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

Before Change

                                          tf.argmax(preds_adv, axis=-1))))
  // Return the adv_x with worst accuracy
  adv_xs = tf.stack(adv_xs)
  accs = tf.stack(accs)
  return tf.gather(adv_xs, tf.argmin(accs))

After Change

    return tf.nn.softmax_cross_entropy_with_logits_v2(
        labels=y, logits=preds)

  all_xents = tf.map_fn(
      _compute_xent,
      transformed_ims,
      parallel_iterations=1) // Must be 1 to avoid keras race conditions

  // Return the adv_x with worst accuracy

  // all_xents is n_total_samples x batch_size (SB)
  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