48a54c32a55896620b0523d5a3a68b1826f957cc,art/attacks/virtual_adversarial.py,VirtualAdversarialMethod,generate,#VirtualAdversarialMethod#Any#,38

Before Change

                kl_div1 = entropy(np.transpose(preds[batch_index_1:batch_index_2]), np.transpose(preds_new))

                d_new = np.zeros_like(d)
                array_iter = np.nditer(d, op_flags=["readwrite"], flags=["multi_index"])
                for x in array_iter:
                    x[...] += self.finite_diff
                    preds_new = self.classifier.predict((val + d)[None, ...], logits=False)
                    kl_div2 = entropy(preds[ind], preds_new[0])

After Change

                kl_div1 = entropy(np.transpose(preds[batch_index_1:batch_index_2]), np.transpose(preds_new))

                d_new = np.zeros_like(d)
                for w in range(d.shape[1]):
                    for h in range(d.shape[2]):
                        for c in range(d.shape[3]):
                            d[:, w, h, c] += self.finite_diff
                            preds_new = self.classifier.predict(batch + d, logits=False)
                            kl_div2 = entropy(np.transpose(preds[batch_index_1:batch_index_2]), np.transpose(preds_new))
                            d_new[:, w, h, c] = (kl_div2 - kl_div1) / (self.finite_diff + tol)
                            d[:, w, h, c] -= self.finite_diff
                d = d_new

            // Apply perturbation and clip
            x_adv[batch_index_1:batch_index_2] = np.clip(batch + self.eps * self._normalize(d), clip_min, clip_max)

Instances