Question regarding loss calculation about tcav

[lkreiskoether]

Hi again,

do we have any updates regarding this problem? In this context I thought that one possible fix for that could be to just change to a op for the loss calculation which expects class probabilities as input. So something like this could work for that:

with graph.as_default():
  self.y_input = tf.placeholder(tf.int64, shape=[None])

  self.pred = tf.expand_dims(self.ends['prediction'][0], 0)
  self.loss = tf.reduce_mean(
      tf.keras.losses.categorical_crossentropy(
          y_true=tf.one_hot(
              self.y_input,
              self.ends['prediction'].get_shape().as_list()[1]),
          y_pred=self.pred))
self._make_gradient_tensors()

What do you think?

Thanks in advance!

[BeenKim]

Hello,

There isn't much deep reasons for doing this logit vs class probability. Since the one is just normalized version of the other, I think either would get you pretty much the same answer. :) Please reopen if see further issues.