) / n).all())