def test_loss_sensitivity(self):
# Get MNIST
(x_train, y_train), (_, _), _, _ = load_mnist()
x_train, y_train = x_train[:NB_TRAIN], y_train[:NB_TRAIN]
# Get classifier
classifier = self._cnn_mnist_k([28, 28, 1])
classifier.fit(x_train, y_train, batch_size=BATCH_SIZE, nb_epochs=2)
l = loss_sensitivity(classifier, x_train, y_train)
self.assertGreaterEqual(l, 0)
def evaluate_robust(self, data_loader):
data_iter = iter(data_loader)
examples, labels = next(data_iter)
examples, labels = examples.cpu().numpy(), labels.cpu().numpy()
labels_one_hot = np.eye(self.nb_classes)[labels]
losses = []
# Compute loss with implicit batching
batch_size = 256
for batch_id in range(
int(np.ceil(examples.shape[0] / float(batch_size)))):
batch_index_1, batch_index_2 = batch_id * batch_size, (
batch_id + 1) * batch_size
batch = examples[batch_index_1:batch_index_2]
batch_labels = labels_one_hot[batch_index_1:batch_index_2]
loss = loss_sensitivity(self.classifier, batch, batch_labels)
losses.append(loss * batch.shape[0])
res = sum(losses) / examples.shape[0]
return res