4887ef8baecbf5315ec0f235e56a4f93cd05aad7,cleverhans/attacks_tf.py,,spm,#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#,1947

Before Change

    sampled_angles = np.random.choice(angles, n_samples)
    transforms = zip(sampled_dxs, sampled_dys, sampled_angles)

  adv_xs = []
  accs = []

  // Perform the transformation
  for (dx, dy, angle) in transforms:
    adv_xs.append(_apply_transformation(x, dx, dy, angle, batch_size))
    preds_adv = model.get_logits(adv_xs[-1])

    // Compute accuracy
    accs.append(tf.count_nonzero(tf.equal(tf.argmax(y, axis=-1),
                                          tf.argmax(preds_adv, axis=-1))))

After Change

  all_xents = tf.stack(all_xents) // SB

  // We want the worst case sample, with the largest xent_loss
  worst_sample_idx = tf.argmax(all_xents, axis=0)  // B

  batch_size = tf.shape(x)[0]
  keys = tf.stack([
      tf.range(batch_size, dtype=tf.int32),
      tf.cast(worst_sample_idx, tf.int32)
  ], axis=1)
  transformed_ims_bshwc = tf.einsum("sbhwc->bshwc", transformed_ims)
  after_lookup = tf.gather_nd(transformed_ims_bshwc, keys)  // BHWC
  return after_lookup

Instances