Only input-output Jacobian of the first image in a batch has value

[Yujun-Shi]

Hi there, I added the following lines to the original code to print the value of the input-output jacobian, and found that only the input-output jacobian of the first image in a batch has value during training. (My tensorflow version is 1.7.0)

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  for i in range(num_of_iterations + num_of_post_processing_iterations):
    # some code
    if i % 100 == 0:
      # shape: (10, 500, 784)
      Jacobian_numpy = Jacobian.eval(feed_dict={x: X_to_train, y_: Y_to_train, keep_prob: 1.0})
      # shape: (500, 10, 784), which is (N, output_dim, input_dim)
      Jacobian_numpy = Jacobian_numpy.transpose((1,0,2))
      print(Jacobian_numpy[0]) # has some value for every entry of the matrix
      print(Jacobian_numpy[1]) # all 0 and the same for any batch index >= 1

Although the parameters of the network could of course still be regularized in this situation (only using one image for every step to regularize jacobian), it's not exactly consistent with the paper.

[TakeByEarn]

i guess this is a simplification of the average jacobian_loss, author use the first data of jacobian loss instead of use tf.reduce_mean( tf.reduce_sum( jacobian_loss)), he want to use one singlee of jacobian to apporximate the acerage jacobian average loss.

[TakeByEarn]

if you calculate all the data's jacobian loss ,then get average jacobian loss, it woule take much more time to do calculation.